EP4791901A1 - Methods of determining response of a tumor to dna-damaging agents or to agents inhibiting or impairing dna repair - Google Patents

Methods of determining response of a tumor to dna-damaging agents or to agents inhibiting or impairing dna repair

Info

Publication number
EP4791901A1
EP4791901A1 EP24786486.1A EP24786486A EP4791901A1 EP 4791901 A1 EP4791901 A1 EP 4791901A1 EP 24786486 A EP24786486 A EP 24786486A EP 4791901 A1 EP4791901 A1 EP 4791901A1
Authority
EP
European Patent Office
Prior art keywords
loh
genomic dna
copy number
therapy
bin
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
EP24786486.1A
Other languages
German (de)
French (fr)
Inventor
Diether Lambrechts
Tom Venken
Pieter BUSSCHAERT
Liselore LOVERIX
Ignace VERGOTE
Toon VAN GORP
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
KU Leuven
Vlaams Instituut voor Biotechnologie
Original Assignee
KU Leuven
Vlaams Instituut voor Biotechnologie
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by KU Leuven, Vlaams Instituut voor Biotechnologie filed Critical KU Leuven
Publication of EP4791901A1 publication Critical patent/EP4791901A1/en
Pending legal-status Critical Current

Links

Classifications

    • C—CHEMISTRY; METALLURGY
    • C12—BIOCHEMISTRY; BEER; SPIRITS; WINE; VINEGAR; MICROBIOLOGY; ENZYMOLOGY; MUTATION OR GENETIC ENGINEERING
    • C12Q—MEASURING OR TESTING PROCESSES INVOLVING ENZYMES, NUCLEIC ACIDS OR MICROORGANISMS; COMPOSITIONS OR TEST PAPERS THEREFOR; PROCESSES OF PREPARING SUCH COMPOSITIONS; CONDITION-RESPONSIVE CONTROL IN MICROBIOLOGICAL OR ENZYMOLOGICAL PROCESSES
    • C12Q1/00—Measuring or testing processes involving enzymes, nucleic acids or microorganisms; Compositions therefor; Processes of preparing such compositions
    • C12Q1/68—Measuring or testing processes involving enzymes, nucleic acids or microorganisms; Compositions therefor; Processes of preparing such compositions involving nucleic acids
    • C12Q1/6876—Nucleic acid products used in the analysis of nucleic acids, e.g. primers or probes
    • C12Q1/6883—Nucleic acid products used in the analysis of nucleic acids, e.g. primers or probes for diseases caused by alterations of genetic material
    • C12Q1/6886—Nucleic acid products used in the analysis of nucleic acids, e.g. primers or probes for diseases caused by alterations of genetic material for cancer
    • C—CHEMISTRY; METALLURGY
    • C12—BIOCHEMISTRY; BEER; SPIRITS; WINE; VINEGAR; MICROBIOLOGY; ENZYMOLOGY; MUTATION OR GENETIC ENGINEERING
    • C12Q—MEASURING OR TESTING PROCESSES INVOLVING ENZYMES, NUCLEIC ACIDS OR MICROORGANISMS; COMPOSITIONS OR TEST PAPERS THEREFOR; PROCESSES OF PREPARING SUCH COMPOSITIONS; CONDITION-RESPONSIVE CONTROL IN MICROBIOLOGICAL OR ENZYMOLOGICAL PROCESSES
    • C12Q2600/00—Oligonucleotides characterized by their use
    • C12Q2600/106—Pharmacogenomics, i.e. genetic variability in individual responses to drugs and drug metabolism
    • C—CHEMISTRY; METALLURGY
    • C12—BIOCHEMISTRY; BEER; SPIRITS; WINE; VINEGAR; MICROBIOLOGY; ENZYMOLOGY; MUTATION OR GENETIC ENGINEERING
    • C12Q—MEASURING OR TESTING PROCESSES INVOLVING ENZYMES, NUCLEIC ACIDS OR MICROORGANISMS; COMPOSITIONS OR TEST PAPERS THEREFOR; PROCESSES OF PREPARING SUCH COMPOSITIONS; CONDITION-RESPONSIVE CONTROL IN MICROBIOLOGICAL OR ENZYMOLOGICAL PROCESSES
    • C12Q2600/00—Oligonucleotides characterized by their use
    • C12Q2600/118—Prognosis of disease development
    • C—CHEMISTRY; METALLURGY
    • C12—BIOCHEMISTRY; BEER; SPIRITS; WINE; VINEGAR; MICROBIOLOGY; ENZYMOLOGY; MUTATION OR GENETIC ENGINEERING
    • C12Q—MEASURING OR TESTING PROCESSES INVOLVING ENZYMES, NUCLEIC ACIDS OR MICROORGANISMS; COMPOSITIONS OR TEST PAPERS THEREFOR; PROCESSES OF PREPARING SUCH COMPOSITIONS; CONDITION-RESPONSIVE CONTROL IN MICROBIOLOGICAL OR ENZYMOLOGICAL PROCESSES
    • C12Q2600/00—Oligonucleotides characterized by their use
    • C12Q2600/156—Polymorphic or mutational markers

Landscapes

  • Chemical & Material Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Organic Chemistry (AREA)
  • Proteomics, Peptides & Aminoacids (AREA)
  • Engineering & Computer Science (AREA)
  • Immunology (AREA)
  • Pathology (AREA)
  • Analytical Chemistry (AREA)
  • Zoology (AREA)
  • Genetics & Genomics (AREA)
  • Wood Science & Technology (AREA)
  • Physics & Mathematics (AREA)
  • Biotechnology (AREA)
  • Microbiology (AREA)
  • Molecular Biology (AREA)
  • Hospice & Palliative Care (AREA)
  • Biophysics (AREA)
  • Oncology (AREA)
  • Biochemistry (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Measuring Or Testing Involving Enzymes Or Micro-Organisms (AREA)

Abstract

This disclosure relates to methods of determining the response of a tumor to DNA-damaging agents or to agents inhibiting or impairing DNA repair (such as inhibitors of the poly ADP ribose polymerase or to platinum-based agents). In particular such methods include computer-assisted classification of a test genomic DNA sample based on BRCA gene mutation status, a BRCA domain mutation feature, and a loss- of-heterozygosity feature, wherein the classifier has been trained on set of the corresponding statuses and features of reference genomic DNA samples of subjects known to respond to DNA-damaging agents or to agents inhibiting or impairing DNA repair.

Description

METHODS OF DETERMINING RESPONSE OF A TUMOR TO DNA-DAMAGING AGENTS OR TO AGENTS INHIBITING OR IMPAIRING DNA REPAIR
FIELD OF THE INVENTION
This disclosure relates to methods of determining the response of a tumor to DNA-damaging agents or to agents inhibiting or impairing DNA repair (such as inhibitors of the poly ADP ribose polymerase or to platinum-based agents). In particular such methods include computer-assisted classification of a test genomic DNA sample based on BRCA gene mutation status, a BRCA domain mutation feature, and a loss- of-heterozygosity feature, wherein the classifier has been trained on set of the corresponding statuses and features of reference genomic DNA samples of subjects known to respond to DNA-damaging agents or to agents inhibiting or impairing DNA repair.
BACKGROUND OF THE INVENTION
It is generally known that different patients having the same type of cancer respond differently to the same therapy. This is true for chemotherapy as well as for the more recent immunotherapies including immune checkpoint inhibitors. This led cancer researchers worldwide to identification of biochemical or genetic markers that can predict response to a particular anti-cancer therapy or that can identify whether a patient on a particular anti-cancer therapy is responding to that therapy, preferably early after initiation of that therapy.
Such successful prediction or early identification of response to therapy, when available, not only improves the outcome for the patients, but also results in health-economic benefits.
One field of genetic testing in the context of predicting therapy response relates to testing of genomic scarring or homologous recombination (repair) deficiency (HRD; homologous recombination repair (HRR) is the process of normal DNA repair, especially of repair of breaks in both DNA strands) testing in the context of predicting response to therapies including poly-ADP-ribose polymerase inhibitors (PARPi), platinum-based therapies, etc.
A number of HRD causes have been identified including mutations or lowered expression of the BRCA1 and/or BRCA2 genes and other genes such as RAD51C and PALB2. HRD is, however, not always linked to identifiable gene mutations. Across cancers, HRD occurs at a frequency of about 6%. Rates can be as high as 30% for ovarian cancer, and intermediate for breast, pancreatic and prostate cancer (12-13%).
A number of genomic scarring parameters have been introduced including loss of heterozygosity (LOH) or percentage LOH (see e.g. W02011/106541; W02011/160063; WO2013/096843; Abkevich et al. 2012, Br J Cancer 107:1776-1782), fractional LOH (FLOH; e.g. Wang et al. 2012, Clin Cancer Res 18:5806-5815), telomeric allelic imbalance (TAI; see e.g. WO 2012/027224; WO2013/130347; Birkbak et al. 2012, Cancer Discov 2:366-375), large scale transitions (LST; see e.g. WO2013/182645; WO 2015/086473; Popova et al. 2012, Cancer Res 72:5454-5462), and combinations thereof (see e.g. WO2014/165785; WO2016/025958; Timms et al. 2014, Breast Cancer Res 16:475). These documents linked diagnostic test outcome to prediction of outcome of treatment, in particular to outcome of treatment with platinumbased therapeutic modalities.
Few commercial assays testing a patient's HRD status are available and include the Myriad Genetics MyChoice® CDx test (considered as the gold standard in the field; based on LOH, TAI, LST, and status of BRCA1 and BRCA2) and the FoundationOne® CDx test (based on copy number alterations in a set of genes (including HRR genes) and %LOH). According to myriad.com/oncology/mychoice-cdx/ (version 02 August 2022), however, methods based on %LOH determination are significantly less successful in determining HRD than the Myriad Genetics MyChoice® CDx test, which is confirmed by Timms et al. 2020 (J Clin Oncol 38(15 suppl):1586). ). In W02022/150063, a LOH score is determined based on a ratio of the number of LOH SNP loci to the number of non-homozygous SNP loci.
A comprehensive overview of clinical assays applied in determining HRD status is provided in Stover et al. 2020 (Gynecologic Oncology 159:887-898).
SUMMARY OF THE INVENTION
This disclosure in one aspect relates to methods of predicting the response of a subject having cancer to a therapy including a DNA damaging agent or including an agent inhibiting or impairing DNA repair, comprising:
- optionally, isolating test genomic DNA from a test genomic DNA sample obtained from the subject;
- detecting or determining in a test genomic DNA obtained from the subject (or detecting or determining in the test genomic DNA) the presence of mutations in the BRCA1 and BRCA2 genes, and classifying the subject as BRCA mutant if a mutation is present in the BRCA1 and/or BRCA2 gene, or as BRCA wild-type (if a mutation is absent in the BRCA1 and/or BRCA2 gene); and locating for a BRCA mutant subject the mutation to the BRCA1 Really Interesting Gene (RING) protein domain, the BRCA1 DNA-binding domain (DBD) protein domain, the BRCA1 C-terminal domain (BRCT) protein domain, other domains of BRCA1, the BRCA2 RAD51-binding protein, the BRCA2 DBD protein domain or other domains of BRCA2, therewith providing a BRCA domain mutation feature;
- detecting or determining the loss of heterozygosity (LOH) status of a set of N bins of the test genomic DNA sequence, therewith providing a set of N LOH/bin features; wherein the N bins are of equal size smaller than 1.5 Mbp; wherein the N bins were selected from a genome-wide set of bins for their highest proportional contribution or most significant contribution to response to said therapy based on (analysis or assessment of) a set of reference genomic DNA samples including genomic DNA samples obtained from subjects responding to said therapy and including genomic DNA samples obtained from subjects not responding to said therapy; - computer-implemented scoring of the set of N LOH/bin features and BRCA domain mutation feature of a BRCA mutant genomic DNA test sample with a data classification model trained to predict response to said therapy on the set of reference genomic DNA samples (in particular BRCA mutant reference genomic DNA samples) including genomic DNA samples obtained from subjects responding to said therapy and including genomic DNA samples obtained from subjects not responding to said therapy, wherein the training is based on the sets of N LOH/bin features and BRCA domain mutation features of the reference genomic DNA samples (in particular BRCA mutant reference genomic DNA samples); or computer-implemented scoring of the set of N LOH/bin features of a BRCA wild-type genomic DNA test sample with a data classification model trained to predict response to said therapy on the set of reference genomic DNA samples (in particular BRCA wild-type reference genomic DNA samples) including genomic DNA samples obtained from subjects responding to said therapy and including genomic DNA samples obtained from subjects not responding to said therapy, wherein the training is based on the sets of N LOH/bin features of the reference genomic DNA samples (in particular BRCA wildtype reference genomic DNA samples),
- predicting the subject to be a responder to said therapy when the trained data classification model scored the test genomic DNA sample as most closely resembling reference genomic DNA from subjects responding to said therapy; or predicting the subject to be a non-responder to said therapy when the trained data classification model scored the test genomic DNA sample as most closely resembling reference genomic DNA from subjects not responding to said therapy.
In one embodiment, such methods can further be comprising:
- detecting or determining a copy number feature of the test genomic DNA wherein the copy number feature is selected from a gene-specific copy number feature, an ASCAT segment copy number feature and a sigminer copy number feature;
- incorporating the detected or determined copy number feature in the computer-implemented scoring of a BRCA mutant genomic DNA test sample or BRCA wild-type genomic DNA test sample wherein the data classification model is trained including the selected copy number feature as additional feature. More in particular, the gene-specific copy number feature, when incorporated, is selected from amplification, deletion, LOH, focal amplification, focal deletion, focal LOH, and minimal copy number of the gene; the gene is selected from CEP43, ARID1B, PARP3, RBI, ARID1A, NF1, PABPC1, RNF19A, DNMT3A, CCNE1, PCM1, DECR1, BRCA1, TP53BP1, FANCL, EMSY, PALB2, TNKS2, ATR, CHEK1, CDK12, TBL1XR1, TCEA1, and EZH2; the ASCAT segment copy number feature, when incorporated, is selected from fraction LOH, fraction deletions, ploidy, tumor ploidy (psi) of the sample, and goodness-of-fit; and the sigminer copy number feature, when incorporated, is selected from the number of breakpoints per 10 Mbp, the absolute copy numbers of segments, the copy number change points between segments, the lengths of oscillating copy number segment chains, and the number of copy number events per chromosome.
In a further embodiment, said gene-specific copy number feature and ASCAT segment copy number feature are incorporated in the computer-implemented classification of a BRCA mutant genomic DNA test sample or BRCA wild-type genomic DNA test sample, and the data classification model is trained including the gene-specific copy number feature and ASCAT segment copy number feature as additional features.
In one embodiment, such methods can further be comprising:
- detecting or determining the copy number (CN) status of a set of M bins of the test genomic DNA sequence and/or the allelic imbalance (Al) status of a set of P bins of the test genomic DNA sequence, therewith providing a set of M CN/bin and/or P Al/bin features; wherein the bins are of equal size smaller than 1.5 Mbp; wherein the M and/or P bins were selected from a genome-wide set of bins for their highest proportional contribution or most significant contribution of their associated CN/bin and/or Al/bin features to response to said therapy based on a set of reference genomic DNA samples including genomic DNA samples obtained from subjects responding to said therapy and including genomic DNA samples obtained from subjects not responding to said therapy;
- incorporating the detected or determined M CN/bin and/or P Al/bin features in the computer- implemented scoring of a BRCA mutant genomic DNA test sample or BRCA wild-type genomic DNA test sample wherein the data classification model is trained including the M CN/bin and/or P Al/bin features as additional feature(s).
In any of the above methods, the LOH status of a bin is based on the copy number of one or more single nucleotide polymorphisms (SNPs) of interest, and is detected or determined as the mean of the LOH positive SNPs relative to all SNPs of interest within the bin, wherein a SNP is LOH positive when the copy number of the minor allele of the SNP is zero. Alternatively, the LOH status of a bin is based on the copy number of a nucleotide position in the bin, and is detected or determined as the copy number of a segment of the genomic DNA with constant copy number comprising the nucleotide position in the bin, wherein a bin is LOH positive when the copy number of the minor allele of the nucleotide position in the bin is zero. More in particular, the copy number is detected or determined based on untargeted or targeted sequencing of the genomic DNA. More in particular relating to all above methods, the trained data classification model is trained to predict a 2-year progression-free survival response to therapy including a DNA damaging agent or an agent inhibiting or impairing DNA repair.
The disclosure further relates to methods of predicting the response of a subject having cancer to therapy including a DNA damaging agent or including an agent inhibiting or impairing DNA repair, comprising:
- optionally, isolating test genomic DNA from a test genomic DNA sample obtained from the subject;
- detecting or determining a gene-specific copy number feature of a test genomic DNA obtained from the subject (or detecting or determining in the test genomic DNA) wherein the gene is selected from CEP43, NF1, PCM1 and TCEA1/FTIIS or is selected from CEP43, PCM1 and TCEA1/FTIIS;
- predicting the response of a subject having cancer to said therapy based, or based in part, on the detected or determined gene-specific copy number feature.
The disclosure further relates to methods of prognosing survival or determining or assessing survival probability of a subject or patient having cancer or a tumor upon receiving treatment including a DNA damaging agent or including an agent inhibiting or impairing DNA repair, comprising:
- predicting, according to any of the relevant methods described hereinabove, the subject to be a responder to the treatment ;
- prognosing the subject to have, or likely to have an increased survival or increased survival probability if predicted to be a responder to the treatment; or
- prognosing the subject not to have, or likely not to have an increased survival or increased survival probability if predicted to be a non-responder to the treatment.
The disclosure further relates to a DNA damaging agent or therapy comprising a DNA damaging agent, or an agent inhibiting or impairing DNA repair or therapy comprising an agent inhibiting or impairing DNA repair for use in treating a tumor or cancer, comprising:
- predicting, according to a method of any of the relevant methods described hereinabove, the subject to be a responder to the treatment, treatment regimen or therapy;
- administering the treatment, treatment regimen or therapy to the subject.
The disclosure further relates to computer programs having instructions which when executed cause a computing or data processing system or device to perform any of the above-described methods, or to perform a step of any of the above-described methods. The disclosure further relates to computing or data processing systems or devices, machine readable media, or diagnostic kits comprising a means for carrying out or performing any of the above-described methods, or for carrying out or performing a step of any of the above-described methods.
In particular to any of the above-described methods, any use of or therapies with the DNA damaging agents or of the agents inhibiting or impairing DNA repair, any of the computer programs, or any of the computing or data processing systems or devices, machine readable media, or diagnostic kits, the DNA damaging agent or an agent inhibiting or impairing DNA repair is an inhibitor of poly ADP ribose polymerase or is a platinum-based agent.
BRIEF DESCRIPTION OF THE DRAWINGS
FIGURE 1. Distinguishing HRD-positive and HRD-negative samples by means of (A) fractional loss-of- heterozygosity (FLOH, calculated by the ASCAT algorithm), and by means of (B) the scarHRD-score, with indication of AUC and P-values. Analysis was performed on 192 samples of the PAOLA-1 trial for which the Myriad myChoice HRD status ("positive" or "negative") was known.
FIGURE 2. ROC curves of Myriad myChoice positive versus Myriad myChoice negative samples. ROC curves obtained by means of LOH as determined by number of analyzing LOH regions of intermediate size, also known as "HRD-LOH" (A, "median AUC"), by means of analyzing LOH regions of intermediate size and large-scale transition (LST) regions (B, "median AUC"), and by means of analyzing LOH regions of intermediate size, large-scale transition (LST) regions and telomeric allelic imbalance (TAI) regions (C, "median AUC"). In all cases, this was compared with the LOH+LST+TAI analysis via the scarHRD algorithm ("AUC original score"). Analysis was performed on 192 samples of the PAOLA-1 trial (same as for Figure 1) for which the Myriad myChoice HRD status ("positive" or "negative") was known. All combined "median" scores were generated using logistic regression and 10-fold cross-validation.
FIGURE 3. Flowchart of the exemplary sequencing and analysis pipeline. In total, 192 FFPE samples from the PAOLA1 trial with known Myriad myChoice status were processed. GATK: Genome Analysis Toolkit; dedup: deduplicate.
FIGURE 4. ROC curves of Myriad myChoice positive versus Myriad myChoice negative samples. HRD score: as determined with scarHRD algorithm; PAM (Partitioning Around Medoids algorithm) score: classification based on mean LOH/bin, on mean LOH/chromosome and on mean genome-wide LOH features. Set of 82 Myriad myChoice HRD-negative and 110 Myriad myChoice HRD-positive samples.
FIGURE 5. Survival curve analysis of olaparib arm (A) and placebo arm (B). Survival probabilities were calculated with respect to classification using the Myriad myChoice HRD status or the PAM derived score (see Figure 4) based on mean LOH/bin, on mean LOH/chromosome and on mean genome-wide LOH features. HRD+: homologous recombination deficient or HRD positive; HRD-: homologous recombination proficient or HRD negative.
FIGURE 6. Survival curve analysis of olaparib arm (A) and placebo arm (B). Survival probabilities were calculated with respect to classification using the Myriad myChoice HRD status or the PAM derived score mean LOH/bin features only. HRD+: homologous recombination deficient or HRD positive; HRD-: homologous recombination proficient or HRD negative. Bin-size: 1 Mbp (1000 kbp).
FIGURE 7. Distinguishing HRD-positive and HRD-negative samples by means of Random Forest (RF) based classification of the LOH/bin features. Analysis was performed on 192 samples of the PAOLA-1 trial for which the Myriad myChoice HRD status ("positive" or "negative") was known.
FIGURE 8. Survival curve analysis of olaparib arm (A) and placebo arm (B). Survival probabilities were calculated with respect to classification using the Myriad myChoice HRD status or the Random Forest based classification of the LOH/bin features. HRD+: homologous recombination deficient or HRD positive; HRD-: homologous recombination proficient or HRD negative; PFS: progression free survival.
FIGURE 9. Survival curve analysis of olaparib arm (A,C) and placebo arm (B,D) as in Figure 6. (A) and (B) bin-size: 0.75 Mbp (750 kbp). (C) and (D) bin-size: 1.25 Mbp (1250 kbp).
FIGURE 10. Survival curve analysis of olaparib arm (A,C,E,G) and placebo arm (B,D,F,H) as in Figure 8. (A) and (B) bin-size: 0.5 Mbp (500 kbp). (C) and (D) bin-size: 0.75 Mbp (750 kbp). (E) and (F) bin-size: 1.25 Mbp (1250 kbp). (G) and (H) bin-size: 2 Mbp (2000 kbp).
FIGURE 11. Survival curve analysis of olaparib arm (A, C) and placebo arm (B, D) as in Figure 6. (A) and (B) random forest classifier model trained on/with all LOH-bins (genome-wide). (C) and (D) random forest classifier trained on/with top-100 most significant LOH-bins.
FIGURE 12. Results of a random forest classifier trained for predicting 2-year progression-free survival (PFS) status based on multiple individual combinations of the indicated features on the training dataset and subsequently applied to the test dataset. Subjects in the training and test datasets/cohorts were not separated based on the BRCA gene mutation status. (A) provides for the % HRD prediction of the model. (B) provides the hazard ratios (HR) with confidence intervals of olaparib-treated subjects vs the placebo group. (C) provides the hazard ratios (HR) with confidence intervals (indicated as horizontal lines per combination of features) of HRD + subjects vs HRD- subjects. Vertical lines in (B) indicate HR status as defined by Myriad status. Vertical lines in (C) indicate expected mean HR values for olaparib-treated subjects and the placebo group.
FIGURE 13. Results of a random forest classifier trained for predicting 2-year progression-free survival (PFS) status based on multiple individual combinations of the indicated features on the training dataset and subsequently applied to the test dataset. Subjects in the training and test datasets/cohorts were separated based on the BRCA gene mutation status, and the BRCA domain mutation feature was included in the classifier training only for BRCA mutant subjects. Provided are the hazard ratios (HR) with confidence intervals of olaparib-treated subjects vs the placebo group (left) and the hazard ratios (HR) with confidence intervals (indicated as horizontal lines per combination of features) of HRD + subjects vs HRD- subjects (right). Vertical lines indicate HR status as defined by Myriad status (left) or expected mean HR values for olaparib-treated subjects and the placebo group (right).
FIGURE 14. Results of a random forest classifier trained for predicting 2-year progression-free survival (PFS) status based on multiple individual combinations of the indicated features applied to the test dataset. See brief description of Figure 13 for further explanation.
FIGURE 15. Design of the training and validation cohorts.
FIGURE 16. Kaplan-Meier Estimates of Progression-Free Survival (PFS), according to the PARPi response test and Myriad myChoice CDx PLUS test in the validation cohort (n=198). Progression-free survival (PFS) was defined as time from randomisation until disease progression or death. Samples were classified as positive or negative according to the PARPi response test as developed and described in the current disclosure. HRD status positive was defined as GIS >42 and/or presence of tumor BRCAm with Myriad myChoice HRD PLUS assay. HRD status negative was defined as GIS <42 in the absence of a tumour BRCAm with Myriad test.
(A) Overview of PFS in function of time with positive vs negative status in olaparib and placebo subgroups according to the PARPI response test. HRD status positive tumours include BRCAm based on the pathogenic variants in the functional BRCA domains.
(B) Overview of PFS in function of time with HRD positive vs negative status in olaparib and placebo subgroups according to Myriad test. HRD status positive tumours including BRCAm according to Myriad test.
(C) PFS in function of time with the PARPi response test positive BRCAwt tumours.
(D) PFS in function of time with HRD status positive BRCAwt tumours according to Myriad test.
(E) PFS in function of time with the PARPi response test positive BRCAm tumours.
(F) PFS in function of time with HRD status positive BRCAm tumours according to Myriad test.
DETAILED DESCRIPTION
Existing homologous recombination (repair) deficiency (HRD) tests assess either the cause (e.g. screening homologous recombination repair (HRR) gene panel for presence of mutations), the result of HRD (i.e., genomic scarring), or some combination of both. Whereas HRD testing has been linked with predicting outcome of some cancer therapies, such predictive results are still not optimal and in general, mutations in most HRR genes, although informative about a patient's general status, have not yet solidly been assigned such predictive power (Miller et al. 2020, Annals Oncol 31:1606-1622). Most assays to identify HRD are relying on multiple different features (e.g. LOH, LST and TAI), and those including a %LOH feature (e.g. W02011106541 refers to determination of % LOH and a sample was defined as having high LOH when >35% of the genome in cells of the sample has LOH) were reported to underperform (see Introductory section hereinabove).
An objective of earlier work was to design a relatively simple assay to determine the LOH status of the DNA in a tumor as a proxy for determining the HRD status of that tumor. A further objective, one perceived as difficult, if not impossible to reach in view of all previous work done in the field, was that such assay should nevertheless perform at least equally well in determining HRD status as the current gold standard assay implying LOH, LST, and TAI features. In particular, prior methods for determining LOH-status of the DNA in a tumor sample (i) only included regions of LOH of a size larger than 1.5 Mb (as shorter regions of LOH were reported not to correlate with HR-deficiency) and more in particular LOH- regions having a size of larger than 1.5 Mb and shorter than a full chromosome (e.g. W02011160063A2, WO2013096843A1), although as supported by published data and in practice the low-end size threshold more likely is 15 Mb (Abkevich et al. 2012, Br J Cancer 107:1776-1782; WO2013096843A1); or (ii) included calculation of fractional or percentage LOH in the DNA of a tumor sample, reported to underperform (e.g. Timms et al. 2020, J Clin Oncol 38(15 suppl) : 1586). It was surprisingly found that a trained machine learning algorithm could reliably determine LOH or HRD status of the DNA of a tumor based on the LOH status of genomic DNA bins of equal size. Two different machine learning algorithms led to equally reliable LOH or HRD status determination, both further equally reliable as the current gold standard assay implying LOH, LST, and TAI features. Moreover further surprising, the bin size was smaller than 1.5 Mb(p), as equal results were obtained with exemplary bin sizes of 1.25 Mbp (1250 kbp), 1 Mbp (1000 kbp), 0.75 Mbp (750 kbp) and 0.5 Mbp (500 kbp), this when relying on the whole genome as source of the bins (see WO2024/083971).
In a further development it was furthermore surprisingly found that not the whole genome is required as source of bins to arrive at equally reliable determination of LOH or HRD status of the DNA of a tumor. Instead, a set of the most relevant bins (relevant in terms of determining LOH or HRD status) was shown to be sufficient. Advantages of reducing the number of bins to be analyzed include simplification of translation of the methodology into a diagnostic kit, and rendering it more feasible for the methodology to be applicable to liquid biopsies in addition to invasive tumor biopsies.
Yet further developments, building further on the above, lead to the current disclosure. In particular, the LOH feature as detected or determined on genomic DNA sequences of 2Mbp or shorter, or of 1.5 Mbp or shorter was combined with other features of the test genomic DNA (as obtained from a subject having cancer or a tumor) and such combinations were in particular found to improve prediction of response (of the subject having the cancer or tumor) to inhibitors of the poly ADP ribose polymerase (PARP inhibitors) or to platinum-based agents, details of which are incorporated further below as well as in Example 8. This disclosure therefore relates to different aspects and embodiments.
In one aspect are covered methods of predicting the response of a subject having cancer to therapy including a DNA damaging agent or an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinumbased agent), comprising:
- optionally, isolating test genomic DNA from a test genomic DNA sample obtained from the subject;
- detecting, determining, assessing or analyzing in a test genomic DNA obtained from the subject (or detecting, determining, assessing or analyzing in the test genomic DNA) the presence of mutations in the BRCA1 and BRCA2 genes, and classifying the subject as BRCA mutant if a mutation is present in the BRCA1 and/or BRCA2 gene, or as BRCA wild-type;
- detecting, determining, assessing or analyzing the loss of heterozygosity (LOH) status of a set of N bins of the test genomic DNA sequence (of a BRCA wild-type or BRCA mutant subject), therewith providing a set of N LOH/bin features; wherein the N bins are of equal size smaller than 1.5 Mbp; wherein the N bins are or were selected (such as by computer-implemented selection) from a genome-wide set of bins for their highest proportional contribution or most significant contribution to (prediction of) response to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent) as detected or determined based on a set of reference genomic DNA samples including genomic DNA samples obtained from subjects responding to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent) and including genomic DNA samples obtained from subjects not responding to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinumbased agent);
- detecting, determining, assessing or analyzing in the test genomic DNA of a BRCA mutant subject the presence of a mutation in the BRCA1 Really Interesting Gene (RING) protein domain, the BRCA1 DNA- binding domain (DBD) protein domain, the BRCA1 C-terminal domain (BRCT) protein domain, other domains of BRCA1, the BRCA2 RAD51-binding protein, the BRCA2 DBD protein domain or other domains of BRCA2, the location of the mutation providing a BRCA domain mutation feature;
- computer-implemented classification, inferencing or scoring of a BRCA mutant genomic DNA test sample based on its set of N LOH/bin features and based on its BRCA domain mutation feature with a data classification algorithm or model trained to predict response to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent) on the set of reference genomic DNA samples (in particular a set of BRCA mutant reference genomic DNA samples) including genomic DNA samples obtained from subjects responding to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent) and including genomic DNA samples obtained from subjects not responding to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum- based agent), wherein the training is based on the sets of N LOH/bin features and BRCA domain mutation features of the reference genomic DN A samples (in particular the set of BRCA mutant reference genomic DNA samples) (resulting in a trained data classification algorithm or model); or, alternatively, inferencing or scoring of the set of N LOH/bin features and the BRCA domain mutation feature of a BRCA mutant test genomic DNA relative to/with the trained data classification algorithm or model; or computer-implemented classification, inferencing or scoring of a BRCA wild-type genomic DNA test sample based on its set of N LOH/bin features with a data classification algorithm or model trained to predict response to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent) on the set of reference genomic DNA samples (in particular a set of BRCA wild-type reference genomic DNA samples) including genomic DNA samples obtained from subjects responding to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent) and including genomic DNA samples obtained from subjects not responding to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent), wherein the training is based on the sets of N LOH/bin features of the reference genomic DNA samples (in particular a set of BRCA wild-type reference genomic DNA samples); or, alternatively, inferencing or scoring of the set of N LOH/bin features of a BRCA wild-type test genomic DNA relative to/with the trained data classification algorithm or model;
- predicting the subject to be a responder to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent) when the trained data classification algorithm or model classified, inferenced or scored the test genomic DNA sample as most closely resembling reference genomic DNA from subjects responding to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent); or predicting the subject to be a non-responder to therapy including a PARP inhibitor or a platinumbased agent when the trained data classification algorithm or model classified, inferenced or scored the test genomic DNA sample as most closely resembling reference genomic DNA from subjects not responding to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent).
Alternatively phrased are covered methods of predicting the response of a subject having cancer to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent), comprising:
- optionally, isolating test genomic DNA from a test genomic DNA sample obtained from the subject;
- detecting, determining, assessing or analyzing in a test genomic DNA obtained from the subject (or detecting, determining, assessing or analyzing in the test genomic DNA) the presence of mutations in the BRCA1 and BRCA2 genes, and classifying the subject as BRCA mutant if a mutation is present in the BRCA1 and/or BRCA2 gene, or as BRCA wild-type; and locating for a BRCA mutant subject the mutation to the BRCA1 Really Interesting Gene (RING) protein domain, the BRCA1 DNA-binding domain (DBD) protein domain, the BRCA1 C-terminal domain (BRCT) protein domain, other domains of BRCA1, the BRCA2 RAD51-binding protein, the BRCA2 DBD protein domain or other domains of BRCA2, therewith providing a BRCA domain mutation feature (for a BRCA mutant subject);
- detecting, determining, assessing or analyzing the loss of heterozygosity (LOH) status of a set of N bins of the test genomic DNA sequence (of a BRCA wild-type or BRCA mutant subject), therewith providing a set of N LOH/bin features; wherein the N bins are of equal size smaller than 1.5 Mbp; wherein the N bins are or were selected (such as by computer-implemented selection) from a genome-wide set of bins for their highest proportional contribution or most significant contribution to (prediction of) response to said therapy as detected or determined based on a set of reference genomic DNA samples including genomic DNA samples obtained from subjects responding to said therapy and including genomic DNA samples obtained from subjects not responding to said therapy;
- computer-implemented classification, inferencing or scoring of the set of N LOH/bin features and the BRCA domain mutation feature of a BRCA mutant genomic DNA test sample with a data classification algorithm or model trained to predict response to said therapy on the set of reference genomic DNA samples (in particular a set of BRCA mutant reference genomic DNA samples) including genomic DNA samples obtained from subjects responding to said therapy and including genomic DNA samples obtained from subjects not responding to said therapy, wherein the training is based on the sets of N LOH/bin features and BRCA domain mutation features of the reference genomic DNA samples; or computer-implemented classification, inferencing or scoring of the set of N LOH/bin features of a BRCA wild-type genomic DNA test sample with a data classification algorithm or model trained to predict response to said therapy on the set of reference genomic DNA samples (in particular a set of BRCA wildtype reference genomic DNA samples) including genomic DNA samples obtained from subjects responding to said therapy and including genomic DNA samples obtained from subjects not responding to said therapy, wherein the training is based on the sets of N LOH/bin features of the reference genomic DNA samples,
- predicting the subject to be a responder to said therapy when the trained data classification algorithm or model classified, inferenced or scored the test genomic DNA sample as most closely resembling reference genomic DNA from subjects responding to said therapy; or predicting the subject to be a non-responder to said therapy when the trained data classification algorithm or model classified, inferenced or scored the test genomic DNA sample as most closely resembling reference genomic DNA from subjects not responding to said therapy. In this alternative phrasing "said therapy" thus refers to a therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent
In particular to this aspect, mutations in the BRCA genes (BRCA1 and/or BRCA2) or in the recited domains of the BRCA genes (see Example 8 for more details) are relative to reference wild-type BRCA gene sequences. More in particular such reference sequences are e.g. the reference sequence of the BRCA1 transcript with GenBank accession no. NM_007294.3 and the reference sequence of the BRCA2 transcript with GenBank accession no. NM_000059.3, both on Human Genome build hgl9. More in particular, relating to the BRCA domain mutations, these mutations are mutations in the DNA sequences corresponding to the protein domains. For the sake of clarity, the BRCA domain mutation feature referred to herein is a feature different from, or independent of the BRCA gene status (mutant or wildtype) itself, as is obviated in Example 8, and see further. The BRCA gene status is a criterion to decide how the further analysis of the (test) genomic DNA is conducted.
Loss of heterozygosity (LOH) can be defined as the absence of one parent's contribution to the genetic material at a specific site in the genome. As such, a locus is LOH positive, or has LOH positive status, when one of the parental copies of the locus is missing, i.e. when the paternal or maternal copy of the locus is missing. A locus can be a SNP, or can comprise a SNP, or can be a genomic region comprising or covering a SNP, such as a bin referred to in the above methods. In particular, the locus or SNP is a locus or SNP of interest.
In one embodiment relating to the above method, LOH positive status is assigned to a SNP or locus in general when the minor allele copy number of the SNP or locus is zero - wherein LOH can be copy number neutral LOH (minor allele copy number = 0; major copy number allele = 2), copy number gain LOH (minor allele copy number = 0; major copy number allele > 2), or copy number deletion LOH (minor allele copy number = 0; major copy number allele < 2). With mean or average LOH as referred to herein is meant the arithmetic or mathematical mean or average of the number of SNPs or loci assessed in a defined region as having LOH positive status divided by the total number of SNPs or loci assessed in that same defined region. In the two extremes, none of the SNPs or loci are LOH positive and the mean or average LOH equals zero (0), or all of the SNPs of loci are LOH positive and the mean or average LOH equals one (1). In any of the above, "LOH positive" can be exchanged for e.g. having LOH status, or having LOH positive status; and "LOH negative" can be exchanged for e.g. not having LOH status, or having LOH negative status.
An advantage of reducing the size of the regions of interest (e.g. bins) is that the copy number of such region of interest (e.g. bins) can be reliably detected or determined based on the copy number as detected or determined for the segments of genomic DNA with constant copy number: the smaller the size of the region of interest (e.g. bin), the larger the chance that there will be no switch of copy number within the region of interest (and the larger the chance LOH of e.g. a short genomic DNA region, such as a bin as defined herein, will be either 0 or 1). As such, the copy number of a small enough region of interest (e.g. bin) will with very few exceptions be identical to, be equal to, or be the same as the copy number of a segment of genomic DNA with constant copy number. This then logically extrapolates to all SNPs, loci or nucleotides (nucleotide positions) of interest present within the region of interest (e.g. bin). Deriving copy number of an individual SNP, locus or nucleotide (position) of interest from a segment of genomic DNA with constant copy number comprising the individual SNP, locus or nucleotide (position) of interest may additionally reduce possible noise-induced errors occurring by individual detection or determination of the copy number of the SNP, locus or nucleotide (position) of interest.
In one embodiment relating to the above methods, LOH as referred to herein is detected or determined based on log ratios (logR) and B allele frequencies (BAFs). LogR refers to log-transformed copy numbers derived from sequencing depth or SNP array data; BAF to the allelic imbalance of SNPs.
In one embodiment to the above method, the LOH status of a bin is based on the copy number of one or more single nucleotide polymorphisms (SNPs) of interest (and located or present in the bin) and is detected or determined as the mean or average of the LOH positive SNPs relative to all SNPs of interest within the bin, wherein a SNP is LOH positive or has a LOH positive status when the copy number of the minor allele (of the SNP) is zero.
In one embodiment thereto, the (minor allele) copy number of a SNP of interest equals or is the (same as the) copy number of a segment of the (test) genomic DNA with constant (minor allele) copy number comprising the SNP of interest.
In a further or further additional embodiment to the above methods, the LOH status of a bin is based on the (minor allele) copy number of a nucleotide position (of interest; or locus of interest) in the bin, and is detected or determined as being equal to or as being the (same as the) (minor allele) copy number of a segment of the (test) genomic DNA with constant (minor allele) copy number comprising the nucleotide position (of interest) in the bin and wherein the bin is LOH positive or has a LOH positive status when the copy number of the minor allele of the nucleotide position (of interest) in the bin is zero. In particular, the nucleotide position (of interest) can be any nucleotide position in the bin (i.e. not limited to a SNP of interest). In particular, the nucleotide position (of interest) is a reference position. Furthermore in particular, the reference position in or within each bin is the same for each bin. Furthermore in particular, the reference position in or within the bin can be the center position of the bin, or can be a position within a stretch of nucleotides of or spanning 5% of the bin size on either side from the center position (e.g. if the bin size is 1 Mb, then the reference position could be within 50kb from either side of the 500 kb midpoint of the bin). In a further or further additional embodiment to the above methods, segments of the (test) genomic DNA with constant copy number are determined or delineated based on the copy number of a plurality of SNPs or nucleotides of interest (at least 1 SNP/nucleotide of interest in or within the segment and at least 1 SNP/nucleotide of interest outside the segment). In a particular or particular further embodiment thereto, the plurality of SNPs or nucleotides of interest is distributed over or across the (test) genomic DNA present in the sample. In a further limiting embodiment thereto, the plurality of SNPs or nucleotides of interest is distributed over or across the (test) genomic DNA excluding centromeric regions. In a further particular embodiment thereto, such segment of genomic DNA with constant copy number is detected or determined by the ASCAT algorithm or the asmultipcf algorithm (Van Loo et al. 2010, Proc Natl Acad Sci USA 107:16910-16915 or Ross et al. 2021, Bioinformatics 37:1909-1911, respectively). Further in particular, the plurality of SNPs or loci or nucleotides of interest is distributed over or across the (test) genomic DNA corresponding to the N selected bins.
Likewise, and if not covered by the bins, SNPs or loci or nucleotides of interest distributed over or across the (test) genomic DNA corresponding to the BRCA genes or BRCA domains can be utilized in order to assess the BRCA mutation and BRCA domain mutation status of a (test) genomic DNA sample.
Thus, in formulating the above method, the step:
- detecting, determining, assessing or analyzing the LOH status of a set of N bins of the test genomic DNA sequence, therewith providing a set of N LOH/bin features can be alternatively phrased as:
- detecting, determining, assessing or analyzing the LOH status of each of the N bins as the mean or average of the LOH positive SNPs relative to all SNPs within each of the N bins, wherein a SNP is LOH positive or has a LOH positive status when the copy number of the minor allele (of the SNP) is zero, therewith providing a set of LOH/bin features; or can be alternatively phrased as:
- detecting, determining the LOH status of each of the N bins by identifying the copy number of a nucleotide position (of interest) in the bin as, as being, or as being equal to/identical to/ the (same as the) copy number of a segment of the genomic DNA with constant copy number comprising the nucleotide position (of interest) in the bin, and wherein a bin is LOH positive or has a LOH positive status when the copy number of the minor allele of the nucleotide position (of interest) in the bin is zero, therewith providing a set of N LOH/bin features.
In a further alternative, the LOH status of a segment of genomic DNA with constant copy number is detected or determined (such segment being LOH positive or having a LOH positive status when the copy number of the minor allele of the segment is zero), and the LOH status of any nucleotide position of interest in or within a bin (including SNPs, or nucleotides or loci of interest) comprised in the segment is then defined, detected or determined as being the (same as the), identical to, or equal to the LOH status of the segment.
The above method may comprise a further step of detecting, determining or delineating segments of constant copy number in the (test) genomic DNA based on the copy number of at least 1 SNP (or nucleotide or locus of interest) or of a plurality of SNPs (or nucleotides or loci of interest).
As indicated in the above method, the N bins were selected, or pre-selected, from a genome-wide set of bins. The selection was more in particular performed by a computer-run or computer-implemented selection test, more in particular a permutation or re-randomization test, to single out the N bins that contribute most to response to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent) as detected or determined based on a set of reference genomic DNA samples including genomic DNA samples obtained from subjects responding to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent) and including genomic DNA samples obtained from subjects not responding to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent). When analyzing a set of such reference genomic DNA samples, it became obvious that, compared to other bins, certain bins contribute to therapy response at a higher frequency throughout the set of reference genomic DNA samples. As such, a ranking of the genomewide set of bins is constructed from null or low contribution to therapy response to the highest contribution to therapy response. From this ranking, either the N bins with the highest proportional contribution (top N bins) or with the most significant contribution (significant N bins, e.g. based on P- value) can be selected. Herein the applied selection threshold is determining the number N, or phrased alternatively, the number or value of N can be varied by varying the applied threshold. In particular, the numeric or absolute value of N is minimally 1 (/V=l) whereas the maximum numeric or absolute value of N may depend on the size of the bin. The maximum numeric or absolute value of N can e.g. be defined such that the N bins are covering less than 100% of the full genome, e.g. 75% (or approximately 75%), 60% (or approximately 60%), 50% (or approximately 50%), 40% (or approximately 40%), 30% (or approximately 30%), 20% (or approximately 20%), 15% (or approximately 15%), 10% (or approximately 10%), 5% (or approximately 5%), 4% (or approximately 4%), 3% (or approximately 3%), 2% (or approximately 2%), or 1% (or approximately 1%) of the full genome. For example, for a bin size of 1 Mbp, 2877 bins were delineated in the full genome (see Example 1). For a 20% coverage of the genome, this would amount to N = 575 bins, for a 10% coverage N would be approximately 285, for a 5% coverage N would be approximately 145, etc. Alternatively, N is an integer with a value of 1 (or 5 or 10) to 1000, 1 (or 5 or 10) to 900, 1 (or 5 or 10) to 800, 1 (or 5 or 10) to 700, 1 (or 5 or 10) to 600, 1 (or 5 or 10) to 500, 1 (or 5 or 10) to 400, 1 (or 5 or 10) to 300, 1 (or 5 or 10) to 200, 1 (or 5 or 10) to 100, or 1 (or 5 or 10) to 50; or with a value of 1, 5, 10, 20, 30, 40 ,50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 325, 350, 375, 400, 425, 450, 475, 500, 525, 550, 575, 600, 650, 700, 750, 800, 850, 900, 950, or 1000.
The computer-implemented classification, inferencing or scoring of a test genomic DNA/test genomic DNA sample referred to in the above method relies on a trained classifier or classification algorithm or model. More in particular, the classifier is/was trained for predicting response to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent) starting from/based on a reference set of genomic DNA samples (including from subjects responding and not responding to the therapy). While training the classifier or classification algorithm or model, it was observed/determined to be advantageous to first assess or detect or determine the BRCA status of the genomic DNA of a genomic DNA sample. Indeed, categorizing the genomic DNA as BRCA wild-type or BRCA mutant prior to the classification enhanced the classification power of the algorithm or model.
Another factor contributing to this classification power was the addition of a BRCA domain mutation feature to the classifier (logically only for the BRCA mutant test genomic DNA samples). A classifier or classification algorithm or model trained for predicting response to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent), starting from a reference set of genomic DNA samples (including from subjects responding and not responding to the therapy), and using the LOH of the (selected) N bins feature of the reference genomic DNA samples in combination with the BRCA domain mutation feature (in BRCA mutant samples), or using the LOH of the (selected) N bins feature of the reference genomic DNA samples alone (in BRCA wild-type samples) as parameters was a first classifier or classification algorithm or model (or test for predicting response to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent)) determined to perform better than a similar classifier or classification algorithm or model trained on the same reference set of genomic DNA samples but using the standard LOH-HRD + TAI + LST features, even when applying to these standard features the separation of BRCA wild-type and BRCA mutant samples prior to classification, and even when adding the BRCA domain mutation feature (for BRCA mutant genomic DNA samples).
The above-described training of the computer-implemented classifier or classification algorithm or model was tested further for purposes of further increasing its power of/in predicting response to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent). Features further increasing (although not necessarily extensively increasing) this discriminating power were identified and include some copy number features of the genomic DNA (of the reference set of genomic DNA samples, or of a test genomic DNA sample). More in particular, any of or any combination of a gene-specific copy number feature, an ASCAT segment copy number feature and a sigminer copy number feature were determined to be valuable, and can be optionally added as single feature or as any combination of these features to the basic method of the first aspect as described above. Such combinations include e.g. gene-specific copy number feature and ASCAT segment copy number feature; gene-specific copy number feature and sigminer copy number feature; ASCAT segment copy number feature and sigminer copy number feature; and gene-specific copy number feature, ASCAT segment copy number feature, and sigminer copy number feature.
Thus, in one embodiment to the above methods of predicting the response of a subject having cancer to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent), such methods are further comprising:
- detecting or determining a copy number feature of the (test) genomic DNA wherein the copy number feature is selected from a gene-specific copy number feature, an ASCAT segment copy number feature and a sigminer copy number feature;
- incorporating the detected or determined copy number feature (or combination thereof as outlined above) in the computer-implemented classification, inferencing or scoring of a BRCA mutant genomic DNA test sample or BRCA wild-type genomic DNA test sample wherein the data classification algorithm or model is trained including the selected copy number feature (or combination thereof) as additional feature(s).
Related to the "gene-specific copy number feature", this in particular is related to the copy number of any individual gene of the following genes, or of any combination of the following genes: ABCB1, ARID1A, ARID1B, ATM, ATR, AXIN1, ARD1, BRCA1, BRCA2, BRIP1, CCNE1, CDK12, CEP43, CEP89, CHEK1, CHEK2, DECR1, DNMT3A, ECT2, EMSY, EZH2, FANCL, LEPROTL1, NAALADL2, NCEH1, NF1, NLGN1, PABPC1, PALB2, PARP1, PARP2, PARP3, PARP4, PCM1, PDCD10, PLD1, POLQ, PTEN, RAD51B, RAD51C, RAD51D, RAD54L, RBI, RNF19A, SERPIN 11, SERPINI2, SLC35G5, SLFN11, SPATA16, TAS2R43, TBL1XR1, TCEA1, TNIK, TNKS, TNKS2, TP53BP1, VPS13B, WDR49, WRN, XRCC5, XRCC6, and ZBBX.
More in particular, such genes are selected from, or are comprising one or more of CEP43, ARID1B, PARP3, RBI, ARID1A, NF1, PABPC1, RNF19A, DNMT3A, CCNE1, PCM1, DECR1, BRCA1, TP53BP1, FANCL, EMSY, PALB2, TNKS2, ATR, CHEK1, CDK12, TBL1XR1, TCEA1, and EZH2.
More in particular, such genes are selected from, or are comprising one or more of CEP43, NF1, PCM1 and TCEA1/FTIIS; or are selected from, or comprising one or more of CEP43, PCM1 and TCEA1/FTIIS.
The copy number feature of the gene-specific copy number feature is in particular selected from amplification, deletion, LOH, focal amplification, focal deletion, focal LOH, and minimal copy number of a gene. Gene-specific copy number related features were detected or determined for a candidate gene based on its overlap with ASCAT segments. Deletion and amplification events are defined based on the total copy number being smaller or larger than 2; LOH events are defined based on the minor allele having copy number 0. Focal events are defined as events smaller than 25% of the respective chromosome arm. When multiple ASCAT segments are overlapping with a gene, the minimum copy number is retained; similarly, (focal) amplification, (focal) deletion or (focal) LOH of a gene is based on at least one overlapping ASCAT segment having this status. More in particular, the copy number feature is selected from amplification, deletion, LOH, and minimal copy number of a selected gene. Furthermore in particular, the gene-specific copy number feature is selected from, or comprising one or more of the genes and associated copy number features as listed in Table 5 (see Example 8).
Similar as for the BRCA genes, if not covered by the bins, e.g. SNPs, loci or nucleotides of interest distributed over or across the (test) genomic DNA corresponding to these genes can be utilized or used as basis for detecting, determining, analyzing or assessing their copy number feature, and the copy number can be analyzed, detected, determined or assessed e.g. by applying the ASCAT algorithm or the asmultipcf algorithm.
As discussed hereinabove (as basis for detecting or determining LOH status of e.g. a SNP), the genomic DNA can be divided in segments of the same copy number, such as by applying e.g. the ASCAT algorithm. The same or similar algorithm can analyse, detect, determine, or assess specific segment copy number features including the number of segments (n_segs), the aberrant cell fraction (rho), the fraction amplified, the fraction LOH, the fraction deletions, the ploidy, the tumor ploidy (psi) of the sample, and the goodness-of-fit parameter. More in particular, the ASCAT segment copy number feature is selected from, or is comprising one or more of the fraction LOH, fraction deletions, ploidy, tumor ploidy (psi) of the sample, and goodness-of-fit parameter as determined by ASCAT or similar tools or algorithms. These features can be assessed, detected, determined or analyzed on a genome-wide scale, or can be based on the information as derived from the set of N selected bins (selected as described herein).
Wang et al. 2021 (PLoS Genet 17(5): el009557) disclosed a copy number signature analysis framework (the "sigminer" or signature miner tool) which lead to 5 copy number (CN) signatures and 3 single base substitution (SBS) signatures based on a set of 80 predefined copy number components. The "sigminer copy number feature" as referred to herein relates to a selection from these 80 predefined copy number components that can be determined, analyzed or assessed using the sigminer package in R (cran.r- project.org/package=sigminer), more in particular using the method "Wang" in the "sig_tally" function of the sigminer package. Thus, the sigminer copy number feature is selected from, or is comprising one or more of these 80 predefined copy number components (see Example 8 for more details). More in particular, the sigminer copy number signature is selected from, or comprising one or more of the features number of breakpoints per 10 Mbp (even more in particular with the frequency of this value = 1), absolute copy numbers of segments (even more in particular with the frequency of this value = 1 ,3, or >8), copy number change points between segments (even more in particular with the frequency of this value = 4, or >8), length of oscillating copy number segment chains/length of chains of oscillating copy number (even more in particular with the frequency of this value = 3), and number of copy number events per chromosome (even more in particular with the frequency of this value=19), as determined via the sigminer package or similar tool or algorithm. Furthermore in particular, the sigminer copy number signature is selected from, or comprising one or more of the features as listed in Table 6 (see Example 8). This copy number feature can be assessed, detected, determined or analyzed on a genomewide scale, or can be based on the information as derived from the set of N selected bins.
The above-described training (for predicting the response of a subject having cancer to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinumbased agent)) of the computer-implemented classifier or classification algorithm or model was tested yet further for purposes of further increasing its power of/in predicting response to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent). Other features were identified that can be optionally added to any of the methods described above. The optional further features include the copy number (CN) status and/or the allelic imbalance (Al) status of a set of bins of the genomic DNA (of the reference set of genomic DNA samples, or of a test genomic DNA sample). More in particular, any of or any combination of copy number (CN) status and/or the allelic imbalance (Al) status can be included. More in particular a set of M bins related to CN and/or a set of P bins related to Al status can be selected, or pre-selected, in a way similar as the N bins for LOH status were selected. Obviously, the (absolute) number N can be the same as the (absolute) number M and/or P, but can also be different. Furthermore, the sequence of the bins themselves can be the same, but are likely to be different, or different in part, as the ranking most likely will not be completely the same for the LOH/bin, CN/bin and Al/bin features.
Thus, in one embodiment to the above methods for predicting the response of a subject having cancer to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent), such methods are further comprising:
- detecting or determining the copy number (CN) status of a set of M bins of the test genomic DNA sequence and/or the allelic imbalance (Al) status of a set of P bins of the test genomic DNA sequence, therewith providing a set of M CN/bin and/or P Al/bin features; wherein the M bins and/or P bins are of equal size smaller than 1.5 Mbp; wherein the M bins and/or P bins each (and separately for CN and Al status, respectively) were selected (such as by computer-implemented selection) from a genome-wide set of bins for their highest proportional contribution or most significant contribution of their associated CN/bin and/or Al/bin features (respectively) to (prediction of) response to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent) based on/using a set of reference genomic DNA samples including genomic DNA samples obtained from subjects responding to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent) and including genomic DNA samples obtained from subjects not responding to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent);
- incorporating the detected or determined set of M CN/bin and/or set of P Al/bin features in the computer-implemented classification, inferencing or scoring of a BRCA mutant genomic DNA test sample or BRCA wild-type genomic DNA test sample wherein the data classification algorithm or model is trained including the sets of M CN/bin and/or sets of P Al/bin features (of the reference genomic DNA samples) as additional feature(s).
In particular, the numeric or absolute value of M or P is minimally 1 (/V=l) whereas the maximum numeric or absolute value of M or P may depend on the size of the bin. The maximum numeric or absolute value of M or P can e.g. be defined such that the M or P bins are covering less than 100% of the full genome, e.g. 75% (or approximately 75%), 60% (or approximately 60%), 50% (or approximately 50%), 40% (or approximately 40%), 30% (or approximately 30%), 20% (or approximately 20%), 15% (or approximately 15%), 10% (or approximately 10%), 5% (or approximately 5%), 4% (or approximately 4%), 3% (or approximately 3%), 2% (or approximately 2%), or 1% (or approximately 1%) of the full genome. For example, for a bin size of 1 Mbp, 2877 bins were delineated in the full genome (see Example 1). For a 20% coverage of the genome, this would amount to M or P = 575 bins, for a 10% coverage M or P would be approximately 285, for a 5% coverage M or P would be approximately 145, etc. Alternatively, M or P is an integer with a value of 1 (or 5 or 10) to 1000, 1 (or 5 or 10) to 900, 1 (or 5 or 10) to 800, 1 (or 5 or 10) to 700, 1 (or 5 or 10) to 600, 1 (or 5 or 10) to 500, 1 (or 5 or 10) to 400, 1 (or 5 or 10) to 300, 1 (or 5 or 10) to 200, 1 (or 5 or 10) to 100, or 1 (or 5 or 10) to 50; or with a value of 1, 5, 10, 20, 30, 40 ,50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 325, 350, 375, 400, 425, 450, 475, 500, 525, 550, 575, 600, 650, 700, 750, 800, 850, 900, 950, or 1000.
In one particular embodiment to any of the above methods, the trained data classification algorithm or model is thus in particular trained to predict a 2-year progression-free survival response to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent). In a further embodiment relating to any of the above methods, the set of N LOH/bin (or M CN/bin or P Al/bin) features is maintained during classification or during computer-implemented data classification, meaning that individual features of the set are not combined, merged, added or collapsed into a larger bin. Thus individual features of a set of N LOH/bin (or M CN/bin or P Al/bin) features are not combined, merged, added or collapsed into a bin larger than a single bin but smaller than a full chromosome or smaller than the genome.
In a further embodiment relating to any of the above methods, centromere regions of chromosome(s) are excluded from the set of N LOH/bin (or M CN/bin or P Al/bin) features.
In a further embodiment relating to any of the above methods, sex chromosomes (X-chromosome and/or Y-chromosome) are excluded from the set of N LOH/bin (or M CN/bin or P Al/bin) features.
In another embodiment relating to any of the above methods, the copy number or mutational status of a SNP, bin, locus, nucleotide, gene (or gene domain), or segment of interest is detected or determined based on, or starting from, sequencing of the (test or reference) genomic DNA comprising the SNP, locus, gene (or gene domain), or segment of interest.
In one particular embodiment, the sequencing is targeted (e.g. involving linear, non-linear or PCR amplification and/or capture by hybridization of a genomic DNA region comprising a SNP, bin, locus, gene (or gene domain), or segment of interest) or untargeted (e.g. whole genome, exome, transcriptome) sequencing. Nucleotide sequencing is possible by any known means including classical Sanger sequencing, massive parallel sequencing and (nano)pore sequencing.
In a further particular embodiment, the sequencing is at low coverage, at high coverage, at lx to lOx coverage, at lx to 500x coverage, at lOx coverage, at 20x coverage, at 30x coverage, at 40x coverage, is at 50x to 500x coverage, is at 50x coverage, at 60x coverage, at 70x coverage, at 80x coverage, at 90x coverage, at lOOx coverage, at llOx coverage, at 120x coverage, at 130x coverage, at 140x coverage, at 150x coverage, at 160x coverage, at 170x coverage, at 180x coverage, at 190x coverage, or at 200x coverage; or wherein the sequencing is at 50x to lOOOx coverage, at 300x coverage, at 400x coverage, at 500x coverage, at 600x coverage, at 700x coverage, at 800x coverage, at 900x coverage, or at lOOOx coverage. Alternatively, the sequencing is shallow sequencing, deep sequencing, ultradeep sequencing, or maximum-depth sequencing.
In a further particular embodiment, the sequencing is paired-end sequencing such as 2x50 bp reads to 2x200 bp reads (e.g. 2x50 bp reads, 2x60 bp reads, 2x70 bp reads, 2x80 bp reads, 2x90 bp reads, 2x100 bp reads, 2x110 bp reads, 2x120 bp reads, 2x130 bp reads, 2x140 bp reads, 2x150 bp reads, 2x151 bp reads, 2x160 bp reads, 2x170 bp reads, 2x180 bp reads, 2x190 bp reads, 2x200 bp reads). In particular, the sequenced stretch of the genomic DNA is comprising one or more of the plurality of SNPs, nucleotides, bins, loci, genes (or gene domains), or segments of interest.
In a further particular embodiment, the sequencing reads are aligned to a reference genome (such as the human hgl8, hgl9, NCBI/hgl8, GRCh37/hgl9, or GRCh38.pl4 reference genome).
In a further particular embodiment, the copy number of a SNP, bin, locus, nucleotide, gene (or gene domain), or segment of interest is detected or determined based on, or starting from, minimum quality sequencing data (e.g. minimum base quality and/or minimum mapping quality).
In a further particular embodiment, detection or determination of the copy number of a SNP, bin, locus, nucleotide, gene (or gene domain), or segment of interest is involving an algorithm or bio-informatic algorithm such as run or executed on a computer. Such algorithm can e.g. be based on a hidden Markov algorithm or on circular binary segmentation.
In a further particular embodiment, detection or determination of the copy number of a SNP, bin, locus, nucleotide, gene (or gene domain), or segment of interest is involving an analytical method compensating for contamination present in the DNA sample (e.g. if the DNA sample is obtained from a tumor sample, DNA from non-tumor cells is a "contaminant" expected to be present in the DNA sample). Numerous methods and kits are available enabling extraction of (genomic) DNA from cells or tissues, or from liquid samples (e.g. cfDNA).
In a further particular embodiment to any of the above methods, the SNP, bin, locus, nucleotide, gene (or gene domain), or segment of interest, or the genomic DNA comprising a SNP, bin, locus, nucleotide, gene (or gene domain), or segment of interest is captured from the genomic DNA sample by means of (hybridization to) an oligonucleotide probe. Alternatively, a plurality of SNPs, bins, loci, nucleotides, genes (or gene domains), or segments of interest, or a plurality of SNPs, bins, loci, nucleotides, genes (or gene domains), or segments of interest-comprising genomic DNA molecules, is captured by or captured by means of (hybridization to) a library of oligonucleotide probes (capture library / nucleic acid capture), or captured by or captured by means of (hybridization to) a plurality of oligonucleotide probes. This process is also known as hybridization capture or target enrichment. The process may include prior shearing, such as mechanical shearing, of the input material (e.g. DNA extracted from formalin fixed paraffin embedded (FFPE) tumor sample). The genomic DNA sample from a tumor or cancer cell may in principle be any type of biological sample comprising tumor genomic DNA including fresh or processed (e.g. FFPE, fresh frozen) tumor biopsy samples, circulating tumor DNA (ctDNA, isolated from e.g. cfDNA), genomic DNA from circulating tumor cells etc. In the methods described hereinabove, the number of SNPs, bins, loci, nucleotides, genes (or gene domains), or segments to be analyzed can be limited as genome-wide coverage is not required. This renders these methods applicable to analysis or assessment of smaller amounts of input genomic DNA, thus in particular also applicable to analysis or assessment of ctDNA or cfDNA. The tumor or cancer can be of primary origin or of metastatic origin. The tumor or cancer may in particular be a tumor or cancer known to be prone or susceptible to (genetic scarring by means of) any defect or impairment in the homologous recombination repair pathway; or known to be prone or susceptible to losing homologous recombination repair proficiency (or HRD negative status); or known to be prone or susceptible to become homologous recombination repair deficient (or HRD positive). The tumor or cancer may in particular be any of e.g. ovarian, breast, colon, prostate, pancreatic, lung, renal or esophagal origin. A plurality of SNPs, loci, nucleotides, genes (or gene domains), or segments is meant to include more than 1 SNP, locus, nucleotide, gene (or gene domain), or segment, thus 2 or more SNPs, loci, nucleotides, genes (or gene domains), or segments. In particular, between 2 and 200000 (200k) SNPs, loci, nucleotides, genes (or gene domains), or segments are envisaged, or between 2 and 150k, between 2 and 100k , between 2 and 75k, between 2 and 50k, or between 2 and 25k SNPs, loci, nucleotides, genes (or gene domains), or segments; or between 10k and 200k SNPs, loci, nucleotides, genes (or gene domains), or segments, or between 10k and 150k, or between 10k and 100k, or between 10k and 75k, or between 10k and 50k, or between 10k and 25k SNPs, loci, nucleotides, genes (or gene domains), or segments.
In a particular embodiment thereto, such oligonucleotide probe(s) are attached to a solid support such as a sheet, bead, or plate well (of any suitable size or dimension). In a particular embodiment thereto, a spacer is introduced between the solid support and the oligonucleotide probe(s) to support efficient capture of the DNA of interest from the sample such as biological sample.
In a particular embodiment thereto, the individual oligonucleotide probes can have a length of 50 to 150bp, or an average length of 50 bp, 60 bp, 70bp, 80 bp, 90 bp, of 100 bp, of 110 bp, of 120 bp, of 130 bp, of 140 bp, or of 150 bp.
In a particular embodiment thereto, the capture library is optionally comprising or further comprising one or a set of gene-specific oligonucleotide probes, or one or a set of dedicated gene-specific oligonucleotide probes. More in particular, such gene-specific oligonucleotides are targeting genes or gene domains of a gene or combination of genes as specifically outlined hereinabove.
In a further particular embodiment, the capture oligonucleotides comprise molecular barcodes, molecular indexes, or unique molecular identifiers (UMIs).
In a further particular embodiment to any of the above methods, the size of the size-defined bins is equal to or lower than/smaller than 2000 kb(p) (2 Mb(p)), or equal to or lower than/smaller than 1500 kb(p) (1.5 Mb(p)), e.g. about 50 kb(p) to 2000 kb(p), about 50 kb(p) to 1500 kb(p), e.g. about 50 kb(p) to 1400 kb(p), e.g. approximately 50 kb(p), e.g. approximately 100 kb(p), e.g. approximately 200 kb(p), e.g. approximately 300 kb(p), e.g. approximately 400 kb(p), e.g. approximately 500 kb(p), approximately 550 kb(p), approximately 600 kb(p), approximately 650 kb(p), approximately 700 kb(p), approximately 750 kb(p), approximately 800 kb(p), approximately 850 kb(p), approximately 900 kb(p), approximately 950 kb(p), approximately 1000 kb(p), approximately 1050 kb(p), approximately 1100 kb(p), approximately 1150 kb(p), approximately 1200 kb(p), approximately 1250 kb(p), approximately 1300 kb(p), approximately 1350 kb(p), approximately 1400 kb(p), approximately 1450 kb(p); and wherein the size of one or more bins may deviate from the size of the majority of the bins depending on chromosome size (e.g. left-over sequence smaller than chosen bin-size after dividing a chromosome or chromosomearm in bins of the chosen size). A size-defined bin can cover either no SNP, locus, nucleotide, gene (or gene domain), or segment of interest, one SNP, locus, nucleotide, gene (or gene domain), or segment of interest, two SNPs, loci, nucleotides, genes (or gene domains), or segments of interest, or three or more SNPs, loci, nucleotides, genes (or gene domains), or segments of interest. The reference to the bin size can exchangeably be expressed as number of bases (b) or as number of basepairs (bp) depending on when referring to single- or double-stranded nucleic acids.
The terms "computer-implemented data classification model/algorithm analysis" or "computer- implemented classification with a data classification model/algorithm" are used herein interchangeably and refer to use of a machine-learning based data classification model or algorithm or to use of a machine-learning model or algorithm for data classification. In particular in the current context, predicting response to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent)of a subject having cancer or tumor based on the described features of a test genomic DNA sample obtained from the subject can involve an algorithm or bio-informatic algorithm such as run or executed on a computer; in particular the algorithm is a machine learning algorithm or model, more in particular the algorithm is a data classification algorithm or data classifier algorithm. Many data classification algorithms exist and include Random Forest (RF), Support Vector Machines (SVM), K-Nearest Neighbors (KNN), and Linear Discriminant Analysis (LDA) (as explained in e.g. Chen et al. 2020, J Big Data 7, 52), and further algorithms involving regression or logistic regression analysis, gradient boosting based analysis, analyses based on multi layer perceptron (MLP), decision trees (DT), adaboost (Ada) (Turgut et al. 2018 doi: 10.1109/EBBT.2018.8391468), (nearest) shrunken centroid classifier algorithms, etc.
Prior to detecting or determining the status of interest of a test item (alternatively: inferencing or scoring a test item) with the help of a computer-implemented data classification algorithm or model, the machine learning model or algorithm must be trained with a set of training items of which the status of interest is known, or must be trained with a set of reference items. In particular in the current context, a set of reference genomic DNA samples was used. More in particular, a set of reference genomic DNA samples of an exemplary clinical trial study was used, wherein the exemplary clinical trial studied the effect of adding the PARP inhibitor olaparib to bevacizumab therapy of ovarian cancer, the beneficial effect of such addition being more pronounced/substantial in patients with a HRD-positive tumor. More in particular, the clinical trial is the PAOLA-1 trial (Ray-Coquard et al. 2019, N Engl J Med 381: 2416-2428; ClinicalTrials.gov identifier NCT02477644). More in particular, the data classification model or algorithm is trained for classifying a reference set of genomic DNA samples (including genomic DNA samples obtained from subjects responding to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent) and including genomic DNA samples obtained from subjects not responding to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent) as predicted responder or predicted non-responders to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent) based on/using the above described combination of input features. More in particular, such trained data classification model or algorithm is capable of predicting response of a subject having cancer to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent) by classifying, inferencing or scoring test genomic DNA or a test genomic DNA sample obtained from the subject based on/using the same combination of input features as applied during the training of the data classification model or algorithm. Such prediction is based on classifying, inferencing or scoring, by the trained data classification model or algorithm, of a test genomic DNA (sample) as most closely resembling reference genomic DNA of responders to therapy, or as most closely resembling reference genomic DNA of non- responders to therapy. It is moreover a plausible assumption that such trained data classification model or algorithm trained for prediction of response to therapy with one member of a class of therapeutic agents will likewise be able to predict response to therapy with another member of the same class of therapeutic agents. For example, a data classification model or algorithm trained for prediction of response to therapy with an agent inhibiting DNA repair (e.g. an inhibitor of PARP) will plausibly be able to predict response to therapy with a DNA damaging agent (e.g. a platinum-based or platinum-containing agent) as both agents provoke the same effect i.e. DNA damage. In addition, the standard test relying on LOH-HRD, TAI and LST was originally described as applicable to DNA damaging agents including platinum-based agents, anthracyclines, topoisomerase I inhibitors and PARP inhibitors (see WO2014/165785).
In a particular embodiment to any of the above methods, a reference genomic DNA sample is assigned as obtained from a subject having cancer responding to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent), or is assigned as obtained from a subject having cancer responding to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent), based on e.g. observation of clinical outcome of the therapy such as e.g. the subject reaching, or not reaching, a defined period of e.g. progression-free survival when on the therapy. Clinical outcomes of a therapy, or alternatively efficacy of a therapy, are discussed in some more detail hereinafter.
In a particular embodiment to any of the above methods, a reference genomic DNA sample is assigned as being LOH or HRD positive, or as being LOH or HRD negative, based on a method, kit or assay known or available in the art. In particular LOH or HRD status of a reference genomic DNA sample can have been detected or determined by means of a HRD-LOH score. It can alternatively have been detected or determined by means of score based on and including HRD-LOH status, large scale transition (LST) status and telomeric allelic imbalance (TAI) status (such as in the Myriad®myChoice® test), combined leading to the detection or determination of a HRD positive (homologous recombination repair deficient) status or a HRD negative (homologous recombination repair proficient) status of a DNA sample. The HRD-LOH score in particular is based on the number of LOH regions of intermediate size (longer than 15 Mb but shorter than the whole chromosome: Abkevich et al. 2012, Br J Cancer 107:1776-1782) occurring in the DNA of a tumor sample. Large-scale state transitions (LST) are chromosomal breaks between adjacent genomic regions longer than 10 Mb that are increased in BRCA1/2 mutant cancers (Popova et al., Cancer Res 72:5454-5462). Telomeric-allelic imbalance (TAI) is allelic imbalance extending to the subtelomere but not crossing the centromere and is enriched in cancers with BRCA1/2 deficiency and response to platinum chemotherapy (Birkbak, et al. 2012, Cancer Discov 2:366-375). As confirmed in the Examples herein, fractional LOH or percentage LOH are parameters that are less reliable in detecting or determining LOH or HRD status of a (test or) reference genomic DNA sample; this is in line with previous observations reported by Timms et al. 2020 (J Clin Oncol 38(15 suppl):1586) such that 19%-61% of patients identified as positive by myChoice® HRD would have been missed by %LOH. Based on a set of LOH and/or HRD positive and -negative reference genomic DNA samples, an LOH positive and -negative reference genomic DNA profile can be made/is made by means of classification (such as by means of machine learning assisted classification) as is subject of this disclosure. A test genomic DNA sample in general will have (or will be inferenced or scored as having) an LOH status profile more or most closely resembling the LOH positive and/or HRD positive reference genomic DNA profile or more or most closely resembling the LOH negative and/or HRD negative reference genomic DNA profile and can that way be classified as having either LOH positive status or LOH negative status. An alternative outcome may be that a (test) genomic DNA cannot be classified as LOH positive or negative or as having a LOH positive or negative status in case the (test) genomic DNA does not closely resemble LOH positive reference genomic DNA and does not closely resemble LOH negative reference genomic DNA. In such cases, the LOH status of the (test) genomic DNA is assigned as unknown, uncertain, not defined, undefined, or not determinable. It has been shown before that the overall LOH status of tumor-associated DNA is a footprint linked to defects in a cell's homologous recombination repair machinery, such as defects in the homologous recombination repair genes BRCA1 and/or BRCA2. In particular, one of the initial HRD scores was build based on the number of LOH regions of intermediate size (longer than 15 Mb but shorter than the whole chromosome: Abkevich et al. 2012, Br J Cancer 107:1776-1782) occurring in the DNA of a tumor sample, and is sometimes referred to as the "HRD-LOH" score. It is thus accepted in the field that detecting or determining LOH status is a proxy or alternative for detecting or determining HRD-status.
The presence of HRD in a tumor cell furthermore is at the basis of methods identifying cancer patients that are expected to respond well to therapies (further) inhibiting or impairing DNA repair or otherwise damaging DNA, therewith exacerbating the defects in HRD leading to LOH, such as with inhibitors of poly ADP ribose polymerase (PARP)(e.g. olaparib, rucaparib, niraparib, talazoparib, veliparib, pamiparib), of type I topoisomerases (e.g. camptothecins such as topotecan, irinotecan and belotecan; and non- camptothecins such as indenoisoquinoline, phenthridines and indolocarbazoles), of type II topoisomerases (e.g. doxorubicin, daunorubicin, and other anthracycline antibiotics (also acting as DNA damaging agent by intercalating in DNA)), of dihydrofolate reductase (DHFR; e.g. methotrexate) or with DNA damaging agents or treatments. DNA damaging treatments include ionizing radiation, irradiation, radiotherapy, or UV radiation. DNA damaging agents further include platins or platinum-containing chemotherapeutics/coordination complexes of platinum, such as cisplatin, carboplatin, oxaliplatin, nedaplatin, triplatin tetranitrate, phenanthriplatin, picoplatin, satraplatin (overall also referred to herein as platinum-based agents). DNA damaging agents yet further non-exhaustively include alkylating agents (e.g. cyclophosphamide), antimetabolites (e.g. 5-fluorouracil, capecitabine, floxuridine, gemcitabine; purine analogs such as 6-mercaptopurine, 8-azaguanine, fludarabine, and cladribine) and cytotoxic antibiotics (e.g. bleomycin).
Any of the above methods may optionally comprise a clinician or a medical professional (e.g. nurse or doctor) taking or obtaining a (test) tumor sample or (test) sample comprising a tumor or cancer cell from a subject or patient; and/or a clinician, medical professional, laboratory technician or laboratory professional obtaining or isolating the (test) genomic DNA from a tumor sample or (test) sample comprising a tumor or cancer cell; and/or a clinician, medical professional, laboratory technician or laboratory professional detecting, determining or assessing any of the above described features of the (test) genomic DNA. In particular, General Data Protection Regulations (GDPRs) are complied with.
Any of the above methods may optionally comprise a clinician or a medical professional (e.g. nurse or doctor), or laboratory technician or laboratory professional, providing a (test) tumor sample or (test) sample comprising a tumor or cancer cell (obtained from a subject or patient) directly or indirectly to a test laboratory (such as independent accredited test laboratory); and/or a clinician or a medical professional, or laboratory technician or laboratory professional, providing the (test) genomic DNA from a tumor sample or (test) sample comprising a tumor or cancer cell (obtained from a subject or patient) directly or indirectly to a test laboratory (such as independent accredited test laboratory); and/or a clinician or a medical professional (e.g. nurse or doctor), or laboratory technician or laboratory professional receiving (in tangible or intangible form) the results of the detection or determination or assessment of any of the above described features of the (test) genomic DNA as detected, determined or assessed by a test laboratory (such as independent accredited test laboratory); wherein such test laboratory (such as independent accredited test laboratory; or one or its technicians or professionals) is able to perform or is capable of performing a method of determining or assessing any of the above described features as described hereinabove, or is able of or capable of performing an analysis with a diagnostic kit as described hereinafter. In particular, GDPRs are complied with.
As described hereinabove, this disclosure relates to methods of predicting the response of a subject or patient having cancer or a tumor to therapy or treatment including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent). Such methods may optionally comprise a clinician or a medical professional (e.g. nurse or doctor) making such prediction or assessment after having detected, determined or assessed the above-described features of the (test) genomic DNA or after having received (in tangible or intangible form) the results of the detection, determination or assessment of the above-described features of the (test) genomic DNA as detected or determined by a test laboratory (cf. supra). Such methods may optionally comprise a test laboratory (cf. supra) making such prediction or assessment after having detected, determined or assessed the abovedescribed features of the (test) genomic DNA, or after having received (in tangible or intangible form) the results of the detection, determination or assessment of the above-described features of the (test) genomic DNA as detected or determined by an individual or entity not linked to the test laboratory.
The invention in a further aspect relates to a DNA damaging agent or therapy comprising a DNA damaging agent, and/or to an agent inhibiting or impairing DNA repair or therapy comprising an agent inhibiting or impairing DNA repair - or to a treatment or treatment regimen comprising a DNA damaging agent and/or comprising an agent inhibiting or impairing DNA repair - for use in (the manufacture of a medicament for) treating a tumor or cancer, comprising:
- predicting, according to any of the relevant methods described hereinabove, the subject to be a responder to the treatment, treatment regimen or therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent); or alternatively, predicting the subject to be a responder to the treatment, treatment regimen or therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent) when the trained data classification algorithm or model as described hereinabove classified, inferenced or scored the test genomic DNA sample as most closely resembling reference genomic DNA from subjects responding to treatment, treatment regimen or therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent);
- administering the agent, therapy, treatment or treatment regimen to the subject.
Alternatively, the invention relates to methods for treating a subject or patient having cancer or a tumor, such methods comprising administering a DNA damaging agent or therapy comprising a DNA damaging agent to the subject or patient having the cancer or tumor, and/or administering an agent inhibiting or impairing DNA repair or therapy comprising an agent inhibiting or impairing DNA repair to the subject or patient having the cancer or tumor, comprising:
- predicting, according to any of the relevant methods described hereinabove, the subject to be a responder to the treatment, treatment regimen or therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent); or alternatively, predicting the subject to be a responder to the treatment, treatment regimen or therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent) when the trained data classification algorithm or model as described hereinabove classified, inferenced or scored the test genomic DNA sample as most closely resembling reference genomic DNA from subjects responding to treatment, treatment regimen or therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent);
- administering the treatment, treatment regimen or therapy to the subject.
By administering the DNA damaging agent (or therapy comprising it) and/or the agent inhibiting or impairing DNA repair (or therapy comprising it) to the subject or patient having the cancer or tumor, the subject or patient is being treated. In particular, a therapeutically effective amount or dose of the DNA damaging agent and/or including an agent inhibiting or impairing DNA repair is administered to the subject or patient as (part of a) treatment or as part of a treatment regimen. When the DNA damaging agent and/or including an agent inhibiting or impairing DNA repair is combined with a further anticancer agent, then the amount or dose of the DNA damaging agent and/or including an agent inhibiting or impairing DNA repair may in itself not be sufficient to result in treatment, but then is sufficient in combination with the further anticancer agent. Such combinations of treatment modalities may e.g. decrease possible adverse effects or side effects of (one of) the individual modalities. "Treatment"/"treating" refers to any rate of reduction, delaying or retardation of the progress of the disease or disorder, or a single symptom thereof, compared to the progress or expected progress of the disease or disorder, or single symptom thereof, when left untreated. This implies that a therapeutic modality on its own may not result in a complete or partial response (or may even not result in any response), but may, in particular when combined with other therapeutic modalities (such as, but not limited thereto: surgery, radiation, etc.), contribute to a complete or partial response (e.g. by rendering the disease or disorder more sensitive to therapy). More desirable, the treatment results in no/zero progress of the disease or disorder, or single symptom thereof (i.e. "inhibition" or "inhibition of progression"), or even in any rate of regression of the already developed disease or disorder, or single symptom thereof. "Suppression/suppressing" can in this context be used as alternative for "treatment/treating". Treatment/treating also refers to achieving a significant amelioration of one or more clinical symptoms associated with a disease or disorder, or of any single symptom thereof. Depending on the situation, the significant amelioration may be scored quantitatively or qualitatively. Qualitative criteria may e.g. by patient well-being. In the case of quantitative evaluation, the significant amelioration is typically a 10% or more, a 20% or more, a 25% or more, a 30% or more, a 40% or more, a 50% or more, a 60% or more, a 70% or more, a 75% or more, a 80% or more, a 95% or more, or a 100% improvement over the situation prior to treatment. The time-frame over which the improvement is evaluated will depend on the type of criteria/disease observed and can be determined by the person skilled in the art.
A "therapeutically effective amount or dose" refers to an amount of a therapeutic agent to treat, inhibit or prevent a disease or disorder in a subject (such as a mammal). In the case of cancers, the therapeutically effective amount of the therapeutic agent may reduce the number of cancer cells; reduce the primary tumor size; inhibit (i.e., slow down to some extent and preferably stop) cancer cell infiltration into peripheral organs; inhibit (i.e., slow down to some extent and preferably stop) tumor metastasis; inhibit, to some extent, tumor growth; and/or relieve to some extent one or more of the symptoms associated with the disorder. To the extent the drug may prevent growth and/or kill existing cancer cells, it may be cytostatic and/or cytotoxic. For cancer therapy, efficacy in vivo can, e.g., be measured by assessing the duration of survival (e.g. overall survival), time to disease progression (TTP), response rates (e.g., complete response and partial response, stable disease), length of progression-free survival (PFS), duration of response, and/or quality of life.
The term "effective amount (or dose)" or "therapeutically effective amount (or dose)" may depend on the dosing regimen of the agent/therapeutic agent or composition comprising the agent/therapeutic agent (e.g. medicament or pharmaceutical composition). The effective amount will generally depend on and/or will need adjustment to the mode of contacting or administration. The effective amount of the agent or composition comprising the agent is the amount required to obtain the desired clinical outcome or therapeutic effect without causing significant or unnecessary toxic effects (often expressed as maximum tolerable dose, MTD). To obtain or maintain the effective amount, the agent or composition comprising the agent may be administered as a single dose or in multiple doses. The effective amount may further vary depending on the severity of the condition that needs to be treated; this may depend on the overall health and physical condition of the subject or patient and usually the treating doctor's or physician's assessment will be required to establish what is the effective amount. The effective amount may further be obtained by a combination of different types of contacting or administration.
The aspects and embodiments described above in general may comprise the administration of one or more therapeutic compounds to a subject (such as a mammal) in need thereof, i.e., harboring a tumor, cancer or neoplasm in need of treatment. In general a (therapeutically) effective amount of (a) therapeutic compound(s) is administered to the mammal in need thereof in order to obtain the described clinical response(s).
"Administering" means any mode of contacting that results in interaction between an agent (e.g. a therapeutic compound or immunotherapeutic compound or including an agent) or composition comprising the agent (such as a medicament or pharmaceutical composition) and an object (e.g. cell, tissue, organ, body lumen) with which said agent or composition is contacted. The interaction between the agent or composition and the object can occur starting immediately or nearly immediately with the administration of the agent or composition, can occur over an extended time period (starting immediately or nearly immediately with the administration of the agent or composition), or can be delayed relative to the time of administration of the agent or composition. More specifically the "contacting" results in delivering an effective amount of the agent or composition comprising the agent to the object.
Another aspect of the invention relates to methods of prognosing survival or determining or assessing survival probability of a subject or patient having cancer or a tumor upon receiving treatment or a treatment regimen comprising a DNA damaging agent or therapy and/or comprising an agent inhibiting or impairing DNA repair or therapy comprising inhibiting or impairing DNA repair (e.g. treatment, treatment regimen or therapy comprising e.g. a PARP inhibitor or a platinum-based agent), comprising: - predicting, according to any of the relevant methods described hereinabove, the subject to be a responder to the treatment, treatment regimen or therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent); or alternatively, predicting the subject to be a responder to the treatment, treatment regimen or therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent) when the trained data classification algorithm or model as described hereinabove classified, inferenced or scored the test genomic DNA sample as most closely resembling reference genomic DNA from subjects responding to treatment, treatment regimen or therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent);
- prognosing the subject or patient to have, or likely to have an increased survival or increased survival probability if predicted to be responder to the treatment, treatment regimen or therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent); or
- prognosing the subject or patient not to have, or likely not to have an increased survival or increased survival probability if predicted to be a non-responder to the treatment, treatment regimen or therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent).
Such increased survival or increased survival probability is a consequence of the increased likelihood of response of the subject or patient to the treatment, treatment regimen, or therapy. Such increased survival or increased survival probability is in comparison with survival or survival probability of a reference subject or patient, such as the average survival or survival probability of subjects or patients with the same tumor or cancer not receiving the treatment, treatment regimen, or therapy. Survival can be expressed as progression-free survival, relapse-free survival, overall survival, etc., and can be determined e.g. by the Kaplan-Meier method or analysis.
Any such methods may optionally comprise a clinician or a medical professional (e.g. nurse or doctor), or a test laboratory (cf. supra) determining the survival probability.
Any of the above methods (of/for predicting response to therapy or treatment including a DNA damaging agent or including an agent inhibiting DNA repair; of/for prognosing survival or determining or assessing survival probability of a subject or patient having cancer or a tumor upon receiving treatment, treatment regimen or therapy comprising a DNA damaging agent or including an agent impairing DNA repair) in one particular embodiment are in vitro methods, or methods on biological samples having been obtained from a subject or patient. In a further particular embodiment, the subject or patient is a treatment-naive subject or patient. In any of the above, a subject or patient in general is a mammalian species having cancer or a tumor or diagnosed with cancer. The mammalian species in general is a higher species including primates, cattle (e.g. cows, sheep, goats, pigs), horses, and pets (e.g. dogs, cats). In one embodiment the subject or patient is a human subject or patient.
The invention further relates to systems, such as computer systems, data-processing systems, computer or machine readable media, computer programs, or diagnostic kits for use in performing any of the above-mentioned methods (of/for predicting response to therapy or treatment including a DNA damaging agent or including an agent inhibiting DNA repair; of/for prognosing survival or determining or assessing survival probability of a subject or patient having cancer or a tumor upon receiving treatment, treatment regimen or therapy comprising a DNA damaging agent or including an agent impairing DNA repair). In one embodiment, such (computer/data-processing) systems or kits are comprising a data classification algorithm or model, or a computer or machine readable medium comprising a data classification algorithm or model; or are comprising a computer program for execution of a data classification algorithm or model; wherein the data classification algorithm or model is comprising two arms, wherein a first or one arm (of the algorithm or model) is trained to predict the response of a BRCA mutant subject or patient having cancer or a tumor to a treatment or treatment regimen comprising comprising a DNA damaging agent or therapy and/or comprising an agent inhibiting or impairing DNA repair or therapy comprising inhibiting or impairing DNA repair (e.g. a treatment, treatment regimen or therapy comprising a PARP inhibitor or a platinum-based agent) wherein the training is (minimally, at least) based on the sets of N LOH/bin features and BRCA domain mutation features of the reference genomic DNA samples; and wherein a second or another arm (of the algorithm or model) is trained to predict the response of a BRCA wild-type subject or patient having cancer or a tumor to a treatment or treatment regimen comprising comprising a DNA damaging agent or therapy and/or comprising an agent inhibiting or impairing DNA repair or therapy comprising inhibiting or impairing DNA repair (e.g. a treatment, treatment regimen or therapy comprising a PARP inhibitor or a platinum-based agent), wherein the training is (minimally, at least) based on the sets of N LOH/bin features of the reference genomic DNA samples. Training of the arms of the data classification algorithm or model can further be expanded based on additional features as described hereinabove (e.g. copy number features, sigminer components or features, M top or significant CN-bins, P top or significant Al- bins,...).
Alternatively, or additionally, such (diagnostic) kits are comprising a set of immobilized oligonucleotides suitable for capturing from a genomic DNA sample the genomic DNA required for the detection or determination of one or more of the above-described input features; in particular, the number of immobilized oligonucleotides is ranging from lk to 200k.
Related to computer programs, this disclosure covers any computer program that in particular is having the instructions which when executed cause a computer system or device, or a computing system or device, to perform any of the methods (or (at least) a step of such methods) subject of this disclosure. More in particular to perform any herein disclosed method (or (at least) a step of such method) of/for predicting response to therapy or treatment including a DNA damaging agent or including an agent inhibiting DNA repair, or any herein disclosed method (or (at least) a step of such method) of/for prognosing survival or determining or assessing survival probability of a subject or patient having cancer or a tumor upon receiving treatment, treatment regimen or therapy comprising a DNA damaging agent or including an agent impairing DNA repair. In particular, the at least one step of any of the methods performed by the computer program is the step of computer-implemented scoring of the selected set or sets of features of the test genomic DNA with a trained data classification model. One or more other steps that can be performed by the computer program include the step of predicting response to the therapy; the step of detecting or determining the loss of heterozygosity (LOH) status of a set of N bins of the test genomic DNA; the step of detecting or determining the copy number (CN) status of a set of M bins of the test genomic DNA sequence and/or the allelic imbalance (Al) status of a set of P bins of the test genomic DNA sequence.
Related to computer or computing systems or to data-processing systems, this disclosure covers any such system comprising a means for carrying out or performing any herein disclosed method (or (at least) a step of such method) of/for predicting response to therapy or treatment including a DNA damaging agent or including an agent inhibiting DNA repair, or any herein disclosed method (or (at least) a step of such method) of/for prognosing survival or determining or assessing survival probability of a subject or patient having cancer or a tumor upon receiving treatment, treatment regimen or therapy comprising a DNA damaging agent or including an agent impairing DNA repair. In particular, the at least one step of any of the methods performed by the computer, computing system or data-processing system is the step of computer-implemented scoring of the selected set or sets of features of the test genomic DNA with a trained data classification model. One or more other steps that can be performed by the computer, computing system or data-processing system include the step of predicting response to the therapy; the step of detecting or determining the loss of heterozygosity (LOH) status of a set of N bins of the test genomic DNA; the step of detecting or determining the copy number (CN) status of a set of M bins of the test genomic DNA sequence and/or the allelic imbalance (Al) status of a set of P bins of the test genomic DNA sequence.
While training the classifier or classification algorithm or model for predicting response to therapy including a DNA damaging agent or including an agent inhibiting DNA repair starting from/based on a reference set of genomic DNA samples (including samples from subjects responding and not responding to the therapy), some gene-specific copy number features were determined as possible additional feature for training the classifier or classification algorithm or model and for subsequent assessment of a test genomic DNA sample; all as described hereinabove. Some of the listed genes appear not to have been implied in the testing of response to therapy including a DNA damaging agent or including an agent inhibiting DNA repair. In particular, this relates to any individual gene of CEP43, NF1, PCMl and TCEA1/FTIIS or of CEP43, PCMl and TCEA1/FTIIS.
The current disclosure therefore in a further aspect relates to methods of predicting the response of a subject having cancer to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. a PARP inhibitor or a platinum-based agent), comprising: - optional ly, isolating test genomic DNA from a test genomic DNA sample obtained from the subject;
- detecting or determining a gene-specific copy number feature of a (test) genomic DNA obtained from the subject (or detecting or determining a gene-specific copy number feature of the (test) genomic DNA) wherein the gene is selected from CEP43, NF1, PCMl and TCEA1/FTIIS or is selected from CEP43, PCMl and TCEA1/FTIIS.
The copy number feature of the gene-specific copy number feature is in particular selected from amplification, deletion, LOH, focal amplification, focal deletion, focal LOH, and minimal copy number of a gene. Gene-specific copy number related features were detected or determined for a candidate gene based on its overlap with ASCAT segments. Deletion and amplification events are defined based on the total copy number being smaller or larger than 2; LOH events are defined based on the minor allele having copy number 0. Focal events are defined as events smaller than 25% of the respective chromosome arm. When multiple ASCAT segments are overlapping with a gene, the minimum copy number is retained; similarly, (focal) amplification, deletion or LOH of a gene is based on at least one overlapping ASCAT segment having this status.
More in particular, the copy number feature is selected from amplification, deletion, LOH, and minimal copy number of a selected gene. More in particular, the copy number feature of the selected gene is selected from those as indicated in Table 5 (see Example 8).
In one embodiment, such methods can further comprise a step of including the detected or determined gene-specific copy number feature into a or into any method of predicting the response of a subject having cancer to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. a PARP inhibitor or a platinum-based agent).
In an alternative embodiment, such method can further comprise a step of predicting the response of a subject having cancer to therapy including a DNA damaging agent or including an agent inhibiting DNA repair (e.g. PARP inhibitor or a platinum-based agent based), or based in part, on the detected or determined gene-specific copy number feature.
In a further embodiment, such methods are including the detection or determination of a further genespecific copy number feature wherein the gene is selected from ABCB1, ARID1A, ARID1B, ATM, ATR, AXIN1, BARD1, BRCA1, BRCA2, BRIP1, CCNE1, CDK12, CEP89, CHEK1, CHEK2, DECR1, DNMT3A, ECT2, EMSY, EZH2, FANCL, LEPR0TL1, NAALADL2, NCEH1, NLGN1, PABPC1, PALB2, PARP1, PARP2, PARP3, PARP4, PDCD10, PLD1, POLQ, PTEN, RAD51B, RAD51C, RAD51D, RAD54L, RBI, RNF19A, SERPINI1, SERPINI2, SLC35G5, SLFN11, SPATA16, TAS2R43, TBL1XR1, TNIK, TNKS, TNKS2, TP53BP1, VPS13B, WDR49, WRN, XRCC5, XRCC6, and ZBBX; or wherein the gene is selected from ABCB1, ARID1A, ARID1B, ATM, ATR, AXIN1, BARD1, BRCA1, BRCA2, BRIP1, CCNE1, CDK12, CEP89, CHEK1, CHEK2, DECR1, DNMT3A, ECT2, EMSY, EZH2, FANCL, LEPR0TL1, NAALADL2, NCEH1, NF1, NLGN1, PABPC1, PALB2, PARP1, PARP2, PARP3, PARP4, PDCD10, PLD1, POLQ, PTEN, RAD51B, RAD51C, RAD51D, RAD54L, RBI, RNF19A, SERPINI1, SERPINI2, SLC35G5, SLFN11, SPATA16, TAS2R43, TBL1XR1, TNIK, TNKS, TNKS2, TP53BP1, VPS13B, WDR49, WRN, XRCC5, XRCC6, and ZBBX. More in particular, the copy number feature of the additional gene is selected from amplification, deletion, LOH, focal amplification, focal deletion, focal LOH, and minimal copy number. More in particular, the copy number feature of the selected gene is selected from those as indicated in Table 5 or comprising one or more of the genes and associated copy number features as indicated in Table 5 (see Example 8).
Computer / computer or computing system / data-processing system
A computer, computer or computing system, or data-processing system as mentioned herein may utilize one or more subsystems. A computer or computer or computing system, or data-processing system may be a single apparatus comprising the one or more subsystems (e.g. internal components), or may be multiple apparatuses each being a subsystem, and optionally, each comprising one or more own subsystems. Desktops, laptops, mainframe servers, tablets, mobile phones etc. all are eligible as computer, computer or computing system, or data-processing system. The subsystems are usually interconnected and include a (central) processor (single-core processor, multi-core processor on a same integrated chip, or multiple processing units on a single circuit board or networked) capable of executing instructions, an input/output (I/O) controller, and a storage device (external, internal, peripheral, cloud, any medium readable by a computer or computer system). Input devices include keyboards, scanners, a computer mouse, camera, microphone, etc. In particular, the input device is a data collection or data generating device (which by itself may comprise a computer or computer or computing system, or data- processing system), such as a polynucleotide sequencing device (whether automated or not). Collected or generated data are fed to a computer, computer or computing system, or data-processing system designed to analyze the collected or generated data; this may be an ordinary on which data analyzing software is installed (on a storage device) or which is capable of accessing data analyzing software (e.g. installed in or transmitted from a network) and whereby the processor of the computer/computing/data-processing system is instructed by the data analysis software on how to process the collected or generated data fed to the computer/computing/data-processing system, and how to display these via a display adapter to an output device. Output devices are further subsystems and comprise printers, monitors, computer readable medium. Input and output devices are usually connected to a computer/computing/data-processing system via input/output ports to one another or via a network.
The specific combination of hardware and software allows implementation of e.g. analysis of data generated by a polynucleotide sequencing device, or of e.g. detection or determination of the copy number of a SNP, or of e.g. detecting or determining the LOH or HRD status of a bin of genomic DNA, or of e.g. detecting or determining the LOH or HRD status of a genomic DNA sample such as by a trained or machine learned data classification algorithm or model, or of e.g. predicting the response of a subject having cancer or a tumor to treatment or therapy including a DNA damaging agent (see above, such as PARP inhibitor or platinum-based agent) such as by a trained or machine learned data classification algorithm or model. Different software packages (proprietary or open source) can be run on a computer or computer system to achieve the desired degree of data analysis. Output of one computerized data analysis can be the input of a subsequent computerized data analysis step, hence creating an analysis pipeline. Software components can be written in different codes (e.g. Java, C, C++, Perl, Python) as long as the computer processor is able to execute the functions of the software component.
The methods of this disclosure may be computer-implemented methods, or methods that are assisted or supported (in part) by a computer or by a computer/computing/data-processing system. For instance, information reflecting the DNA sequencing analysis can be provided in user readable format by at least one/another processor. The same or a further processor may be calculating, as outlined herein, the LOH or HRD status of a test genomic DNA (such as relative to a control, standard or reference) or predicting response to therapy from the information received. The one or more processors may be coupled to random access memory operating under control of or in conjunction with a computer operating system. The processors may be included in one or more servers, clusters, or other computers or hardware resources, or may be implemented using cloud-based resources. The operating system may be, for example, a distribution of the LinuxTM operating system, the UnixTM operating system, or other open- source or proprietary operating system or platform. Processors may communicate with data storage devices, such as a database stored on a hard drive or drive array or such as a computer or machine readable medium, to access or store program instructions other data. Processors may further communicate via a network interface, which in turn may communicate via the one or more networks, such as the Internet or other public or private networks, such that a query or other request may be received from a client, or other device or service. Such computer-implemented methods (or such methods that are assisted or supported by a computer/computing/data-processing system) may be provided as a kit or as part of a kit. The bioinformatics software required to perform (part of) the computer-implemented methods, i.e. a computer program product, may also be part of a kit, or may be provided as an individual product. A computer product may also consist of a computer or machine readable medium (in any form such as disks (hard, soft disks etc.), cards (memory cards etc.), tapes, sticks (memory, USB sticks etc.), microchips, DVDs, CDs, etc.) which is storing any of the instructions, computer program, or bioinformatics software enabling a computer system to perform at least one of the analysis of the herein described methods and/or to perform at least one calculation as described herein. Furthermore, the computer/computing/data-processing system can be set up or configurated such that it is compliant with GDPR. Other Definitions
The present invention is described with respect to particular embodiments and with reference to certain drawings but the invention is not limited thereto but only by the claims. Any reference signs in the claims shall not be construed as limiting the scope. The drawings described are only schematic and are nonlimiting. In the drawings, the size of some of the elements may be exaggerated and not drawn on scale for illustrative purposes. Where the term "comprising" is used in the present description and claims, it does not exclude other elements or steps. Where an indefinite or definite article is used when referring to a singular noun e.g. "a" or "an", "the", this includes a plural of that noun unless something else is specifically stated. Furthermore, the terms first, second, third and the like in the description and in the claims, are used for distinguishing between similar elements and not necessarily for describing a sequential or chronological order. It is to be understood that the terms so used are interchangeable under appropriate circumstances and that the embodiments of the invention described herein are capable of operation in other sequences than described or illustrated herein. Unless specifically defined herein, all terms used herein have the same meaning as they would to one skilled in the art of the present invention. Practitioners are particularly directed to Sambrook et al., Molecular Cloning: A Laboratory Manual, 4th ed., Cold Spring Harbor Press, Plainsview, New York (2012); and Ausubel et al., current Protocols in Molecular Biology (Supplement 100), John Wiley & Sons, New York (2012), for definitions and terms of the art. The definitions provided herein should not be construed to have a scope less than understood by a person of ordinary skill in the art.
In referring to genes or proteins herein, no distinction is made in the annotation. Thus, whereas for example the human BRCA1 gene would be referred to as the BRCA1 gene, the mRNA as BRCA1 mRNA, and the protein as BRCA1, such distinction is not, or not always, made hereinabove or hereinafter.
Sometimes, terms or words herein are between brackets. As will be clear from the context some of these bracketed terms or words are optional further clarification of the term or word. For instance "(test) genomic DNA sample" obviously refers to a genomic DNA sample, and the genomic DNA sample can thus be optionally clarified by referring to a test genomic DNA sample. Another example is "(further) inhibiting or impairing DNA repair", basically meaning inhibiting or impairing DNA repair, and optionally referring to further inhibiting or impairing DNA repair.
It is to be understood that although particular embodiments, specific configurations as well as materials and/or molecules, have been discussed herein for methods etc. according to the present invention, various changes or modifications in form and detail may be made without departing from the scope and spirit of this invention. The following examples are provided to better illustrate particular embodiments, and they should not be considered limiting the application. The application is limited only by the claims. The content of the documents cited herein are incorporated by reference. EXAMPLES
EXAMPLE 1. Materials and methods
Targeted sequencing. Remaining DNA samples extracted from formalin-Fixed Paraffin-Embedded (FFPE) tissues were obtained from the PAOLA1 trial (Ray-Coquard et al. 2019, N Engl J Med 381: 2416-2428) through the Arcagy-Gineco (Paris, France) driven ENGOT HRD European initiative (Pujade-Lauraine et al. 2021, Int J Gynecol Cancer 31(Suppl 3):A208); samples from a subgroup (n=192) of the total PAOLA-1 trial population (n=806) were used herein. Specific genomic regions enriched using a capture-based probe design were sequenced (100 ng DNA input). This capture-based probe design included probes targeting a genome wide panel of single nucleotide polymorphisms (SNPs), sequenced at median target coverage of ~150x. In total 96,861 SNPs were selected based on population frequency with a uniform distribution across the genome. Centromere regions were excluded. The SNPs were covered by 92,723 specifically designed probes. The average size of oligonucleotide probes was 120 bp. Capture libraries were sequenced on an Illumina Novaseq (paired-end, 2x150 bp reads).
Bioinformatics pipeline.
Mapping. Paired-end reads were processed with Burrows-Wheeler Aligner (BWA, version 0.4.1; version 0.7.15 in relation to Example 8) using the hgl9 human reference genome. Illumina adapter sequences were trimmed from the reads using TrimGalore (version 0.4.1). Duplicated reads were filtered and realigned locally around indels with the genomic analysis toolkit GATK (version 3.6 - Broad Institute). Sequencing metrics were collected with Picard (version 2.5.0).
Copy number. Allelecount (version 4.2.1) was applied to determine the count coverage of each allele in the HRD genome wide panel regions. Minimum base quality and minimum mapping quality were set to 15. Copy numbers of target and antitarget regions were inferred using CNVkit (version 0.9.9). Next, the results from Allelecount and CNVkit were integrated as input for the allele-specific copy number analysis of tumors (ASCAT) tool (version 2.5.2; Van Loo et al. 2010, Proc Natl Acad Sci USA 107:16910-16915).
Loss-of-heterozygosity features. The ASCAT output was used to derive loss-of-heterozygosity (LOH) features based on the allele-specific copy number values. For each SNP, ASCAT reports the copy number value for both the major and minor allele. LOH was defined as the copy number of the minor allele being 0. Next, the genome was divided into bins of 1000 kb in size and the mean LOH across all SNPs within each bin was calculated for each sample, resulting in 2877 bin-specific LOH (or LOH/bin) features per sample. In addition to the bin features, also 23 chromosome specific features were created by calculating the mean LOH for each chromosome (LOH/chromosome features). Finally, a genome-wide LOH (LOH/genome feature) was calculated from the mean LOH across all SNPs per sample. Using this procedure, we obtained 2901 LOH features in total per sample, with each feature representing a local genomic region with averaged LOH values across multiple SNPs (LOH/bin) or features covering an entire chromosome (LOH/chromosome) or the entire genome (LOH/genome). All LOH calculations were performed in R (version 3.6.3).
Statistical analysis. Based on the 2877 bin-specific LOH features or the 2901 LOH features/sample (LOH/bin features + LOH/chromosome features + LOH/genome feature), an exemplary nearest shrunken centroid analysis was performed using the PAM (Partitioning Around Medoids) package in R (version 3.6.3) to check the power of this approach in distinguishing Myriad myChoice positive from Myriad myChoice negative samples. Using leave-one-out cross-validation, the best LOH features within each training set were selected and next a PAM score for each test sample was determined from the Myriadpositive status class probabilities. Whereas the Myriad myChoice HRD status was available for the samples, the actual Myriad myChoice HRD scores were not available and were approximated using the scarHRD algorithm (version 0.1.1; an R package which determines the levels of homologous recombination deficiency (telomeric allelic imbalance, loss of heterozygosity, number of large-scale transitions)). Receiver operation characteristic (ROC) for the PAM and HRD scores with respect to Myriad myChoice status were constructed using the pROC package (version 1.18.0). As an alternative to PAM, the Random Forest learning algorithm was used.
EXAMPLE 2. HRD classification based on fractional loss-of-heterozygosity (FLOH)
Wang et al. 2012 (Clin Cancer Res 18: 5806-5815) proposed fractional LOH (FLOH) as feature to predict response of breast cancer to chemotherapy. FLOH was defined as the proportion of inferred LOH calls within the total number of SNPs. These authors, however, did not compare the FLOH feature to the Myriad MyChoice HRD classification. The ASCAT tool provides a feature for determining FLOH.
The performance of FLOH (via ASCAT tool feature) in classifying a subset of 192 samples of the PAOLA1 trial (HRD-positive: n=110; HRD-negative: n=82) was compared with the LOH+LST+TAI HRD score as approximated by the scarHRD algorithm. Figure 1 indicates that, in contrast to the scarHRD score (Figure IB), the FLOH feature calculated using ASCAT (Figure 1A) is not suited for distinguishing HRD-positive and HRD-negative samples.
These results appear to support the previously reported underperformance of the %LOH feature in distinguishing HRD-positive and HRD-negative samples (e.g. Timms et al. 2020, J Clin Oncol 38(15 suppl):1586).
EXAMPLE 3. HRD classification based on number of LOH regions of intermediate size
Abkevich et al. 2012 (Br J Cancer 107:1776-1782) defined a homologous recombination deficiency (HRD) score as the number of loss of heterozygosity regions of intermediate size regions observed in a tumour sample. This score was determined using 10-fold cross-validation for the same subset of 192 samples of the PAOLA1 trial as used in Example 2, and compared with the LOH+LST HRD score and LOH+LST+TAI HRD score as approximated by the scarHRD algorithm. As indicated in Figure 2A, the AUC value of the LOH-only score (AUC = 0.92) is close to the AUC value of the LOH+LST-score (AUC = 0.95; Figure 2B) and to the AUC value of the LOH+LST+TAI-score (AUC = 0.95; Figure 2C). The AUC value for the LOH+LST+TAI HRD score as approximated by the scarHRD algorithm = 0.93 ("AUC original score").
EXAMPLE 4. HRD classification based on LOH/bin, LOH/chromosome, and LOH/genome features
In searching for an alternative method for determining the HRD status of genomic DNA of a tumor or cancer cell relying on LOH features only, but performing as well as the combined LOH + LST + TAI features, it was found that the mean LOH/bin information, the mean LOH/chromosome and the mean LOH/genome information could be used efficiently in determining the HRD status of genomic tumor or cancer DNA. Moreover, for the mean LOH/bin information, the bin size could be lowered to under 1.5 Mb and data classification of the mean LOH/bin information, the mean LOH/chromosome and the mean LOH/genome information could be used efficiently in determining the HRD status of genomic tumor or cancer DNA.
Furthermore, these alternative methods relying on mean LOH/bin information, the mean LOH/chromosome and the mean LOH/genome information were found to predict as reliably as based on the Myriad myChoice status (itself based on LOH + LST + TAI) the response of high-grade serous ovarian cancer patients who received PARP inhibitor treatment (olaparib).
A cohort of 192 FFPE high-grade serous ovarian cancer samples from the PAOLA1 trial (identical to the samples used in Examples 2 and 3), composed of 132 patients administered with olaparib and 60 patients who received a placebo was examined. The Myriad myChoice status was determined previously for each sample as part of the PAOLA-1 trial. The Myriad myChoice HRD score was in contrast not available but approximated herein by the scarHRD algorithm (LOH (based on number of regions of LOH of intermediate size) + telomeric allelic imbalance (TAI) + large-scale transitions (LST)).
We herein used PAM as exemplary classification algorithm on the cohort of 192 samples to correlate LOH features with Myriad myChoice status. Figure 3 shows the exemplary pipeline that was applied on all samples. The PAM score constitutes the class probabilities for Myriad-positive status and is extracted using leave-one-out cross-validation from the PAM analysis. A combination of LOH-features was fed to the classification algorithm: mean LOH/bin, mean LOH/chromosome, and mean genome-wide LOH (see above).
Figure 4 demonstrates the performance of predicting HRD status using the estimated scarHRD score ("HRD score": LOH (based on number of regions of LOH of intermediate size)+LST+TAI) or using the herein developed alternative method ("PAM score": mean LOH/bin features + mean LOH/chromosome features + and mean genome-wide LOH feature) with receiver-operator characteristic (ROC) curves. We found that the current alternative method based on LOH-only features (mean LOH/bin, mean LOH/chromosome, and mean genome-wide LOH) reached an AUC = 0.91, which closely resembles the scarHRD score (LOH (based on number of regions of LOH of intermediate size)+LST+TAI); AUC = 0.93) in predicting Myriad myChoice positive or negative status.
Encouraged by this unexpected positive result, we next assessed the survival curves of samples stratified by scarHRD or the current alternative method. Since the alternative data classifier score is a probability from 0 (HRD negative) to 1 (HRD positive), we herein simply used an exemplary threshold of 0.5 to stratify HRD positive and HRD negative samples. Figure 5A shows the results for the patients treated with olaparib, with HRD positive patients (according to Myriad myChoice test) displaying a longer survival compared to HRD negative patients (according to Myriad myChoice test). The survival curves of patients stratified by the current alternative LOH-only based data HRD classifier score (mean LOH/bin, mean LOH/chromosome, and mean genome-wide LOH features) are as good as identical. Figure 5B shows the results from patients in a control group administered with a placebo. All patients display shorter survival times compared to the olaparib treated patients and again a resemblance between the current alternative LOH-only based data classifier HRD score (mean LOH/bin, mean LOH/chromosome, and mean genome-wide LOH features) and scarHRD score (LOH (based on number of regions of LOH of intermediate size)+LST+TAI features) can be observed.
In conclusion, these results demonstrate that the current alternative LOH-only features and analysis as outlined herein can be effective predictors of HRD status in high-grade serous ovarian cancer patients who received platinum therapy, and this in a way fully comparable to the gold standard (LOH+LST+TAI) HRD score.
EXAMPLE 5. HRD classification based on LOH/bin features
In searching for a further alternative method for determining the HRD status of genomic DNA of a tumor or cancer cell relying on LOH features only, but performing as well as the combined LOH + LST + TAI features, it was found that, compared to Example 4, the mean LOH/bin information is sufficient, and that the mean LOH/chromosome and the mean LOH/genome information could be omitted, for efficiently in determining the HRD status of genomic tumor or cancer DNA.
Furthermore, these alternative methods relying on mean LOH/bin information were found to predict as reliably as based on the Myriad myChoice status (itself based on LOH + LST + TAI) the response of highgrade serous ovarian cancer patients who received PARP inhibitor treatment (olaparib). This was in particular demonstrated in Figure 6 assessing the survival curves of samples stratified by scarHRD or the alternative method of Example 4 (results shown based on 10-fold cross-validation) but limited to using the mean LOH/bin features. Figure 6A shows the results for the patients treated with olaparib, with HRD positive patients (according to Myriad myChoice test) displaying a longer survival compared to HRD negative patients (according to Myriad myChoice test). The survival curves of patients stratified by the alternative LOH-only based data HRD classifier score (mean LOH/bin features only; mean LOH/chromosome and mean genome-wide LOH features omitted) are as good as identical. Figure 6B shows the results for the placebo group. The bin-size applied for the results depicted in Figure 6 is IMbp (1000 kbp). The same alternative method of Example 4 but limited to mean LOH/bin features was reiterated with an exemplary alternative bin size of 0.75 Mbp (750 kbp) for which Kaplan-Meier curves are depicted in Figure 9 A/B (A: olaparib-treated group; B: placebo group); and with an exemplary alternative bin size of 1.25 Mbp (1250 kbp) for which Kaplan-Meier curves are depicted in Figure 9 C/D (C: olaparib- treated group; D: placebo group).
Additionally, the hazard ratio's (HRs) for the HRD-positive groups were determined. Hazard is defined as the slope of the survival curve. The hazard ratio compares two treatments. In case of the hazard ratio being 2.0, then the rate of deaths in one treatment group is twice the rate in the other group. Herein, the hazard ratio was determined for the olaparib-treated HRD-positive subject group versus the control (placebo) HRD-positive subject group, a low hazard ratio implies a beneficial result of olaparib treatment. Results are summarized in Table 1.
Table 1. Survival hazard ratio's determined for HRD-positive subjects based on PAM-analysis of mean LOH/bin for different bin sizes.
From Figures 6 and 9, and Table 1, it is obvious that analyses based on alternative bin sizes (all <1.5 Mbp (1500 kbp)) lead to comparable results.
In conclusion, these results demonstrate that a further alternative LOH-only features and analysis as outlined herein can be effective predictors of HRD status in high-grade serous ovarian cancer patients who received PARP inhibitor therapy, and this in a way fully comparable to the gold standard. EXAMPLE 6. HRD classification based on LOH/bin features using Random Forest (RF) classifier
Example 5 relied on the PAM algorithm as classifier of the mean LOH/bin features only. The same analysis (on the mean LOH/bin features only; same sample set) was performed using an alternative classifier, i.e. the Random Forest data classification algorithm (Breiman 2001, Machine Learning 45:5-32; Breiman and Cutler's Random Forests for Classification and Regression Version 4.7-1.1 Date 2022-01-24: https://cran.r-project.org/web/packages/randomForest/randomForest.pdf).
As an illustration of this alternative machine learning approach, the random forest classifier was trained on the Myriad myChoice status. The random forest classifier consisted of 5000 trees. At each node, 45 LOH bins were randomly selected as predictors and the minimum node size was set to 10. These parameters are exemplary and can be varied. LOH status of a bin was determined based on the copy number of the minor allele of the bin midpoint nucleotide position wherein this copy number was derived from the segment of the genomic DNA with constant copy number comprising the bin midpoint nucleotide position.
Within the training data set, 10-fold cross-validation was used for classification metrics, and a model was subsequently fitted to the full training set to obtain classification metrics on the validation set. ROC curves were constructed similar to the PAM method described earlier, and yielded an AUC value of 0.93 (i.e. identical to the AUC value obtained with the scarHRD score, see Example 4).
Figure 7 indicates that the RF classifier can distinguish between Myriad myChoice HRD positive and negative samples (as well as the scarHRD algorithm, see Figure IB). Furthermore, survival probabilities of olaparib- or placebo-treated ovarian cancer patients based on the RF classification of the LOH/bin features are near identical to the survival probabilities based on the Myriad myChoice HRD status, as illustrated in Figure 8 (A: olaparib-treated group; B: placebo group; results shown for 10-fold cross- validation). The bin-size applied forthe results depicted in Figure s is IMbp (1000 kbp). The same method RF classifier method was re-iterated with an exemplary alternative bin size of 0.5 Mbp (500 kbp) for which results are depicted in Figure 10 A/B (A: olaparib-treated group; B: placebo group); with an exemplary alternative bin size of 0.75 Mbp (750 kbp) for which results are depicted in Figure 10 C/D (C: olaparib-treated group; D: placebo group); with an exemplary alternative bin size of 1.25 Mbp (1250 kbp) for which results are depicted in Figure 10 E/F (E: olaparib-treated group; F: placebo group); and with an exemplary alternative bin size of 2 Mbp (2000 kbp) for which Kaplan-Meier curves are depicted in Figure 10 G/H (G: olaparib-treated group; H: placebo group).
Additionally, the hazard ratio's (HRs) for the HRD-positive groups were determined (cfr. Example 5), results are summarized in Table 2.
Table 2. Survival hazard ratio's determined for HRD-positive subjects based on RF-analysis of mean LOH/bin for different bin sizes.
From Figures 8 and 10, and Table 2, it is obvious that analyses based on alternative bin sizes of <=2 Mb(p) and <1.5 Mb (1500 kbp)) lead to comparable results.
EXAMPLE 7. Further simplification of HRD classification based on LOH/bin features
The analyses as outlined in the foregoing Examples relied on whole-genome analysis, including dividing the whole genome in bins for purposes of determining the mean LOH/bin feature. In a next step, it was assessed whether the number of required bins could be lowered without compromising reliability of determination of LOH or HRD status of the DNA of a tumor compared to the whole-genome approach.
To select a lower number of bins, the dataset was split 50/50 into a training set and a test set. Genome wide profiles for genomic instability, in particular LOH, were drawn for all samples of a reference set of genomic DNA samples (the training cohort of the PAOLA-1 trial) and heterogeneity of LOH patterns between HRD positive and HRD negative samples was observed. Significantly frequent LOH regions contributing to HRD were identified via a permutation test across the genome for samples of the training cohort (permutation P<0.05). Hundred LOH regions (number randomly chosen with simplification in mind) with the largest weight in the model were selected.
The training cohort was used to train the random forest classifier model with all LOH bins (genome-wide; bin size 1 Mbp (1000 kbp)), thus resulting in the RF classifier trained on all LOH bins. Subsequently the random forest classifier model was separately retrained on the training cohort with only the top-100 most important LOH-bins, thus resulting in the second RF classifier trained on the top-100 LOH bins.
The test cohort was then analyzed separately both with the RF classifier trained on all LOH bins and with the second RF classifier trained on the top-100 LOH bins, and the results were compared. Results are depicted in Figure 11 in the same manner as in Figures 8 and 10. As a test dataset/cohort was used instead of 10-fold cross-validation, the shape of the Kaplan-Meier curves in Figure 11 differs from the shape of these curves in Figures 8 and 10. Independent thereof, both the RF classifier trained on all LOH bins and the second RF classifier trained on the top-100 LOH bins perform equally well, and this was further underscored by determining the similar hazard ratios (cfr. Examples 5 and 6) applied to HRD- positive and HRD-negative subjects as determined either by a RF classifier trained on all LOH bins or by a RF classifier trained only on the top-100 LOH bins, as summarized in Table 3. Table 3. Hazard ratios for a an olaparib vs placebo Cox proportional hazards model, applied to HRD- positive and HRD-negative subjects as determined by a RF classifier trained on all LOH bins and by a RF classifier trained only on the top-100 LOH bins.
EXAMPLE 8. Further improvement of sample classification based on mean LOH/bin features
Methods used herein are described in Example 1 and further expanded in subsections of Example 9.
As outlined in the previous Examples, an alternative method for determining the HRD status of genomic DNA of a tumor or cancer cell relying on LOH features only, but performing as well as the combined LOH + LST + TAI features, was developed for efficiently determining the HRD status of a test genomic tumor or cancer DNA sample. In a first improvement, such methods for determining the HRD status of genomic DNA of a tumor or cancer cell relying on the (mean) LOH/bin information could be simplified by reducing the number of bins needed for analysis. Building further on this, the next aim was to develop a method for determining the HRD status of genomic DNA of a tumor or cancer cell performing better compared to the combined LOH + LST + TAI features, and compared to the method relying on mean LOH/bin (either genome-wide or most significant LOH bins), with a special focus on improving prediction of response of a subject to PARP inhibitor (PARPi) therapy (as compared to focus on determination of the LOH or HRD status of a genomic DNA sample of a subject having cancer or a tumor).
To this end, several additional features, more in particular input features, where combined with different LOH/bin features and subjected to random forest classifier analysis. Any candidate random forest model was trained to predict exceptional response to PARPi, exceptional response or long term response being defined as progression free survival (PFS) of >2 years (progression after halting PARPi therapy) vs PFS <2 years (progression before halting PARPi therapy), abbreviated as 2yPFS. Any final random forest model is therefore also referred to herein as the "PARPi response test". Patients with a PFS value censored before 2 years were assigned to <2 years (for more information on patient and study design, see Example 9.1; for more information on determination of PFS, see Example 9.2).
Input features
Given the large number of possible input features/variables in addition to the LOH/bin feature, preliminary random forest models were built per feature or per feature set to estimate the importance of individual additional variables/features in predicting response to PARPi; variable/feature importance was assessed in the training cohort using P-values calculated by the ranger package for the corrected impurity measure. Likewise, a preliminary random forest classifier was trained for predicting response to PARPi with the LOH status of genome-wide bins, of which the top-/V most important bins based on a permutation measure were retained (see Example 7 exemplifying /V=100) as key features for the final random forest model/PARPi response test.
Based on this feature selection, a general/final random forest classifier was then trained to predict 2yPFS, also referred to as the PARPi response test as already indicated above. Training was done only on patients in the training set/cohort receiving olaparib, i.e., patients for whom long term response to PARPi could be assessed. Ten-fold cross validation was used first to measure performance in the training set/cohort. A final random forest model/PARPi response test was then fit to the full training set/cohort and applied to the test set/cohort.
Feature sets retained for further combination analysis included:
1. LOH-related features:
- all LOH bins (genome-wide; bin size 1 Mbp; see Example 5);
- all significant LOH bins (selected based on P-values < 0.05 of the corrected impurity measure);
- top-/V LOH bins (selected based on the ranks of a permutation measure; see Example 7; importance of individual LOH regions was identified via a permutation test for samples of the training cohort, the "N" LOH bins with the largest weight in the model were selected);
- all significant LOH bins + significant CN bins + significant Al bins ("significant LOH/CN/AI bins"; each selected based on P-values < 0.05 of the corrected impurity measure;);
- top-/V LOH bins + top-/V CN bins + top-/V Al bins ("top-/V LOH, CN, Al bins"; all selected based on the ranks of a permutation measure);
-top-/V LOH bins + significant CN bins + significant Al bins ("top-/V bins & significant CN/AI bins") ;
2. As mutations in the BRCA genes are known drivers for development of HRD in cancer cells, the "BRCA status feature" (i.e., mutated or not mutated, see Example 9.4) was incorporated in preliminary random forest models, with amending the HRD status (as determined with any model/classifier) from negative or uncertain to positive in case the BRCA status was positive (i.e. when mutations were detected in a BRCA gene). The "BRCA status feature" was retained to be included in the final RF model/PARPi response test;
3. Labidi-Galy et al. 2023 (Ann Oncol 34:152-162) recently mentioned a larger impact on clinical outcome of mutations in specific functional domains of the BRCA genes in the full cohort of the PAOLA-1 trial. The same regions of interest were offered as input features to the preliminary random forest model in case a pathogenic BRCA variant was present. More in particular, 2-year PFS was analysed in patients with BRCAl/2m in the tumour sample according to mutation location in functional domains of BRCA1 Really Interesting New Gene (RING) domain, amino acids (AA): 8-96; BRCA1 DNA-binding domain (DBD), AA:452-1092; BRCA1 C-terminal domain of BRCA1 (BRCT), AA: 1646-1736 and 1750-1855; or BRCA1 Other Location (locations in BRCA1 different from the previous specifically defined BRCA1 domains); BRCA2 RAD51-binding domain, AA: 900-2000; BRCA2 DBD, AA: 2459-3190; or BRCA2 Other Location (locations in BRCA2 different from the previous specifically defined BRCA2 domains), thus yielding the feature referred to herein as "BRCA domain mutation" feature. In more detail, the presence of a mutation located in either one of 7 functional domains of BRCA1 or BRCA2 protein was separately withheld as potential feature for incorporation in the final RF model/PARPi response test; this feature with 7 possible values is referred to hereinabove and hereinafter as "BRCA domain mutation feature". These 7 functional domains/values are listed in Table 4.
Table 4. Functional BRCA1 and BRCA2 protein domains
Mutations were evaluated on transcripts NM_007294.3 (BRCA1) and NM_000059.3 (BRCA2) on human reference genome hgl9.
4. Furthermore, a random forest classifier was trained on 2-year PFS status using copy-number features specifically related to genes (hereinabove and hereinafter referred to as "gene-specific copy number features") and using copy-number features related to ASCAT segmentation (hereinabove and hereinafter referred to as "ASCAT segment copy number features"), and the importance of the single variables was assessed in the training set using P-values calculated by the ranger package from the corrected impurity measure. ASCAT & gene-specific copy number features with P-value < 0.05 in this preliminary random forest model described were retained as "significant ASCAT/gene" features (or, alternatively, "characteristic ASCAT/gene" of "distinctive ASCAT/gene" features).
In more detail, some general features related to segmentation of the genome, the "ASCAT segment copy number features" were evaluated, i.e., the total number of segments of each sample, the fraction of LOH (fraction LOH), deletions (fraction deletions) and amplifications derived from the summed segment sizes, the ASCAT optimization parameters psi (optimizing ploidy) and rho (optimizing tumor purity), and resulting goodness-of-fit and calculated ploidy. The ASCAT profiles (genome-wide allele-specific copy- number profiles) provide a basis for all genome-wide computations from which gains, losses, copy- number-neutral events and LOH can be determined (Van Loo et al. 2010, PNAS 107:16910-16915). ASCAT merges the segmentation and copy-number estimation of a sample, taking into account both ploidy of tumour cells and non-aberrant cell admixture. Via aggregation of ASCAT profiles across the samples with exceptional PFS and lower PFS rates, differences in number of segments, genome-wide fraction of LOH, genome-wide fraction of deletions, genome-wide fraction of amplifications and tumour ploidy were identified between the subgroups. Via variable importance analysis of a separate random forest model within the training cohort, we determined that samples with 2yPFS show significant differences between a number of these features. Setting as exemplary threshold a P-value < 0.05 in the preliminary 2yPFS random forest model, 5 global ASCAT features were withheld for the current example: fraction LOH, fraction deletions, ploidy, optimization parameter psi (for optimizing ploidy) of the sample, and goodness-of-fit, all as determined by the ASCAT algorithm.
Furthermore, "gene-specific copy number features" were considered starting from a total of 62 genes with possible effect on PARPi response. These features were tested in a temporary random forest trained for 2-year PFS using only the ASCAT segment and gene-specific copy number features on samples of the training cohort. Seven copy-number features were determined for each of the 62 candidate genes, thus amounting to 62 x 7 = 434 individual candidate features. The copy number features are amplification (amp), deletion (del), LOH, focal amplification (focal amp), focal deletion (focal del), focal LOH, and minimal copy number (min CN) of the gene. Gene-specific copy number related features were determined for a candidate gene based on its overlap with ASCAT segments. Deletion and amplification events are defined based on the total copy number being smaller or larger than 2; LOH events are defined based on the minor allele having copy number 0. Focal events are defined as events smaller than 25% of the respective chromosome arm. When multiple ASCAT segments are overlapping with a gene, the minimum copy number is retained; similarly, (focal) amplification, deletion or LOH of a gene is based on at least one overlapping ASCAT segment having this status.
The 62 candidate genes selected for this analysis are ABCB1, ARID1A, ARID1B, ATM, ATR, AXIN1, BARD1, BRCA1, BRCA2, BRIP1, CCNE1, CDK12, CEP43, CEP89, CHEK1, CHEK2, DECR1, DNMT3A, ECT2, EMSY, EZH2, FANCL, LEPR0TL1, NAALADL2, NCEH1, NF1, NLGN1, PABPC1, PALB2, PARPI, PARP2, PARP3, PARP4, PCM1, PDCD10, PLD1, POLQ, PTEN, RAD51B, RAD51C, RAD51D, RAD54L, RBI, RNF19A, SERPINI1, SERPINI2, SLC35G5, SLFN11, SPATA16, TAS2R43, TBL1XR1, TCEA1, TNIK, TNKS, TNKS2, TP53BP1, VPS13B, WDR49, WRN, XRCC5, XRCC6, ZBBX.
Setting as exemplary threshold a P-value < 0.05 for corrected impurity measure in the preliminary random forest model, 38 gene-specific copy number features were withheld, as well as 5 of the ASCAT segment copy number features (see above). The 38 gene-specific copy number features relate to 24 individual genes and are listed in Table 5. Table 5. 38 gene-specific copy number features relate to 24 individual genes withheld for possible inclusion in the final RF classification model/PARPi response test. The most important gene-specific copy number features based on this training cohort appear to be located at the genes CEP43, ARID1B, PARP3, RBI, ARID1A, NF1, TNKS2 and DNMT3A.
5. Wang et al. 2021 (PLoS Genet 17(5): el009557) disclosed a copy number signature analysis framework (the "sigminer" or signature miner tool) which led to 5 copy number (CN) signatures and 3 single base substitution (SBS) signatures based on a set of 80 predefined copy number components (sigminer package in R: cran.r-project.org/package=sigminer).
The sigminer R package was used to extract copy number alteration features from ASCAT-segmented genome-wide copy number profiles. A matrix of components (referred to hereinabove and hereinafter as "sigminer components") according to Wang et al. 2021 was generated by applying the method "Wang" in the "sig_tally" function of the sigminer package. Based on RF classifier model training, significant "Wang components" (based on P-values < 0.05 of the corrected impurity measure) were retained for further analysis. In more detail, based on histograms of 8 copy number characteristics (i.e., number of breakpoints per 10 Mbp or BP10MB, number of breakpoints per chromosome arm or BPArm, absolute copy numbers of segments or CN, copy number change points between segments or CNCP, lengths of oscillating copy number segment chains or OsCN, segment sizes or SS, number of chromosomes with >50% copy number alterations or NC50, and the number of copy number events per chromosome or BoChr; these copy number characteristics are referred to hereinabove and hereinafter as "sigminer copy number characteristics"), several components per characteristic were defined. Each component is in itself a range of possible values for the corresponding characteristic — the combined value ranges of the individual components together cover the value range of the copy number characteristic. A total of 80 components was defined (BP10MB characteristic: 7 components; BPArm characteristic: 14 components; CN characteristic: 7 components; CNCP characteristic: 7 components; OsCN characteristic: 7 components; SS characteristic: 8 components; NC50 characteristic: 7 components; BoChr characteristic: 23 components). These 80 components are referred to as the "sigminer components".
Setting as exemplary threshold a P-value < 0.05 in the preliminary random forest model (see above), 8 sigminer components related to 5 sigminer copy number characteristics were retained. These 8 sigminer components are listed in Table 6.
Table 6. 8 significant sigminer components withheld for inclusion in the final RF classification model
Evaluation of candidate classification models
The results of an initial further analysis incorporating training of a random forest classifier on 2-year PFS status including multiple individual combinations of the above-described features on the full training dataset/cohort and subsequent applied to the full test dataset/cohort (see higher) are presented in Figure 12. Figure 12A provides for the % HRD prediction of the model (expecting ~50% HRD+ subjects within the train and test cohorts). Figure 12B provides the hazard ratios (HR) with confidence intervals (indicated as horizontal lines per combination of features) of olaparib-treated subjects vs the placebo group (HRD+ subjects are expected to have a low HR (<1) whereas HRD- subjects are expected to have a "neutral" HR near 1.0). Figure 12C provides the hazard ratios (HR) with confidence intervals (indicated as horizontal lines per combination of features) of HRD+ subjects vs HRD- subjects (HRD+ subjects are expected to have a low HR (<1) as having an increased chance or responding to olaparib whereas HRD- subjects are expected to have a "neutral" HR near 1.0). Vertical lines in Figures 12B and 12C indicate HR status as defined by Myriad status. It can be concluded from Figure 12A that inclusion of "BRCA domains" as a feature or variable in the model appeared to be advantageous in terms of better estimating the % HRD-positive subjects within the datasets. Furthermore, and ideally, the hazard ratio confidence intervals should not overlap when assessing HR of olaparib vs placebo for HRD+ versus HRD- subjects (confident distinction between olaparib responders and non-responders) and/or when assessing HR of HRD+ vs HRD- (confident distinction between HRD+ and HRD- samples). Figure 12B (HR olaparib vs placebo) indicates that both conditions (no overlap) are met for the combinations "significant LOH/CN/AI bins & BRCA status", "top-100 LOH bins & significant CN/AI bins & BRCA status", "top-160 LOH, CN, Al bins & BRCA status", "top-100 LOH, CN, Al bins & BRCA status", and "top-120 LOH, CN, Al bins & BRCA domains" in the training dataset, but this was not reflected in the test dataset.
In a further step in developing a method of improved prediction of response of a subject to PARP inhibitor (PARPi) therapy, tumor samples were divided according to their BRCA-status prior to further analysis with either one of the RF models as summarized in Figure 12. And only for BRCA-mutant samples was the "BRCA domain" feature/variable incorporated in the RF model. Results obtained for the training dataset are depicted in Figure 13. By including this step, the number of models resulting in no overlap of the hazard ratio confidence intervals of both HR olaparib vs placebo and HR HRD+ vs HRD- has significantly increased, indicating that division of the samples according to BRCA-status prior to further classification is beneficial; these models include "significant LOH/CN/AI bins & BRCA status", "top-100 LOH bins & significant CN/AI bins & BRCA status", "top-160 LOH, CN, Al bins & BRCA status", "top-100 LOH, CN, Al bins & BRCA status", "top-160/-100 LOH, CN, Al bins & significant ASCAT/gene features & BRCA domains", "top-160/-100 LOH bins & significant ASCAT/gene features & BRCA domains", "top- 160/-100 LOH bins & significant ASCAT/gene features & significant Wang components & BRCA domains", "top-160/-100 LOH bins & significant CN/AI bins & significant ASCAT/gene features & significant Wang components & BRCA domains", "top-160/-100 LOH bins & significant CN/AI bins & significant ASCAT/gene features & BRCA domains", "top-160/-100 LOH bins & significant CN/AI bins & BRCA domains", and "top-80/-90/-100/-110/-120/-130/-140/-150/-160/-170/-180 LOH, CN, Al bins & BRCA domains". Importantly, even the "standard combination" of LOH, LST and TAI ("HRD, LST, TAI") topped up with the BRCA domain feature was not able to fully resolve the HR confidence intervals of either one of HR olaparib vs placebo and HR HRD+ vs HRD-.
In a subsequent step, the test dataset was analyzed with these RF models, and results are depicted in Figure 14.
The RF models including as variables "top-160/-100 LOH bins & significant ASCAT/gene features & BRCA domains", "top-160 LOH bins & significant ASCAT/gene features & significant Wang components & BRCA domains" re-occurred in producing no overlap of the hazard ratio confidence intervals of both HR olaparib vs placebo and HR HRD+ vs HRD-; the combination "top-100 LOH bins & BRCA domains" was a newly identified combination; and the "standard combination" topped up with the BRCA domain feature again failed.
As an example "PARPi response test", the RF model including as variables "top-100 LOH bins & significant ASCAT/gene features & BRCA domains" including division of the samples according to their BRCA-status prior to running the RF model was subjected to further detailed analysis, in particular its capability to predict a 24-month PFS as PARPi response for patients treated with olaparib.
Amongst the 198 samples of the PAOLA-1 training cohort, 96 (48.5%) samples are classified as HRD- positive with this PARPi response test. According to the status of the Myriad HRD test 52.5% of this cohort was HRD-positive (104/198). The HR for positive prediction of 2yPFS was 0.33 with this PARPi response test (95% Cl 0.187-0.571) vs HR 0.40 with Myriad test (95% Cl 0.235-0.672). 76.9% versus 39.6% of patients had no disease progression with olaparib versus placebo, respectively with this PARPi response positive test at 2 years. Of the Myriad test positive patients 73.5% showed no disease progression at 2yPFS when treated with olaparib versus 38.7% with placebo.
In PARPi response test negative tumours, there was no difference in 2yPFS: 30.2% versus 23.3% of patients had no disease progression (HR 1.01; 95% Cl 0.646-1.585) with olaparib versus placebo, respectively. Likewise, 24.2% versus 16.2% of HRD negative samples with Myriad had showed no disease progression (HR 1.01; 95% Cl 0.614-1.664). When considering BRCAwt samples only, the 2yPFS was 66.0% with olaparib and 47.6% with placebo (HR 0.49; 95% Cl 0.209-1.141) following the positive group according to the PARPi response test. In Myriad HRD positive BRCAwt tumours, 2yPFS was 66.3% versus 60% (HR 0.54; 95% Cl 0.248-1.195). Via the approach of the PARPi response test, 4 BRCAm samples were considered as negative in the 2yPFS prediction. When considering only BRCAm tumours, the 2yPFS was 82.6% with olaparib and 35.0% with placebo (HR 0.26; 95% Cl 0.122-0.539) with the PARPi response positive samples. For Myriad HRD positive BRCAm tumours, 2yPFS was 77.9% versus 28.6% (HR 0.32; 95% Cl 0.156-0.647).
After optimalisation in the training cohort, the PARPi response test was independently repeated on the 198 samples of the test/validation cohort. Within this subgroup the PARPi response test classified 94 (47.5%) samples as positive. Following the Myriad HRD test status 56.7% of this cohort was considered HRD positive (112/198). The HR for positive prediction of 2yPFS was 0.32 with the PARPi response test (95% Cl 0.181-0.566) vs HR 0.40 with Myriad test (95% Cl 0.248-0.661). This translates into 78.1% versus 32.1% of patients had no disease progression with olaparib versus placebo, respectively with the PARPi response positive test at 2 years. Of the Myriad test positive patients 69.2% showed no disease progression at 2yPFS when treated with olaparib versus 32.4% with placebo (Figure 16A). In both the PARPi response test negative and Myriad negative tumours, there was no difference in 2yPFS: 23.9% versus 28.5% of patients had no disease progression with olaparib versus placebo, respectively with PARPi response test (HR 1.01; 95% Cl 0.637-1.610). 25.0% versus 30.6% of HRD negative samples with Myriad had showed no disease progression (HR 1.11; 95% Cl 0.630-1.960) (Figure 16B).
Regarding the BRCAwt samples, the 2yPFS was 70.8% with olaparib and 22.2% with placebo (HR 0.18; 95% Cl 0.068-0.487) for the PARPi response positive samples (Figure 16C). For Myriad HRD positive BRCAwt tumours, 2yPFS was 53.6% versus 29.4% (HR 0.35; 95% Cl 0.173-0.700) (Figure 16D). Via the approach of the PARPi response test, 3 BRCAm samples were considered as negative in the 2yPFS prediction.
When considering only BRCAm tumours, the 2yPFS was 82.5% with olaparib and 36.8% with placebo (HR 0.43; 95% Cl 0.209-0.868) with the PARPi response positive samples (Figure 16E). For Myriad HRD positive BRCAm tumours, 2yPFS was 79.3% versus 35.0% (HR 0.48; 95% Cl 0.241-0.971) (Figure 16F).
Whereas the "gene-specific copy number features" and "ASCAT segment copy number features" were combined in the final RF model/PARPi response, it is clear that either one of them can be included separately.
EXAMPLE 9. Methods relating to Example 8.
9.1. Patient and study design (PAOLA-1) As part of the ENGOT European HRD initiative, samples of the international phase III, randomised, double-blind PAOLA-l/ENGOT-ov25 trial (n=806) (NCT02477644) were shared. The PAOLA-1 trial evaluated first-line standard therapy, including bevacizumab, in advanced ovarian cancer with the addition of maintenance olaparib or placebo (Ray-Coquard et al. 2019, N Engl J Med 381: 2416-2428; ClinicalTrials.gov identifier NCT02477644; Pujade-Lauraine et al. 2021, Int J Gynecol Cancer 31 (Suppl 3):A208). During the trial, 468 samples were tested with the Myriad test, the remaining content of extracted DNA from formalin-fixed paraffin-embedded tumour tissue (FFPE) was shared with academic centres afterwards. 50-100 ng DNA of each sample was received in 96-well plates transported at -80°C. For the current analysis we used the extended PFS and OS (5yPFS and 5yOS) of PAOLA-1 at data cut off March 2022 (Ray-Coquard et al. 2023, Ann Oncol 34:681-392). For the development of our classifiers, the sample set was subsequently split into two equal cohorts of 198 patients, i.e., a training and validation cohort. Both cohorts were balanced according to treatment arm, BRCA status, initial response according to RECIST (before start maintenance treatment) and 2 years PFS (2yPFS) (Eisenhauer et al. 2009, Eur J Cancer 45:228-247). To evaluate the training set, the 198 samples of the training set were split internally by repeated 10-fold cross validation. The second half of the set was set aside to measure final model performance.
More information on the design of the training and validation cohorts is provided in Figure 15. From the total PAOLA-1 samples available (n= 468) only the samples with optimal ASCAT genome-wide copy number profiles were used to train and validate the new classifier model. In 72 samples, the ASCAT profiles were considered suboptimal; the most common reason for exclusion was low tumour content, resulting in a noisy profile. In 17 of these samples a pathogenic BRCA mutation could be identified. The remaining tumour samples with optimal profiles were divided into two cohorts and balanced for 2yPFS. The random forest based PARPi response classifier was trained on the samples with response data on olaparib (with bevacizumab) treatment.
9.2. Determination of progression-free survival (PFS), scarHRD score, ROC for the HRD scores, and hazard ratios (HR)
Investigator-assessed PFS, the primary endpoint of the PAOLA-1 trial, was defined as the time from randomization to investigator-assessed disease progression according to Response Evaluation Criteria in Solid Tumors vl.l (RECIST) or death (Eisenhauer et al. 2009, Eur J Cancer 45:228-247). Patients were censored as of their last tumour assessment in case of documented progression or death. For this post hoc analysis, duration of investigator-assessed PFS during PAOLA-1 was used to define the outcome subgroups. Patients with progression-free survival >2 years (progression after discontinuation of olaparib) were classified as the exceptional benefit subgroup; patients with disease progression within 2 years were classified as progression during PARPi maintenance treatment. Myriad test HRD scores were not available and therefore were approximated using the scarHRD package (version 0.1.1) for each sample. Receiver operating characteristic (ROC) for the HRD scores with respect to Myriad status were constructed using the pROC package (version 1.18.0). Kaplan-Meier curves are used to estimate PFS. Hazard ratios (HR) and associated 95% confidence intervals (Cl) were calculated using Cox proportional hazards regression. All analyses were not prespecified and are of exploratory intent.
9.3. Genome-wide loss-of-heterozygosity (LOH), allelic imbalance (Al), and copy number (CN) features The genome was divided into bins of 1000 kbp in size using the QDNAseq package vl.24.0 in R, excluding blacklisted regions. The LOH status of each bin is derived from ASCAT segmentation data based on overlap of segments with the midpoint of the bin. In case the minor allele of the corresponding segment has copy number zero according to ASCAT, the LOH status of the bin is set to positive. Similarly, in case the minor allele differs from the major allele, the Al status of the bin is set to positive. Finally, the total copy number of the bin is also registered. As such LOH/bin, Al/bin and CN/bin features were collected, and these are also referred to as "LOH bins", "Al bins" and "CN bins".
9.4. Gene mutational status
In general, for determining the mutational status of BRCA or other genes, the genes were sequenced for coding exons and flanking intronic sequences. All loss-of-function variants (frameshift, nonsense, canonical splice site), as well as missense variants already classified as pathogenic or likely pathogenic in public databases (BRCA-exchange, ClinVar, COSMIC, Seshat) were considered as clinically relevant mutations. A variant allele frequency of 5% was considered as a minimum threshold. The same public databases in combination with in-house clinical databases were used to classify missense variants as neutral variants or variants of unknown significance. Variants of unknown significance were classified as somatic BRCA wild type (BRCAwt) in further results.

Claims

1. A method of predicting the response of a subject having cancer to a therapy including a DNA damaging agent or including an agent inhibiting or impairing DNA repair, comprising: detecting in a test genomic DNA obtained from the subject the presence of mutations in the BRCA1 and BRCA2 genes, and classifying the subject as BRCA mutant if a mutation is present in the BRCA1 and/or BRCA2 gene, or as BRCA wild-type; and locating for a BRCA mutant subject the mutation to the BRCA1 Really Interesting Gene (RING) protein domain, the BRCA1 DNA- binding domain (DBD) protein domain, the BRCA1 C-terminal domain (BRCT) protein domain, other domains of BRCA1, the BRCA2 RAD51-binding protein, the BRCA2 DBD protein domain or other domains of BRCA2, therewith providing a BRCA domain mutation feature; detecting the loss of heterozygosity (LOH) status of a set of N bins of the test genomic DNA sequence, therewith providing a set of N LOH/bin features; wherein the N bins are of equal size smaller than 1.5 Mbp; wherein the N bins were selected from a genome-wide set of bins for their highest proportional contribution or most significant contribution to response to said therapy based on a set of reference genomic DNA samples including genomic DNA samples obtained from subjects responding to said therapy and including genomic DNA samples obtained from subjects not responding to said therapy; computer-implemented scoring of the set of N LOH/bin features and BRCA domain mutation feature of a BRCA mutant genomic DNA test sample with a data classification model trained to predict response to said therapy on the set of reference genomic DNA samples including genomic DNA samples obtained from subjects responding to said therapy and including genomic DNA samples obtained from subjects not responding to said therapy , wherein the training is based on the sets of N LOH/bin features and BRCA domain mutation features of the reference genomic DNA samples; or computer-implemented scoring of the set of N LOH/bin features of a BRCA wild-type genomic DNA test sample with a data classification model trained to predict response to said therapy on the set of reference genomic DNA samples including genomic DNA samples obtained from subjects responding to said therapy and including genomic DNA samples obtained from subjects not responding to said therapy, wherein the training is based on the sets of N LOH/bin features of the reference genomic DNA samples, predicting the subject to be a responder to said therapy when the trained data classification model scored the test genomic DNA sample as most closely resembling reference genomic DNA from subjects responding to said therapy; or predicting the subject to be a non-responder to said therapy when the trained data classification model scored the test genomic DNA sample as most closely resembling reference genomic DNA from subjects not responding to said therapy.
2. The method according to claim 1 further comprising: detecting a copy number feature of the test genomic DNA wherein the copy number feature is selected from a gene-specific copy number feature, an ASCAT segment copy number feature and a sigminer copy number feature; incorporating the detected copy number feature in the computer-implemented scoring of a BRCA mutant genomic DNA test sample or BRCA wild-type genomic DNA test sample wherein the data classification model is trained including the selected copy number feature as additional feature.
3. The method according to claim 2 wherein the gene-specific copy number feature is selected from amplification, deletion, LOH, focal amplification, focal deletion, focal LOH, and minimal copy number of the gene; the gene is selected from CEP43, ARID1B, PARP3, RBI, ARID1A, NF1, PABPC1, RNF19A, DNMT3A, CCNE1, PCM1, DECR1, BRCA1, TP53BP1, FANCL, EMSY, PALB2, TNKS2, ATR, CHEK1, CDK12, TBL1XR1, TCEA1, and EZH2; the ASCAT segment copy number feature is selected from fraction LOH, fraction deletions, ploidy, tumor ploidy (psi) of the sample, and goodness-of-fit; and the sigminer copy number feature is selected from the number of breakpoints per 10 Mbp, the absolute copy numbers of segments, the copy number change points between segments, the lengths of oscillating copy number segment chains, and the number of copy number events per chromosome.
4. The method according to claim 2 wherein the gene-specific copy number feature and ASCAT segment copy number feature are incorporated in the computer-implemented classification of a BRCA mutant genomic DNA test sample or BRCA wild-type genomic DNA test sample, wherein the data classification model is trained including the gene-specific copy number feature and ASCAT segment copy number feature as additional features.
5. The method according to any one of claims 1 to 4 further comprising: detecting the copy number (CN) status of a set of M bins of the test genomic DNA sequence and/or the allelic imbalance (Al) status of a set of P bins of the test genomic DNA sequence, therewith providing a set of M CN/bin and/or P Al/bin features; wherein the bins are of equal size smaller than 1.5 Mbp; wherein the M and/or P bins were selected from a genome-wide set of bins for their highest proportional contribution or most significant contribution of their associated CN/bin and/or Al/bin features to response to said therapy based on a set of reference genomic DNA samples including genomic DNA samples obtained from subjects responding to said therapy and including genomic DNA samples obtained from subjects not responding to said therapy; incorporating the detected M CN/bin and/or P Al/bin features in the computer-implemented scoring of a BRCA mutant genomic DNA test sample or BRCA wild-type genomic DNA test sample wherein the data classification model is trained including the M CN/bin and/or P Al/bin features as additional feature(s).
6. The method according to any one of claims 1 to 5 wherein the LOH status of a bin is based on the copy number of one or more single nucleotide polymorphisms (SNPs) of interest, and is detected as the mean of the LOH positive SNPs relative to all SNPs of interest within the bin, wherein a SNP is LOH positive when the copy number of the minor allele of the SNP is zero.
7. The method according to any one of claims 1 to 5 wherein the LOH status of a bin is based on the copy number of a nucleotide position in the bin, and is detected as the copy number of a segment of the genomic DNA with constant copy number comprising the nucleotide position in the bin, wherein a bin is LOH positive when the copy number of the minor allele of the nucleotide position in the bin is zero.
8. The method according to claim 6 or 7 wherein the copy number is detected based on untargeted or targeted sequencing of the genomic DNA.
9. The method according to any one of claims 1 to 8 wherein the trained data classification model is trained to predict a 2-year progression-free survival response to therapy including a DNA damaging agent or an agent inhibiting or impairing DNA repair.
10. A method of predicting the response of a subject having cancer to therapy including a DNA damaging agent or including an agent inhibiting or impairing DNA repair, comprising: detecting a gene-specific copy number feature of a test genomic DNA obtained from the subject wherein the gene is selected from CEP43, NF1, PCMl and TCEAl/FTIIS or is selected from CEP43, PCMl and TCEA1/FTIIS; predicting the response of a subject having cancer to said therapy based, or based in part, on the detected gene-specific copy number feature.
11. A method of prognosing survival or determining or assessing survival probability of a subject or patient having cancer or a tumor upon receiving treatment including a DNA damaging agent or including an agent inhibiting or impairing DNA repair, comprising: predicting, according to any one of claims 1 to 12, the subject to be a responder to the treatment ; prognosing the subject to have, or likely to have an increased survival or increased survival probability if predicted to be a responder to the treatment; or prognosing the subject not to have, or likely not to have an increased survival or increased survival probability if predicted to be a non-responder to the treatment.
12. A DNA damaging agent or therapy comprising a DNA damaging agent, or an agent inhibiting or impairing DNA repair or therapy comprising an agent inhibiting or impairing DNA repair for use in treating a tumor or cancer, comprising: predicting, according to a method of any one of claims 1 to 12, the subject to be a responder to the treatment, treatment regimen or therapy; administering the treatment, treatment regimen or therapy to the subject.
13. A computer program having instructions which when executed cause a computing or data processing system or device to perform a method according to any one of claims 1 to 11, or to perform a step of a method according to any one of claims 1 to 11.
14. A computing or data processing system or device, machine readable medium, or diagnostic kit comprising a means for carrying out or performing a method according to any one of claims 1 to 11, or to perform a step of a method according to any one of claims 1 to 11.
15. The methods according to any one of claims 1 to 11, the DNA damaging agent or therapy for use according to claim 12, the computer program according to claim 13, or the computing or data processing system or device, machine readable medium, or diagnostic kit according to claim 14 wherein the DNA damaging agent or an agent inhibiting or impairing DNA repair is an inhibitor of poly ADP ribose polymerase or is a platinum-based agent.
EP24786486.1A 2023-10-10 2024-10-09 Methods of determining response of a tumor to dna-damaging agents or to agents inhibiting or impairing dna repair Pending EP4791901A1 (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
EP23202688 2023-10-10
PCT/EP2024/078341 WO2025078404A1 (en) 2023-10-10 2024-10-09 Methods of determining response of a tumor to dna-damaging agents or to agents inhibiting or impairing dna repair

Publications (1)

Publication Number Publication Date
EP4791901A1 true EP4791901A1 (en) 2026-08-19

Family

ID=88315375

Family Applications (1)

Application Number Title Priority Date Filing Date
EP24786486.1A Pending EP4791901A1 (en) 2023-10-10 2024-10-09 Methods of determining response of a tumor to dna-damaging agents or to agents inhibiting or impairing dna repair

Country Status (3)

Country Link
EP (1) EP4791901A1 (en)
CN (1) CN122349571A (en)
WO (1) WO2025078404A1 (en)

Family Cites Families (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130079423A1 (en) 2010-02-24 2013-03-28 Myriad Genetics, Incorporated Diagnostic methods involving loss of heterozygosity
EP3327148B1 (en) 2010-06-18 2021-03-17 Myriad Genetics, Inc. Methods of predicting the status of brca1 and brca2 genes in a cancer cell
ES2704303T3 (en) 2010-08-24 2019-03-15 Dana Farber Cancer Inst Inc Procedures for the prediction of a response against cancer
JP6325453B2 (en) 2011-12-21 2018-05-16 ミリアド・ジェネティックス・インコーポレイテッド Methods and materials for assessing loss of heterozygosity
ES2687024T3 (en) 2012-02-23 2018-10-23 Children's Medical Center Corporation Procedures for predicting the cancer response
CA2867434C (en) 2012-06-07 2021-10-12 Institut Curie Methods for detecting inactivation of the homologous recombination pathway (brca1/2) in human tumors
DK2981624T3 (en) 2013-04-05 2020-03-02 Myriad Genetics Inc METHODS OF ASSESSMENT OF HOMOLOGICAL RECOMBINATION DEFICIENCY AND PREDICTION OF RESPONSE TO CANCER TREATMENT
CA2931181C (en) 2013-12-09 2023-01-24 Institut Curie Methods for detecting inactivation of the homologous recombination pathway (brca1/2) in human tumors
JP6663350B2 (en) * 2014-01-16 2020-03-11 クロヴィス・オンコロジー，インコーポレーテッド Use of PARP inhibitors to treat patients with breast or ovarian cancer showing loss of heterozygosity
DK3686288T3 (en) 2014-08-15 2023-05-22 Myriad Genetics Inc METHODS AND MATERIALS FOR THE ANALYSIS OF HOMOLOGOUS RECOMBINATION DEFICIENCY
CN114096681B (en) 2021-01-10 2024-06-21 行动基因(智财)有限公司 Homologous recombination deletion detection method and reagent set
US20260139317A1 (en) 2022-10-19 2026-05-21 Vib Vzw Method of Determining Loss of Heterozygosity Status of a Tumor

Also Published As

Publication number Publication date
WO2025078404A1 (en) 2025-04-17
CN122349571A (en) 2026-07-07

Legal Events

Date Code Title Description
STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: UNKNOWN

STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: THE INTERNATIONAL PUBLICATION HAS BEEN MADE

PUAI Public reference made under article 153(3) epc to a published international application that has entered the european phase

Free format text: ORIGINAL CODE: 0009012

STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: REQUEST FOR EXAMINATION WAS MADE