WO2023091967A1

WO2023091967A1 - Systems and methods for personalized treatment of tumors

Info

Publication number: WO2023091967A1
Application number: PCT/US2022/079985
Authority: WO
Inventors: Jason Liwei Guo; Shamik MASCHARAK; Deshka S. FOSTER; Jeffrey A. NORTON; Michael T. Longaker
Original assignee: The Board Of Trustees Of The Leland Stanford Junior University
Priority date: 2021-11-16
Filing date: 2022-11-16
Publication date: 2023-05-25

Abstract

Systems and methods for personalized treatment of tumors in accordance with embodiments of the invention are illustrated. One embodiment includes a method for determining prognosis for tumors, including obtaining a histological stain of a tumor, extracting a matrix ultrastructure of the tumor from the histological stain, obtaining an image of the tumor, where the image of the tumor is at sufficient resolution to demarcate individual cells, annotating each cell in the image with the respective cell type, determining cellular interactions between cells in the image based on the annotated cell types, providing the matrix ultrastructure and the cellular interactions to a prognosis model, obtaining an estimated survival time from the prognosis model.

Description

Systems and Methods for Personalized Treatment of Tumors

CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] The current application claims the benefit of and priority under 35 U.S.C. § 119(e) to U.S. Provisional Patent Application No. 63/264,159 entitled “Systems and Methods for Personalized Treatment of Tumors” filed November 16, 2021 , and U.S. Provisional Patent Application No. 63/383,256 entitled “Desmoplastic Stromal Signatures Predict Patient Outcomes in Pancreatic Ductal Adenocarcinoma” filed November 11 , 2022. The disclosures of U.S. Provisional Patent Application Nos. 63/264,159 and 63/383,256 are hereby incorporated by reference in its entirety for all purposes.

GOVERNMENT RIGHTS LICENSE

[0002] This invention was made with Government support under contract GM136659 awarded by the National Institutes of Health. The Government has certain rights in the invention.

FIELD OF THE INVENTION

[0003] The present invention generally relates to the diagnosis and treatment of tumors, namely the analysis and development of treatment protocols based on analyzed properties of desmoplastic tissue surrounding tumors.

BACKGROUND

[0004] Tumors are masses of abnormal cells. Tumors can be benign or malignant. Malignant tumors are often referred to as cancer. Cancer causes millions of deaths per year. Fibroblasts are a type of cell which synthesizes the extracellular matrix and collagen, produces the structural framework (desmoplasia), and plays a role in wound healing.

SUMMARY OF THE INVENTION

[0005] Systems and methods for personalized treatment of tumors in accordance with embodiments of the invention are illustrated. One embodiment includes a method for determining prognosis for tumors, including obtaining a histological stain of a tumor, extracting a matrix ultrastructure of the tumor from the histological stain, obtaining an image of the tumor, where the image of the tumor is at sufficient resolution to demarcate individual cells, annotating each cell in the image with the respective cell type, determining cellular interactions between cells in the image based on the annotated cell types, providing the matrix ultrastructure and the cellular interactions to a prognosis model, obtaining an estimated survival time from the prognosis model.

[0006] In another embodiment, the method includes providing an adjuvant therapy when the estimated survival time is low.

[0007] In a further embodiment, the method includes providing an adjuvant therapy when the estimated survival time is less than two years.

[0008] In still another embodiment, the prognosis model is a machine learning model trained using a training data set comprising matrix ultrastructures and determined cellular interactions from patients annotated with known survival times associated with the patients.

[0009] In a still further embodiment, annotating each cell in the image includes obtaining a plurality of biomarker concentrations for each cell, determining cell types by projecting biomarker concentrations onto a manifold containing cell types mapped to biomarker concentrations.

[0010] In yet another embodiment, obtaining the plurality of biomarker concentrations includes performing a co-detection by indexing (CODEX) assay of the tumor.

[0011] In a yet further embodiment, the plurality of biomarker concentrations includes concentrations for biomarkers selected from the group consisting of: aSMA, Vimentin, S100A4 (FSP1 ), PDGFRa, FAP, IL-6, IL-1 , MGP, YAP. CXCL-12 (SDF-1 ), CD26, CD56, HLA Class I, Collagen I, Collagen IV, Fibronectin, PD-1 , PD-L1 , CTLA-4, Ki67, CD4, CD8, CD11 c, CD20, CD31 , CD45, CD68, HLA DR, Pan Cytokeratin, and E-cadherin.

[0012] In another additional embodiment, determining cell interactions includes extracting cell type vectors f_a and fb representing the presence (1 ) or absence (0) of cell type a and b at each point in a sample of the tumor, generating an adjacency matrix, j, to identify neighboring points, calculating a plurality of single-cell interaction scores by applying the dot product of paired cell type vectors with the adjacency matrix (f_a-j-/b), and calculating normalized patient-level cell interaction scores by summing the single-cell interaction scores for a given cell type by the total number of cells in the sample of the tumor.

[0013] In a further additional embodiment, the adjacency matrix is generated using a k-nearest neighbors method where k is between 10 and 100.

[0014] In another embodiment again, the method further includes providing patient age, sex, and disease stage to the prognosis model.

[0015] In a further embodiment again, a system for determining prognosis for tumors includes a tissue analyzer, a processor, and a memory, the memory containing a tumor classification application that configures the processor to receive a histological stain of a tumor, extract a matrix ultrastructure of the tumor from the histological stain, obtain an image of the tumor from the tissue analyzer, where the image of the tumor is at sufficient resolution to demarcate individual cells, annotate each cell in the image with the respective cell type, determine cellular interactions between cells in the image based on the annotated cell types, provide the matrix ultrastructure and the cellular interactions to a prognosis model, and obtain an estimated survival time from the prognosis model.

[0016] In still yet another embodiment, the tumor classification application further configures the processor to recommend an adjuvant therapy when the estimated survival time is low.

[0017] In a still yet further embodiment, the tumor classification application further configures the processor to recommend adjuvant therapy when the estimated survival time is less than two years.

[0018] In still another additional embodiment, the prognosis model is a machine learning model trained using a training data set comprising matrix ultrastructures and determined cellular interactions from patients annotated with known survival times associated with the patients.

[0019] In a still further additional embodiment, to annotate each cell in the image the tumor classification application further configures the processor to obtain a plurality of biomarker concentrations for each cell from the tissue analyzer, and determine cell types by projecting biomarker concentrations onto a manifold containing cell types mapped to biomarker concentrations. [0020] In still another embodiment again, the tissue analyzer performs a co-detection by indexing (CODEX) assay of the tumor.

[0021] In a still further embodiment again, the plurality of biomarker concentrations includes concentrations for biomarkers selected from the group consisting of: aSMA, Vimentin, S100A4 (FSP1 ), PDGFRa, FAP, IL-6, IL-1 , MGP, YAP. CXCL-12 (SDF-1 ), CD26, CD56, HLA Class I, Collagen I, Collagen IV, Fibronectin, PD-1 , PD-L1 , CTLA-4, Ki67, CD4, CD8, CD11 c, CD20, CD31 , CD45, CD68, HLA DR, Pan Cytokeratin, and E- cadherin.

[0022] In yet another additional embodiment, to determine cell interactions, the tumor classification application further directs the processor to extract cell type vectors f_a and fb representing the presence (1 ) or absence (0) of cell type a and b at each point in a sample of the tumor, generate an adjacency matrix, j, to identify neighboring points, calculate a plurality of single-cell interaction scores by applying the dot product of paired cell type vectors with the adjacency matrix ( _a-j-/b), and calculate normalized patient-level cell interaction scores by summing the single-cell interaction scores for a given cell type by the total number of cells in the sample of the tumor.

[0023] In a yet further additional embodiment, the adjacency matrix is generated using a k-nearest neighbors method where k is between 10 and 100.

[0024] In yet another embodiment again, the tumor classification application further directs the processor to provide patient age, sex, and disease stage to the prognosis model.

[0025] Additional embodiments and features are set forth in part in the description that follows, and in part will become apparent to those skilled in the art upon examination of the specification or may be learned by the practice of the invention. A further understanding of the nature and advantages of the present invention may be realized by reference to the remaining portions of the specification and the drawings, which forms a part of this disclosure.

BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The description and claims will be more fully understood with reference to the following figures and data graphs, which are presented as exemplary embodiments of the invention and should not be construed as a complete recitation of the scope of the invention.

[0027] FIG. 1 illustrates a personalized treatment system in accordance with an embodiment of the invention.

[0028] FIG. 2 illustrates a personalized treatment device in accordance with an embodiment of the invention.

[0029] FIG. 3 is a flow chart for a personalized treatment process in accordance with an embodiment of the invention.

[0030] FIG. 4 is a table illustrating a guide for tissue classifications in accordance with an embodiment of the invention.

[0031] FIG. 5 shows differences in cell types in tumors having low survival time and high survival time.

[0032] FIGs. 6A and 6B illustrate the strength of correlation between different cell type interactions to survival and recurrence of the cancer, respectively.

DETAILED DESCRIPTION

[0033] Cancer is a serious disease which is responsible for millions of deaths per year. For many types of cancers, especially post-metastasis, treatment protocols are meant to extend life as long as possible, but are not currently able to cure the disease. Conventional medical techniques are capable of identifying various cancerous tissues and broadly classifying cancers into different types. Cancer types are typically determined by the type of tissue in which the cancer originates (histological type), and the primary site in the body in which the cancer developed. However, conventional classifications are often overbroad, leading patients to be told there is wide range of survivability. Aside from a large time window for expected survival is distressing to patients, these overbroad categorizations may further group treatment protocols where granularity in classifying the specific tumor would provide differential treatment plans between patients. For example, a patient with 2 months to live may opt for much more aggressive treatments in contrast with a patient who has 5 years to live.

[0034] Systems and methods described herein can more accurately predict survivability for a patient with a given tumor based on the spatial organization of tissue, including the scar tissue, that develops around the tumor. Fibroblasts which form the scar tissue are typically thought of as merely structural. However, fibroblasts are in fact immunomodulatory, and as shown below, significantly impact the growth and lethality of the tumor itself. In many situations, particular fibroblast growths can promote tumor growth and even metastasis. Fibroblast interaction with surrounding cells can also be clinically informative. Fibroblast tissue that is formed by the tumor and cellular interactions related to the fibroblast can be used to classify tumors according to survivability (i.e. risk stratification). By identifying the particular fibroblast activity and/or stromal cell activity specific to a patient’s tumor, a more accurate prognosis can be delivered. Further, the accurate prognosis and understanding of the specific spatial form of the tumor can be used to develop a personalized treatment plan that can increase the lifespan of the patient compared to conventional treatment protocols based on conventional classifications.

[0035] In numerous embodiments, specific checkpoint inhibitors can be utilized to alter the tumor and/or fibroblast activity of the patient. In various embodiments, a specific set of mechanomodulator drugs can be introduced. In many embodiments, adjuvant therapy can be introduced earlier and/or more aggressively than standard in response to particular fibroblast activity to produce more positive patient outcomes. For example, if estimated survival time is less than 2 years, more aggressive therapies may be immediately introduced.

[0036] By way of example, pancreatic ductal adenocarcinoma (PDAC) remains the only major cancer with a rising death rate in the United States and is projected to be the second leading cause of cancer deaths in the next decade. Many PDAC patients recur within 2 years after curative intent pancreatectomy and either preoperative and/or postoperative chemotherapy. Accordingly, the overall 5-year survival rate remains roughly 8-10%. PDAC and other solid tumors are defined by desmoplasia, an extensive fibrotic reaction that results from malignant cell crosstalk with surrounding stromal tissue. While, univariate morphological properties of the desmoplasia, such as collagen fiber orientation and density, have been correlated with differential patient outcomes, these approaches fail to capture the full geometric complexity and heterogeneity of ECM patterning observed in PDAC desmoplasia. Thus, the impact of comprehensive desmoplastic architecture on PDAC patient outcomes remains unknown. [0037] PDAC is further characterized by a complex tumor microenvironment (TME) with spatially defined tumor, stromal, and immune cell populations that drive overall disease progression and patient outcomes. There has been recent discussion of various cell phenotypes and biomarkers which may affect prognosis, however these analyses do not incorporate the prognostic value of spatial cell organization, which includes a myriad of cell-cell interactions, spatially defined cell communities, and extracellular matrix (ECM) architectures. Systems and methods described herein account for cell- and matrix-based spatial organization to predict differential patient outcomes in cancers like PDAC. Systems for personalized treatment of tumors are discussed in further detail below.

Personalized Treatment Systems

[0038] Turning now to FIG. 1 , a personalized treatment system in accordance with an embodiment of the invention is illustrated. System 100 includes a tissue analyzer 110. In numerous embodiments, the tissue analyzer is capable of spatially measuring biomarkers in a given tissue sample which can be used to identifying individual cell types. In many embodiments, Co-Detection by indEXing (CODEX) technology is used to automatically measure a large number of biomarkers. However, individual assays can be used, and/or any other biomarker measurement technologies as appropriate to the requirements of specific applications of embodiments of the invention. In various embodiments, the tissue analyzer is capable of producing histological stains of tumors as image data. More than one tissue analyzer may be used in a given system in order to perform different analyses. [0039] The system 100 further includes a personalized treatment device 120. In many embodiments, the personalized treatment device is a computing platform capable of performing personalized treatment processes as discussed herein. Personalized treatment devices can obtain data generated by tissue analyzers in order to quantify cellular interactions within a tissue sample, and determine an ultrastructure of the desmoplasia for the tissue sample. In numerous embodiments, histological stains are processed by personalized treatment devices in order to identify the ultrastructure. Example systems and methods for identifying ultrastructures are described in U.S. Patent Application No. 17/597,833 titled “Systems and Methods for Analyzing, Detecting, and Treating Fibrotic Connective Tissue Network Formation” filed July 27, 2022, the entirety of which is incorporated by reference herein.

[0040] Personalized treatment devices can use spatially recorded biomarkers to identify individual cell types in a given tissue sample. Based on the spatial arrangement of the individual cells, cell interactions can be computed as discussed in further detail below. Tissue analyzers can utilize cell interaction data and the ultrastructure to classify the risk of the tumor and/or produce an estimated survival time for the patient.

[0041] An output device 130 is included in system 100. Output devices are any device that can display a treatment plan and/or any other data produced by the system to a user. As can be readily appreciated, any or all of the tissue analyzer, personalized treatment device, and output device can be integrated into the same hardware platform as appropriate to the requirements of specific applications of embodiments of the invention. Tissue analyzers, personalized treatment devices, and output devices can be connected via a network 140. Networks can be wired, wireless, or any combination thereof. However, as can readily be appreciated, any number of different system architectures can be used without departing from the scope or spirit of the invention. For example, a personalized treatment device may directly obtain an ultrastructure and/or a tissue sample pre-labeled with cell types instead of obtaining them from tissue analyzers.

[0042] Turning now to FIG. 2, a personalized treatment device in accordance with an embodiment of the invention is illustrated. Personalized treatment device 200 includes a processor 210. Processor 210 can be any logic circuitry capable of performing entropybased treatment processes as appropriate to the requirements of specific applications of embodiments of the invention. For example, processor 210 can be implemented using a central processing unit (CPU), graphics processing unit (GPU), application-specific integrated circuit (ASIC), field-programmable gate array (FPGAs), and/or any other logic processing circuitry or combination thereof. Controller 200 further includes an input/output (I/O) interface 220. I/O interfaces can enable communication with other devices.

[0043] The controller 200 further includes a memory 230. The memory can be implemented using volatile memory, nonvolatile memory, or any combination thereof. The memory 230 contains a tumor classification application 232. In many embodiments, the tumor classification application directs the processor to carry out various personalized treatment processes such as those discussed herein. In many embodiments, the memory 232 variously contains at different points in time ultrastructure data 234, cell interaction data 236, and patient medical records 238. While particular system and computing architectures are illustrated in FIGs. 1 and 2, and discussed above, as can be readily appreciated, any number of different hardware architectures can be used as appropriate to the requirements of specific applications of embodiments of the invention.

Personalized Treatment Processes

[0044] Personalized treatment processes are capable of analyzing desmoplastic tissue surrounding a tumor in order to determine an effective, personalized treatment for the patient. In many embodiments, biomarkers for cells in and around the tumor are spatially analyzed to measure cell interactions. A machine learning model can be trained on a training data set that contains many different sample ultrastructures and associated cell interactions which have been annotated with patient outcomes. The trained machine learning model can then be presented with future patient samples to provide a classification.

[0045] Turning now to FIG. 3, a personalized treatment process in accordance with an embodiment of the invention is illustrated. Process 300 includes obtaining (310) histological stains of a tumor. In many embodiments, Masson’s Trichome is used to stain the samples. A matrix ultrastructure architecture is extracted (320) from the histological stains. In many embodiments, the image of the stains are normalized using any of a number of different normalization methods. In some embodiments, a red/green/blue histogram normalization and color deconvolution is applied where each pure stain is characterized by absorbances within three RGB channels. An ortho-normal transformation applied to said images can produce individual images corresponding to each color’s individual contribution to the image. As applied to trichome images, this can produce deconvoluted blue images corresponding to connective tissue fibers. Noise reduction of deconvoluted images can be achieved using an adaptive Winer filter which preferentially smooths regions with low variance, thereby preserving sharp fiber edges. Smoothed images can be binarized and processed through erosion filters with diamondshaped structuring elements to select fiber-shaped objects. A fiber network extracted from the filtered image can be used to measure fiber length, width, persistence, alignment, any/or any other physical fiber properties. In numerous embodiments, alternative stains can be used to produce similar results.

[0046] Cell types and locations within the tumor are identified (330). In numerous embodiments, an image of a tissue sample having at least cellular level resolution is annotated with the cell type of each cell. There are many different ways to identify cell types ranging from inefficient (e.g. manually) to very efficient. In many embodiments, CODEX is used to measure the presence of biomarkers.

[0047] These biomarkers can be spatially analyzed in the context of the image data in order to identify each individual cell in the sample. Biomarkers that can be used to identify cell types include (but not limited to): aSMA, Vimentin, S100A4 (FSP1 ), PDGFRa, FAP, IL-6, IL-1 , MGP, YAP. CXCL-12 (SDF-1 ), CD26, CD56, HLA Class I, Collagen I, Collagen IV, Fibronectin, PD-1 , PD-L1 , CTLA-4, Ki67, CD4, CD8, CD11 c, CD20, CD31 , CD45, CD68, HLA DR, Pan Cytokeratin, and E-cadherin. Individual cell types can be determined for each cell based on these biomarkers in accordance with the identification table in FIG. 4.

[0048] In various embodiments, cell types can be identified by anchor-based transferring measured biomarker values onto a premade CODEX manifold. CODEX manifolds can be generated by sampling a large number of CODEX samples. Cells in the CODEX samples are gated for marker expression, and individual protein channels are normalized to a Gaussian distribution. Protein expression is then batch corrected between specimens using mutual nearest neighbors, and projected onto a manifold by applying a dimensionality reduction via principal component analysis and a subsequent uniform manifold approximation and projection. As can be readily appreciated, any number of different methods can be used to label individual cells as appropriate to the requirements of specific applications of embodiments of the invention.

[0049] Interactions between cells are determined (340) based on the spatial arrangement of different cell types. In order to extract cell interaction data from the image and biomarker data, for all possible cell type pairings, cell-cell spatial interactions for pair of cells in the labeled tissue sample can be quantified first by extracting cell type vectors f_a and fb representing the presence (1 ) or absence (0) of cell type a and b at each point in the sample. Following this, an adjacency matrix, j, can be extracted to identify neighboring points. In numerous embodiments, a k-nearest neighbors approach is applied to extract the adjacency matrix. In various embodiments, neighboring points within k=20 nearest neighbors of each other are identified. However, k can be varied from 10-100 without departing from the scope or spirit of the invention. Single-cell interaction scores can be calculated using the dot product of paired cell type vectors with the adjacency matrix (i.e. fa-j-fb). To calculate normalized patient-level interaction scores representing the interactions between classes of cells (rather than individual cell pairs), the single-cell interaction scores (i.e. fa-j-fb) for the patient are summed and divided by the number of patient-specific cells. The normalized patient-level interaction scores reflect a transformation of a large amount of cellular interactions into a relatively small set of representative values which can be more efficiently processed. In many embodiments, it is these normalized patient-level interaction scores which are stored as the cellular interaction data and used for further processing. However, in various embodiments, the entire set of cell-level cellular interactions can be used.

[0050] In numerous embodiments, patient medical records are obtained (350). Patient medical information from the records, e.g. sex, age, disease stage, etc. can be combined into an input vector along with cell interaction data and the matrix ultrastructure architecture. In numerous embodiments, patient medical records are electronic medical records (EMRs, sometimes electronic health records- EHRs). These data are provided (360) to a prognosis model. In many embodiments, the prognosis model is a machine learning model. In various embodiments, the supervised machine learning model is trained using annotated training data which includes sets of tissue image data and cellular interaction data which are annotated with patient outcomes and treatments they underwent. In various embodiments, medical records for the patients are included in the annotated training data. As can be readily appreciated, any number of different models can be selected including (but not limited to) artificial neural networks, generalized additive models, k-nearest neighbor models, linear discriminant analysis, random forests, and support vector machines. The machine learning model can classify the based on the cellular interaction data and/or the image data. By incorporating this spatial information, a more accurate prognosis of the patient as well as a personalized treatment plan are obtained (370) from the machine learning model. In numerous embodiments, only a treatment plan is obtained. In some embodiments, only an estimated remaining lifespan is obtained. In numerous embodiments, the treatment plan indicates an earlier application adjuvant therapy in response to low estimated remaining lifespan and/or a high-risk classification. The patient’s tumor can then be treated (380) in accordance with the obtained treatment plan.

[0051] Turning now to FIG. 5, two example tumor samples annotated with cell types in accordance with an embodiment of the invention are illustrated. As can be seen, different spatial arrangements of cell types indicate different survival times. In the illustrated embodiment, the patient on the left likely has more than 2 years left to live, where the patient on the right is unlikely to survive longer than 2 years. FIGs. 6A-B illustrate the strength of correlation between different cell type interactions to survival and recurrence of the cancer, respectively.

[0052] Although specific systems and methods of personalized treatment of tumors are discussed above, many different methods can be implemented in accordance with many different embodiments of the invention. For example, in numerous embodiments, a profile of other non-fibroblast cells and their interactions can be provided to the machine learning model and/or otherwise used to refine the treatment plan. By way of non- exhaustive specific examples, if low immune cell interactions are present, it can indicate a need for immunotherapy. If high proliferating tumor cell interactions are identified, it can indicate a need for antiproliferative drugs.

[0053] It is therefore to be understood that the present invention may be practiced in ways other than specifically described, without departing from the scope and spirit of the present invention. Thus, embodiments of the present invention should be considered in all respects as illustrative and not restrictive. Accordingly, the scope of the invention should be determined not by the embodiments illustrated, but by the appended claims and their equivalents.

Claims

What is claimed is:

1 . A method for determining prognosis for tumors, comprising: obtaining a histological stain of a tumor; extracting a matrix ultrastructure of the tumor from the histological stain; obtaining an image of the tumor, where the image of the tumor is at sufficient resolution to demarcate individual cells; annotating each cell in the image with the respective cell type; determining cellular interactions between cells in the image based on the annotated cell types; providing the matrix ultrastructure and the cellular interactions to a prognosis model; and obtaining an estimated survival time from the prognosis model.

2. The method of claim 1 , further comprising providing an adjuvant therapy when the estimated survival time is low.

3. The method of claim 1 , further comprising providing an adjuvant therapy when the estimated survival time is less than two years.

4. The method of claim 1 , wherein the prognosis model is a machine learning model trained using a training data set comprising matrix ultrastructures and determined cellular interactions from patients annotated with known survival times associated with the patients.

5. The method of claim 1 , wherein annotating each cell in the image comprises: obtaining a plurality of biomarker concentrations for each cell; and determining cell types by projecting biomarker concentrations onto a manifold containing cell types mapped to biomarker concentrations.

6. The method of claim 5, wherein obtaining the plurality of biomarker concentrations comprises performing a co-detection by indexing (CODEX) assay of the tumor.

7. The method of claim 5, wherein the plurality of biomarker concentrations comprises concentrations for biomarkers selected from the group consisting of: aSMA, Vimentin, S100A4 (FSP1 ), PDGFRa, FAP, IL-6, IL-1 , MGP, YAP. CXCL-12 (SDF-1), CD26, CD56, HLA Class I, Collagen I, Collagen IV, Fibronectin, PD-1 , PD-L1 , CTLA-4, Ki67, CD4, CD8, CD11c, CD20, CD31 , CD45, CD68, HLA DR, Pan Cytokeratin, and E- cadherin.

8. The method of claim 1 , wherein determining cell interactions comprises: extracting cell type vectors f_a and fb representing the presence (1 ) or absence (0) of cell type a and b at each point in a sample of the tumor; generating an adjacency matrix, j, to identify neighboring points; calculating a plurality of single-cell interaction scores by applying the dot product of paired cell type vectors with the adjacency matrix ( _a-j-/b); and calculating normalized patient-level cell interaction scores by summing the single-cell interaction scores for a given cell type by the total number of cells in the sample of the tumor.

9. The method of claim 8, wherein the adjacency matrix is generated using a k- nearest neighbors method where k is between 10 and 100.

10. The method of claim 9, further comprising providing patient age, sex, and disease stage to the prognosis model.

11. A system for determining prognosis for tumors, comprising: a tissue analyzer; a processor; and a memory, the memory containing a tumor classification application that configures the processor to: receive a histological stain of a tumor; extract a matrix ultrastructure of the tumor from the histological stain; obtain an image of the tumor from the tissue analyzer, where the image of the tumor is at sufficient resolution to demarcate individual cells; annotate each cell in the image with the respective cell type; determine cellular interactions between cells in the image based on the annotated cell types; provide the matrix ultrastructure and the cellular interactions to a prognosis model; and obtain an estimated survival time from the prognosis model.

12. The system of claim 11 , wherein the tumor classification application further configures the processor to recommend an adjuvant therapy when the estimated survival time is low.

13. The system of claim 11 , wherein the tumor classification application further configures the processor to recommend adjuvant therapy when the estimated survival time is less than two years.

14. The system of claim 11 , wherein the prognosis model is a machine learning model trained using a training data set comprising matrix ultrastructures and determined cellular interactions from patients annotated with known survival times associated with the patients.

-15-

15. The system of claim 11 , wherein to annotate each cell in the image the tumor classification application further configures the processor to: obtain a plurality of biomarker concentrations for each cell from the tissue analyzer; and determine cell types by projecting biomarker concentrations onto a manifold containing cell types mapped to biomarker concentrations.

16. The system of claim 15, wherein the tissue analyzer performs a co-detection by indexing (CODEX) assay of the tumor.

17. The system of claim 15, wherein the plurality of biomarker concentrations comprises concentrations for biomarkers selected from the group consisting of: aSMA, Vimentin, S100A4 (FSP1 ), PDGFRa, FAP, IL-6, IL-1 , MGP, YAP. CXCL-12 (SDF-1 ), CD26, CD56, HLA Class I, Collagen I, Collagen IV, Fibronectin, PD-1 , PD-L1 , CTLA-4, Ki67, CD4, CD8, CD11c, CD20, CD31 , CD45, CD68, HLA DR, Pan Cytokeratin, and E- cadherin.

18. The system of claim 1 , wherein to determine cell interactions, the tumor classification application further directs the processor to: extract cell type vectors f_a and fb representing the presence (1 ) or absence (0) of cell type a and b at each point in a sample of the tumor; generate an adjacency matrix, j, to identify neighboring points; calculate a plurality of single-cell interaction scores by applying the dot product of paired cell type vectors with the adjacency matrix (f_a-j-/b); and calculate normalized patient-level cell interaction scores by summing the singlecell interaction scores for a given cell type by the total number of cells in the sample of the tumor.

19. The system of claim 18, wherein the adjacency matrix is generated using a k- nearest neighbors method where k is between 10 and 100.

-16-

20. The system of claim 19, wherein the tumor classification application further directs the processor to provide patient age, sex, and disease stage to the prognosis model.

-17-