IT202100012299A1 - Identification of genes and proteins linked to a disease using artificial intelligence - Google Patents

Description

Title: ?Identification of genes and proteins linked to a disease using artificial intelligence?.

DESCRIPTION

Technical field

The present invention is developed in the technical field of scientific research on proteins associated with pathologies.

State of the art

Basic scientific research leads to the discovery of new treatments and the genesis of pathologies by looking for particular proteins or compounds involved.

For research purposes can? be useful to identify proteins having different types of link with a disease. For example, the presence of a protein, or an interaction with a test protein, in a given biological sample, such as a biological fluid, can be a marker of the existence of a pathological condition. There? allows you to develop diagnostic methods of pathology, but also to understand more? precisely the nature and dynamics of the genesis and development of the pathology. This understanding can find utility in the treatment of pathology. In addition, a protein that was used in the cure ? always considered to be related to the disease.

The starting point for new research? represented by previous research, carried out around the world. The research results are collected in scientific articles, which describe for example the involvement of proteins, linked to one or more? different genes, in the development or treatment of a disease.

Prior art problem

A researcher who is interested in identifying a protein or compound linked to a disease initially relies on his own knowledge, reference texts or scientific articles for the selection of proteins deemed appropriate.

In this way the researcher identifies candidate proteins to satisfy his research objectives. Then tests, of known technique, are carried out to verify if the candidate proteins are actually linked to the pathology. The tests can for example be tests suitable for verifying the presence of the protein in a biological sample, for example antigen tests, in particular ELISA tests. In other research contexts, the tests can be tests suitable for verifying the theoretical effectiveness of the use of the protein in treating the disease.

By theoretical effectiveness is meant in particular the ability? of health technology in question to produce improvements in health when used in the most? favourable.

However, the literature on the subject extremely large and uneven. For example, different articles, also based on the cultural background of the authors and their geographical origin, may use different nomenclature for the same gene, the same protein or the same pathology.

In fact, the vastness of this literature does not allow a researcher to make the most of the work already? accomplished by others, who may have already? identified other proteins linked to the same or a similar pathology. In addition, the researcher is often not able to identify a protein which, although it has never been recognized as connected to a certain pathology, is actually? a candidate with high potential?, due to his affinity? with other proteins already? identified in related contexts.

Summary of the invention

An object of the present invention ? to facilitate the search and selection of entities? biological processes, such as genes and proteins, linked to a particular pathology.

AND? therefore the purpose of the invention is also to improve the quality? and the speed? of scientific research.

AND? moreover, the purpose of the invention is to facilitate an operator's access to information on entities? potentially relevant to its activity, but difficult to find from a direct search in the literature.

This and other purposes are achieved by a method for identifying entities? biological characteristics linked to a pathology, given by proteins or nucleic acids, and by a software engine of recommendation with artificial intelligence by entity? biological, according to any one of the joined claims.

The invention foresees that an input list of entities? biological proteins, selected from proteins and nucleic acids, such as protein-coding genes, is provided to a neural network, with the optional addition of disease names. Are entities included in the input list? that are relevant to the pathology, for example entity? already known to the operator, in his personal knowledge or on the basis of literature at his disposal, to be connected to the pathology or similar pathologies. Does the neural network have an overall list of entities available? biological, memorized and processed in the training phase of the neural network with millions of documents (papers or scientific references).

A neural network prediction function assigns scores in the overall list based on the input list that ? been provided. Based on the scores awarded, entities are selected biological to be recommended, expressed for example as names of publicly recognized genes, which can be tested by the operator, possibly after analysis of the targeted literature on the entities? recommended.

The neural network? equipped with a training function that corrects the prediction function, based on the information contained in a learning database. Does the learning database contain scientific articles relevant to the entities? biologicals of interest.

Advantageously, the learning performed by the neural network makes it possible to exploit the information contained in a vast literature, thus to subsequently select a limited number of entities? biologics for operator consideration.

Advantageously, the entities recommended biological entities can include entities? already found as connected to the same pathology, and therefore with test results easily verifiable in the literature and transferable to the contingent case, or entity? not yet recognized as related to the disease, but which are highly related to other entities? already considered, and which therefore represent good candidates for this new research.

BRIEF DESCRIPTION OF THE FIGURES

Figure 1 schematically shows a neural network model, used in one embodiment of the entity recommendation software. biological products according to the invention.

DETAILED DESCRIPTION

Is an artificial intelligence software recommendation engine for entities described here? organic. It also describes a method for locating entities? biological characteristics related to a given pathology, which makes use of the recommendation engine.

The invention can also be provided in the form of a computer-readable data carrier, in which the recommendation engine is contained.

The entities proteins of interest for the invention are proteins and nucleic acids, in particular genes, and in the continuation of the description the expression ?entity? organic? will be always used in this sense. Preferred examples are protein coding genes, protein complexes, cancer antigens, other molecules involved in human body pathways, fragments of known nucleic acids, messenger RNA, etc.

The recommendation engine comprises an input interface and an output interface.

The input interface ? configured to receive a list of inputs, containing the indication of one or more? entity organic. By entering the input list, the user queries the recommendation engine, i.e. requests a recommendation. The input list ? generated by the user. As more detailed below, gi? during the generation of the input list, suggestions can be provided by the recommendation engine to disambiguate the existing aliases in the community? scientific. ? Is the input list expected to include entities? pertinent to the research or to the specific pathology chosen, in any case within the knowledge of the user.

The indications included in the input list are preferably textual and can be suggested by the system. In other words, each entity ? represented by a string of characters. In particular, in the case of genes ? a representation by nomenclature Gene Symbol HGNC Hugo Gene Nomenclature Committee (https://www.genenames.org/) is preferred. In the case of entities different from the genes pu? the corresponding internationally recognized standard be used.

Optionally, the input interface ? configured to also receive the indication of one or more? disease names. Where the input interface allows both of these types of data to be entered, it is not? however, it is necessary that the user, or operator, inserts them all, but the recommendation engine can? work even only with partial data that omit, for example, the pathology and include only the entity? organic.

The indication of one or more? pathologies preferably includes the pathology for which the entity is to be identified? related biologicals. Alternatively, ? it is permissible that pathologies related to it are indicated.

In the preferred embodiment, the input interface comprises a transcoding function, configured to recognize as equivalent different nomenclatures, or different forms of textual representation of the same entities? organic. For this purpose, the transcoding function has a correspondence table in memory which, for at least some entities, biological, indicates different textual representations of each entity.

Optionally, the correspondence table also designates text representation forms of a protein with text representation forms of a gene that codes for the protein as equivalent.

The transcoding function, therefore, allows to recognize inputs that are supplied with different possible forms of representation that can be used by the user, that have a recognized reference in the literature, so that have been implemented in the correspondence table.

In addition, the input interface can? include a writing suggestion function that makes use of the transcoding function. The writing suggestion function ? configured, while a user inserts in the input interface at least one? entity? biological of the input list, to suggest to the user a complete name of an entity? biological partially written, or a? entity? biological equivalent to? entity? written by the user according to the correspondence table.

The output interface ? configured to provide one or more outputs? entity recommended organic products, i.e. a list of outputs. Are all entities preferably excluded from the output list? biologics contained in the input list.

Below, in table 1 ? indicated an example of input list supplied to the recommendation engine, with entity? biology related to diabetes. In table 2 ? indicated the output list obtained.

Table 1

Table 2

Received the indication of one or more? entity recommended organic, ? It is possible for the user to select at least one of them for his own research. This selection can take place following a? further consultation, by the user, of the literature on the entities? recommended organics. For this purpose, the output list ? preferably accompanied by hypertext links that redirect to texts of scientific literature relevant to the entities? recommended. Preferably, the hypertext links select and search the major scientific literature sites for the suggested references (for example the series of genes, in which pathology they are involved, in which pathways, etc.).

In the preferred embodiment, since the recommendation is supplied as a gene symbol, it becomes possible for the researcher to carry out known art tests aimed at verifying whether at least one protein encoded by one of the entities recommended biologics whether or not it is related to the disease.

In alternative embodiments, if the recommendation is rendered in the form of protein names, the tests are performed directly on the entities? recommended organics.

The tests on the connection of the protein with the pathology can be suitable tests to verify if, at least in conjunction with a temporal phase of the pathology, the protein is present in at least one biological sample of a subject with the pathology, or if in the presence of the pathology the protein has a concentration other than a predetermined concentration detectable in non-pathological conditions, or whether the protein permits diagnosis of disease when used as a reagent in a diagnostic kit, or whether the protein is effective in the context of disease treatment.

The entities biological signals to be recommended by the output interface are identified by a neural network integrated in the recommendation engine, globally indicated in figure 1 with the reference 100.

In detail, the neural network 100 comprises a trainable embedding function, configured to vectorically represent the entities organic.

In the event that the input list contains indications of entity? in textual form, the embedding function preferably comprises a conversion function from textual form to integer form, indicated in the figure with 101 (input genesymbol list, or list of Gene Symbols in input), and a conversion function from numbers integers 101 to vectors 102.

In the illustrated embodiment, the embedding function also comprises a list uniforming function, configured to bring the input list to a predetermined length, indicated in figure 1 as max_length (maximum length). There? ? obtainable, for example, by adding elements of null value to the input list until the pre-established length is reached. Preferably, the null value elements are added to the level where the entities? biological are in the form of whole numbers 101.

For the purpose of converting from text form to integer form 101, the neural network has in memory an coding dictionary, in which text strings are associated with integers. Instead, the conversion function from integers 101 to vectors 102 ? trainable type.

In the preferred embodiments, the vector representation is a representation with vectors 102 having a number of dimensions greater than fifty, indicated in figure 1 as emb_dim (embedding dimensions). The number of dimensions affects how other neural network functions work, described below. In general, a low number of dimensions results in lower quality recommendations, because Not ? Is it possible to take into account all the modalities? interactions between different proteins. On the other hand, too many dimensions require more training data, and risk increasing the sensitivity too much. of the neural network to any isolated training data technically distant from the others. Once the number of dimensions has been selected, algorithms known to experts in the sector can be used for the vectorial conversion of the embedding.

The embedding function ? used on entities? biological input list. However, the same embedding function ? also used for the vectorial representation of entities? biological with which the neural network is confronted in other processes, for example in? training, described below. In detail, does the neural network have an overall list of entities in its memory? biological, and the embedding function allows the vectorial representation of each of them.

Similarly, the embedding function can? also operate on the indications of pathologies, which are also the subject of a comprehensive list. Also the pathologies, therefore, have a form of representation in integer numbers 103 (input disease list, or list of pathologies in input) and in vectors 104. However, in the following for simplicity? will you do? reference to entities only? organic.

In addition, the embedding function ? preferably configured to operate the reverse transformation if necessary, or to convert entities? biology from vector representation 102 to integer representation 101 or textual.

The neural network then comprises a trainable prediction function, indicated globally with 110. The prediction function 110 ? configured to process the input list, precisely in the form of vectorial representation 102 obtained downstream of the embedding function. Processing the input list leads to assigning a score to the entities? biologicals of the overall list.

So, the neural network ? configured to select from the overall list one or more? entity biologics to recommend, based on the assigned score, which are provided to the output interface. For example, do you choose a predetermined number of entities? biological (in figure n_genes, or number of genes), given by the entity? higher scoring organic crops, preferably provided they are not included in the input list.

In the preferred embodiment, the trainable prediction function 110 ? composed of the following layers, or layers:

- an average layer 111 (average layer, indicated in figure 1 also as emb average, or embedding media), configured to average the vectorial representations 102 of the entities? biological present in the input list, and if present also the vectorial representations 104 of the pathologies,

- if the input of pathologies is also envisaged, a concatenation layer 112 (concatenate layer, also indicated in the figure as user embedding layer, or user embedding layer), configured to concatenate the averages 111 of the entity vectorial representations 102, 104 ? biological and pathological

- a series of fully connected layers 113 (fully connected network), with decreasing size and activation function ReLU, and

- an output layer 114 (output layer) with SoftMax activation function (in the figure, SoftMax Output Layer).

Each of these layers and activation functions are known to those skilled in the art, and are therefore not explored in detail herein. Furthermore, other prediction function models which can achieve substantially equivalent results for the purposes of the invention can be obtained by a person skilled in the art, such as multilayer perceptrons, convolutional neural networks, or recurrent neural networks.

The neural network includes a training function, configured to train the prediction function. Thanks to the training, therefore, the prediction function assigns scores, useful for the selection of the entities? to be recommended, which reflect correlations between the entities? biological, and possibly also with pathologies, which can be obtained from training data.

Pi? in detail, the training function ? configured to obtain learning lists of entities? biological, and optionally pathologies, from a learning database, and to modify the prediction function and the embedding function, based on the learning lists. Therefore, the prediction function, receiving a specific same list of inputs, ? configured to attribute different scores to the same entity? of the overall list, depending on whether the input is provided before or after the modification made by the training function.

Preferably, the training function ? configured to change the prediction function, as well as? the embedding function, by gradient descent, with mode? known to the technicians of the sector. A suitable gradient descent function? given by Stocastic Batch Gradient Descent, with variable learning rate.

According to one aspect of the invention, the learning database comprises a plurality of? of texts of scientific literature containing references to entities? organic. Preferably such texts belong to the medical sector. An example of texts are scientific articles. Various databases of this type are available on the internet and accessible to the public for free, others are licensed for payment.

In some embodiments, the neural network includes a text mining function. The text mining feature ? configured to analyze the texts of the database, cos? to identify in each text the entity? biological data to which reference is made, and optionally pathologies or other data which are deemed useful. In other embodiments, the text mining function can not be expected, as some public databases, usable as a learning database, offer the texts already? labeled to contain entity references? biological without needing to consult the entire text. In some public databases similar labels are obtained precisely through text mining functions, which therefore are not? need to repeat by the recommendation engine.

An optional phase, following the identification of the references to the entities? biological in the texts, ? given by the exclusion of texts that contain an excessive number of references to entities? biological, in particular above a pre-established maximum number of references. The maximum number of references pu? be, for example, between 5 and 50. Ci? allows you to exclude from? learning texts that are actually? simple catalogs of genes or proteins, as well as? texts of a too general nature, rather than specific to a pathology or a family of pathologies.

As anticipated, the training function ? configured to obtain from the learning database learning lists of entities? organic. One of the operations carried out for this purpose, before or after text mining and the elimination of texts with too many references, consists in grouping the plurality of texts into subsets. of texts in the database.

In particular, the neural network comprises a grouping function which has at least one grouping criterion in memory.

The grouping function ? configured to group in each subset texts of the learning database that are uniform on the basis of one or more? of the grouping criteria.

An example of a preferred grouping criterion ? given by at least one pathology referred to in the text. As commented, the reference to the pathology pu? be itself obtained by the function of text mining, or pu? be pre-indexed using tags included in the starting database.

There? in general it also applies to the other possible grouping criteria. In other words, the grouping function ? configured to verify the consistency of each text with each pre-selected grouping criterion by means of text mining and/or by consulting the labels pre-assigned to the text.

Other examples of grouping criteria are given by the presence in a text of quotations that point to one or more? other texts, from a common collection in which several texts are contained, for example a particular journal that contains articles or of which more than one have been published? editions, by an author of the various texts, by a publisher who oversaw their publication, by a title that refers to a topic covered, by a period of publication of the texts, and by the presence of keywords.

In some embodiments, the learning lists are generated with reference to a single grouping criterion.

In other embodiments, the learning lists are generated with reference to a combination of grouping criteria, by means of the logical operator ?And?. For example, a combined policy can be given by the combination of a publisher with a presence of keywords in the text. Another combined criterion can? be the presence of citations between different texts, and a common period of publication of the texts.

In still other embodiments, more? grouping criteria, or combinations of grouping criteria, can be used in parallel. For use in parallel we mean here that a first plurality? of learning lists are obtained through one or more? first criteria, and a second plurality? of learning lists are obtained through one or more? second criteria. Obviously, further pluralities can also be contemplated? after the second.

In these embodiments, the same text can fall into different lists. For example, the first plurality? of lists can? be originated by dividing the database into subsets on the basis of the sole presence of references to pathologies, and the second plurality? of lists can? be originated by re-dividing the same database into subsets on the basis of the collections to which the texts belong (in a completely independent way to the other subdivision carried out). Each text will belong then to one of the subsets originating from the first criterion, and to one of the subsets originating from the second criterion.

In this way, starting from the same database, increases the number of lists available with entity? biological systems that are related to each other by context.

Indeed, thanks to all these possible modalities? of grouping, the presence of two texts in the same subset ? related to an affinity? of content, such that the entity? biological processes linked to the texts, even where the grouping criterion is not specifically the pathology being treated, appear to belong to the same conceptual area. There? is reflected in the property? municipalities or similar structures of the various entities? organic. In summary, these entities they have in common the fact that they have been cited in the same context.

The training function ? then configured to create, for each subset, a respective learning list, made up of entities? to which the texts of the subset refer.

Thanks to the described learning, when the prediction function receives the input list, the entities? that are recommended are those that, in the texts of the learning database, have a high degree of correlation with the entities? of the input list. So where the input list contains entities? related to a given research or a pathology, the entities? biological recommended frequently have functionality? useful in correlation to the same research or pathology. There? can? be verified also for entity? recommended that have never been described in the texts of the learning database in connection with the specific pathology of interest.

Once the learning lists are created, an example of a training function can be configured, for each learning list, for:

- select an entity in the learning list? biological test,

- provide the prediction function, as an input list, the deprived learning list of the entity? of tests,

- receive from the prediction function one or more? entity recommended biological results,

- check if the entity? of tests ? included among the entities result, and based on what? generate a feedback composed of correction gradients for the prediction function, and optionally for the embedding function.

In this process, it should be noted that the learning lists are subjected by the input interface and the neural network to the same processing as each input list. Therefore, the transcoding function allows the prediction and training functions to recognize correlations independently of the different and non-uniform nomenclatures with which the entities can be represented. biology in the texts of the learning database.

Furthermore, the learning lists undergo the same layers as the prediction function. The embedding function, as well as? various layers of the prediction function (including the series of fully connected layers 113), offer different vector representations that evolve during training. Therefore, if at the beginning of the training the vectorial representations of the entities? behaviors can be randomly distributed in the vector space, do the correction gradients modify the vector representation of the entities for each training list that is supplied? biological involved. In particular, the vector relative to the entity? biological test approaches in one or more? dimensions to the vectors relative to the other entities? of the same training list.

Thanks to the presence of the media layer, the representative vectors of all the entities? biological data provided in a training list are combined to form a single vector of averages. It should be noted that this average vector, in the vectorial representation internal to the neural network, is connected to the complex of entities? biology referred to in all texts of a particular subset of the learning database, obtained with a particular value of the grouping criterion.

So, does the average vector synthetically represent the entities? biology related to a disease, a group of texts with reciprocal citations, a collection of texts, an author, and/or a period of publication.

Similarly, once the training is complete, the input list supplied by the user is transformed into an average vector, which ? interpretable by the neural network as representative of the conceptual area for which the user is asking the query. Advantageously, with the selection, for the input list, of entity? biological related to a disease, the entity? that will be recommended are similar in vector to the average vector obtained, and therefore also pertinent to the pathology, and potentially useful for its diagnosis, understanding or treatment.

Obviously a technician of the branch can? make numerous modifications equivalent to the variations set out above, without thereby departing from the scope of protection defined by the appended claims.

Claims (9)

1. Method for identifying entities? biological, selected from proteins and nucleic acids or parts thereof, linked to a pathology, including: - generate an input list of entities? biological, selected among proteins and nucleic acids, pertinent to the dynamics or to a treatment of the pathology, - provide the input list to an entity-based artificial intelligence recommendation software engine? biological, having a neural network (100), - said neural network (100) having in memory an overall list of entities? biological, selected among proteins and nucleic acids, - process the input list, by the neural network (100), through a trainable prediction function (110), in such a way as to assign a score to the entities? biological of the overall list, - provide in output, by the recommendation engine, one or more? entity recommended biologicals, extracted from the overall list based on the score attributed by the neural network (100), e - test at least one of the entities? recommended biological agents, or a protein encoded by it, to verify its presence in conjunction with the pathology or a theoretical efficacy in the treatment of the pathology, wherein the neural network (100) comprises a training function, configured to derive learning lists of entities? biological proteins, selected among proteins and nucleic acids, from a learning database, and to modify the prediction function based on the learning lists, and in which the learning database includes a plurality? of texts of scientific literature containing references to entities? biological, selected among proteins and nucleic acids. 2. Method according to claim 1, wherein the training function ? configured to obtain learning lists using: - group the plurality into subsets? of texts in the database, the texts of each subset being uniform for at least one grouping criterion, preferably a pathology, a presence of reciprocal citations, a publication collection, an author, a publisher, a title, a publication period and/or a keyword presence, e - create for each subset a respective learning list composed of entities? to which the texts of the subset refer. 3. Method according to claim 2, wherein the neural network (100) comprises a text mining function, configured to analyze the texts of the database, so as to identify in each text at least one entity? biological to which it refers. 4. A method according to any claim from 1 to 3, wherein the neural network (100) comprises an embedding function, configured to vectorically represent each entity? biological, in which preferably the vector representation (102) ? a representation with a number of dimensions greater than fifty. 5. Method according to any claim from 1 to 4, wherein the training function ? configured to modify the prediction function (110) by gradient descent. The method according to any claim from 1 to 5, comprising: - provide in input to the recommendation engine, in addition to the input list, also the indication of one or more? pathologies, in which preferably the texts of scientific literature contain, in addition to said references to entities? biological, also references to pathologies. The method according to any claim from 1 to 6, wherein the trainable prediction function comprises: - a media layer (111) configured to mediate vector representations of entities? biological (102) present in the input list, and optionally vector representations of indications of pathologies (104), - optionally, a concatenation layer (112) configured to concatenate the averages (111) of the vector representations of entities? biological and pathological (102, 104), - a series of fully connected layers (113), with decreasing size and ReLU activation function, and - an output layer (114) with SoftMax activation function. 8. Method according to any claim from 1 to 7, wherein the neural network (100) comprises a transcoding function, configured to recognize as equivalent different forms of textual representation of the same entity? biological, for at least some entities? organic. 9. Artificial intelligence software recommendation engine for entities? biological, selected from proteins and nucleic acids, including: - an input interface, configured to receive an input list of entities? biological, selected among proteins and nucleic acids, - a neural network (100), having in memory an overall list of entities? biological networks, selected from proteins and nucleic acids, the neural network (100) being configured for: - process the input list using a trainable prediction function (110), so as to assign a score to the entities? biological of the overall list, e - select from the overall list one or more? entity organic products to be recommended, based on the score assigned, - an? output interface, configured to output one or more? entity biologics to recommend that have been selected by the neural network (100), wherein the neural network (100) comprises a training function, configured to derive learning lists of entities? biologicals, selected among proteins and nucleic acids, from a learning database, and to modify the prediction function based on the learning lists, in which the learning database includes a plurality? of texts of scientific literature containing references to entities? biological, selected among proteins and nucleic acids.