Methods and systems for predicting gene expression perturbations

Abstract

The present invention provides a computer-implemented method for predicting gene expression perturbations. According to an embodiment of the present invention aimed at improving the quality of predictions, the method includes the steps of generating a knowledge graph KG (104) from domain knowledge, the KG (104) describing relationships including associations, similarities, and / or interactions between a plurality of entities, the entities including at least some genes and perturbation agents, training a machine learning (ML) model (118) to predict perturbed gene expression from pre-perturbation gene expression data (106) and learned embeddings of the KG (104) entities, and using the trained ML model (118) to provide gene expression data obtained from a genetic sample from a subject and predict the response of the genetic sample in terms of changes in gene expression resulting from applying one or more perturbation agents to the genetic sample.

Description

[Technical field]

[0001] The present invention relates to methods, systems, and computer readable media for predicting gene expression perturbations. [Background technology]

[0002] Gene expression profiling is the simultaneous measurement of the activity of multiple genes expressed by a cell. At a macro level, expression profiling provides information about the physiological state of a cell, tissue, or organism that can be used to detect and diagnose a disease or condition. Comparative studies have shown that in drug response prediction tasks, gene expression profiles may consistently provide the highest predictive power among individual profiling datasets (for reference, see Costello, J., Heiser, L., Georgii, E. et al. A community effort to assess and improve drug sensitivity prediction algorithms. Nat Biotechnol 32, 1202-1212 (2014). https: / / doi.org / 10.1038 / nbt.2877). Prediction and analysis of changes in gene expression profiles under exposure to external perturbagens may provide insight into how organisms respond to changes in environmental factors or therapeutic treatments (for reference, see Mannheimer, JD, Prasad, A. & Gustafson, DL Predicting chemosensitivity using drug perturbed gene dynamics. BMC Bioinformatics 22, 15 (2021). https: / / doi.org / 10.1186 / s12859-020-03947-y).

[0003] A wide range of statistical and machine learning methods have been employed to address the problem of response prediction. Classical statistical methods analyze control and perturbed gene expression profiles to identify genes that experience significant changes in expression levels. Various studies have used linear and nonlinear methods, most notably regression models, kernel methods, neural networks, and ensemble methods. Recent supervised machine learning models offer the opportunity to predict gene perturbations using multi-omics data and multi-task learning approaches that leverage information across cell lines or tissues, and perturbant similarities (see for reference Xia, F., Allen, J., Balaprakash, P. et al. A cross-study analysis of drug response prediction in cancer cell lines, Briefings in Bioinformatics (2021). https: / / doi.org / 10.1093 / bib / bbab356).

[0004] However, existing models still lack the simultaneous integration of multi-omics data with prior domain knowledge of molecular interactions. In biomedical applications, omics data modalities such as transcriptomic gene expression profiles may reflect the patient's individualized context, which determines how molecular interactions work in a specific case, to predict the impact of drug exposure. For example, if a drug molecule is designed to control some protein by inhibiting its expression, but the corresponding gene is already suppressed in the patient's tissue for some reason, the drug is unlikely to work as expected. Drug effects may also be modulated by other drugs or environmental factors through a series of molecular interactions that should be considered by the model. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Costello, J., Heiser, L., Georgii, E. et al. A community effort to assess and improve drug sensitivity prediction algorithms. Nat Biotechnol 32, 1202-1212 (2014). https: / / doi.org / 10.1038 / nbt.2877 [Non-Patent Document 2] Mannheimer, JD, Prasad, A. & Gustafson, DL Predicting chemosensitivity using drug perturbed gene dynamics. BMC Bioinformatics 22, 15 (2021). https: / / doi.org / 10.1186 / s12859-020-03947-y [Non-Patent Document 3] Xia, F., Allen, J., Balaprakash, P. et al. A cross-study analysis of drug response prediction in cancer cell lines, Briefings in Bioinformatics (2021). https: / / doi.org / 10.1093 / bib / bbab356 [Non-Patent Document 4] https: / / arxiv.org / pdf / 1703.06103.pdf Summary of the Invention [Problem to be solved by the invention]

[0006] It is therefore an object of the present invention to improve and further develop methods and systems for predicting gene expression perturbations such that the quality of the predictions is improved. [Means for solving the problem]

[0007] According to the present invention, the above object is achieved by a computer-implemented method for predicting gene expression perturbations, the method comprising the steps of: generating a knowledge graph KG from domain knowledge, the KG describing relationships including associations, similarities, and / or interactions between a plurality of entities, the entities including at least some genes and perturbation agents; training a machine learning (ML) model to predict perturbed gene expression from pre-perturbation gene expression data and learned embeddings of the KG entities; and using the trained ML model to provide gene expression data obtained from a subject-derived genetic sample and predict the response of the genetic sample in terms of changes in gene expression resulting from applying one or more perturbation agents to the genetic sample. Furthermore, the above object is achieved by a system for predicting gene expression perturbations, and a tangible non-transitory computer-readable medium, as defined in the independent claims.

[0008] According to an embodiment, the present invention provides a method and system configured to predict a response to a perturbation agent presented to a cell, tissue, or organism. In the context of the present invention, the term "perturbation agent" should be understood in the broadest sense and may include, among others, for example, drug substances, chemical compounds, environmental factors, etc. The response may be quantified as a change in gene expression levels and may be used for downstream analysis of biomedical applications. An embodiment of the present invention provides a method for incorporating prior domain knowledge in the form of heterogeneous graphs describing associations, similarities, and / or interactions between entities (e.g., drugs, genes, diseases). More specifically, predicting gene expression perturbations may be provided by using drug and gene embeddings derived from graph representation learning based on experimentally validated properties and topological relationships of the entities in a network of heterogeneous molecular interactions. By leveraging these known relationships to learn perturbation agent and gene embeddings and using the entity embeddings to predict perturbed gene expression, additional structure and knowledge about the underlying biological mechanisms that improve the accuracy of the predictions are introduced.

[0009] According to embodiments of the present invention, it is recognized that the quality of predictions can be significantly improved by integrating prior knowledge as a relational KG (heterogeneous graph) in the process of gene expression level prediction, or in particular to search for reusable drugs for neoantigen combination therapy. In contrast to prior art pipelines that analyze perturbed gene expression profiles, embodiments of the present invention consider the changes in gene expression that occur when cells are exposed to various perturbants and contextualize it with domain knowledge about the interactions between genes, perturbants, and other entities (e.g., information about (types of) drug-target interactions). Properties of the entities (e.g., information about the properties of drugs, such as the encoded molecular structure of the drug / perturbant) are treated as node features in the knowledge graph and are therefore included in the predictive model. This complements the information captured only from the graph topology (relationships) and improves the quality of predictions. Since gene expression assays are measured directly from the RNA of a specific patient, the predictive output provided by embodiments of the present invention is personalized.

[0010] Note that in general, the methods described herein rely on there being enough information to first construct the KG and gene expression data. Clearly, the granularity of the domain KG (number of entities, entity features, relationship types, etc.) may affect predictions. Furthermore, embodiments of the present invention assume that genes and perturbants from gene expression assays are also represented in the KG.

[0011] According to one embodiment of the present invention, entities in the knowledge graph may include associated attributes describing the respective entities, which may provide entity features including gene ontology GO annotations, expressed protein sequences of genes, and / or Simplified Molecular Input Line Entry System (SMILES) strings of drug molecules.

[0012] According to one embodiment of the present invention, learning of embeddings of KG entities can be performed by (known) node representation learning methods (e.g., using DISTMULT factorization, DeepWalk followed by skip-gram) that capture the graph topology into node features while solving the link prediction problem.

[0013] According to one embodiment of the present invention, the ML model may be trained sequentially by first learning embeddings of the KG entities and then training downstream models to make final predictions of gene expression based on the learned embeddings of the KG.

[0014] Alternatively, it may be provided that the ML model is trained end-to-end by simultaneously learning the embeddings of KG entities and the prediction of gene expression. In this context, it is noted that the learning in the end-to-end implementation is not guaranteed to converge to the optimum for a given loss function. Also, exploration of more robust modular architectures, losses, and deep metric learning / negative sampling methods may bring improvements here.

[0015] According to one embodiment of the present invention, it may be provided that the contextual representation of KG entities is obtained by training an ML model, for example an R-GCN (Relational Graph Convolutional Network) model, to predict the correct edges between entities in the KG (link prediction). In this context, a loss function J that encourages the ML model to score observable positive KG triplets higher than negative ones, to optimize the cross-entropy loss, LP It may be provided that is used.

[0016] According to one embodiment of the present invention, a loss function J is used to optimize a regression loss based on the distance between the ground truth and predicted values ​​of gene expression levels to predict the perturbation effect in a drug-gene pair. rcan be used to predict perturbation effects in drug-gene pairs. Ground truth includes actual (i.e., known or measured) gene expression values. Depending on the implementation, ground truth may include perturbed gene expression values ​​or changes in gene expression levels caused by a perturbant.

[0017] According to one embodiment of the present invention, the loss function J LP and the loss function J r The loss function J is determined as a weighted combination of J = αJ LP +(1-α)J r It can be provided that the parameters of the ML model are learned by optimizing α∈(0,1), where α∈(0,1) is a hyperparameter that controls the tradeoff between preserving the structure of the KG and being able to accurately predict perturbed gene expression levels.

[0018] According to one embodiment of the present invention, the method may be configured to find drugs that alter the expression of desired neo-antigens. In this context, the method may comprise the steps of specifying a set of desired neo-antigens and providing a patient-specific pre-treatment gene expression assay of the patient's patient-specific tumor tissue. The trained ML model may then be used to infer, from the pre-treatment gene expression assay of several candidate drugs, the post-treatment expression of genes that generate one or more neo-antigens of the specified set of desired neo-antigens. In this context, it may further be provided to assign a score to each drug of several candidate drugs based on the inferred perturbation of genes associated with the expression of the desired neo-antigens. The candidate drugs may then be ranked according to the assigned scores and the ranked drugs may be provided as an output, for example, to assist an expert in making a decision regarding a combination therapy with a neo-antigen.

[0019] According to one embodiment of the present invention, the method may be configured to perform drug reuse. In this context, it may be provided that the trained ML model is used to predict the gene perturbation profile of a set of drugs for a pre-treatment gene expression assay of a patient's patient-specific tumor tissue. Then, the predicted gene perturbation profiles for pairs of drugs in the set of drugs may be compared. Based on this, each drug pair may be assigned a score based on the similarity of predicted gene expression perturbations for a set of predefined genes. The drug pairs may be ranked according to the assigned scores, and the ranked list may be provided as an output indicating the possibility of reuse of the listed drugs.

[0020] According to an alternative embodiment of the present invention, it may be provided that the predicted gene perturbation profile is compared with a predefined target gene expression profile. Based on this, each drug may be assigned a score based on the similarity between the respective predicted gene perturbation profile and the predefined target gene expression profile. The drugs may be ranked according to the assigned scores, and the ranked list may be provided as an output indicating the repurposing potential of the listed drugs.

[0021] There are several ways in which the teaching of the present invention can be designed and further developed in an advantageous manner. For this purpose, reference is made on the one hand to the dependent claims and on the other hand to the following description of preferred embodiments of the invention as illustrated by the figures, by way of example. In connection with the description of the preferred embodiments of the invention by means of the figures, preferred embodiments and further developments of the teaching are generally described. [Brief description of the drawings]

[0022] [Figure 1] FIG. 1 is a schematic diagram illustrating a high-level workflow of a method for predicting gene expression perturbations, according to one embodiment of the present invention. [Diagram 2]FIG. 1 is a schematic diagram illustrating the system architecture of a computing server for predicting gene expression perturbations according to one embodiment of the present invention. [Diagram 3] FIG. 2 is a schematic diagram illustrating an implementation of an experiment encoder for drug-gene pairs, according to one embodiment of the present invention. [Figure 4] FIG. 1 is a schematic diagram illustrating a system architecture implemented to find drugs that alter the expression of desired neo-antigens, according to one embodiment of the present invention. [Diagram 5] FIG. 1 is a schematic diagram illustrating a system architecture implemented for drug reuse, according to one embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0023] The present invention provides a method and system for estimating the effect of a perturbation agent (e.g., a drug or other small molecule) on gene expression. According to an embodiment, the present invention is based on the intuition that the effect of a perturbation agent not only depends on its known relationship with its primary target, but may also be influenced by the activity of several off-target genes. Thus, by combining the entire network of possible biological interactions with gene expression input, a more comprehensive evaluation of the effect of the perturbation agent is possible.

[0024] According to one aspect of the present invention, the method / system is configured to utilize prior domain knowledge about relevant entities (including, for example, genes, perturbagens, diseases, etc.) and their relationships (particularly associations, similarities, and / or interactions) organized in the form of a knowledge graph to predict changes in gene expression given current gene expression levels when some perturbation agent is applied.

[0025] FIG. 1 is a schematic diagram illustrating a high-level workflow of a method / system 100 for predicting gene expression perturbations, according to an embodiment of the present invention.

[0026] The method includes a training phase during which, according to the illustrated architecture, the following inputs are used: 1. A data source 102 containing data on relationships between perturbants, genes, diseases, and potentially other entities to construct a biomedical knowledge graph (KG) 104 relevant to the domain of interest. In one embodiment, the relationships are converted to triplets (h, r, t), where h and t are entities (constituting the nodes of the KG 104) and r is the relationship type between the respective entities (i.e., the relationships constitute the edges of the KG 104). In addition, attributes describing the related entities can be associated with the nodes of the KG 104 to provide additional entity features (e.g., Gene Ontology (GO) annotations, expressed protein sequences of genes, Simplified Molecular Input Line Entry System (SMILES) strings of drug molecules, etc.). 2. A dataset 106 of high throughput gene expression assays 108 from ad-hoc cultured cells (cell lines) 110 or patient tissue. According to a preferred embodiment, this dataset 106 includes gene expression from perturbed samples (i.e., post-treatment gene expression values ​​obtained after applying a perturbation agent to a sample or exposing the sample to a perturbation agent), gene expression from control samples, and the perturbation agent. Additionally, the dataset 106 may include information regarding the experimental details of the gene expression assay 108.

[0027] After the training phase, when making predictions, the inputs are one or more gene expression assays and the perturbant for which perturbation is suspected.

[0028] Figure 2 illustrates a system architecture of a computing infrastructure according to one embodiment of the present invention. Like reference numbers refer to like or similar components throughout the figures.

[0029] According to the illustrated embodiment, the method, performed alone or in combination by one or more processors of the computing server 130, includes generating a knowledge graph 104 from input data sources 102 representing domain knowledge, as shown in step (1). Relevant domain knowledge features are assigned to node entities of the knowledge graph 104.

[0030] Furthermore, the method includes an encoding procedure performed by the graph-based encoder 112, which converts the node features into representations in a latent space (entity embeddings), as shown in step (2).

[0031] Further, the method includes a data matching procedure performed by a data matching module 114 configured to perform a data match between entities in the knowledge graph 104 and entities in the gene expression assay dataset 106. In one embodiment, at least perturbants and genes are matched (i.e., perturbant (drug) and gene entities in the KG are mapped to corresponding perturbant (drug) and gene entities in the gene expression assay by entity identifier or name).

[0032] Furthermore, the method includes a prediction procedure performed by the decoder 116 configured to predict the perturbed gene expression levels, as shown in step (4).

[0033] According to one embodiment of the present invention, the method includes a training loop in which node embeddings are learned and gene expression changes caused by perturbation factors are predicted, for example, by minimizing the distance to actual values ​​derived from the gene expression assay dataset 106 used for training. The node embeddings can be learned by any known node representation learning method, for example, while solving the link prediction problem of the knowledge graph 104. Essentially, any method that searches for node representations and captures the graph topology in the node features can be used to compute the embeddings of the entities. One way to do this is to learn node embeddings that can be used to solve the link prediction problem. The link prediction can be used to "verify" that the node embeddings have a certain quality in a predefined sense. In particular, it can be provided to implement the relational graph convolution model R-GCN (see https: / / arxiv.org / pdf / 1703.06103.pdf, which is incorporated herein by reference) and score the edges of the knowledge graph using DistMult factorization. Then, based on that scoring, a cross-entropy loss can be introduced for each potential edge, i.e., between any two nodes. Node representations can be taken into account while computing the scores of adjacent edges: a trained model with "good" node embeddings should score true edges of the correct type in a graph higher than false edges (i.e., non-existent edges or edges of the wrong type).

[0034] Each machine learning model 118 can be trained end-to-end, i.e., the node embeddings and gene expression predictions are trained simultaneously, or the models 118 can be trained sequentially, i.e., the node embeddings are first trained to be optimal for the link prediction task in the KG 104, and then the downstream models are trained to make the final predictions 120 of gene expression values.

[0035] According to one embodiment of the present invention, the method includes a testing / inference phase. During this phase, perturbations in gene expression are predicted from gene expression levels measured before perturbation. Here, perturbations may include application of perturbing agents (e.g., compounds, drug substances, environmental agents, etc.) to samples (e.g., cell lines, tissues, or organisms). Entity embeddings may be used to capture the context learned in the training phase. For example, for a given gene expression profile, the output of the system may consist of predicted gene expression levels after exposure to (or application of) one or more selected perturbing agents. This may include a detailed cross-sectional overview of each gene and perturbing agent.

[0036] Use Case 1: Predicting the individual effect of drug treatment According to an embodiment of the present invention, the system architecture shown in Figure 2 can be configured to predict the individual effect of drug treatment. In this context, the drug is considered as a perturbation agent and the changes in gene expression in the patient's tissue are considered as a proxy for the patient's response / clinical outcome to a treatment or other intervention. The predicted post-treatment gene expression values ​​can be mined to build features for downstream analysis, for example as part of a pipeline of a system implementing a patient stratification or drug response prediction task.

[0037] Below is a description of a possible implementation of a machine learning method proposed according to an embodiment of the present invention to predict gene expression changes due to treatment taking into account a priori knowledge of relationships between entities as well as the entire treatment context such as drug exposure time, dosage, pre-treatment gene expression across cell lines or tissues of an individual patient. The terminology and functionality of the modules used in such a method as well as the respective data flows correspond to those depicted in FIG. 2.

[0038] Gene expression assays A variety of transcriptomics techniques can be used to generate gene expression assays for analysis. 106 Typically, gene expression levels are expected to be provided by DNA microarrays, or sequence-based techniques such as RNA-Seq.

[0039] Pretreatment of gene expression The raw data from gene expression assays 106 are assumed to be pre-processed in a uniform manner. Typical pre-processing steps, generally indicated at 107, may include scale conversion, missing value management, replicate handling, batch effect correction, and / or filtering, etc. Pre-processing steps may be performed before performing any analysis. Furthermore, it is assumed that all gene expression values ​​in the ground truth 117 have been pre-processed in the same manner to make them comparable.

[0040] Domain Knowledge Graph With respect to creating the domain knowledge graph 104, information about the domain entities and relationships can be used to construct a graph in the form of a knowledge graph G=(V,E), where V=V 1 ∪...∪V |K| is the set of vertices / nodes in the graph, E=E 1 ∪...∪E |R| is the set of graph edges. Each entity of type k is a vertex

[0041]

number

[0042] and each relationship of type r between two entities is represented by an edge e r ∈E rThere should be |K|≧2 types of entities, i.e., at least genes and drugs, and |R|≧1 relationship types, i.e., at least one type of gene-drug interaction. No isolated vertices are allowed. In the context of the knowledge graph 104, a data vector is a triplet d=(v,r,v'), where v,v'∈V,r∈R.

[0043] Each vertex v ∈ V has entity features

[0044]

number

[0045] may be assigned a real-valued vector of

[0046]

number

[0047] is an encoder trained to create a non-contextual representation of an entity of type k, and θ k are its parameters, and x v is an attribute assigned to vertex v. The attributes associated with a graph vertex v may include a representation of the molecule, which may be an expressed protein sequence or GO annotation in the case of a gene entity, or a fingerprint, ChemProp, SMILES string, or other molecular descriptor in the case of a drug compound. In the absence of features, we use the unique one-hot vectors of the nodes in the graph G to find f V can be set.

[0048] Graph Encoder According to one embodiment of the present invention, the graph encoder 112 may be configured to compute a contextualized representation of the knowledge graph entities. To this end, it operates on the graph G and computes a real-valued embedding h for each entity v∈V. v =h(f V;ψ), and the real-valued embedding h for each relation type r ∈ R r =h(f r One can train a Relational Graph Convolutional Network (R-GCN) whose goal is to learn a set of features (v,r,v') that are the sum of the features of the KG104 and the features of the KG104, respectively. The goal of the R-GCN in the training phase is to be able to classify any triplet (v,r,v') using the embeddings of its constituents as true if this triplet is present in the KG104 and false otherwise.

[0049] Data Match The data match module 114 may be configured to match drug perturbants and genes from the gene expression assays 106 with corresponding drugs and genes in the KG 104 (e.g., by corresponding identifiers or names). In this context, it may be provided that potential drug-gene pairs are also generated from the gene expression assays 106.

[0050] Experimental Encoder The experimental encoder 115 encodes the KG node embeddings learned for each possible drug-gene pair into individual vectors Z v,v' =g(h v ,h v' ,m c ), wherein h v and h v' denotes the embedding of the corresponding drug node and gene node, and m c represents pre-processed additional information capturing the context of a particular experiment (e.g., drug dose, drug exposure, cell line or tissue). Figure 3 shows a schematic of a corresponding implementation of the experiment encoder 115 for a drug-gene pair.

[0051] Predictors of perturbation effects The predictor 116 of the perturbation effect of a drug on a gene can be, for example, either regression or classification. In this disclosure, the predicted gene expression value for a drug node and gene node pair is

[0052]

number

[0053] In the formula, f Φ (.) is trained to estimate the perturbed gene expression for each queried drug-gene pair. To fit the model, a supervised machine learning training paradigm can be employed.

[0054] Loss function According to an embodiment of the present invention, the following loss functions may be considered: 1) To obtain a contextualized representation of KG104 entities, an R-GCN (Relational Graph Convolutional Network) model can be trained to predict the presence of certain types of edges between entities in the graph. In one embodiment, a cross-entropy loss can be provided that encourages the model to score observable positive KG triplets higher than negative ones.

[0055]

number

[0056] Here, S=S pos ∪S neg And S pos is the set of positive triplets, S neg Let V be the set of negative triplets, x=0 for negative triplets and x=1 for positive triplets, s(.) is a scoring function, v,v'∈V. As a scoring function, one can use factorization (e.g., DISTMULT, RESCAL). Finding a solution to the optimization problem can be done numerically by performing stochastic gradient descent. 2) To predict the perturbation effect in a drug-gene pair, a regression loss can be optimized based on the distance between the ground truth gene expression values ​​117 and the predicted gene expression, e.g., RMSE (root mean square error).

[0057]

number

[0058] where v, v' are the nodes of linked drugs and genes, Q is the set of available drug-gene combinations, and y v,v' is the actual perturbed gene expression value,

[0059]

number

[0060] indicates predicted gene expression. 3) In one embodiment, it may be provided to apply l1 or l2 regularization to the loss function during training in order to impose soft constraints on the model parameters.

[0061] Model Training The model can be trained by optimizing the loss function according to one of the following approaches: 1) For the sequential training procedure, contextualized representations of knowledge graph entities are first learned by optimizing equation (1) above. In a subsequent step, the learned representations are fixed for downstream computations and equation (2) above is optimized. 2) For an end-to-end implementation, the model parameters are learned by optimizing a loss, which is determined as a weighted combination of the loss functions (1) and (2). J = αJ LP +(1-α)J r In the formula, α ∈ (0,1) denotes a hyperparameter.

[0062] Alpha J for learning entity embeddings LP +(1-α)J rBy combining these two goals, we are able to a) preserve the KG structure and b) accurately predict perturbed gene expression levels (an end-to-end approach). The trade-off between these two goals is controlled by the parameter α, which allows us to extract representations of entities that reflect both the measured gene expression levels and a wealth of prior knowledge about the interactions between them.

[0063] Note that it is possible to use inductive learning methods to learn embeddings of graph nodes and make subsequent predictions for entities that were not present in the initial knowledge graph.

[0064] Use Case 2: Discovery of drugs that alter the expression of a desired neoantigen According to an embodiment, the present invention provides a method for finding agents that alter the expression of a desired neo-antigen. Figure 4 shows a schematic of the corresponding system architecture. Again, like reference numbers indicate like or similar components or functions to the previous figures.

[0065] Cancer cells often contain DNA alterations, accompanied by the expression of peptides called neoantigens, which are not present in normal healthy cells and provide targets for cancer immunotherapy. Since neoantigens are generated by genetic mutations, their expression can be altered by drugs through mechanisms similar to normal genes. This opens up an attractive avenue to predict drugs that promote the expression of desired antigens in individual tumors, thus providing opportunities for repurposing existing drugs to be used in combination therapy with neoantigens. Pre- and post-treatment RNA sequencing data can be used to quantify transcript expression and validate predictions. In one embodiment, the present invention focuses on solving the task of prioritizing drugs according to their potential to enhance neoantigen expression in a particular patient.

[0066] Input data In the above mentioned case of finding drugs that enhance the expression of one or more neo-antigens, the input for model training may include a domain-specific KG, such as KG104 mentioned above, and pre- and post-treatment gene expression assays. Genes and drugs for follow-up prediction should be represented in the KG. The gene expression assay should include data on genes involved in the expression of the desired neo-antigens in tumor tissue.

[0067] At the time of prediction, patient-specific pre-treatment gene expression assays and a list of candidate drugs are expected as input.

[0068] Machine learning models The machine learning model 118 is trained to predict gene expression perturbations induced by various agents, as described above in connection with use case 1. Upon prediction, the trained model 118 may be used to infer post-treatment expression of genes that generate desired neo-antigens from pre-treatment gene expression assays 142 in patient-specific tumor tissue of patient 140, as generally shown at 144. The desired neo-antigens may be extracted from a neo-antigen database 147. In one embodiment, the changes in gene expression may be treated as continuous variables. According to another embodiment, the changes in gene expression levels may be discretized (e.g., increased or decreased) and treated as categorical variables to solve classification problems instead of regression.

[0069] Ranking of drug candidates According to one embodiment, a method for finding drugs that alter expression of a desired neo-antigen includes a drug ranking process, as generally indicated at 146. In one embodiment, during the ranking process, it may be provided to assign each drug a score based on the predicted perturbation of genes associated with the expression of the desired neo-antigen. The output 148 may then be analyzed by an expert to draw conclusions about potential drugs for use in combination therapy with the neo-antigen, as generally indicated at 150.

[0070] Use case 3: Drug reuse According to an embodiment, the present invention provides a method for reusing medicines. Figure 5 shows a schematic representation of the corresponding system architecture and provides an overview of the predictive workflow in this context. Again, like reference numbers indicate similar or similar components or functions to the previous figures.

[0071] Traditional "in silico" approaches for drug repurposing rely on having a notion of structural similarity, i.e., two compounds are more likely to have the same pharmacological properties if they share many chemical substructures. Network-based methods draw conclusions by exploring pathways between drug and disease modules in networks of protein-protein interactions. Another paradigm is connectivity mapping, which considers that compounds will cause similar effects if they affect cellular gene expression in a similar way. The latter can be used to find new drug indications without making clear a priori hypotheses about the mechanism of action. However, individually, all approaches tend to show limited usefulness for various reasons.

[0072] According to an embodiment, the present invention provides a method implemented as a tool for drug repurposing that predicts drug compounds with similar effects on gene expression and takes into account the KGs of drug-gene-disease interactions and the drug molecular structures associated with the corresponding entities of the KGs. The gene expression profile of the patient is used to allow personalized repurposing of FDA approved drugs that could potentially provide the desired transcriptome-level response.

[0073] Input data In the above mentioned case of drug repurposing, the input for model training may include a KG of drug-gene-disease interactions, such as KG104 mentioned above, and pre- and post-treatment gene expression assays. Genes and drugs for follow-up prediction should be represented in the KG. Drug entities in the KG should include the encoded drug's molecular structure or drug properties (e.g., solubility, permeability, metabolic stability, transporter effects, or physicochemical properties), which may be assigned as entity attributes.

[0074] At prediction time, the inputs are pre-treatment gene expression assays from patient tissue and a list of drugs being explored.

[0075] Machine learning models The machine learning model 118 is trained to predict the perturbed gene expression in the same manner as described above in connection with use case 1. Pre-treatment gene expression assays are used to predict gene expression changes for the set of explored drugs, generally as shown at 144. During the training phase, the KG node embedding takes into account both the structure of the drug-gene-disease graph and the drug-related attributes of the corresponding drug nodes.

[0076] Ranking drugs by perturbation similarity According to one embodiment, the predicted gene perturbation profiles are compared for pairs of drugs in the drug database 143. As generally indicated at 160 / 161, the drug pairs can be scored and ranked based on the changes in expression in a predefined set of one or more considered genes. For example, this set of genes can include genes belonging to a particular disease module (in practice, disease modules are used to organize and prioritize disease-related genes identified by high-throughput analysis, to provide an overview of disease mechanisms by performing pathway analysis, and to help identify new disease genes, biomarkers, or therapeutic targets). State-of-the-art methods can be employed for the calculation of similarity scores between gene expression profiles and for hypothesis testing of statistically significant differences in drug-induced effects.

[0077] Ranking drugs by similarity to target gene perturbations Alternatively or additionally, as generally shown at 162 / 163, drugs may also be ranked according to their likelihood of causing a desired genetic perturbation (e.g., increasing or suppressing a desired gene) to identify repurposed candidates by comparing the predicted perturbation profile to a predefined target gene expression profile 145 for the set of drugs explored.

[0078] The above mentioned use cases can be suitably used for the discovery of drug perturbation profiles, with particular interest for personalized drug response prediction, or "in silico" hypothesis generation to find FDA approved drug candidates for repurposing. The results can be validated, for example, in subsequent pharmacoepidemiological analyses on health-related real-world data (RWD).

[0079] As detailed above, the present invention provides a method and system for predicting perturbed gene expression using prior domain knowledge. According to an embodiment, the method may include an "offline" phase and an "online" phase.

[0080] In the offline phase, domain knowledge data may be collected from structured public and private data sources, including data on genes, perturbagens, and other possible entities and their relationships. Additionally, the offline phase may include extracting relationship types and entity characteristics that can be used to learn entity embeddings, such as data on types of interactions between genes, perturbagens, and other possible entities, attributes of genes (such as protein sequences or GO annotations) and perturbagens (such as fingerprints, ChemProp, SMILES strings in the case of drug molecules). In one embodiment, the offline phase may further include collecting / providing control and perturbed gene expression data, e.g., obtained from patient tissue.

[0081] The online phase may include the following steps, where the numbering of the individual steps corresponds to the numbers in brackets shown in Figures 4 and 5. (1) Generate a knowledge graph (KG) 104 from the entities and relationships, which should include at least the interactions between genes and perturbations. (2) Train an ML model 118 to learn embeddings of the KG 104 entities. (3) Train an ML model 118 to predict perturbed gene expression from the pre-perturbation gene expression data and the learned embeddings of KG entities. (4) Predict the perturbed gene expression for new unknown samples / unknown perturbation conditions (output of Use Case 1 above). (5) Creating features for downstream analysis from the predicted perturbed gene expression levels (use phase). This step is optional and therefore not shown in Figures 4 and 5. (6) Rank the perturbations according to their desired effect on gene expression. (7a) List drugs that may alter the expression of the desired neoantigen (output of Use Case 2 above). (7b) List drugs that can potentially be reused (output of Use Case 3 above).

[0082] Many modifications and other embodiments of the inventions described herein will come to mind to one skilled in the art to which this invention pertains having the benefit of the teachings presented in the foregoing descriptions and the associated drawings. It is to be understood, therefore, that the invention is not to be limited to the specific embodiments disclosed, and that modifications and other embodiments are intended to be included within the scope of the appended claims. Although specific terms are employed herein, they are used in a generic and descriptive sense only and not for purposes of limitation. [Explanation of symbols]

[0083] 100 Methods / Systems 102 Data Sources 104 Knowledge Graph 106 Gene Expression Assays 108 High-Throughput Gene Expression Assays 110 Ad-hoc cultured cells (cell lines) 112 Graph-Based Encoders 114 Data Match Module 115 Experimental Encoder 116 Predictors 117 ground truth gene expression values 118 Machine Learning Models 120 Final Prediction 130 Computing Server 140 patients 142 Gene Expression Assays 143 Drug Database 145 Target Gene Expression Profile 147 New Antigen Database 148 Output

Claims

1. 1. A computer-implemented method for predicting gene expression perturbations, comprising: generating a knowledge graph (KG) (104) from domain knowledge, the KG (104) describing relationships, including associations, similarities, and / or interactions, between a plurality of entities, the entities including at least some genes and perturbations; training a machine learning (ML) model (118) to predict perturbed gene expression from pre-perturbed gene expression data (106) and learned embeddings of entities of said KG (104); providing gene expression data obtained from a genetic sample from a subject, and using the trained ML model (118) to predict a response of the genetic sample with respect to a change in gene expression resulting from application of one or more perturbations to the genetic sample; A method comprising:

2. the entities of the KG (104) include associated attributes that describe each entity; 2. The method of claim 1, wherein the associated attributes provide entity features including Gene Ontology GO annotations, expressed protein sequences of the genes, and / or Simplified Molecular Input Line Entry System (SMILES) strings of drug molecules.

3. learning embeddings of the entities of the KG (104) by solving a link prediction problem.

3. The method of claim 1 or 2, further comprising:

4. The ML model (118), First, learn the embeddings of the entities in the KG (104); Next, a downstream model is trained to make a final prediction of gene expression based on the learned embeddings of the KG (104). The method according to claim 1 , wherein the first and second inputs are sequentially trained by:

5. 4. The method of claim 1, wherein the ML model (118) is trained end-to-end by simultaneously learning the embeddings of entities of the KG (104) and the predictions of gene expression.

6. obtaining a contextual representation of KG (104) entities by training the ML model (118) to predict edges between entities of the KG (104); 6. The method of claim 1, further comprising:

7. To optimize the cross-entropy loss, a loss function J is used that encourages the ML model (118) to score observable positive KG (104) triplets higher than negative ones. LP Steps to use 7. The method of claim 6, further comprising:

8. To predict the perturbation effect in drug-gene pairs, we use a loss function J that optimizes the regression loss based on the distance between the ground truth (117) and predicted gene expression values. r Steps to use 8. The method of claim 6 or 7, further comprising:

9. Loss function J LP and the loss function J r The loss function J is determined as a weighted combination of J=αJ LP +(1-α)J r learning parameters of the ML model (118) by optimizing where α ∈ (0,1) is a hyperparameter that controls the tradeoff between preserving the structure of the KG (104) and being able to accurately predict perturbed gene expression levels. The method of claim 8.

10. Specifying a set of desired neo-antigens; providing a patient-specific pre-treatment gene expression assay (142) of patient-specific tumor tissue of a patient (140); using said trained ML model (118) to predict, from said pre-treatment gene expression assays (142) of several candidate drugs, post-treatment expression of genes that generate one or more neo-antigens of said specified set of desired neo-antigens; 10. The method of claim 1, further comprising:

11. assigning a score to each of said number of candidate agents based on said predicted perturbation of said genes associated with expression of said desired neo-antigen; ranking the drugs according to their assigned scores; providing the ranked medications as an output; 11. The method of claim 10, further comprising:

12. using the trained ML model (118) to predict a gene perturbation profile for a set of agents for a pre-treatment gene expression assay (142) of patient-specific tumor tissue of a patient (140); comparing the predicted genetic perturbation profiles for pairs of drugs in the set of drugs; assigning a score to each drug pair based on the similarity of predicted gene expression perturbations for a predefined set of genes; ranking the drug pairs according to their assigned scores; 12. The method of claim 1, further comprising:

13. using the trained ML model (118) to predict a gene perturbation profile for a set of agents for a pre-treatment gene expression assay (142) of patient-specific tumor tissue of a patient (140); comparing said predicted gene perturbation profile to a predefined target gene expression profile (145); assigning a score to each agent based on the similarity of the respective predicted gene perturbation profile to the predefined target gene expression profile (145); ranking the drugs according to their assigned scores; 13. The method of any one of claims 1 to 12, further comprising:

14. 1. A system for predicting gene expression perturbations, the system comprising one or more processors and a memory storing instructions that, when executed by the one or more processors, cause the system to: generating a knowledge graph (KG) (104) from domain knowledge, the KG (104) describing relationships, including associations, similarities, and / or interactions, between a plurality of entities, the entities including at least some genes and perturbations; training a machine learning (ML) model (118) to predict perturbed gene expression from the pre-perturbed gene expression data (106) and learned embeddings of entities of the KG (104); providing gene expression data obtained from a genetic sample from a subject, and using the trained ML model (118) to predict a response of the genetic sample with respect to a change in gene expression resulting from application of one or more perturbations to the genetic sample; A system that allows the user to:

15. 1. A tangible, non-transitory computer readable medium having instructions that, when executed by one or more processors, alone or in combination, provide for the performance of a method for predicting gene expression perturbations, the method comprising: generating a knowledge graph (KG) (104) from domain knowledge, the KG (104) describing relationships, including associations, similarities, and / or interactions, between a plurality of entities, the entities including at least some genes and perturbations; training a machine learning (ML) model (118) to predict perturbed gene expression from pre-perturbed gene expression data (106) and learned embeddings of entities of said KG (104); providing gene expression data obtained from a genetic sample from a subject, and using the trained ML model (118) to predict a response of the genetic sample with respect to a change in gene expression resulting from application of one or more perturbations to the genetic sample; 4. A tangible, non-transitory computer readable medium, comprising:

Images