Graph convolution networks to identify and quantify gene and cancer specific transcriptome signatures of cancer driver events

Abstract

This disclosure describes a Machine Learning (ML) framework, including graph convolution neural networks (GCN) for identifying gene expression signatures related to cancer driver events. The model is trained to identify the TP53 mutational state of cancer samples from gene expression, utilizing comprehensive curated graph structures of gene interactions. A quantitative score is generated to rank the severity of a driver event in each sample. Very high AUG results on unseen data across several tumor types are achieved with this method. There is a strong correlation with protein function. The Signatures in Transcriptome Associated with Mutated Protein (STAMP) model can also predict driver events in many combinations of important cancer genes / pathways and several tumor types, based on well-established annotations from the literature. The STAMP model therefore can identify and quantify driver events, which may lead the way to improved targeted therapy selection and prioritization in cancer patients.

Description

GRAPH CONVOLUTION NETWORKS TO IDENTIFY AND QUANTIFY GENE AND CANCER SPECIFIC TRANSCRIPTOME SIGNATURES OF CANCER DRIVER EVENTSCROSS REFERENCE TO RELATED APPLICATIONS

[0001] This application claims benefit of and priority to US Provisional Patent Application 63 / 516,961 filed on August 1, 2023, which is incorporated by reference in its entirety.BACKGROUND OF THE INVENTION

[0002] The present invention relates to methods and systems for identifying cancer driver events.

[0003] The identification and drug targeting of cancer causing (driver) genetic alterations has seen immense improvement in recent years, with many new targeted therapies developed. However, identifying, prioritizing, and treating genetic alterations is insufficient for the majority of cancer patients. Current clinical practice relies mainly on DNA level mutational analysis, which in many cases fails to identify treatable driver events. The use of transcriptomics, a complex and highly informative representation of cellular and tumor state, would enhance diagnostics and treatment successes.

[0004] Accordingly, a need arises for techniques that provide a method to identify cancer driver genes in mRNA (transcriptomics).SUMMARY OF THE INVENTION

[0005] Aspects of the disclosure relate to systems and methods for identifying and scoring a sample with genes.

[0006] This disclosure describes a method and a system for classifying a gene as mutant or wild type and also assigning to the gene a score associated with the classification. The method further comprises receiving a set of curated data on human genes, wherein the curated data are RNA expression data. The curated data are divided into a first, a second, and a third dataset. A classifier rained on the first dataset to identify a datum as mutant or wild-type. The classifier is validated on the second dataset and tested on the third dataset. The method further comprises receiving a new datum and applying the trained classifier on the received new datum to classify the received new datum as mutant or wild-type. The method calculates a score associated with the received new datum and reports the classification and the score associated with the received new datum to a user device.

[0007] The classifier may comprise, for example, a graph convolutional network, a random forest model, or a support vector machines model. The score may be an inverse of an activation function of a next-to-last layer of the classifier. The score may be used to predict the efficacy of a medication on a cell line or on a tumor. The efficacy prediction may comprise estimating an IC50 of the medication. The method may further comprise identifying an active molecular pathway and a medication associated with the active molecular pathway. The method may further comprise diagnosing a patient with a condition based on the score. The method may further comprise proposing a treatment for the patient based on diagnosis and on the score.

[0008] A method is described for applying a trained classifier to propose a treatment of a medical condition. The method receives a datum of a patient’s genes and applies the trained classifier to classify the received datum as mutant or wild-type. The method further calculates a score associated with the received datum. The method predicts efficacy of a medication on a cell line oron a tumor associated with the received datum. The method further reports the classification, the calculated score, the medication, and the predicted efficacy to a user device.

[0009] The classifier may comprise, for example, a graph convolutional network, a random forest model, or a support vector machines model. The score may be an inverse of an activation function of a next-to-last layer of the classifier. The method may further comprise identifying an active molecular pathway and a medication associated with the active molecular pathway.

[0010] A system is described for calculating a score comprising, for example, a processor and computer memory. The computer memory comprises computer instructions, which, when executed by the processor, perform a method comprising a series of steps. One of the steps may include receiving a set of curated data on human genes, wherein the curated data are RNA expression data. The curated data is divided into a first dataset, a second dataset, and a third dataset. A classifier is trained on the first dataset to identify a datum as mutant or wild-type. The classifier is validated on the second dataset and tested on the third dataset. The system receives a new datum and applies the trained classifier on the new datum to classify the new datum as mutant or wildtype. In addition, the system calculates a score associated with the received new datum. The score and the classification associated with the new datum may be reported to a user device.

[0011] The classifier may comprise a graph convolutional network, a random forest model, or a support vector machines model. The score may be an inverse of an activation function of a next- to-last layer of the classifier. The score may be used to predict the efficacy of a medication on a cell line or on a tumor. The efficacy prediction may comprise estimating an IC50 of the medication. The system may further comprise identifying an active molecular pathway and a medication associated with the active molecular pathway. The system may further comprise diagnosing apatient with a condition based on the score. The system may further comprise proposing a treatment for the patient based on diagnosis and on the score.BRIEF DESCRIPTION OF THE DRAWINGS

[0012] So that the manner in which the above recited features of the present invention can be understood in detail, a more particular description of the invention, briefly summarized above, may be had by reference to embodiments, some of which are illustrated in the appended drawings. It is to be noted, however, that the appended drawings illustrate only typical embodiments of this invention, and the invention may be implemented in other equally effective embodiments.

[0013] Figure 1 illustrates a workflow diagram of the overall methods of this disclosure.

[0014] Figure 2A-2F illustrates the results of driver event identification for various samples.

[0015] Figure 3A-3E illustrates validation of the models.

[0016] Figure 4A-4I illustrates a matrix of genes and tumor types for which the model’s predictions were excellent.

[0017] Figure 5A-5H illustrates the results of applying the model to identify driver genes.

[0018] Figure 6A-6F illustrates an example of the STAMP model stratifying response to treatment in clinical cohorts.

[0019] Figure 7A-7F illustrates an example of the STAMP model stratifying response to treatment in clinical cohorts.

[0020] Figure 8 illustrates down-sampling analysis for ELR (MSigDB).

[0021] Figure 9A-9C illustrates a demonstration of TP53’s mutational status and protein function prediction for Signatures in Transcriptome Associated with Mutated Protein (STAMP) and the “pseudo” mutant or WT samples concepts.

[0022] Figure 10 illustrates a multivariable analysis for coxph model of TCGA’s pan cancer cohort.

[0023] Figure 11 A- HE illustrates further survival analysis based on the mutational or non- mutational state.

[0024] Figure 12A-12I illustrates additional SBS analysis.

[0025] Figure 13A-13F illustrates additional SBS analysis.

[0026] Figure 14A-14K illustrates a comparison of STAMP to TP53_PROF.

[0027] Figure 15A-15D illustrates applying and validating STAMP trained on various pathways in several tumor type cohorts.

[0028] Figure 16A-16I illustrates a further GDSC analysis.

[0029] Figure 17 illustrates the results of different graph structures used to train STAMP on Breast cancer for TP53 mutational status.

[0030] Figure 18A- 18B illustrates little correlation between the score and the IC50 when identified by amplification status.

[0031] Figure 19 illustrates a computing device, which may be used to implement the teachings of the present invention.

[0032] Other features of the present embodiments will be apparent from the Detailed Description that follows.DETAILED DESCRIPTION

[0033] In the following detailed description of the preferred embodiments, reference is made to the accompanying drawings, which form a part hereof, and within which are shown by way of illustration specific embodiments by which the invention may be practiced. It is to be understood that other embodiments may be utilized and structural changes may be made without departing from the scope of the invention. Electrical, mechanical, logical, and structural changes may be made to the embodiments without departing from the spirit and scope of the present teachings. The following detailed description is therefore not to be taken in a limiting sense, and the scope of the present disclosure is defined by the appended claims and their equivalents.

[0034] The present disclosure relates to methods and systems for identifying driver events in mRNA (transcriptomics).

[0035] Cancer diagnosis and treatment have seen major breakthroughs in recent years with the emergence of precision medicine. Driver alterations in cancer genes are identified and are targeted by novel therapies, leading to significant improvement in clinical outcomes. However, only a small fraction of cancer patients benefits from such treatments, and many questions have previously remained unanswered. Three such questions stand in the foundations of the inventions described in the present specification. First, less frequent or less characterized variants in important cancer genes occur in a substantial percentage of cancer patients. Such variants are mostly obscured in their effect - whether they contribute to tumorigenesis (driver mutations) or not (passenger mutations). Second, some mutations may cause a spectrum of partial defects in protein activity, further challenging attempts to identify and prioritize their impact. Third, for some patients, mutational analysis identifiesa paucity of mutational events or none at all. This suggests that analyses of other molecular levels such as epigenetics, transcriptomics and protein-protein interactions are required to complement mutation effect classification.

[0036] Gene expression, measured through sequencing of mRNA, captures simultaneously an image of the activity of all genes in a sample, thus allowing a rich and complex representation of the tumor state. Transcription profiles were also shown to hold valuable information regarding genetic aberrations, clinical traits, and therapeutics. The high complexity of transcriptomics is a significant challenge for its analysis, and arriving at accurate, unbiased conclusions is far from trivial. Several approaches were developed to address this complexity, such as pathway analysis and network analysis. Machine learning (ML) and deep learning (DL) models can be used to address this challenge by focusing the analysis to produce valuable insights. In conventional ML, prior analysis of the data retrieves relevant features for learning (a process termed feature selection). In the field of DL, algorithmic architectures are developed to target specific complex tasks. For example, Convolutional Neural Networks (CNN) successfully learn from images by using pixel spatial proximity to assume similarity in pixels’ content.

[0037] Graph structures such as Funcoup and HumanNet v2 contain human-genome representation maps that define various interactions between genes. These graphs accumulate data from Metabolic or Signaling pathway information, complex and physical protein-protein interactions (PPIs), Genomic contexts and co-expression patterns into a single graph structure. A Graph Convolution Neural Network (GCN) is a DL architecture that uses graph structures and graph nodes proximity to focus the neural network’s learning in a similar manner to CNN using image pixels’ spatial proximity. Thus, it may serve as an ideal architecture to focus learning from transcriptomics by integrating established biological knowledge in the form of graphs. Others have examined the ability of a GCN to predict agene’s masked expression level using the expression values and connectivity maps of its graph neighbors. A similar approach may be applied to identify driver event signatures from gene expression.

[0038] Multiple ML and DL models for cancer genomics have been developed in recent years. Some works use mutational or transcriptome data from cell line studies to predict their response to various drugs. One recent approach used an innovative architecture with a two-branched neural network on cell line’s mutational atlas to predict their response to treatment, and to even identify novel drug synergies. Driver event predictions from DNA mutational data is by itself a blooming field of study, and highly accurate models have been developed in recent years. Similar works using only transcriptome data are not as frequent. Some of these works showed indications that training a model to predict driver events could benefit from a gene specific and tumor specific approach. Despite the reduction in samples size, the advantage of such focus may be explained by the reduced complexity of the learned trait. Whether this approach is beneficial for transcriptome -based models is yet to be studied. TP53, the most frequently mutated and most studied gene in human cancer, is a natural candidate for exploring a gene specific approach of cancer driver predictions.

[0039] The present specification describes a method named Signatures in Transcriptome Associated with Mutated Protein models (STAMP), a comprehensive supervised deep learning (DL) method to obtain driver event signatures from RNA-Seq samples using a GCN 101, an overview of which is illustrated in FIG. 1. FIG. 1 depicts a workflow diagram presenting the strategy used for the development of STAMP. At Step 1 of FIG. 1, the data 103 used as input for the models are prepared for learning. This data 103 comprises gene connectivity graphs 105, such as a curated biological database (e.g., HumanNet v2 or Funcoup). Other databases (e.g. The Cancer Genome Atlas = TCGA) may also be referenced. For comparison, at Step 2 of FIG. 1, features are selected (107)for use with a machine learning method (109) such as random forest (RF) and elastic net regression(ELR) to predict a mutational state. In the STAMP method, a graph convolution neural network 101 may instead be used to classify the mutational state. At Step 3 of FIG. 3, these methods are combined to predict various metrics such as survival rate, single base signatures, and classify variants (111). In addition, the metrics or scores output by the model(s) may be used for treatment prioritization (113). FIG. 1 describes an overview of the various methods used to validate the analysis, and the potential utilities that can be obtained from STAMP.

[0040] In an embodiment, the machine learning methods used to predict a mutational state may include random forest (RF), elastic net regression (EER), convolutional neural networks (CNN), support vector machines (SVM), logistic regression (ER), naive Bayes (NB), k-nearest neighbor (KNN), amongst others.

[0041] A gene- and tumor-specific approach is adapted, first trained and validated on TP53 and then expanded to several other important cancer genes and pathways. STAMP predictions may be more significantly correlated with molecular and clinical traits of p53 compared to mere mutational analysis. Models trained for genes and tumor types for which relevant targeted therapy exists are shown to strongly correlate with drug response in an independent cell line database. Hence, STAMP successfully identifies signatures of transcriptome data for driver events, improving mutational analysis and possibly leading to the identification of new, undetected driver events. It enables potential utilities for therapeutics prioritization using gene expression information.Framework for STAMP models

[0042] Aspects of this present invention included training machine learning models to predict TP53’s mutational state based on RNA-Seq samples from the Cancer Genome Atlas (TCGA) 115 (FIG. 1). TP53 Mutational status label was defined in a binary manner, as mutated (M, positive)for samples with any TP53 mutation, and Wild Type (WT, negative) otherwise. A Pan-Cancer model (n = 9,077 samples with both mutational and RNA-seq data) was trained in parallel to a tumor specific Breast Cancer (BRCA, n = 1,042) model, the tumor type with the highest number of samples in TCGA. Each dataset was randomly divided in a 60-20-20 manner to train, validate, and test sets, respectively. Division was performed before model creation, so that on each dataset (pan-cancer or BRCA), the same partition was done for all models. The training set was used to train the models, the validation set was used for hyperparameter tuning (including identification of the optimal feature selection method) and the test set was used to evaluate the models by their ability to generalize on unseen data. Two conventional learning models were used for the analysis, Random Forest (RF) and Elastic Net Regression (EER). Each conventional model was run using all features (all genes), and then using feature selection methods.

[0043] Feature selection

[0044] Gene expression samples comprise of almost 20,000 genes as features, while only a few hundreds to a few thousands of samples with transcriptome data currently exist in publicly available cancer databases. Training on such sample sizes (e.g., on all 20,000 genes) often leads to overfitting. To address this challenge, a feature selection process can be applied, as it allows the reduction of redundant, noisy or irrelevant features. A reduced feature space also allows a computationally more efficient, and therefore faster, learning process. Feature selection processes can generally be divided into two approaches: (i) Data-driven and (ii) Prior knowledge-based. Data driven feature selection identifies specific features (genes) correlated with the studied trait using the investigated dataset. By contrast, a knowledge -based approach refers to the use of prior established knowledge for selecting the features more relevant to the question at hand.

[0045] For purposes of the present invention, two feature selection methods were used: a data- driven approach using Differential Expression (DE) and a prior knowledge approach using pathway gene sets curated from the literature. DE was used in the RNA-Seq training cohort to examine the correlations of each gene with TP53 mutational state. The 150 most positively correlated genes and the 150 most negatively correlated genes were selected (n = 300 genes). For the second approach, TP53 pathway related gene sets were extracted from the Molecular Signatures Database (MSigDB. n = 311 genes).

[0046] Trained conventional models performed very well on the validation set. ELR with MSigDB feature selection performed best (AUC = 93.1% on BRCA; AUC = 94.1% on pan-cancer). ELR with DE feature selection had slightly worse results (AUC = 92.7% on BRCA; AUC = 93.2% on pan-cancer). RF performed better with DE compared to RF with MSigDB (AUC = 81.8% on BRCA; 84.1% on pan-cancer), but generally had worse results then ELR (See Table 1 for full description of AUC on the validation set). Hence, ELR with MSigDB was selected for further analyses. RF with DE is also further examined to maintain variability in the methods compared.

[0047] Table 1. AUC scores for Validation set on TCGA Breast. Comparison of models and feature selection methods.

[0048] Since the positive label ratio (i.e., the percentage of samples with mutated TP53) varies between the cohorts (7%-92.8%, see Table 2, below), several measures were used in the final step of comparison on the test set, to allow more comprehensive comparison between the unbalanceddata sets. Mainly, Accuracy, AUC and Fl were calculated and compared for the different models and cohorts. The models generalized well on the test set. From the conventional learning models, ELR with MSigDB performed best on both cohorts, by all measures (AUC = 93.99%, ACC = 88.82%, Fl = 84.2% on pan-cancer; AUC = 97.2%, ACC = 93.3%, Fl = 89.71% on BRCA. see FIG. 2a-c for full comparison).

[0049] FIG. 2 illustrates the performance of STAMP models on different cohorts.

[0050] FIGS 2A-2B — Test set ROC curves for: GCN (green - 201), ELR with MSigDB feature selection (red - 203) and RF with DE feature selection (purple - 205) on the Pan-cancer (a) and Breast Cancer (b) cohorts in learning TP53 mutational state. GCN is best for pan-cancer (AUC = 94.95%), ELR is best for Breast Cancer (AUC = 97.2%).

[0051] FIG. 2C — Accuracy, AUC and Fl metrics for the three models (GCN in green - 207, ELR in red - 209, and RF in purple - 211). Comparison by several metrics is important in cohorts with classification groups imbalance.

[0052] FIG. 2D — Validating STAMP models on the independent METABRIC cohort of breast cancer tumors. GCN (213) performed best (AUC = 90.22%); RF (215) also performed well (AUC = 89.28%); ELR (217) performed markedly worse (AUC = 84.12%). The pan-cancer GCN model (219) performed worse than the Breast cancer GCN model (AUC = 83.99% vs. 90.22%, respectively).

[0053] FIG 2E — Down-sampling analysis on Breast Cancer portrays the boundaries for applying STAMP. A model was trained for various combinations of Cohort size (x axis) and number of samples with positive label (y axis). Fl scores are shown (blue - 221 - low; yellow - 223 - medium; red - 225 - high).

[0054] FIG. 2F — Expanding STAMP models (GCN

[0227] in green, ELR

[0229] in red and RF

[0231] in purple) on all eligible tumor types (x axis, ordered by log of cohort size, as shown by green barplot). Most models for tumor types that passed the down-sampling criteria achieved high AUC on the test set (y axis). Abbreviations: GCN- Graph Convolution Network; ELR- Elastic Net Regression; RF- Random Forests; MSigDB- Molecular Signatures Database; DE- Differential Expression; AUC- Area Under the (ROC) Curve;

[0055] Table 2. Tumor type cohorts in TCGA. Cohort size, positive label ratio, exclusion by downsampling criteria

[0056] Graph Convolution Neural Network integrates prior biological knowledge into model architecture

[0057] Graph Convolution Network (GCN) is a graph based, deep learning architecture that allows both biologically curated and data driven feature extraction processes to be integrated in the same model. Genes are included in the model as graph nodes, and gene interactions of different types form the graph edges (interactions include protein-protein interactions, co-expression, common pathways and more. See the Methods section later in this specification). Edges are then used as constraints on the neural network’s weight multiplications of each layer, allowing modification of features in non-initial layers. This creates features that are both data driven (modified across the neural network’s layers) and rely on the biologically curated graphs. The GCN, compared to the conventional learning models, also reduces variability in the feature selection process. That is, one graph structure can generate relevant sub-graphs for any query gene or tumor type, whereas DE analysis requires re-establishing for every new query, and MSigDB would require finding new sets when addressing new genes. Graphs used by the GCN architecture can vary. For the following analysis, two graph structures were used and compared: HumannetV2 and Funcoup. GCN was applied using TP53’s first degree neighbor genes (n = 344 for HumannetV2; n = 1,425 for Funcoup), and a grid search is used for hyperparameter tuning (see Methods). While most analyses were performed with a relatively small number of layers (1-4), a similar experiment (including grid search for hyperparameter tuning) was performed with a GCN architecture of 20 layers.

[0058] The best results on the validation were achieved using GCN, and outperformed conventional learning methods (AUC = 94.3% on BRCA, second best was EER-MSigDB with93.1%; AUC = 94.8% on pan-cancer, second best was ELR-MSigDB with 94.1%, Table 1). In some experiments, GCN with 20 layers failed to converge into successful learning and was therefore omitted from further analysis (AUC = 50% for both cohorts, Table 1). on the test set, GCN generalized well for both cohorts. When comparing to the conventional models, it performed slightly better in the Pan-Cancer test set by AUC, but not by the other performance measures (AUC = 94.95%, second best was ELR-MSigDB with 93.99%, FIGS. 2A-2B. Comparison by other metrics is shown in FIG. 2C). On the BRCA cohort it performed worse than the conventional models (AUC 91.91%, ELR MSigDB was better with 97.2%, Figure 2A-2C). GCN’s better performance for pan-cancer may hint to the advantage deep neural networks are expected to gain as more data accumulate.

[0059] STAMP models accurately predicted mutational state in an independent cohort of breast tumors

[0060] Results on both the validation and the test sets confirmed successful learning of TP53 mutational patterns in gene expression for various TCGA tumor type cohorts. However, this analysis does not indicate generalizability to other independent cancer databases, which could be more challenging. To examine whether STAMP models can generalize on samples outside TCGA, the METABRIC database, consisting of 2,000 breast tumors, was used. This examination not only validates STAMP on independent data but also tests the generalizability of the models that were trained on RNA-Seq data in predicting for Micro-Array gene expression samples. TCGA also applies a unique normalization procedure, not used for METABRIC, further challenging the attempted generalization.

[0061] The models predicted with very high AUC scores on the METABRIC cohort, where GCN had the remarkable AUC score of 90.22% (FIG. 2D). RF with DE had next best performance withAUC = 89.28%. Interestingly, the ELR model with MsigDB which performed best on the TCGA Breast cancer test set, had significantly worse performance on METABRIC (AUC = 84.12%). This may indicate that the ELR model represents overfitting towards TCGA data, as it fails to reproduce a higher AUC score on this independent dataset compared to GCN or RF.

[0062] The METABRIC cohort allowed us to investigate another open question, of the pan-cancer compared to tumor-type specific models. The GCN model was trained on TCGA pan-cancer to predict for METABRIC. The pan-cancer model performed significantly worse than the tumor-type specific model (AUC = 83.99% vs. 90.22%, respectively). This good performance on METABRIC likely reflects a more accurate measure (compared to the TCGA left-out test set) of the model’s potential to predict on unseen data. Hence, based on these experiments, GCN is the best performing architecture for predicting mutational status from gene expression samples, and tumor type specific models are better than the pan-cancer approach.

[0063] Down-sampling explores criteria for tumor types and genes relevant for learning

[0064] To apply this framework on further cancer specific and gene specific cohorts, the boundaries for successful learning need to be explored in cases of smaller sample size or of lower positive / negative label ratio. A down sampling analysis was therefore performed on TP53 Breast Cancer cohort. Breast cancer was used due to its high sample size (n = 1,042), its relative abundance of TP53 mutations (32.6% mutated samples) and the fact that the models showed good performance on this cohort. After excluding a validation set (20% of samples), both number of samples and positive label ratio in the training set were gradually reduced using random sampling. GCN and ELR were trained for each pair of downgraded values, to determine the sample size and positive label ratio required for successful learning. The Fl score on the test set was used for comparison. For the GCN model, learning did not lose power if the percentage of mutated sampleswas 10% or above and the number of samples was 300 or above. Good results were also achieved with only 100 samples, if the mutated percentage was kept at 30% (FIG. 2E). Comparable boundaries were examined for ELR, which seems to be better for tumor types with a small number of samples, as shown in FIG. 8. Six tumor types were excluded from further analysis under criteria implicated by these results (see Methods and Table 2).

[0065] FIG. 8 illustrates down-sampling analysis for ELR (MSigDB). The Fl score (shaded from 0

[0801] to 0.8

[0803] ) is given for models trained in various scenarios of sample size and positive label ratio. ELR seems to perform better than GCN for a smaller sample size and for a smaller positive label ratio (compare to FIG. 2D).

[0066] Extending methodology for all eligible tumor types

[0067] TCGA tumor types and gene cohorts were selected with sufficient power for training, as defined by the criteria described above. Models were trained and tuned on all relevant tumor type cohorts in TCGA (n = 17). Test set AUC comparison for this expanded analysis is given in FIG. 2F. The results show that learning the genetic trait of TP53 mutational state from RNA-Seq samples is applicable across a varying set of tumor types. RF model was generally worse than the other models across tumor types.

[0068] Validating STAMP ’s correlation with p53 functionality through genetic and clinical traits

[0069] FIG. 9 illustrates conceptually how STAMP is trained.

[0070] FIG. 9A — To train STAMP, TCGA tumor samples are first divided by their TP53 mutational status. The TP53 mutational status as Mutated (red - 901) or WT (blue - 903) is reported.

[0071] FIG. 9B — STAMP forms a binary prediction for all TCGA samples. These predictions are identified as Deleterious (D - 905) or non-deleterious (ND - 907). In addition, a linear score is calculated to quantify the effect of dysfunction severity.

[0072] FIG. 9C — Samples that have a mutation in TP53 (i.e., colored red

[0901] in FIG. 9A) but are predicted as non-deleterious by STAMP (i.e., colored blue

[0907] in FIG. 9B) can be considered “Pseudo WT”, colored here in yellow (909). Their STAMP prediction implies an RNA signature similar to samples with WT TP53. Similarly, samples with WT TP53 (blue

[0903] in FIG. 9A), predicted as D by STAMP (red

[0905] in FIG. 9B) can be considered “Pseudo Mutant”, colored here in turquoise (911).

[0073] As described, samples in TCGA were divided by a binary label based on their TP53 mutational status. The STAMP classifier was then trained to predict the mutational status based on transcriptome profiles (FIG. 9A). A prediction is given to each sample, indicating its expected p53 functional effect, as either Deleterious (D, mutant-like RNA patterns - 905) or Non- Deleterious (ND, Wt-like RNA patterns - 907), as shown in FIG. 9b. The transcriptome profiles of certain samples contradict their actual TP53 status at the DNA level, making the model’s predictions different from the original mutational state. These differences could be due to false predictions, but could alternatively indicate a true biological behavior in certain samples, opposite to their mutational state (“Pseudo Mutant”

[0911] or “Pseudo Wt”

[0909] samples, FIG. 9c). Hence, the analyses conducted in the following sections examine whether the STAMP classification (D / ND) better correlates with genetics and clinical traits that have distinct behavior in mutant from Wt TP53, compared to mutational state. Such correlation could suggest that the RNA based models would identify driver events in a protein or pathway, which are undetected by the gene’s mutational analysis.

[0074] STAMP ’S predictions form a quantification score that stratifies survival of patients by TP53 dysfunction severity

[0075] Most clinical annotations address the effect of mutations on protein function as either aberrant or functional. However, quantifying the level of dysfunction caused by each mutation could enable the ranking of mutations in a tumor and hence allow gene targeting prioritization. To address this purpose, the model’s prediction scores were modified into a score for quantifying driver event effect (FIG. 9A-9C). This score may be calculated by inverting the sigmoid activation function responsible for the binary classification. The score is validated in the analyses presented below.

[0076] To validate STAMP'S predictions and the produced score, survival data was retrieved from TCGA and the results are reported in FIG. 3A-3E. Patients with mutated TP53 in the pan-cancer cohort have shown significantly shorter overall survival (OS) compared to patients with Wt TP53 (FIG. 3a, p = 1.2e-41, log-rank test). As STAMP forms a score to each tumor based on its TP53 status, whether such a score could rank tumors for their survival prognosis was tested.

[0077] FIG. 3A-3E illustrates the validation of the STAMP models.

[0078] FIG. 3A — Survival analysis on TCGA pan-cancer cohort divided by TP53 mutational state (M in red

[0301] , Wt in blue

[0303] , log test p = 1.2e-41. See FIG. I la for a similar division by STAMP’S predictions).

[0079] FIG. 3B — TCGA Pan-cancer cohort divided into quartiles based on STAMP’S score. STAMP’S TP53 model predictions stratify tumors for their survival outcome.

[0080] FIG. 3C — Boxplots presenting association of SBS4 in LU AD to TP53’s actual mutational state (left, M in red

[0305] and Wt in blue

[0307] ), compared with STAMP’S TP53 model predictions(right, D prediction in red

[0309] and ND in blue

[0311] ). STAMP strengthens enrichment for SBS4(p = 3e-05 vs 0.00033).

[0081] FIG. 3D — Correlation of STAMP’S score with the prevalence of SBS4 attributed mutations in LU AD is also statistically significant (Spearman p = 2.16e-07).

[0082] FIG. 3E — Comparison of STAMP’S quantitative score between missense mutations predicted to be deleterious (D) (313) and non-deleterious (ND) (315) as defined by a previously published algorithm TP53_PROF (Wilcoxon p = 1.2e-l l).

[0083] FIG. 3A-3E Abbreviations: M- Mutant; Wt- Wildtype; Q1-Q4 Quartile 1 to Quartile 4; D- Deleterious; ND- Non-deleterious; and LU AD- Lung adenocarcinoma.

[0084] STAMP’S score for all TCGA samples was divided by quartiles, forming 4 groups of patients with distinct prediction values. Statistical differences were than compared between those groups using the log test on the cox regression model. Remarkably, survival prognosis was statistically distinct between all 4 groups of patients (p-values are given for adjacent quartiles: QI to Q2, p = 2e-06; Q2 to Q3, p = 2e-15; Q3 to Q4, p = 0.006, FIG. 3B). A multi-variable analysis was performed, showing statistically significant distinction when controlling for tumor type, age, and mutation burden (FIG. 10). The effect STAMP has on survival was also examined in several tumor types where division by TP53’s mutational status was statistically significant in survival data. Results of this analyses are presented in detail in FIGS. 9A-9C-15A-15D.

[0085] Single Base Signature (SBS) analysis validates advantage of gene expression models in predicting p53 functionality over mutational state

[0086] Some mutational processes can generate a characteristic signature of base pair substitutions across the DNA sequence. Single Base Signature (SBS) 4, for example, is associated withsmoking. It is enriched in Lung adenocarcinoma (LU AD) and was shown to correlate with TP53 mutation occurrence. The smoking related SBS4, as well as several other signatures, are used here as a novel, strongly supported validation of the TP53 STAMP models. First, the Wilcoxon Rank Sum Test was used to compare mutant and Wt TP53 (the model’s label) for their SBS4 enrichment in LUAD, showing statistically significant enrichment for mutant TP53 samples (p = 0.00033, FIG. 3C, left). SBS enrichment was next examined based on STAMP’S predictions. If the model improves the identification of p53 activity compared to mutational analysis, the enrichment for SBS4 should also be enhanced. Indeed, enrichment for SBS4 in LUAD was higher when separating samples by STAMP predictions, compared to separation by mutational state (p = 3e-05 vs. p = 0.00033, FIG. 3Cc). Since the sample size remains constant in both analyses, lower p-value indicates stronger effect size. SBS4 enrichment was also strongly correlated with STAMP’S score (p = p = 2.16e-07, spearman correlation coefficient test. FIG. 3D). The correlation of STAMP with APOBEC and HR-DDR signatures was also examined in pan-cancer or in relevant tumor types, with similar results. These analyses are presented in detail in FIGS. 10 and 11 A-l IE and elsewhere in this specification. The correlation with SBS enrichment emphasizes the potential of the STAMP RNA-based models in identifying protein function patterns, as captured by related biological processes.

[0087] STAMP identifies patterns of deleterious and non-deleterious missense variants in TP 53

[0088] In an embodiment, the binary label STAMP may be trained on separated TP53 Wt samples from mutated ones, regardless of the type or effect of mutations. Thus, examining STAMP’S correlation with different mutation types and mutation effects can serve as another validation for the models. Accordingly, it was next examined whether STAMP could identify, among TP53 mutated tumors, those with a driver mutation from those that carry a passenger variant with noeffect on p53’s activity (pseudo Wt). Missense mutations were the focus of this analysis, as they are frequent in TP53 and are considered more challenging to identify for their effect. TP53_PROF is a recently published gene-specific machine learning algorithm that classifies all TP53 missense variants for their effect on protein function with very high accuracy (96.5%). In this example, TP53_PROF was used to validate STAMP, by comparing STAMP’S score to TP53_PROF’s variant classification. In this analysis, Deleterious (D) and Non-deleterious (ND) annotations refer to TP53_PROF’s variant classification for missense mutations.

[0089] The pan-cancer model was examined, since it contains many more samples and thus allows an improved statistical power. Significantly lower values were given by STAMP’S score for samples with missense mutations predicted as ND, compared to missense mutations predicted as D (FIG. 3E, p = 1.2e-l 1, Wilcoxon Rank Sum Test). Eight more tumor type cohorts with enough ND labeled samples were also examined, with all but one showing significantly lower values for ND samples (FIG. 14A-14K). Samples with truncating mutations were also examined and had similar values to D missense mutated samples (FIG. 14A-14K). Overall, STAMP’S score successfully identifies patterns of TP53 missense mutations effect on protein function.

[0090] Forming a gene-expression based atlas of driver event signatures across important cancer genes and pathways.

[0091] Expanding STAMP to apply on genes other than TP53 is of immense importance, since many available diagnostic biomarkers and treatments are based on targeting specific cancer genes. Many more biomarkers and treatments are in clinical trials and are expected to be available in the future. In addition to other genes, identifying the functionality of important cancer pathways (i.e., whether the entire pathway functions normally) could have great significance for portraying a tumor’s molecular state and for therapeutic considerations. Hence, STAMP’S pipeline wasextended to predict dysfunction of important cancer genes other than TP53, and to predict function of entire cancer pathways. To perform such extensions, instead of using the mutational status as with TP53, genes’ and pathways’ alteration maps from the work of Sanchez-Vega et al. were used as sample labels. The authors used prior knowledge from pathway databases, driver databases, mutational and epigenetic analyses as well as other sources from scientific literature. They integrated these to form comprehensive alteration maps across all TCGA samples, defining binary functional annotations of many important cancer genes, and of entire pathways. These alteration maps can be used as highly accurate and rigorously validated labels for STAMP. Such labeling would enable the training of models to predict dysfunction in the level of the pathway, parallel to training gene specific models.

[0092] FIG. 4A-4I illustrates the creation and validation of an atlas of aberrations across genes, pathways and tumor types.

[0093] FIG. 4A - provides a heatmap illustrating the gene and tumor type intersections for which a STAMP model was trained. Slots in gray (401) are gene-tumor type cohorts that did not pass the down-sampling criteria, as shown in FIG. 2E. Slots are colored in a gradient of red (high) (403), yellow (medium) (405) and blue (low) (407) according to the test set’s AUC achieved by each model.

[0094] FIG. 4B-4E - shows a survival analysis similar to FIG. 3A-3E, for LATS2 in LGG (b-c) and for BRAF in SKCM (d-e).

[0095] FIG. 4F ,4H - illustrates an SBS analysis similar to FIG. 3A-3E, for MSI related SBS44 in COAD, and SBS20 in STAD.

[0096] FIG. 4G, 41 - show that the score correlates with SBS enrichment for these two models.

[0097] FIG. 4A-4I Abbreviations: SBS- Single Base Signature; LGG- Low Grade Glioma;SKCM- Skin Cutaneous Melanoma; MSI- Micro-Satellite Instability; COAD- Colorectal Adenocarcinoma; STAD- Stomach Adenocarcinoma.

[0098] The down-sampling criteria defined in previous sections of this disclosure were used to determine relevant gene and tumor type combinations with sufficient power for training, leading to 296 trained models across 25 tumor types, 10 pathways and 44 genes (FIG. 4A and FIG. 15A). Specifically, models were trained and tuned for cohorts with more than 300 samples and at least 10% of samples with a positive label, or with 100 samples and at least 30% of samples with a positive label. ELR and GCN models were both trained and tuned. The AUC scores for the different genes and for the pathway models are presented in FIG. 4A and in FIG. 15 A, respectively (gray slots represent gene-tumor type combinations that did not pass the down-sampling criteria).

[0099] The newly trained models allowed further inspection of survival and SBS analyses. One example is the effect of the LATS2 gene on survival in Low Grade Glioma (LGG) patients. LATS2, a core component of the Hippo signaling pathway, is dysregulated via epigenetic silencing in several cancer types. In LGG, LATS2 silencing through promoter hypermethylation was shown to significantly affect prognosis. The effect of STAMP’S LATS2 model on survival outcomes was therefore examined. The survival effect was statistically significant based on LATS2 aberration in LGG (the label used for training, log-test p = 2.5e-16, Eigure 4B), an effect enhanced when dividing by STAMP predictions (p = 9.3e-19, Eigure 15B). As with TP53’s survival analysis, STAMP’S score for LATS2 gene on LGG samples was divided by quartiles, and then compared for survival outcome differences. Statistically distinct outcomes were observed between 3 groups (QI to Q2, p = 7.12e-07; Q2 to Q3 had borderline p value of 0.071, Q2 to Q4 was distinct, p = 0.0063; Q3 and Q4 were not distinct, p = 0.68. EIG. 4C and EIG. 15C). A multi-variable analysiswas performed, showing statistically significant distinction when controlling the age and mutation burden features (FIG. 15D). Another statistically significant advantage in survival was observed in TCGA’s Skin Cancer (SKCM) patients with BRAF Mutation (the model’s label, log-test p = 0.037, FIG. 4D), possibly due to improved treatment opportunities. Patients predicted by STAMP as BRAF dysfunctional (i.e. ND = non-deleterious) had enhanced statistically significant survival advantage compared to the label (p = 0.0007 vs. 0.037, FIG. 4E).

[0100] SBS analyses were also performed for non-TP53 models. SBS44, a frequent signature in Colorectal cancer (COAD), is associated with defective DNA mismatch repair and with high microsatellite instability (MSI). High MSI is present in 10-15% of COAD cases and is known to affect prognosis for these patients. The APC gene, a tumor suppressor important in a large portion of COAD cases, is mutated significantly more frequently in MSI Low patients. Therefore, SBS44 enrichment was examined for the APC model in COAD. Indeed, COAD patients labeled with functional APC are enriched for SBS44 (Wilcoxon p = 3.2e-12, FIG. 4F). As with TP53 and its related SBS, the APC model improved the correlation with SBS44 compared to the original label (p < 2.2e-16 vs 3.2e-12, FIG. 4F). STAMP’S score was also strongly correlated with SBS44 enrichment (Spearman’s correlation, p = 4.5e-16, FIG. 4G). SBS20 is another signature associated with high MSI and is frequent in Stomach adenocarcinoma (STAD). In STAD, high MSI appears in a subset of patients, affecting prognosis and response to treatment. The NOTCH pathway activation is correlated with low MSI in gastric cancers. Indeed, the NOTCH pathway aberration, as labeled by Sanchez-Vega et al., was significantly enriched for SBS20 in Stomach Adenocarcinoma (p = 6.3e-08, FIG. 4H). The NOTCH pathway model in STAD improved the correlation with SBS20 compared to the original label (p < 2.2e-16 vs 6.3e-08, FIG. 4H). The score again strongly correlated with SBS20 enrichment (Spearman’s correlation, p = 1.3e-15, FIG. 41).

[0101] Testing STAMP correlation with drug response in the GDSC cell lines database

[0102] In the embodiment described herein, STAMP has been successfully applied to many genes and pathways in a tumor-type specific context. It has also been examined whether STAMP could quantify the effect of gene targeting therapies in cancer. The correlation of RNA model predictions with drugs IC50 in cell lines from the Genomics of Drug Sensitivity in Cancer (GDSC) database was tested.

[0103] The GDSC is the largest public resource for drug sensitivity in cancer cell lines and was therefore used in this analysis. First tested was Nutlin, an MDM2 antagonist that has a strong effect when TP53 is Wt through activating the p53 pathway, and for which TP53 mutant tumors are resistant. Accordingly, Nutlin can be used here as a strong validation of the TP53 model’s predictions. To improve statistical power, Nutlin was tested in the pan-cancer cohort. Five more tumor-type, gene and drug combinations were identified with relevant data in GDSC, with a trained STAMP model (i.e., passing the down-sampling criteria. All selected models also had test set AUC > 80%) and with possible clinical relevance. Those were: Afatinib, targeting ERBB2 in Breast cancer; PLX-4720 and Dabrafenib, both targeting the BRAF V600E mutation in Skin Cancer; KRAS G12C inhibitor, targeting KRAS G12C mutation in Lung Adenocarcinoma, and Lapatinib, targeting EGFR in Glioblastoma. The latter is used as negative control, since Lapatinib administration to Glioblastoma patients with EGFR mutation was shown not to improve clinical outcome. Drug IC50 values were compared to STAMP’S score using Pearson’s Correlation Coefficient test.

[0104] FIG. 5A-5H illustrates the correlation of GDSC cell lines IC50 values with STAMP’S score.

[0105] FIG. 5 A — TP53, pan-cancer model score (y axis) correlates with IC50 of Nutlin of pancancer cell lines (x axis. Pearson correlation coefficient R = 0.5, p < 2.2e-16). Ellipses are shown to compare the region of mutated samples (red ellipse 501, diameter spans from IC50 of 2.5 to 8, score of -7 to 6) and of Wt samples (blue ellipse 503, diameter spans from IC50 of -1 to 5, score of -15 to -2).

[0106] FIG. 5B — ERBB2, Breast cancer model score correlates with IC50 of Afatinib (R = -0.6, p = 8.1e-06). Points are colored by ERBB2 status: Red- amplification (505), turquoise- loss of function (507), green- gain of function (509), purple- neutral (511). red ellipse (513) shows cell lines with amplification status separately examined in f.

[0107] FIGS. 5C-5D — BRAF, SKCM model score correlates with: c IC50 of PEX-4720 (R = - 0.63, p = 5.7e-05); d IC50 of Dabrafenib (R = -0.54, p = 0.0021). Samples are colored by BRAF status: Wt- blue (515), Mutant- red (517).

[0108] FIG. 5E — Similar to 5A, but only Wt TP53 cell lines are examined. Ellipses drawn in a are shown again for comparison. A subset of samples with Wt TP53 shows resistance to Nutlin, similar to most mutated samples (as emphasized by the red ellipse 502). STAMP’S score statistically significantly correlates with IC50 of Nutlin when examining samples with Wt TP53 status only (R = 0.38, p = 2.7e-10).

[0109] FIG. 5F — Similar to 5B, but only cell lines with ERBB2 amplification are shown. Despite the very low sample size, the score statistically significantly correlates with IC50 of Afatinib when examining samples with amplified ERBB2 status only (R = -0.65, p = 0.011).

[0110] FIG. 5G — EGFR, GBM model score and Lapatinib IC50 correlation is tested as negative control. Samples are colored for EGFR status: red- mutated (519), blue- Wt (521). As expected, no statistically significant correlation is observed (R = -0.088, p = 0.68).

[0111] FIG. 5H — IC50 levels of mutated (red) (523) and Wt (blue) (525) EGFR samples are shown in boxplots with no observed significant difference (Wilcoxon sum rank test, p = 0.16), confirming this as a negative control.

[0112] FIG. 5A-5H Abbreviations: GDSC- Genomics of Drug Sensitivity in Cancer; Mut- Mutated; Wt- Wildtype; Amp.- amplified; Loss- loss of function; Gain- gain of function; BRCA- Breast Cancer; SKCM- Skin cutaneous melanoma; GBM- Glioblastoma multiforme.

[0113] For TP53 in the pan-cancer cohort, Nutlin IC50 and the score had a correlation coefficient of R = 0.5 (p-val < 2.2e-16, Figure 5 A), thus representing resistance of tumors with aberrant TP53 activity to Nutlin. Cell lines are colored for TP53 mutational status, emphasizing the model’s success in predicting the aberrant state as well as strongly correlating with Nutlin’ s IC50. For ERBB2 in Breast cancer, correlation coefficient of Afatinib IC50 and the score was R = -0.6 (p = 8.1e-06, FIG. 5B), representing sensitivity of cell lines with increased ERBB2 activity to the drug. Cell lines are colored for ERBB2 status (Amplified, loss of function, gain of function, and neutral). For BRAF in Skin cancer, correlation coefficient for PLX-4720 was R = -0.63 (p = 5.7e-05, FIG. 5C) and for Dabrafenib R = -0.54 (p = 0.0021, FIG. 5D). BRAF cell lines are colored for the presence of V600E mutation (coloring for mutant compared to Wt is shown in FIG. 16A-16B). The correlation of STAMP predictions was also examined for samples with the same genetic profile. Remarkably, even when examining only Wt TP53 cell lines that are expected to be sensitive to Nutlin, the score had a strong correlation with IC50 (FIG. 5E, R = 0.38, p = 2.7e-10).Of note, samples with TP53 Wt that had a score similar to that of mutant TP53 (pseudo-Mutantsamples) also had higher ICsovalues (as shown by the red circle (502) in FIG. 5E). For TP53 mutant-only samples, where Nutlin is not expected to have an effect, the correlation was weak but still statistically significant (R = 0.095, p = 0.032, FIG. 16C). Pseudo Wt TP53 had IC50 values similar to those of true Wt samples (as shown by the blue circle (1601) in FIG. 16C). A similar pattern was observed for Breast cancer cell lines with amplified ERBB2 (samples inside the red circle in FIG. 5B), where the score was able to stratify their Afatinib IC50 values (R=-0.65, p = 0.011, FIG. 5F). When examining only the subset of Skin cancer cell lines with V600E mutation in BRAF, results showed a similar trend (R = -0.35, p = 0.08 for PLX-4720; R = -0.32, p = 0.14 for Dabrafenib, FIG. 16D-16E). The borderline statistical significance values are likely due to the very small sample size. These analyses prove an important function of STAMP in identifying drug susceptibility even amongst samples with an identical genetic profile.

[0114] For the negative control, comparing Lapatinib IC50 in Glioblastoma to the EGFR model’s score was statistically insignificant (R = -0.088, p = 0.68, FIG. 5G). This is in accordance with indistinctive IC50 values of mutated compared to Wt EGFR in the Glioblastoma samples (FIG. 5H, Wilcoxon sum rank test p = 0.16). For KRAS G12C inhibitor in LU AD, the correlation with STAMP’S score was also not significant (R = -0.19, p = 0.2, FIG. 16F. samples colored for presence of G12C mutation). This is in accordance with the KRAS-G12C mutated samples and samples without this mutation having similar, relatively high, IC50 values (FIG. 16G, Wile, p = 0.1). The KRAS G12C inhibitor effect is not significant for samples based on mutation status (any mutation) in KRAS as well (FIG. 16H-16I, Wile, p-val = 0.31). Thus, these analyses serve as a second negative control.

[0115] Of note, MDM2 expression levels might be the explanation for the cell line Nutlin response stratification by the TP53 model shown in FIG. 5E. If true, this would suggest that STAMP’Scontribution to the significance of these results is minimal. To examine this, cell lines from the Nutlin3 cohort with TP53 WT status (as shown in FIG. 5E) and with TP53 mutant status (as shown in Supplementary FIG. 16C) were colored for MDM2 amplification status. As shown in FIG. 18A-18B, coloring cell lines for their MDM2 amplification status did not produce an explanation for the significant stratification achieved by STAMP of response to Nutlin3.Examples and Further Details

[0116] This disclosure describes STAMP, a machine learning algorithm to predict and quantify driver events in a gene specific and cancer-type specific manner based on gene expression information. Conventional and deep learning models were explored along with different strategies to reduce the dimensionality of the RNA-Seq data. Models were trained to predict TP53 mutational status and were then expanded to multiple genes and pathways based on driver event annotations.

[0117] STAMP achieves high accuracy on an unseen TCGA test set and on an independent cohort - METABRIC, proving its utility in predicting mutational type for new patients. The fact that METABRIC samples used for prediction are derived from a microarray platform, while the model is trained on TCGA’s samples derived from RNA-Seq platform, further validates and establishes the model’s robustness. STAMP was shown to have improved prediction compared to mere mutational analysis for the case of TP53, by achieving stronger correlation with relevant SBS and with survival data.

[0118] STAMP compared to ENLIGHT study

[0119] STAMP is an improvement over state-of-the-art tools at predicting response to treatment in clinical cohorts, suggesting potential biomarkers for cancer treatment.

[0120] A major collection of clinical cohorts, containing both mRNA and response to treatment data, was recently made available through the ENLIGHT study. ENLIGHT is a transcriptomic based tool for drug response prediction in cancer patients, and the datasets used for tuning and validating ENLIGHT were used here to examine STAMP for stratification of drug response in cancer patients. Cohorts for which the tested drug targeted a gene or a pathway with a relevant STAMP model were analyzed, and a comparison was made with ENLIGHT’ s score.

[0121] Ten cohorts were identified as relevant for evaluating STAMP, comprising 508 cancer patients, treated with 5 distinct drugs in 6 distinct setups (Table 3 and Table 4). Other than cohorts with well-established drug targets, the PTEN model was also evaluated for a cohort of Breast Cancer patients treated with Anti-PD 1. PTEN knock down in Breast Cancer cell lines was shown to cause significantly higher PD-L1 expression. PTEN alterations were also shown to be associated with significantly lower Overall Response Rate in metastatic triple negative breast cancer patients treated with Anti-PD 1. Hence, the PTEN model was examined as potential biomarker for Anti- PD 1 response in Breast Cancer patients.

[0122] Table 3 Clinical cohorts used to assess STAMP| Overall: n = 846 |

[0123] Table 4 Clinical cohorts excluded from analysis between STAMP and ENLIGHT

[0124] For 5 of the 6 setups, STAMP achieved a significant stratification of response to treatment(Hepatocellular Carcinoma patients treated with Sorafenib, predicted by the RTK-RAS pathway model, Wilcoxon p = 4.5e-07; Breast Cancer patients treated with Sorafenib, predicted by the RTK-RAS pathway model, Wilcoxon p = 0.012; Combined five cohorts of Breast Cancer patients treated with Trastuzumab, predicted by the ERBB2 gene model, Wilcoxon p = 2.6e-04; Head and Neck Cancer patients treated with Cetuximab, predicted by the EGFR gene model, Wilcoxon p = 0.028; Breast Cancer patients treated with Durvalumab, predicted by the PTEN gene model, Wilcoxon p = 2.04e-04; For Breast Cancer patients treated with Lapatinib, predicted by the RTK- RAS pathway model, results were not statistically significant. Wilcoxon p =0.2. FIGS. 6A-6F). All cohorts except the two treated with Sorafenib received additional treatments other than the mentioned targeted therapy (Table 3), thus emphasizing STAMP’S robustness.

[0125] Compared to ENLIGHT, STAMP improved AUC of response to treatment in 3 of the 6 setups (Sorafenib in Hepatocellular carcinoma, Sorafenib in Breast cancer and Anti-PDl in Breast cancer) with up to 22.76% improvement in AUC. In the other setups, STAMP was inferior by only up to 6.15% in AUC (FIGS. 6A-6F). While ENLIGHT’s strength is stated in drugs with more accurate targets, it is noted to be less predictive for drugs with many targets such as Sorafenib. The major advantage STAMP presents for Sorafenib is arguably due to its ability to predict not only gene but also pathway-level dysfunction, which accounts for the multiple targets of Sorafenib.

[0126] The I-SPY2 study is an adaptive clinical trial platform testing drugs for high-risk breast cancer patients in the neoadjuvant setting. One arm of I-SPY2 was already presented in the analysis above (Using the Anti-PDl agent Durvalumab in Breast cancer patients) as it was also used in the validation of ENLIGHT (FIG. 6F). However, other arms of I-SPY2 have tested different treatments on different patients. These include patients treated by chemotherapy combined withHER2 related therapies, as well as with Pembrolizumab (Table 3 and Table 4). These patients were not evaluated by ENLIGHT, presenting an opportunity to further expand the clinical analysis. The ERBB2 gene model for Breast cancer was evaluated, significantly stratifying response on all patients treated by a Her2 related therapy (Wilcoxon p = 5.62e-03, FIG. 7A. see Table 3 and Table 4 for treatments included and cohort size and for excluded samples). ENLIGHT’ s score was also generated for these patients and STAMP showed advantage in AUC (AUC = 59.78% for STAMP, AUC = 57% for ENLIGHT. FIG. 7B). Importantly, when examining LSPY 2 patients with similar Her2 status, STAMP was still able to stratify treatment response (patients with: Her2 positive, Wilcoxon p = 0.068; Her2 negative, Wilcoxon p = 0.038; FIGS. 7C-7D).

[0127] LSPY2 also includes patients with Her2 negative status tested for response to the Anti- PD1 drug Pembrolizumab. This allowed a second evaluation of STAMP’S predictability of response to anti-PDl. Indeed, the STAMP PTEN gene model for Breast cancer successfully stratified response to Pembrolizumab (Wilcoxon p = 0.032, FIG. 7E). STAMP also showed advantage over ENLIGHT (AUC = 65.32% for STAMP, AUC = 62.39% for ENLIGHT, FIG. 7F).

[0128] Overall, STAMP was able to stratify response to treatment in several clinical cohorts, applying both gene specific and pathway specific models for relevant targeted therapies. STAMP’S scores suggest potential important biomarkers that could be used in clinical settings. STAMP improved predictions over ENLIGHT, the current state-of-the-art method, in several cohorts.

[0129] FIG. 6A-6F. STAMP stratifies response to treatment in clinical cohorts.

[0130] FIGS. 6A-6F Boxplots show STAMP’S stratification of patients, and ROC curves comparing STAMP 601 with ENLIGHT 603, for the following clinical setups:

[0131] FIG. 6A — Hepatocellular Carcinoma patients treated with Sorafenib, predicted by the RTK RAS Pathway model.

[0132] FIG. 6B — Breast Cancer patients treated with Sorafenib, predicted by the RTK RAS Pathway model.

[0133] FIG. 6C — Breast Cancer patients treated with Lapatinib, predicted by the RTK RAS Pathway model.

[0134] FIG. 6D — Breast Cancer patients treated with Trastuzumab, predicted by the ERBB2 Gene model.

[0135] FIG. 6E — Head and Neck Cancer patients treated with Cetuximab, predicted by the EGFR Gene model.

[0136] FIG. 6F — Breast Cancer patients treated with Durvalumab, predicted by the PTEN Gene model. STAMP statistically significantly stratified all but the Lapatinib setup, suggesting that STAMP’S prediction could be used as a biomarker for treatment response. The ROC curves show STAMP improved performances over ENLIGHT in 3 of the setups, with up to 22.76% improvement in AUC. Where STAMP did not improve performances, AUC difference was up to 6.15%.

[0137] FIG. 7A-7F. STAMP stratifies I-SPY 2 patients for response to Her2 related treatments, in patients with the same Her2 status, and for response to Anti-PDl.

[0138] FIG. 7A-7B — STAMP stratifies all patients treated with a Her2 related therapy (Wilcoxon p = 5.621e-03), with a small advantage over ENLIGHT (AUC = 59.78% for STAMP, AUC = 57% for ENLIGHT). See also Table 3 and Table 4 for more information.

[0139] FIG. 7C-7D — STAMP stratifies response to treatment in patients both positive and negative Her2 status (Wilcoxon p = 0.068 for Her2 positive, Wilcoxon p = 0.0378 for Her2 negative).

[0140] FIG. 7E-7F — Patients in this arm of I-SPY 2 treated with pembrolizumab allow further validation for the PTEN model predictability of response to Anti-PD 1 (Pembrolizumab, Wilcoxon p = 0.0326), again with an advantage over ENLIGHT (AUC = 65.32% for STAMP, AUC = 62.39% for ENLIGHT).

[0141] The learnt mRNA signatures are designed to provide guidance for genetic profiling that would complement mutational analysis. STAMP can potentially identify “pseudo” samples, i.e., mutated tumors that behave as if the protein has Wt-like activity, or Wt tumors with an aberrant protein’s activity pattern, as illustrated in FIG. 9A-9C.

[0142] In clinical practice, this could assist in identifying driver events based on the gene expression profile, even in samples for which no targetable mutations are found. One of the novelties of this work is that in addition to the binary prediction, STAMP forms an effect quantification score. Such a score can guide driver events prioritization for treatment in samples where many targetable driver mutations occur. Indeed, STAMP’S predictions were shown to correlate with many drugs’ IC50 levels in the cell lines’ database of GDSC. The strong correlation shown for this score with variable drug IC50 values, even for samples with the same genetic profile (e.g., all Wt TP53), demonstrates the potential for prioritization of patients in a manner not feasible by mutational analysis alone. The quantifying score can stratify patients for their expected drug response. Survival outcome stratification was also possible based on STAMP’S TP53 pan-cancer model, which further indicates the validity of the score in measuring effect severity of driver events.

[0143] STAMP may be trained on human tumor RNA-Seq samples derived from TCGA.Examples of validation for drug response predictability from human samples and their corresponding response to treatment are shown in, for example FIG. 6A-6F and FIG. 7A-7F of this disclosure. Hence, STAMP has been validated on cell line data from the GDSC database, which uses microarray for gene expression profiling. The RNA-Seq human based model successfully predicting for micro-array cell line samples comes as another evidence for STAMP’S robustness. Most studies in the literature examine cell line response to treatment to deduce predictions for human profiles, but this disclosure describes an opposite approach. While STAMP models were trained on human tumors, they are tested on cell line response to treatment and show strong correlations across multiple drugs, targeted genes, and tumor types. Since the models were exposed only to human mRNA profiles, it is possible that predictability of human tumor’s response to treatment would be even better than that presented in this study for cell lines.

[0144] GCN appears to be the best model due to its superiority in predicting using METABRIC, as well as on the pan-cancer TCGA test set and on the validation sets. GCN, being a deep learning architecture, is also expected to have a growing advantage in the future as public data on cancer patients accumulate.

[0145] STAMP was trained on TCGA data, which is normalized as z scores based on healthy tissue gene expression profiles. To optimally predict for new data, STAMP requires a similar normalization, which is mostly inapplicable. Here we used a simplified approach, normalizing for each new dataset using all expression profiles available in that set. This simplified approach worked surprisingly well, especially in datasets with a large sample size such as METABRIC, I- SPY2 and GDSC. However, for very small datasets, as some of the ENEIGHT assembled cohorts are, this approach would not work properly. Thus, the normalizing methodology is a current caveatin STAMP’S predictability, and an important direction for future work. Ideally, STAMP would be applicable for even a single sample. Improving the normalization technique is also highly likely to improve STAMP’S current presented achievements.

[0146] While each STAMP model was trained on many tumors to gain a gene specific prediction, they may also be applied on specific patients, to form a genetic profile. Multiple pathways and gene specific models can be applied in parallel on a patient’s RNA-Seq sample, to establish a rich description of the patient’s status. In a clinical framework, such profiling could help direct drug administration and prioritization, including for combination therapy.Methods / Data Sources and availability

[0147] Gene expression data

[0148] RNA-Seq gene expression data of TCGA pan-cancer samples was downloaded. IlluminaHiSeq was used for the RNA Sequencing. The dataset was downloaded from the following link, HiSeqV2.gz - Academic Torrents , uploaded by Xena team, UCSC. The gene expression values are log2(x+l) transformed from RS EM (RSEM: accurate quantification of gene and isoform expression from RNA-Seq data, publicly available).

[0149] TCGAbiolinks package version 2.12.6 in R was also used to download raw gene expression counts, that were used for the differential expression analysis using the Limma package.

[0150] TP53 Mutational status and clinical data downloading and integration

[0151] Pan-cancer data was downloaded from cBioPortal using the ‘Pan-cancer TCGA’ quick select. TCGAbiolinks was also used to extract the clinical and mutational data, allowing a greater number of samples with mutation profiles. Samples with conflicting mutational status were removed from the analysis, leading to 10,258 profiled samples in the final analysis, with 35.68%mutated samples (3,660). Survival data, where available, was also downloaded from cBioPortal. The profiled data was integrated with the RNA-Seq data to produce a final dataset (9,077 samples).

[0152] Pathway and other genes data

[0153] Data used for labeling the training data of genes and pathways other than TP53 was downloaded from supplementary table S4 of Sanchez-Vega et al. This table contains genomic alteration matrices (amplifications, deletions, point mutations, epigenetic silencing, and fusions) for important cancer genes in 9,125 TCGA samples. Alterations were excluded with low recurrence across tumor samples and alterations annotated as likely passenger by OncoKB. A map for the genes’ alteration status was then defined based on the different alterations recognized in that gene. Alteration status for genes in each pathway are integrated to define pathway level alteration status. For the analysis presented in this work, the alteration status in gene and pathway levels were used for labeling.Training algorithms and feature selection

[0154] Differential Expression using Limma package

[0155] The Voom command from Limma package was used to perform differential expression analysis. All genes in the RNA-Seq data of the training set samples were examined and were ranked based on correlation to TP53 mutational status (expression of the TP53 Wt group was compared to that of the TP53 mutant group). The 150 most positively and the 150 most negatively correlated genes were selected (highest / lowest Log2FC).

[0156] MutSigDB set for TP53

[0157] From the Molecular Signature Database the set of genes named“FISCHER_DIRECT_P53_TARGETS_META_ANALYSIS” were selected. This set contains 311 genes directly bound and regulated by p53.

[0158] Random forests

[0159] Random Forests is an ensemble learning method, used in this case for binary classification. This is done by using the bagging technique on decision trees. Bagging means to repeatedly select n (which is a hyperparameter) random samples from the training set, before fitting a decision tree for those samples. In addition to bagging, ‘RF’ also randomly selects a subset of features. The decision tree is thus trained on a subset of samples and on a subset of features from the original data. Each tree is trained as a binary classifier, and the algorithm chooses the class according to a majority tree votes.

[0160] Elastic net Regression

[0161] For elastic net regression, ‘ElasticNet’ from the scikit-learn package in python was used. Elastic net regression is a specific regularization technique applied in this case with logistic regression. It linearly combines the LI and L2 penalties of the lasso and ridge methods respectively. It can “reduce” features by generating zero-valued coefficients, and thus reduce overfitting (the effect of the Lasso penalty). At the same time, it will reduce impact of features less relevant for learning (the effect of the ridge penalty).

[0162] Graph Convolution Networks

[0163] Dutil et al. (Dutil F, Cohen JP, Weiss M, Derevyanko G, Bengio Y. Towards Gene Expression Convolutions using Gene Interaction Graphs. Published online 2018) provides a method to implement GCN on gene expression data. The present specification adopts the genespecific approach used in Dutil et al. to predict a gene’s masked expression levels from transcription patterns of other genes. The convolution operations are used to retrieve information for a node according to its edges. Convolution is approximated to first degree neighbor nodes, meaning only nodes directly connected to the queried gene are used, as suggested by Defferrard et al. (Defferrard M, Bresson X, Vandergheynst P. “Convolutional neural networks on graphs with fast localized spectral filtering”. Adv Neural Inf Process Syst. 2016;(Nips):3844-3852) and by Kipf & Welling (Kipf TN, Welling M. “Semi-supervised classification with graph convolutional networks”. 5th International Conference on Learning Representations, ICLR 2017 - Conference Track Proceedings. 2017.). Both the graph’s adjacency matrix and the samples’ gene expression table are processed before the convolution. The identity matrix is added to the adjacency matrix A to form self-loops for all genes. A is also normalized using D, the diagonal node degree matrix, to neutralize discrepancy that could happen due to the different number of edges each node has. This normalization is done using the spectral propagation rule as proposed by Kipf & Welling (2016). The multiplication for one convolutional layer is: A’X(i)0, where A’ is the normalized adjacency matrix, X(i) is the Ithlayer’s input and 0 is the Ithlayer’s weights. This three matrices multiplication is done in each layer, but the adjacency matrix remains the same for all layers and multiplies X, a matrix of n (nodes) times c (features). Therefore, the number of learned parameters per layer is only c times o, where o is the number of output parameters in the given layer. Expression data is embedded so that each gene is represented by a vector of parameters learned during training. Each convolution layer also contains the following steps: a skip connection is added to preserve both the given node and its neighborhood, and a ReLU activation function is applied. Aggregation of nodes is done to reduce the number of genes after each layer, based on hierarchical clustering. The graph connectivity map is used for clusters’ variance calculation, and Max pooling is performedon resulting clusters. When tested, dropout is used with a 40% chance for each node to be dropped after every convolution layer. In the final layer the remaining nodes are concatenated into a linear layer, followed by an activation function (e.g., sigmoid) that generates the binary classification prediction.

[0164] Division into train, validation, test

[0165] For training each tumor type and gene model, 20% of the data are kept for the test set. 60% of the data are used for training the model and hyperparameter tuning is performed on the remaining 20% used as the validation set. Only the best performing models for each algorithm (RF, ELR, GCN) are examined on the test set.

[0166] Hyperparameter tuning

[0167] A grid search approach may be used to train the different models. For GCN, the following hyperparameters may be tuned: graph structures, number of layers, number of weights in each layer, batch size and learning rate. An example for comparison of two hyperparameters in the grid is shown in FIG. 17. For RF tuned hyperparameters were n_estimator, criterion and maximum features. For ELR learning rate and ratio were tuned.

[0168] Funcoup graph structure

[0169] FunCoup used a Bayesian algorithm and orthology-based information transfer to infer functional associations. Edges in FunCoup have been given confidence scores for supporting different interactions such as metabolic, signaling, complex and physical protein interactions. It integrates data from multiple levels of biological processes, like co-expression, phylogenetic profile similarity, PPI, subcellular co-localization, and genetic interaction profile similarities. It contains 5,505,787 edges covering 16,765 gene nodes.

[0170] HumanNet v2 graph structure

[0171] HumanNetV2 contains functional associations inferred from PPIs, gene co-expression, protein co-occurrence, genomic contexts and more. It integrates these diverse types of omics data using Bayesian framework. HumanNetV2 contains 476,399 edges covering 16,243 gene nodes.Analysis and Validation

[0172] Quantitative score for protein / pathway aberration severity

[0173] The score for the GCN model was produced by extracting the probability scores produced by the model, and by inverting the activation function responsible for the binary pattern of prediction. In an embodiment, where the activation function is a Sigmoid defined as: o (x) = 1 / (1 + exp(-x)). The inverse Logit function may also be used to calculate a score: logit(p) = o-1(p) = ln(p / (l-p))

[0174] Tumor types excluded from the TP53 mutational state analysis

[0175] Tumor types with n < 13 samples in either labeling group (mutated samples or non-mutated samples) were excluded from the analysis.

[0176] Down-sampling Criteria for further analysis of tumor types, genes and pathways

[0177] Tumor types were excluded if they had less than 100 samples (UCS, KICH, ACC, READ) or if they had a mutated (or non-mutated) percentage under 15% and number of samples under 200 (LAML, ESCA).

[0178] METABRIC analysis

[0179] Gene expression samples were downloaded using the MetaGxBreast package, while mutational annotations were downloaded from www.cbioportal.org. To fit the values to TCGAnormalization scores, the Z score was calculated with a normalization over all METABRIC samples, and then GCN models are applied on the normalized score. ROC curves and AUC scores were calculated and visualized in R.

[0180] GCN score compared by mutation types and by TP53_PROF classification

[0181] TP53 mutation status was taken from TCGA annotations as downloaded from cBioPortal. Deletions and insertions were combined for Inframe and frameshift mutations. TP53_PROF classification for deleterious or non-deleterious TP53 missense variants was extracted from supplementary table S5 from Ben Cohen et al. in Briefings in Bioinformatics Vol. 23, Iss. 2, March 2022 (reproduced in the Appendix).

[0182] Mutation type and TP53_PROF comparison

[0183] The TP53 mutation type comparison was performed on pan-cancer data for power analysis considerations, as TP53_PROF predicts only a few samples as non-deleterious for any specific tumor type. Truncating included nonsense and fusion mutations. Splice mutations included splice site, splice region and translation site. Inframe and Frameshift included both insertions and deletions. Missense variants were separated to deleterious (D) or non-deleterious (ND) according to TP53_PROF’s predictions.

[0184] GDSC cell lines drug response analysis

[0185] Gene expression, drug response and mutational data of GDSC was downloaded from the appropriate website. Gene expression is derived using Affymetrix Human Genome U219 Array and is normalized using RMA. To fit to TCGA normalization, Z score was calculated with a normalization over all GDSC cell lines, and then GCN models are applied on the normalized score.Tumor type and gene models were selected if: (i) they had a GCN model trained on them (i.e. ifthey passed the down-sampling criteria), (ii) The corresponding cell lines in GDSC had data regarding response to drugs targeting the query gene. Four genes fit this threshold: ERBB2, BRAF, EGFR and KRAS. For each the tumor type where GCN was trained was selected, and the compound targeting this gene. Five analyses were therefore performed: ERBB2 model in breast cancer (the only model for ERBB2) with the drug Afatinib; BRAF in SKCM with the drugs Dabrafenib and PLX-4720; KRAS in LU AD with KRAS (G12C) inhibitor. EGFR in Glioblastoma (the only model for EGFR) with the drug Lapatinib was used as negative control since Glioblastoma patients with EGFR mutations do not respond to this drug.

[0186] Data availability

[0187] Gene expression data for TCGA is available from:

[0188] Academic Torrents, which is a product of the Institute for Reproducible Research (a U.S. 501(c)3 nonprofit) and by using TCGABiolinks package for differential expression analysis. Clinical and mutational annotations were downloaded from www.cbioportal.org and were compared with TCGABiolinks package annotations. MSigDB sets were downloaded from: the Gene Set Enrichment Site from the Broad Institute . The GCN application code with the graph structures for Funcoup 4 and Humannet V2 graph structures were downloaded from the repository of Dutil et al. (Biorxiv, 2018), available at Bertinus / Gene-Graph- Analysis, which is a public site and provides, amongst other things, a copy of the reference “Analysis of Gene Interaction Graphs as Prior Knowledge for Machine Learning Models" TP53_PROF predictions are available from Ben-Cohen et al, 2022, TP53JPROF: a machine learning model to predict impact of missense mutations in TP53, Briefings in Bioinformatics, Volume 23, Issue 2, March 2022, bbab524.

[0189] GDSC data (gene expression, mutational and clinical) was downloaded from the Genomics of Drug Sensitivity in Cancer site, jointly sponsored by the Wellcome Sanger Institute and Massachusetts General Hospital Cancer Center. This site is a repository for approximately 1000 human cancer cell lines which have been characterized and screened with 100s of compounds and allows a user to find drug response data and genomic markers of sensitivity. METABRIC data (gene expression and clinical) was downloaded using the MetaGxBreast package in R. Mutational data for METABRIC was downloaded from cBioPortal.

[0190] FIG. 10 Multivariable analysis for coxph model of TCGA’s pan cancer cohort, separated to quartiles based on STAMP’S linear score. Survival distinction is statistically significant even when accounting for tumor type, age and mutation burden.

[0191] FIG. 11A-11E shows that survival distinction is enhanced by STAMP’S predictions compared to original mutational state.

[0192] FIG. HA on Pan cancer and FIG. 11B-11E on several tumor type cohorts in TCGA (Pancreatic adenocarcinoma, PA AD; Fiver hepatocellular carcinoma, EIHC; Uterine corpus endothelial carcinoma, UCEC; Colorectal adenocarcinoma, COAD, respectively). For COAD, STAMP did not improve survival distinction. See supplementary information below for further description of this analysis.

[0193] FIG. 12A-12I. SBS analysis for APOBEC and HR-DDR signatures on Breast cancer (BRCA, all signatures) and on Head and neck cancer (HNSC, APOBEC signatures).

[0194] FIG. 12A — STAMP enhances SBS enrichment for a SBS3 in BRCA.

[0195] FIG. 12B - SBS 13 in BRCA.

[0196] FIG. 12C - SBS2 in HNSC.

[0197] FIG. 12D-12F — When separating between true WT (blue - 1201), true mutant (red - 1203), pseudo mutant (green - 1205) and pseudo WT (yellow - 1207), SBS enrichment is observed even between samples with the same actual mutational state. (12D-12F for SBS3 in BRCA, SBS13 in BRCA, SBS2 in HNSC respectively).

[0198] FIG. 12G-12I — STAMP’S linear score is correlated with SBS enrichment (g-i for SBS3 in BRCA, SBS 13 in BRCA, SBS2 in HNSC respectively). See supplementary information below for further description of this analysis.

[0199] FIG. 13A-13F. SBS analysis for APOBEC and HR-DDR signatures on Breast cancer (BRCA, all signatures) and on Head and neck cancer (HNSC, APOBEC signatures).

[0200] FIGS. 13A-13D — STAMP enhances SBS enrichment. (13A-13D) for SBS13 in pan cancer, SBS2 in pan cancer, SBS2 in BRCA, SBS 13 in HNSC respectively).

[0201] FIGS. 13E-13F - similar to FIGS. 12d-12f, but for SBS2 in BRCA and for SBS13 in HNSC, respectively.

[0202] FIG. 14A-14K.

[0203] FIGS. 14A-14H — STAMP’S linear score correlates with missense mutation classification based on TP53_PROF in: Low Grade Glioma (LGG), Colorectal Cancer (COAD), Lung adenocarcinoma (LUAD), Stomach adenocarcinoma (STAD), Head and neck Carcinoma (HNSC), Breast Cancer (BRCA), Uterus corpus endothelial carcinoma (UCEC), Bladder cancer (BLCA), respectively.

[0204] FIG. 141 — STAMP’S linear score for samples based on their TP53 mutation type. Frameshift (red), Truncating (yellow) and Splice (green) variants all behave similar to Deleterious(D) missense variants. Unexpectedly, Inframe (purple) samples also behave like D missense variants. FIGS. 14J-14K are similar to FIG. 141, but in specific tumor type cohorts:

[0205] FIG. 14J - BRCA and

[0206] FIG. 14K — LU AD. In-frame variants seem to show a pattern closer to ND in these tumor type specific cohorts.

[0207] FIG. 15A-15D.

[0208] FIG. 15A — applying STAMP to various pathways in several tumor type cohorts, gray slots (1501) show pathway and tumor type combinations that did not pass the down-sampling criteria.

[0209] FIG. 15B — STAMP’S predictions enhance survival distinction based on LATS2 model in Low Grade Glioma (LGG), compared to the Sanchez-Vega et al.duti annotations (the label used for training, compare this with FIG. 3B).

[0210] FIG. 15C — STAMP’S linear score can form 3 distinct groups in LATS2 LGG cohort, when combining Q3 and Q4 samples together (red - 1503), QI in blue (1505), and Q2 in orange (1507).

[0211] FIG. 15D — multivariable analysis shows STAMP’S LATS2 model in LGG maintains statistical significance in survival analysis when accounting for age and mutation burden.

[0212] FIG. 16A-16I. GDSC analysis.

[0213] FIG. 16A — Skin cancer (SKCM) BRAF model stratifies GDSC cell lines for their PLX- 4720 IC50 values, cell lines are colored by BRAF mutation status (orange 1603, blue 1605).

[0214] FIG. 16B — similar to 16A for Dabrafenib IC50.

[0215] FIG. 16C — TP53 mutated only samples are still stratified by STAMP’S linear score. Most mutated samples have a high IC50, and a high linear score value (red circle 1601, similar to thecircle in FIG. 5A). A small portion of cell lines present with WT like IC50 levels, and also has lower STAMP linear score values (“Pseudo WT” samples given in the blue circle 1601, similar to the circle in FIG. 5A).

[0216] FIG. 16D-16E — PLX-4720 and Dabrafenib respectively, show correlation with STAMP’S score in mutated only samples, with borderline statistical significance (likely due to the small sample size).

[0217] FIG. 16F — KRAS model in Lung adenocarcinoma (LU AD) on KRAS G12C inhibitor is used as negative control, with no significant correlation of IC50 levels to STAMP’S score. Samples with a mutation in BRAF are colored in red (1607), and WT BRAF samples are colored in blue (1609).

[0218] FIG. 16G — the lack of correlation is in accordance with indistinct IC50 levels for BRAF mutated (red - 1611) compared to WT (blue - 1613) cell lines.

[0219] FIG. 16H-16I — similar to FIGS. 16F-16G, only this time samples with G12C mutation in KRAS are colored in red 1615, and the rest of samples are colored in blue 1617.

[0220] FIG. 17. Preliminary investigation of different graph structures to train STAMP on Breast cancer for TP53 mutational status. Many experiments were conducted, implying that funcoup and humannetV2 are preferable on the genemania and on the regnet graphs. Dropout was also examined and did not show improvement in performance. Further tuning of models may also be performed along similar lines, as needed.

[0221] FIG. 18A-18B illustrates the correlation of the score with IC50 for WT and M groups on the pan-cancer dataset

[0222] Computer Hardware

[0223] The present systems and methods may include implementation on a system or systems that provide multi-processor, multi-tasking, multi-process, and / or multi-thread computing, as well as implementation on systems that provide only single processor, single thread computing. Multiprocessor computing involves performing computing using more than one processor. Multi-tasking computing involves performing computing using more than one operating system task. A task is an operating system concept that refers to the combination of a program being executed and bookkeeping information used by the operating system. Whenever a program is executed, the operating system creates a new task for it. The task is like an envelope for the program in that it identifies the program with a task number and attaches other bookkeeping information to it. Many operating systems, including Linux, UNIX®, OS / 2®, and Windows®, are capable of running many tasks at the same time and are called multitasking operating systems. Multi-tasking is the ability of an operating system to execute more than one executable at the same time. Each executable is running in its own address space, meaning that the executables have no way to share any of their memory. This has advantages, because it is impossible for any program to damage the execution of any of the other programs running on the system. However, the programs have no way to exchange any information except through the operating system (or by reading files stored on the file system). Multi-process computing is similar to multi-tasking computing, as the terms task and process are often used interchangeably, although some operating systems make a distinction between the two.

[0224] The present invention may be a system, a method, and / or a computer program product at any possible technical detail level of integration. The computer program product may include a computer readable storage medium (or media) having computer readable program instructions thereon for causing a processor to carry out aspects of the present invention. The computerreadable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device.

[0225] Fig. 19 illustrates a computing device 1900 as it relates to the present disclosure. The computing device 1900 may, for example, perform calculations, execute routines and algorithms, process data, communicate with other devices via a network, and display results. For example, a computing device 1900 may comprise a processor or CPU 1904, a network adapter 1906 for communication with a network 1908. The network 1908 may connect the computing device 1900 to other devices 1950 or to other data (not shown in the figure). The computing device may comprise an input / output device 1902. Such an input / output component 1902 may be an input device, an output device, or both and the computing device 1900 may have several such components. Example input devices 1902 include a keyboard, a mouse, a microphone, a touchpad, a joystick, and the like. Example output devices 1902 include a display, a speaker, a haptic feedback device, and the like. The computing device 1900 may further comprise memory 1910 or a computer readable storage medium 1910. In the computer memory 1910 may reside instructions for carrying out the methods and techniques described elsewhere in this disclosure. The computer memory 1910 may also comprise an operating system 1930 for control of the various parts and components of the computing device 1900. The memory 1910 may also store data, for example training data 1912, testing data 1914, and validation data 1916. The memory 1910 may also comprise algorithms such as a graph convolutional network classifier 1918, other classifier(s) 1920, machine learning algorithms 1922, score calculating algorithms 1924, or other algorithms 1926.

[0226] The computer readable storage medium may be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagneticstorage device, a semiconductor storage device, or any suitable combination of the foregoing. A non-exhaustive list of more specific examples of the computer readable storage medium includes the following: a portable computer diskette, a hard disk, a random access memory (RAM), a readonly memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanically encoded device such as punch-cards or raised structures in a groove having instructions recorded thereon, and any suitable combination of the foregoing. A computer readable storage medium, as used herein, is not to be construed as being transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission media (e.g., light pulses passing through a fiber-optic cable), or electrical signals transmitted through a wire.

[0227] Computer readable program instructions described herein can be downloaded to respective computing / processing devices from a computer readable storage medium or to an external computer or external storage device via a network, for example, the Internet, a local area network, a wide area network and / or a wireless network. The network may comprise copper transmission cables, optical transmission fibers, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives computer readable program instructions from the network and forwards the computer readable program instructions for storage in a computer readable storage medium within the respective computing / processing device.

[0228] Computer readable program instructions for carrying out operations of the present invention may be assembler instructions, instruction-set-architecture (ISA) instructions, machineinstructions, machine dependent instructions, microcode, firmware instructions, state-setting data, configuration data for integrated circuitry, or either source code or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, C++, or the like, and procedural programming languages, such as the “C” programming language or similar programming languages. The computer readable program instructions may execute entirely on the user's computer, partly on the user's computer, as a standalone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet Service Provider). In some embodiments, electronic circuitry including, for example, programmable logic circuitry, field-programmable gate arrays (FPGA), or programmable logic arrays (PL A) may execute the computer readable program instructions by utilizing state information of the computer readable program instructions to personalize the electronic circuitry, in order to perform aspects of the present invention.

[0229] Aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer readable program instructions. These computer readable program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of thecomputer or other programmable data processing apparatus, create means for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0230] These computer readable program instructions may also be stored in a computer readable storage medium that can direct a computer, a programmable data processing apparatus, and / or other devices to function in a particular manner, such that the computer readable storage medium having instructions stored therein comprises an article of manufacture including instructions which implement aspects of the function / act specified in the flowchart and / or block diagram block or blocks.

[0231] The computer readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process, such that the instructions which execute on the computer, other programmable apparatus, or other device implement the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0232] The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of instructions, which comprises one or more executable instructions for implementing the specified logical function(s). In some alternative implementations, the functions noted in the blocks may occur out of the order noted in the Figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flowchart illustration, andcombinations of blocks in the block diagrams and / or flowchart illustration, can be implemented by special purpose hardware -based systems that perform the specified functions or acts, or that carry out combinations of special purpose hardware and computer instructions.Example embodiments

[0233] Example 1 : A method comprising: receiving a set of curated data on human genes, wherein the curated data are RNA expression data; dividing the curated data into a first dataset, a second dataset, and a third dataset; training a classifier using the first dataset to identify a datum as mutant or wild-type; validating the classifier on the second dataset; testing the classifier on the third dataset; receiving a new datum; applying the trained classifier to the received new datum to classify the received new datum as mutant or wildtype; calculating a score associated with the received new datum; and reporting the classification and the score associated with the received new datum to a user device.

[0234] Example 2: The method of Example 1, wherein the classifier comprises a graph convolutional network, a random forest model, or a support vector machines model.

[0235] Example 3: The method of any one of Examples 1-2, wherein the score is an inverse of an activation function of a next-to-last layer of the classifier.

[0236] Example 4: The method of any one of Examples 1-3, further comprising using the score to predict the efficacy of a medication on a cell line or on a tumor.

[0237] Example 5: The method of any one of Examples 1-4, wherein the efficacy prediction comprises estimating an IC50 of the medication.

[0238] Example 6: The method of any one of Examples 1-5, further comprising identifying an active molecular pathway and a medication associated with the active molecular pathway based on the score.

[0239] Example 7: The method of any one of Examples 1-6, further comprising the step of diagnosing a patient with a condition based on the score.

[0240] Example 8: The method of any one of Examples 1-7, further comprising treating the patient based on diagnosis and an efficacy prediction of a medication on a cell line or on a tumor based on the score.

[0241] Example 9: A method of treating a medical condition of a subject in need, the method comprising the steps of: receiving a datum of the subject’s genes; applying the trained classifier to classify the received datum as mutant or wild-type; calculating a score associated with the received datum; predicting an efficacy of a medication on a cell line or on a tumor associated with the received datum; reporting the classification, the calculated score, the medication, and the predicted efficacy to a user device; and using the medication for treating the medical condition of the subject if so indicated by the calculated score.

[0242] Example 10: The method of Example 9, wherein the classifier comprises a graph convolutional network, a random forest model, or a support vector machines model.

[0243] Example 11: The method of any one of Examples 9-10, wherein the calculated score is an inverse of an activation function of a next-to-last layer of the classifier.

[0244] Example 12: A system for calculating a score comprising: a processor; and computer memory, wherein the computer memory comprises computer instructions, which, when executed by the processor perform a method comprising the steps of: receiving a set of curated data on human genes, wherein the curated data are RNA expression data; dividing the curated data into a first dataset, a second dataset, and a third dataset; training a classifier using the first dataset to identify a datum as mutant or wild-type; validatingthe classifier on the second dataset; testing the classifier on the third dataset; receiving a new datum; applying the trained classifier to the received new datum to classify the new datum as mutant or wild-type; calculating a score associated with the received new datum; and reporting the classification and the score associated with the new datum to a user device.

[0245] Example 13: The system of Example 12, wherein the classifier comprises a graph convolutional network, a random forest model, or a support vector machines model.

[0246] Example 14: The system of any one of Examples 12-13, wherein the score is the inverse of the activation function of the next-to-last layer of the classifier.

[0247] Example 15: The system of any one of Examples 12-14, further comprising using the score to predict the efficacy of a medication on a cell line or on a tumor.

[0248] Example 16: The system of any one of Examples 12-15, wherein the efficacy prediction comprises estimating on IC50 of the medication.

[0249] Example 17: The system of any one of Examples 12-16, further comprising identifying an active molecular pathway and a medication associated with the active molecular pathway based on the score.

[0250] Example 18: The system of any one of Examples 12-17, further comprising the step of diagnosing a patient with a condition based on the score.

[0251] Example 19: The system of any one of Examples 12-18, further comprising treating the patient based on diagnosis and an efficacy prediction of a medication on a cell line or on a tumor based on the score.

[0252] Although specific embodiments of the present invention have been described, it will be understood by those of skill in the art that there are other embodiments that are equivalent to thedescribed embodiments. Accordingly, it is to be understood that the invention is not to be limited by the specific illustrated embodiments, but only by the scope of the appended claims.APPENDIXTable S5 Protein variations with TP53 prediction as deleterious (D) or non-deleterious (ND)

Claims

CLAIMSWhat is claimed is:

1. A method comprising: receiving a set of curated data on human genes, wherein the curated data are RNA expression data; dividing the curated data into a first dataset, a second dataset, and a third dataset; training a classifier using the first dataset to identify a datum as mutant or wild-type; validating the classifier on the second dataset; testing the classifier on the third dataset; receiving a new datum; applying the trained classifier to the received new datum to classify the received new datum as mutant or wild- type; calculating a score associated with the received new datum; and reporting the classification and the score associated with the received new datum to a user device.

2. The method of claim 1 , wherein the classifier comprises a graph convolutional network, a random forest model, or a support vector machines model.

3. The method of claim 2, wherein the score is an inverse of an activation function of a next-to-last layer of the classifier.

4. The method of claim 1 , further comprising using the score to predict the efficacy of a medication on a cell line or on a tumor.

5. The method of claim 4, wherein the efficacy prediction comprises estimating an IC50 of the medication.

6. The method of claim 1 , further comprising identifying an active molecular pathway and a medication associated with the active molecular pathway based on the score.

7. The method of claim 1, further comprising the step of diagnosing a patient with a condition based on the score.

8. The method of claim 7, further comprising treating the patient based on diagnosis and an efficacy prediction of a medication on a cell line or on a tumor based on the score.

9. A method of treating a medical condition of a subject in need, the method comprising the steps of: receiving a datum of the subject’s genes; applying the trained classifier to classify the received datum as mutant or wild-type; calculating a score associated with the received datum; predicting an efficacy of a medication on a cell line or on a tumor associated with the received datum; reporting the classification, the calculated score, the medication, and the predicted efficacy to a user device; and using the medication for treating the medical condition of the subject if so indicated by the calculated score.

10. The method of claim 9, wherein the classifier comprises a graph convolutional network, a random forest model, or a support vector machines model.

11. The method of claim 9, wherein the calculated score is an inverse of an activation function of a next-to-last layer of the classifier.

12. A system for calculating a score comprising: a processor; andcomputer memory, wherein the computer memory comprises computer instructions, which, when executed by the processor perform a method comprising the steps of: receiving a set of curated data on human genes, wherein the curated data areRNA expression data; dividing the curated data into a first dataset, a second dataset, and a third dataset; training a classifier using the first dataset to identify a datum as mutant or wildtype; validating the classifier on the second dataset; testing the classifier on the third dataset; receiving a new datum; applying the trained classifier to the received new datum to classify the new datum as mutant or wild-type; calculating a score associated with the received new datum; and reporting the classification and the score associated with the new datum to a user device.

13. The system of claim 12, wherein the classifier comprises a graph convolutional network, a random forest model, or a support vector machines model.

14. The system of claim 13, wherein the score is the inverse of the activation function of the next-to-last layer of the classifier.

15. The system of claim 12, further comprising using the score to predict the efficacy of a medication on a cell line or on a tumor.

16. The system of claim 15, wherein the efficacy prediction comprises estimating on IC50 of the medication.

17. The system of claim 12, further comprising identifying an active molecular pathway and a medication associated with the active molecular pathway based on the score.

18. The system of claim 12, further comprising the step of diagnosing a patient with a condition based on the score.

19. The system of claim 18, further comprising treating the patient based on diagnosis and an efficacy prediction of a medication on a cell line or on a tumor based on the score.

Images