Interpretable deep learning predicts chemoresistance

Abstract

An ensemble of predictive models that elucidate how cancer mutations impact the response to common replication stress-inducing (RSi) agents. The models implement recent advances in deep learning to facilitate multi-drug prediction and mechanistic interpretation. Initial studies in tumor cells identify 41 molecular assemblies that integrate alterations in hundreds of genes for accurate drug response prediction. These cover roles in transcription, repair, cell-cycle checkpoints, and growth signaling, of which 30 are shown by loss-of-function genetic screens to regulate drug sensitivity or replication restart. The model translates to cisplatin-treated cervical cancer patients, highlighting an RTK (receptor tyrosine kinase)-JAK-STAT assembly governing resistance. This invention defines a compendium of mechanisms by which mutations affect therapeutic responses, with implications for precision medicine.

Description

[0001] INTERPRETABLE DEEP LEARNING PREDICTS CHEMORESISTANCE

[0002] CROSS-REFERENCE TO RELATED APPLICATIONS

[0003] This application claims benefit of priority of U.S. Provisional Patent Application Serial No. 63 / 621,149 filed January 16, 2024, which is incorporated by reference herein in its entirety.

[0004] GOVERNMENT SPONSORSHIP

[0005] This invention was made with government support under grant NIGMS P41 GM103504 awarded by the National Institutes of Health and under grant U54 CA274502 awarded by the National Cancer Institute. The government has certain rights in the invention.

[0006] TECHNICAL FIELD

[0007] The present invention relates to a biologically interpretable artificial intelligence model of drug response.

[0008] BACKGROUND

[0009] DNA replication is central to cell division and proliferation, involving closely orchestrated functions among hundreds of proteins (1,2). Although the replication machinery is highly accurate, it faces challenges from both extrinsic and intrinsic factors (3). These challenges can result in stalled replication forks, occurrence of DNA breaks, reduced replication precision, and other factors collectively known as RS (4). Accordingly, cells have evolved a robust RS response which activates DNA damage repair signaling, or alternatively induces cell death, to maintain genome integrity within the cell population (5-9). Due to sustained proliferative signaling and / or defective DNA repair, cancer cells undergo persistent replication stress (10,11) making them strongly RS response dependent. A consequence of this dependency is that replication stress becomes an exploitable therapeutic vulnerability in cancer treatment (12,13).

[0010] A multitude of cancer therapeutics leverage replication stress to eliminate cancer cells, employing a diversity of RSi mechanisms (Fig. 8). Classical chemotherapeutic agents induce RS by directly affecting DNA integrity. Examples include DNA crosslinkers and alkylating agents (e.g., cisplatin, methotrexate), topoisomerase I and II inhibitors (e.g., camptothecin, etoposide, doxorubicin), ribonucleotide reductase inhibitors (gemcitabine, hydroxyurea), and nucleotide analogues (5-fluorouracil) (12,14,15). Other agents elevate RS by targeting DNA polymerases or DNA damage response proteins. Some of these have progressed to advanced stages of preclinical and clinical development, including inhibitors of the DNA polymerase (e.g., CD437), the baseexcision repair factors PARP1 and PARP2 (e.g, olaparib), ATR-CHK1 signaling (e.g, ceralasertib), or the WEE1 checkpoint kinase (e.g, MK-1775) (13,16). Treatment outcomes vary widely due to significant differences in drug sensitivity and resistance across tumors, motivating efforts to better understand response mechanisms. These efforts have led to an expanding catalog of genetic alterations associated with sensitivity or resistance to RSi drugs, including mutations in BRCA1, BRCA2, ATM, or ATR (Fig. 1, Panel A) (17-26). This list is almost certainly not exhaustive, since the vast majority of genetic alterations identified are rare rather than common in tumors. Moreover, it is unclear how these multiple single-gene effects integrate to constitute an overall drug response.

[0011] Currently, two machine learning (ML) methodologies are emerging as particularly useful in understanding the effects of genetic alterations in precision medicine applications (90). The first methodology, known as “interpretable” ML, includes a wide array of approaches that attempt to gain insight into the mechanisms or rationale underlying a model’s predictions (27). In this regard, we and others have developed a series of “visible” neural network (VNN) models which guide the internal architecture of the model using knowledge maps of biological components and functions (28-34, 91). In contrast to conventional “black box” neural networks (35), a prime advantage of VNNs is that they not only learn to predict biomedical outcomes from sparse genetic feature sets but their predictions can be mapped to internal state changes in molecular mechanisms and pathways, allowing for model interpretation. To initiate a proteomics-driven resource of molecular mechanisms in cancer, we recently used affinity purification mass spectrometry to delineate the physical interactions of a broad set of cancer proteins. We integrated these interactions with other ‘omics datasets to create a large cancer protein-protein association network, which was iteratively clustered to generate a map of frequently altered protein assemblies (NeST, Nested Systems in Tumors, Fig. 9 Panel A and Panel B) (36). This map exhibits hierarchical structure, whereby individual cancer proteins converge to form protein complexes and larger assemblies which, in turn, incorporate into broad processes and organelles (Fig. 1, Panel A and Fig. 9, Panel B). The second ML methodology is multi-task learning, whereby several learning tasks are solved at the same time while exploiting commonalities and differences across these tasks (37). Multi-task learning has been previously used to boost the efficacy of drug response and mechanism of action prediction (38,39), although it has not yet been combined with knowledge bases of cancer pathways like NeST or (to our knowledge) applied to understand the response to RSi drugs.

[0012] SUMMARY OF THE INVENTION

[0013] At least one embodiment of the present invention relates to a biologically interpretable artificial intelligence (Al) model of drug response. The model takes an input of genetic alteration information of clinically accessible genes and predicts drug response based on a comprehensive map of cancer protein assemblies. After training the model with Al optimization procedures, the model highlights importance protein assemblies and / or pathways that govern drug response.

[0014] Motivated by the above challenges and advancements, here we describe and evaluate interpretable deep learning models of RSi drug response (Fig. 1 , Panel B). Starting from the genetic alterations detected in a tumor sample, the models predict the sensitivity or resistance to specific RSi drugs. The architecture of these models is based on knowledge of cancer protein complexes in NeST, meaning their predictions can be interpreted mechanistically, and both single-drug and multi-task (multi-drug) models are evaluated. These models identify a constellation of protein assemblies that integrate genetic alterations to predict drug sensitivity or resistance, which we show can be validated by multiple types of genome-wide functional assays and by their utility in predicting responses to previously unseen RSi agents as well as in chemotherapy-treated patients (Fig. 1, Panel B).

[0015] BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1. Overview. Panel A, Genetic alterations that modulate replication stress-inducing (RSi) drug responses are typically rare but converge on multiscale protein assemblies in cancer. Top: Waterfall plot of gene alteration frequencies in The Cancer Genome Atlas (TCGA) cervical squamous carcinoma cohort, showing that most genetic alterations are rare. Alterations include somatic mutations and copy number aberrations of genes covered by clinical cancer gene panels. Middle: Genetic alterations converge on specific multi-protein assemblies in the NeST (Nested Systems in Tumors) hierarchy. Bottom: Waterfall plot showing genetic alteration frequencies of NeST protein assemblies which, by nature, tend to be higher than for individual genes. Panel B, Pipeline for drug response prediction, (i) Acquisition of datasets used in this study. RSi drug response data from Genomics of Drug Sensitivity in Cancer (GDSC) and Cancer Therapeutics Response Portal (CTRP) databases. Tumor cell-line genome sequencing data from Cancer Cell Line Encyclopedia (CCLE). (ii) Construction of Visible Neural Network (VNN) models with architecture guided by NeST. (iii) Interpretation of VNN models to recognize assemblies for which genetic alterations are most important for RSi drug responses. Validation by genome-wide loss- of-function screens, (iv) Translation of VNNs to predict chemotherapy responses in cancer patients.

[0017] Figure 2. Performance of RSi-drug response models. Panel A, Schematic representation of the VNN models implemented in this study for single-drug and multi-drug scenarios. Panel B, Accuracy (odds ratio) of drug response prediction for each of the initial six RSi agents. Multi-drug and single-drug VNNs (greyscale-red, greyscale-orange) are compared against black-box artificial neural networks with a matched number of parameters and layers (ANN, greyscale-blue), DeepCDR (greyscale-purple), and DrugCell (greyscale-cyan). Error bars show the confidence interval of the odds ratio for each model. Performance was assessed on drug responses of held-out cell lines not seen during training. Panel C, Accuracy (odds ratio) of drug response prediction by the multi-drug model, evaluated for additional RSi drugs (n = 6, greyscale-red) and non-RSi drugs (n = 33, greyscale-blue), none of which were seen during training. Error bars show standard deviation of the odds ratio for each drug. Performance was assessed using held-out cell lines. Statistical comparisons conducted with the Mann-Whitney U test.

[0018] Figure 3. Identification of molecular mechanisms important for RSi drug responses. Panel A, System Importance (SI) profiles for prediction of initial six RSi drug responses. Nodes indicate protein assemblies; node sizes denote assembly sizes in numbers of proteins; greyscale-colors mark assemblies of above-threshold importance (Methods) for each drug response. The number of important assemblies is indicated in parentheses for each single-drug model, with a total number of 37 assemblies important for at least one model. Panel B, Venn diagram showing numbers of important assemblies identified by single-drug (n = 37, left) versus multi-drug VNNs (n = 36, right). The union set is defined as RSi assemblies (n = 41). Panel C, UMAP projection of SI profiles of each drug model constructed in this study (points). The HDBSCAN algorithm (hierarchical density-based spatial clustering of applications with noise) is used to group drug models into clusters (greyscale-colors) by similarity of important assemblies (SI profiles); gray models fall outside of stable clusters. Bold font face indicates RSi drug models, light font face indicates non- RSi drug models. Panel D, RSi assemblies (defined in Panel B) are plotted by their average importance scores for models trained on the initial RSi discovery agents (left), the later RSi test agents (middle), or non-RSi drug models (right). Horizontal line in each swarm represents the mean. ***P < 0.01; P-values by one-sided Mann-Whitney U test.

[0019] Figure 4. Evaluation of important assemblies by systematic siRNA screening. Panel A, Assemblies are analyzed for their enrichment for gene knockdowns that affect RS response via the absolute value of the replication restart score |RRS|, defined as the ratio of yH2AX to EdU intensity readouts. RRS screening data by Kavanaugh et al. (49). Panel B, Percent of assemblies validated by enrichment in RRS assay, shown separately forthose important to RSi drug models (left) versus all others (right). Error bars display the standard error of the proportion. ***P < 0.01; P-value by two sample z-test. Panel C, NeST assemblies ranked by their degree of enrichment, from most to least significant. Assemblies with FDR < 0.1 are highlighted in greyscale-red (n = 46). Selected RSi assemblies were labeled. Panel D, Mean importance scores for the 46 RRS-enriched assemblies (highlighted in Panel C), shown separately for RSi (left) and non-RSi drug models (right). ***P < 0.01; P-value by Mann-Whitney U test. Panel E, Venn diagram showing numbers of RSi assemblies validated by enrichment in the RRS assay or a separate chemogenetic drug sensitivity analysis (Fig. 12). Panel F, Validated assemblies in the two functional assays are indicated on the NeST assemblies map.

[0020] Figure 5. Model translation to predict clinical outcomes in cervical and lung cancers. Panel A, Top five cancer types in The Cancer Genome Atlas (TCGA) by largest number of patients treated by cisplatin: CESC, LUSC, HNSC, LUAD and OV. Right barplot shows the fraction of patients of each type that were treated with cisplatin. Panel B, Kaplan-Meier survival curves of progression-free survival (PFS) for TCGA patients with various cancer types that are indicated by different greyscale-colors. Patients with CESC or LUSC were chosen for the cisplatin VNN model validation because they showed the best outcome following cisplatin treatment. Panel C, Kaplan- Meier survival curves of progression-free survival (PFS) for cisplatin-treated and non-ci splatin- treated patients with CESC. Panel D, Kaplan-Meier survival curves of progression-free survival (PFS) for cisplatin-treated and non-cisplatin-treated patients with LUSC. Panel E, Application of the cisplatin VNN to predict patient responses to cisplatin treatment. Panel F, Kaplan-Meier survival curves of progression-free survival (PFS) for patients with CESC stratified by the cisplatin VNN prediction status (sensitive versus resistant, shown by greyscale-orange versus greyscale- blue). **P < 0.05; P-value by log-rank test. Panel G, Kaplan-Meier survival curves of progression- free survival (PFS) for patients with LUSC stratified by the cisplatin VNN prediction status (sensitive versus resistant, shown by greyscale-orange versus greyscale-blue). **P < 0.05; P-value by log-rank test. HR, Hazard Ratio. Panel H, Predictive performance in response to cisplatin of RandomForest, ElasticNet, and the cisplatin VNN models (boxplots of different greyscale-colors) in patients with CESC (left) or LUSC (right). Performance measured using the Concordance index (C-index). Boxplot midline is median, box boundaries show upper and lower quartile, and whiskers show C-indices measured by five respective cross-validation modes of RandomForest, ElasticNet, and the cisplatin VNN models.

[0021] Figure 6. Protein assemblies as biomarkers of cisplatin outcome. Panel A, Evaluation of individual assemblies for prediction of responses to cisplatin in patients with CESC. Panel B, Swarmplot showing the predictive performance (C-index) of genes and assemblies in response to cisplatin in patients with CESC. Assemblies are grouped into the 29 important for cisplatin response in cell lines (right) versus all other assemblies (middle). The top 10 predictive assemblies in patients with CESC are listed in greyscale-orange. ***P < 0.01; P-value by Mann-Whitney U test. Panel C, Interaction network of proteins in NeST: 126. Edges denote Integrated Association Stringency (IAS) scores used to assemble the NeST map (36). Panel D, Kaplan-Meier survival curves of progression-free survival (PFS) for cisplatin-treated patients with CESC stratified by in silico activity of NeST: 126. ***p < 0.01; P-value by log-rank test. HR, Hazard Ratio. Panel E, LEFT: Predictive performance (C-index) of NeST: 126 assembly (top row) and its genes (all other rows except BARD1 and TOP2A) for prediction of CESC cisplatin response. RIGHT: Corresponding frequency at which the assembly or its genes is genetically altered in patients with CESC. Large circles indicate the mean C-index over five models from cross-validation. Error bars denote 95% confidence intervals (CI). C-index and alteration frequency of the top two predictive genes BARDl and TOP2A that are not members of NeST : 126 are shown at bottom, separately.

[0022] Figure 7. RTK-JAK-STAT assembly as an example predicting cisplatin resistance. Panel A, Protein network defining NeST: 89 (RTK-IAK-STAT signaling). Edges denote Integrated Association Stringency (IAS) scores used to assemble the NeST map. Panel B, Oncogenic signaling pathways captured by NeST:89. Panel C, Kaplan-Meier survival curves of progression- free survival (PFS) for cisplatin-treated patients with CESC stratified by in silico activity of NeST:89. HR, Hazard Ratio. ***P < 0.01; P-value by log-rank test. Panel D, LEFT: Predictive performance (C-index) of NeST:89 assembly (top row) and its genes (all other rows except BARD1 and TOP2A) for prediction of CESC cisplatin response. RIGHT: Corresponding frequency at which the assembly or its genes is genetically altered in patients with CESC. Large circles represent the mean C-index over five models from cross-validation. Error bars denote 95% confidence intervals. C-index and alteration frequency of the top two predictive genes BARD1 and T0P2A that are not members of NeST:89 are shown at bottom, separately. Panel E, OncoPrint showing NeST:89 genes (rows) by cisplatin-treated patients with CESC (columns). Patients stratified into binary response subtypes based on NeST:89 status (low activity at left, enriched for cisplatin-sensitive patients; high activity at right, enriched for cisplatin-resistant patients). Genes sorted from top-to-bottom based on alteration frequency in cisplatin-resistant subtype. Panel F, Boolean circuit diagram representing the approximate logic function by which the genetic alteration status of eight NeST:89 genes predicts resistance to cisplatin in the CESC cohort. Inputs from activating or upregulating alterations are directly connected to the OR gates whereas inputs from inactivating or downregulating alterations are first negated to represent an opposite signal.

[0023] Figure 8. Aspects of replication stress addressed by cancer therapeutics. Cancer therapeutics can invoke diverse mechanisms to induce RS. Shown are nine conventional chemotherapeutics disrupting DNA integrity or replication (left), along with three inhibitors targeting DNA damage repair signaling cascades (middle and right).

[0024] Figure 9. Nested systems in tumors (NeST). Panel A, Workflow depicting construction of the NST hierarchy of protein assemblies in tumor cells. APMS data from 61 cancer protein baits were integrated with a compendium of published protein interaction data to produce an integrated protein network. Panel B, Community detection identified nested protein assemblies inside the network. Protein assemblies under mutational selection pressure were identified, producing the NeST hierarchical map.

[0025] Figure 10. VNN schematic. Panel A, The first layer of a VNN incorporates gene-level features, including gene mutations, copy number amplifications (CNA), and copy number deletions (CND). Subsequent layers aggregate gene-level features into assembly-level information, guided by the hierarchical relationships defined by the NeST map. The output state of each gene (g) and assembly (A) is represented by artificial neurons (one neuron per gene, multiple neurons per assembly). Hierarchical gene-to assembly and assembly-to-assembly connections are assigned weights that are optimized during neural network training. Panel B, Position of the assemblies detailed in panel (A) within the greater NeST map. Each node indicates a protein assembly. An example path of information flow, from the neurons of CDK holoenzyme complex to Cell cycle through to model Root, is shown in greyscale-red. Panel C, The additional layer in the multi-task (multi-drug) model enables a single model to be trained for multiple related agents by transforming the state of the Root into multiple outputs, one for each drug of interest. Apart from this final layer, all neural network weights are shared across drugs, yielding potential boosts in predictive power and model interpretation.

[0026] Figure 11. Performance and interpretation of the multi-drug VNN. Panel A, Accuracy of the multi-drug VNN (odds ratio, y-axis) in predicting the cellular response to each of six RSi drugs (x-axis). The multi-drug VNN trained on six RSi drugs (greyscale-red points) is compared to alternate multi-drug VNNs trained on combinations of RSi and non-RSi drugs (gray points). Panel B, UMAP projection of drugs based on the response profiles across cell lines profiled in GDSC and CTRP. Note that all RSi drugs, save for CD437, cluster in the top left quadrant, indicating similarity in their response profiles across cell lines, with gemcitabine and etoposide showing particularly high similarity. Panel C, System importance (SI) profile for the unified multi-drug VNN. The number of important assemblies is indicated in parentheses for multi-drug model.

[0027] Figure 12. Evaluation of assemblies by systematic drug sensitivity screens. Panel A, Assemblies are analyzed for their enrichment for CRISPR / Cas9 gene knockouts that affect sensitivity to an RSi drug in either direction (increased or decreased sensitivity). Genetic screen and scoring by Olivieri et al. (47). Panel B, Percent of assemblies enriched in drug sensitivity screens, shown separately for assemblies important to RSi drug models (left) versus all others (right). Error bars display the standard error of the proportion. ***P < 0.01; P-value by two sample z-test. Panel C, Percent of important assemblies in single-drug VNN models enriched in each individual drug sensitivity screen. Panel D, NeST assemblies ranked by their degree of enrichment in cisplatin screen, from most to least significant. Assemblies with FDR < 0.1 are highlighted in greyscale-red. Panel E, NeST assemblies ranked by their degree of enrichment in etoposide screen, from most to least significant. Assemblies with FDR < 0.1 are highlighted in greyscale-red. Panel F, NeST assemblies ranked by their degree of enrichment in camptothecin screen, from most to least significant. Assemblies with FDR < 0.1 are highlighted in greyscale-red. Panel G, NeST assemblies ranked by their degree of enrichment in CD437 screen, from most to least significant. Assemblies with FDR < 0.1 are highlighted in greyscale-red. Panel H, NeST assemblies ranked by their degree of enrichment in olaparib screen, from most to least significant. Assemblies with FDR < 0.1 are highlighted in greyscale-red. Panel I, NeST assemblies ranked by their degree of enrichment in gemcitabine screen, from most to least significant. Assemblies with FDR < 0.1 are highlighted in greyscale-red. In panels D-I, selected RSi assemblies were labeled. Panel J, Mean importance scores for assemblies enriched in one or more drug sensitivity screens (union from panels D-I, N = 61), shown separately for RSi (left) and non-RSi drug models (right). ***P < 0.01; P -value by Mann-Whitney U test. Panel K, Mean importance scores for assemblies validated in functional screens, shown separately for RSi (left) and non-RSi drug models (right). ***P < 0.01; P -value by Mann-Whitney U test.

[0028] Figure 13. Survival analysis of cisplatin-treated TCGA cohorts. Panels A C, Kaplan-Meier progression-free survival (PFS) plots are shown for cisplatin-treated vs. non-cisplatin-treated patients from the TCGA head-and-neck squamous cell carcinoma (HNSC, Panel A), lung adenocarcinoma (LU AD, Panel B) and ovarian carcinoma (OV, Panel C) cohorts. None of these panels show significant effects of cisplatin on PFS (all P > 0.1). Panel D, Kaplan-Meier PFS plots for HNSC, OV, and LUAD patients stratified by the cisplatin VNN prediction status. Panel E, Kaplan-Meier PFS plots for CESC and LUSC patients stratified by the cisplatin VNN prediction status. Panel F, Swarmplot showing the predictive performance (C -index) of assemblies in response to cisplatin in patients with CESC. Assemblies are grouped into the 83 non-RSi assemblies, 41 RSi assemblies, 30 assemblies validated in at least one screening mode and 15 assemblies validated n both screening modes (from left to right). ***P < 0.01; *P < 0.1; P-value by Mann-Whitney U test. Panel G, Performance comparison of 12 RSi drug models and 33 non- RSi drug models predicting response to cisplatin-treated patients with CESC. Performance measured by concordance index (C-index); **P < 0.05; P-value by Mann-Whitney U test. For all Kaplan-Meier plots, sensitive versus resistant shown respectively by greyscale-blue versus greyscale-orange; **P < 0.05; P-value by log-rank test.

[0029] DETAILED DESCRIPTION

[0030] All publications, patents, and patent applications mentioned in this specification are herein incorporated by reference to the same extent as if each individual publication, patent, or patent application was specifically and individually indicated to be incorporated by reference. In embodiments, the invention provides a method for modeling a replication stressinducing (RSi) drug response in a subject, comprising detecting genetic alterations in a tumor sample of the subject, and predicting the sensitivity or resistance to specific RSi drugs by incorporating information of cancer protein complexes.

[0031] In embodiments, the method implements a visible neural network (VNN) model to use a knowledge map of biological components and functions combined with genetic alterations associated with sensitivity or resistance to RSi drugs. In embodiments, the VNN predicts biomedical outcomes from genetic feature sets, and maps internal state changes in molecular mechanisms and pathways.

[0032] In embodiments, multi-task learning is used to boost the efficacy of drug response and mechanism of action predictions, combined with knowledge bases of cancer protein complexes and pathways, such as NeST, applied to determine a response to RSi drugs. In embodiments, both single-drug and multi-task (multi-drug) models are evaluated. In embodiments, the invention comprises training the model with artificial intelligence optimization procedures. In embodiments, the model is effective to highlight important protein assemblies or pathways that govern drug response.

[0033] In embodiments, inputting genetic alterations of clinically acceptable genes comprises: taking in three binary input features per gene; describing the presence or absence of point mutation, insertion, deletion, copy number amplification, or copy number deletion; incorporating the three binary input features into a first layer of NeST-VNN to represent gene-level alteration features; and integrating the gene-level alteration features with respective protein assemblies into a second layer of the NeST-VNN, wherein the respective protein assemblies are represented by a bank of artificial neurons. In embodiments, the bank of artificial neurons are connected with larger assemblies of neurons, allowing for the flow of genetic information from small focal assemblies to effects on larger-scale assemblies and super assemblies.

[0034] In embodiments, the invention provides a biologically interpretable deep learning model that combines a plurality of protein assemblies and genetic alterations to predict drug sensitivity or resistance responses to replication stress-inducing (RSi) agents in chemotherapy-treated patients. In embodiments, the model associates single gene alterations with drug responses with a map of protein complexes and larger molecular assemblies associated with cancer. In embodiments, the model comprises 41 protein assemblies in which genetic alterations modulate RSi drug responses as described herein.

[0035] In embodiments, the present invention provides a computer system, comprising one or more processors; and memory storing executable instructions that, as a result of execution, cause the one or more processors of the computer system to conduct the methods and models described herein.

[0036] In embodiments, the invention provides from methods of treatment of a subject in need comprising administering an effective amount of a drug or therapy to the subject, wherein the drug or therapy is identified by the systems and methods described herein.

[0037] These and other embodiments and combinations of the embodiments will be apparent to one of ordinary skill in the art upon a review of the detailed description herein.

[0038] Unless defined otherwise, all terms of art, notations and other technical and scientific terms or terminology used herein are intended to have the same meaning as is commonly understood by one of ordinary skill in the art to which the claimed subject matter pertains. In some embodiments, terms with commonly understood meanings are defined herein for clarity and / or for ready reference, and the inclusion of such definitions herein should not necessarily be construed to represent a substantial difference over what is generally understood in the art.

[0039] As used herein, the terms “comprises,” “comprising,” “includes,” “including,” “has,” “having,” “contains”, “containing,” “characterized by,” or any other variation thereof, are intended to encompass a non-exclusive inclusion, subject to any limitation explicitly indicated otherwise, of the recited components. For example, a composition, and / or a method that “comprises” a list of elements (e.g., components, features, or steps) is not necessarily limited to only those elements (or components or steps), but may include other elements (or components or steps) not expressly listed or inherent to the composition and / or method. Reference throughout this specification to “one embodiment,” “an embodiment,” “a particular embodiment,” “a related embodiment,” “a certain embodiment,” “an additional embodiment,” or “a further embodiment” or combinations thereof means that a particular feature, structure or characteristic described in connection with the embodiment is included in at least one embodiment of the present invention. Thus, the appearances of the foregoing phrases in various places throughout this specification are not necessarily all referring to the same embodiment. Furthermore, the particular features, structures, or characteristics may be combined in any suitable manner in one or more embodiments. As used herein, the transitional phrases “consists of’ and “consisting of’ exclude any element, step, or component not specified. For example, “consists of’ or “consisting of’ used in a claim would limit the claim to the components, materials or steps specifically recited in the claim. When the phrase “consists of’ or “consisting of’ appears in a clause of the body of a claim, rather than immediately following the preamble, the phrase “consists of’ or “consisting of’ limits only the elements (or components or steps) set forth in that clause; other elements (or components) are not excluded from the claim as a whole.

[0040] As used herein, the transitional phrases “consists essentially of’ and “consisting essentially of’ are used to define a composition and / or method that includes materials, steps, features, components, or elements, in addition to those literally disclosed, provided that these additional materials, steps, features, components, or elements do not materially affect the basic and novel characterise c(s) of the claimed invention. The term “consisting essentially of’ occupies a middle ground between “comprising” and “consisting of’. It is understood that aspects and embodiments of the invention described herein include “consisting” and / or “consisting essentially of’ aspects and embodiments.

[0041] When introducing elements of the present invention or the preferred embodiment(s) thereof, the articles “a”, “an”, “the” and “said” are intended to mean that there are one or more of the elements. The terms “comprising”, “including” and “having” are intended to be inclusive and mean that there may be additional elements other than the listed elements.

[0042] The term “and / or” when used in a list of two or more items, means that any one of the listed items can be employed by itself or in combination with any one or more of the listed items. For example, the expression “A and / or B” is intended to mean either or both of A and B, i.e. A alone, B alone or A and B in combination. The expression “A, B and / or C” is intended to mean A alone, B alone, C alone, A and B in combination, A and C in combination, B and C in combination or A, B, and C in combination.

[0043] Throughout this application, various embodiments may be presented in a range format. It should be understood that the description in range format is merely for convenience and brevity and should not be construed as an inflexible limitation on the scope of the disclosure. Accordingly, the description of a range should be considered to have specifically disclosed all the possible subranges as well as individual numerical values within that range. For example, description of a range such as from 1 to 6 should be considered to have specifically disclosed subranges such as from 1 to 3, from 1 to 4, from 1 to 5, from 2 to 4, from 2 to 6, from 3 to 6 etc., as well as individual numbers within that range, for example, 1, 2, 3, 4, 5, and 6. This applies regardless of the breadth of the range.

[0044] The terms “quantifying” “determining,” “measuring,” “evaluating,” “assessing,” “assaying,” and “analyzing” are often used interchangeably herein to refer to forms of measurement. The terms include determining if an element is present or not (for example, detection). These terms can include quantitative, qualitative or quantitative and qualitative determinations. Assessing can be relative or absolute. “Detecting the presence of’ can include determining the amount of something present in addition to determining whether it is present or absent depending on the context.

[0045] As used herein, the term “about” a number refers to that number plus or minus 10% of that number. The term “about” a range refers to that range minus 10% of its lowest value and plus 10% of its greatest value.

[0046] In an aspect, the disclosure provides in addition to a method of prognosing or diagnosing a disease or condition, a method of treating or preventing a disease or disorder in a subject in need thereof, further comprising administering an effective amount of a pharmaceutical composition for treatment of the identified disease or conditions. In some embodiments, the disease or condition comprises a cancer, metabolic disease, immune disease (including autoimmune disease), inflammatory condition, congenital disorders of glycosylation, or reproductive health condition. In some embodiments, the disease is cancer, and in embodiments is cervical cancer.

[0047] A non-exhaustive list of cancer types and / or stages that may be identified using machinelearning models described herein include the following: Adrenocortical Carcinoma (TCGA-ACC); Bladder Urothelial Carcinoma (TCGA-BLCA); Brain Lower Grade Glioma (TCGA-LGG); Breast Invasive Carcinoma (TCGA-BRCA); Cervical Squamous Cell Carcinoma and Endocervical Adenocarcinoma (TCGA-CESC); Cholangiocarcinoma (TCGA-CHOL); Colon Adenocarcinoma (TCGA-COAD); Lymphoid Neoplasm Diffuse Large B-cell Lymphoma (TCGA-DLBC); Esophageal Carcinoma (TCGA-ESCA); Gastric Adenocarcinoma (TCGA-GA); Glioblastoma Multiforme (TCGA-GBM); Head and Neck Squamous Cell Carcinoma (TCGA-HNSC); Kidney Chromophobe (TCGA-KICH); Kidney Renal Clear Cell Carcinoma (TCGA-KIRC); Kidney Renal Papillary Cell Carcinoma (TCGA-KIRP); Liver Hepatocellular Carcinoma (TCGA-LIHC); Lung Adenocarcinoma (TCGA-LUAD); Lung Squamous Cell Carcinoma (TCGA-LUSC); Mesothelioma (TCGA-MESO); Ovarian Serous Cystadenocarcinoma (TCGA-OV); Pancreatic Adenocarcinoma (TCGA-PAAD); Pheochromocytoma and Paraganglioma (TCGA-PCPG); Prostate Adenocarcinoma (TCGA-PRAD); Rectum Adenocarcinoma (TCGA-READ); Sarcoma (TCGA-SARC); Skin Cutaneous Melanoma (TCGA-SKCM); Stomach Adenocarcinoma (TCGA- STAD); Testicular Germ Cell Tumors (TCGA-TGCT); Thyroid Carcinoma (TCGA-THCA); Thymoma (TCGA-THYM); Uterine Carcinosarcoma (TCGA-UCEC); Uterine Corpus Endometrial Carcinoma (TCGA-UCS); Uveal Melanoma (TCGA-UVM).

[0048] The terms “subject,” “patient” and “individual” are used interchangeably herein to refer to a vertebrate, preferably a mammal, more preferably a human. Tissues, cells, and their progeny of a biological entity obtained in vivo or cultured in vitro are also encompassed. A “subject,” “patient” or “individual” as used herein, includes any animal that exhibits pain that can be treated with the vectors, compositions, and methods contemplated herein. Suitable subjects (e.g., patients) include laboratory animals (such as mouse, rat, rabbit, or guinea pig), farm animals, and domestic animals or pets (such as a cat or dog). Non-human primates and, preferably, human patients, are included.

[0049] In some embodiments, “administering” comprises administering a therapeutically effective amount to a subject.

[0050] As used herein, the term “amount” refers to “an amount effective” or “an effective amount” of a cell to achieve a beneficial or desired prophylactic or therapeutic result, including clinical results. As used herein, “therapeutically effective amount” refers to an amount of a pharmaceutically active compound(s) that is sufficient to treat or ameliorate, or in some manner reduce the symptoms associated with diseases and medical conditions. When used with reference to a method, the method is sufficiently effective to treat or ameliorate, or in some manner reduce the symptoms associated with diseases or conditions. For example, an effective amount in reference to diseases is that amount which is sufficient to block or prevent onset; or if disease pathology has begun, to palliate, ameliorate, stabilize, reverse or slow progression of the disease, or otherwise reduce pathological consequences of the disease. In any case, an effective amount may be given in single or divided doses.

[0051] An “effective amount” of a therapeutic administration will depend upon the degree of correlation between the glyco-motif profiles and the subject’s glycosylation-related disease or condition, among other variables such as the age, health, and weight, which is within the skill or one of ordinary skill to determine.

[0052] As used herein, the terms “treat,” “treatment,” or “treating” embraces at least an amelioration of the symptoms associated with diseases in the patient, where amelioration is used in a broad sense to refer to at least a reduction in the magnitude of a parameter, e.g. a symptom associated with the disease or condition being treated. As such, “treatment” also includes situations where the disease, disorder, or pathological condition, or at least symptoms associated therewith, are completely inhibited (e.g. prevented from happening) or stopped (e.g. terminated) such that the patient no longer suffers from the condition, or at least the symptoms that characterize the condition.

[0053] As used herein, and unless otherwise specified, the terms "prevent," "preventing" and "prevention" refer to the prevention of the onset, recurrence or spread of a disease or disorder, or of one or more symptoms thereof. In certain embodiments, the terms refer to the treatment with or administration of a compound or dosage form provided herein, with or without one or more other additional active agent(s), prior to the onset of symptoms, particularly to subjects at risk of disease or disorders provided herein. The terms encompass the inhibition or reduction of a symptom of the particular disease. In certain embodiments, subjects with familial history of a disease are potential candidates for preventive regimens. In certain embodiments, subjects who have a history of recurring symptoms are also potential candidates for prevention. In this regard, the term "prevention" may be interchangeably used with the term "prophylactic treatment."

[0054] As used herein, and unless otherwise specified, a "prophylactically effective amount" of a compound is an amount sufficient to prevent a disease or disorder, or prevent its recurrence. A prophylactically effective amount of a compound means an amount of therapeutic agent, alone or in combination with one or more other agent(s), which provides a prophylactic benefit in the prevention of the disease. The term "prophylactically effective amount" can encompass an amount that improves overall prophylaxis or enhances the prophylactic efficacy of another prophylactic agent.

[0055] Trained Algorithms

[0056] Methods and systems as described herein may employ one or more trained algorithms. The trained algorithm(s) may process or operate on one or more datasets comprising information about biomolecules (e.g., biomolecular features), biochemical features (e.g., lectin binding), or any combination thereof. In some embodiments, the datasets comprise structural or sequence information about biomolecules. The one or more datasets may be observed empirically, derived from computational studies, be derived from or contained in one or more databases, or any combination thereof.

[0057] The trained algorithm may comprise an unsupervised machine learning algorithm. The trained algorithm may comprise a supervised machine learning algorithm. The trained algorithm may comprise a semi-supervised machine learning algorithm. The trained algorithm may comprise a classification and regression tree (CART) algorithm. The supervised machine learning algorithm may comprise, for example, a Random Forest, a support vector machine (SVM), a neural network, or a deep learning algorithm. The trained algorithm may comprise a self-supervised machine learning algorithm.

[0058] In some embodiments, a machine learning algorithm (or software module) of a platform as described herein utilizes one or more neural networks. In some embodiments, a neural network is a type of computational system that can learn the relationships between an input dataset and a target dataset. A neural network may be a software representation of a human neural system (e.g. cognitive system), intended to capture “learning” and “generalization” abilities as used by a human. In some embodiments, the machine learning algorithm (or software module) comprises a neural network comprising a CNN. Non-limiting examples of structural components of embodiments of the machine learning software described herein include: CNNs, recurrent neural networks, dilated CNNs, fully-connected neural networks, deep generative models, recurrent neural networks (RNNs), RNNs using long short-term memory (LSTM) units, and Boltzmann machines.

[0059] In some embodiments, a neural network comprises a series of layers termed “neurons.” In some embodiments, a neural network comprises an input layer, to which data is presented; one or more internal, and / or “hidden”, layers; and an output layer. A neuron may be connected to neurons in other layers via connections that have weights, which are parameters that control the strength of the connection. The number of neurons in each layer may be related to the complexity of the problem to be solved. The minimum number of neurons required in a layer may be determined by the problem complexity, and the maximum number may be limited by the ability of the neural network to generalize. The input neurons may receive data being presented and then transmit that data to the first hidden layer through connections’ weights, which are modified during training. The first hidden layer may process the data and transmit its result to the next layer through a second set of weighted connections. Each subsequent layer may “pool” the results from the previous layers into more complex relationships. In addition, whereas conventional software programs require writing specific instructions to perform a function, neural networks are programmed by training them with a known sample set and allowing them to modify themselves during (and after) training so as to provide a desired output such as an output value. After training, when a neural network is presented with new input data, it is configured to generalize what was “learned” during training and apply what was learned from training to the new previously unseen input data to generate an output associated with that input.

[0060] In some embodiments, the neural network comprises ANNs. ANN may be machine learning algorithms that may be trained to map an input dataset to an output dataset, where the ANN comprises an interconnected group of nodes organized into multiple layers of nodes. For example, the ANN architecture may comprise at least an input layer, one or more hidden layers, and an output layer. The ANN may comprise any total number of layers, and any number of hidden layers, where the hidden layers function as trainable feature extractors that allow mapping of a set of input data to an output value or set of output values. As used herein, a deep learning algorithm (such as a DNN) is an ANN comprising a plurality of hidden layers, e.g., two or more hidden layers. Each layer of the neural network may comprise a number of nodes (or “neurons”). A node receives input that comes either directly from the input data or the output of nodes in previous layers, and performs a specific operation, e.g., a summation operation. A connection from an input to a node is associated with a weight (or weighting factor). The node may sum up the products of all pairs of inputs and their associated weights. The weighted sum may be offset with a bias. The output of a node or neuron may be gated using a threshold or activation function. The activation function may be a linear or non-linear function. The activation function may be, for example, a rectified linear unit (ReLU) activation function, a Leaky ReLU activation function, or other function such as a saturating hyperbolic tangent, identity, binary step, logistic, arctan, softsign, parametric rectified linear unit, exponential linear unit, softplus, bent identity, softexponential, sinusoid, sine, Gaussian, or sigmoid function, or any combination thereof.

[0061] The weighting factors, bias values, and threshold values, or other computational parameters of the neural network, may be “taught” or “learned” in a training phase using one or more sets of training data. For example, the parameters may be trained using the input data from a training dataset and a gradient descent or backward propagation method so that the output value(s) that the ANN computes are consistent with the examples included in the training dataset.

[0062] The number of nodes used in the input layer of the ANN or DNN may be at least about 10, 50, 100, 200, 300, 400, 500, 600, 700, 800, 900, 1,000, 2,000, 3,000, 4,000, 5,000, 6,000, 7,000, 8,000, 9,000, 10,000, 20,000, 30,000, 40,000, 50,000, 60,000, 70,000, 80,000, 90,000, 100,000, or greater. In some instances, the number of node used in the input layer may be at most about 100,000, 90,000, 80,000, 70,000, 60,000, 50,000, 40,000, 30,000, 20,000, 10,000, 9,000, 8,000, 7,000, 6,000, 5,000, 4,000, 3,000, 2,000, 1,000, 900, 800, 700, 600, 500, 400, 300, 200, 100, 50, 10, or less. In some instances, the total number of layers used in the ANN or DNN (including input and output layers) may be at least about 3, 4, 5, 10, 15, 20, or greater. In some instances, the total number of layers may be at most about 20, 15, 10, 5, 4, 3, or less.

[0063] In some instances, the total number of learnable or trainable parameters, e.g., weighting factors, biases, or threshold values, used in the ANN or DNN may be at least about 10, 50, 100, 200, 300, 400, 500, 600, 700, 800, 900, 1,000, 2,000, 3,000, 4,000, 5,000, 6,000, 7,000, 8,000, 9,000, 10,000, 20,000, 30,000, 40,000, 50,000, 60,000, 70,000, 80,000, 90,000, 100,000, or greater. In some instances, the number of learnable parameters may be at most about 100,000, 90,000, 80,000, 70,000, 60,000, 50,000, 40,000, 30,000, 20,000, 10,000, 9,000, 8,000, 7,000, 6,000, 5,000, 4,000, 3,000, 2,000, 1,000, 900, 800, 700, 600, 500, 400, 300, 200, 100, 50, 10, or less.

[0064] In some embodiments of a machine learning software module as described herein, a machine learning software module comprises a neural network such as a deep CNN. In some embodiments in which a CNN is used, the network is constructed with any number of convolutional layers, dilated layers or fully-connected layers. In some embodiments, the number of convolutional layers is between 1-10 and the dilated layers between 0-10. The total number of convolutional layers (including input and output layers) may be at least about 1, 2, 3, 4, 5, 10, 15, 20, or greater, and the total number of dilated layers may be at least about 1, 2, 3, 4, 5, 10, 15, 20, or greater. The total number of convolutional layers may be at most about 20, 15, 10, 5, 4, 3, or less, and the total number of dilated layers may be at most about 20, 15, 10, 5, 4, 3, or less. In some embodiments, the number of convolutional layers is between 1-10 and the fully-connected layers between 0-10. The total number of convolutional layers (including input and output layers) may be at least about 1, 2, 3, 4, 5, 10, 15, 20, or greater, and the total number of fully-connected layers may be at least about 1, 2, 3, 4, 5, 10, 15, 20, or greater. The total number of convolutional layers may be at most about 20, 15, 10, 5, 4, 3, 2, 1, or less, and the total number of fully-connected layers may be at most about 20, 15, 10, 5, 4, 3, 2, 1, or less.

[0065] In some embodiments, the input data for training of the ANN may comprise a variety of input values depending whether the machine learning algorithm is used for processing sequence or structural data. In general, the ANN or deep learning algorithm may be trained using one or more training datasets comprising the same or different sets of input and paired output data.

[0066] In some embodiments, a machine learning software module comprises a neural network comprising a CNN, RNN, dilated CNN, fully-connected neural networks, deep generative models and deep restricted Boltzmann machines.

[0067] In some embodiments, a machine learning algorithm comprises CNNs. The CNN may be deep and feedforward ANNs. The CNN may be applicable to analyzing visual imagery. The CNN may comprise an input, an output layer, and multiple hidden layers. The hidden layers of a CNN may comprise convolutional layers, pooling layers, fully-connected layers and normalization layers. The layers may be organized in 3 dimensions: width, height and depth.

[0068] The convolutional layers may apply a convolution operation to the input and pass results of the convolution operation to the next layer. For processing images, the convolution operation may reduce the number of free parameters, allowing the network to be deeper with fewer parameters. In neural networks, each neuron may receive input from some number of locations in the previous layer. In a convolutional layer, neurons may receive input from only a restricted subarea of the previous layer. The convolutional layer's parameters may comprise a set of learnable filters (or kernels). The learnable filters may have a small receptive field and extend through the full depth of the input volume. During the forward pass, each filter may be convolved across the width and height of the input volume, compute the dot product between the entries of the filter and the input, and produce a two-dimensional activation map of that filter. As a result, the network may learn filters that activate when it detects some specific type of feature at some spatial position in the input.

[0069] In some embodiments, the pooling layers comprise global pooling layers. The global pooling layers may combine the outputs of neuron clusters at one layer into a single neuron in the next layer. For example, max pooling layers may use the maximum value from each of a cluster of neurons in the prior layer; and average pooling layers may use the average value from each of a cluster of neurons at the prior layer.

[0070] In some embodiments, the fully-connected layers connect every neuron in one layer to every neuron in another layer. In neural networks, each neuron may receive input from some number locations in the previous layer. In a fully-connected layer, each neuron may receive input from every element of the previous layer.

[0071] In some embodiments, the normalization layer is a batch normalization layer. The batch normalization layer may improve the performance and stability of neural networks. The batch normalization layer may provide any layer in a neural network with inputs that are zero mean / unit variance. The advantages of using batch normalization layer may include faster trained networks, higher learning rates, easier to initialize weights, more activation functions viable, and simpler process of creating deep networks.

[0072] In some embodiments, a machine learning software module comprises a recurrent neural network software module. A recurrent neural network software module may be configured to receive sequential data as an input, such as consecutive data inputs, and the recurrent neural network software module updates an internal state at every time step. A recurrent neural network can use internal state (memory) to process sequences of inputs. The recurrent neural network may be applicable to tasks such as handwriting recognition or speech recognition. The recurrent neural network may also be applicable to next word prediction, music composition, image captioning, time series anomaly detection, machine translation, scene labeling, and stock market prediction. A recurrent neural network may comprise fully recurrent neural network, independently recurrent neural network, Elman networks, Jordan networks, Echo state, neural history compressor, long short-term memory, gated recurrent unit, multiple timescales model, neural Turing machines, differentiable neural computer, and neural network pushdown automata.

[0073] In some embodiments, a machine learning software module comprises a supervised or unsupervised learning method such as, for example, support vector machines (“SVMs”), random forests, clustering algorithm (or software module), gradient boosting, logistic regression, and / or decision trees. The supervised learning algorithms may be algorithms that rely on the use of a set of labeled, paired training data examples to infer the relationship between an input data and output data. The unsupervised learning algorithms may be algorithms used to draw inferences from training datasets to the output data. The unsupervised learning algorithm may comprise cluster analysis, which may be used for exploratory data analysis to find hidden patterns or groupings in process data. One example of unsupervised learning method may comprise principal component analysis. The principal component analysis may comprise reducing the dimensionality of one or more variables. The dimensionality of a given variable may be at least 1, 5, 10, 50, 100, 200, 300, 400, 500, 600, 700, 800, 900, 1,000, 1,100, 1,200, 1,300, 1,400, 1,500, 1,600, 1,700, 1,800, or greater. The dimensionality of a given variables may be at most 1,800, 1,700, 1,600, 1,500, 1,400, 1,300, 1,200, 1,100, 1,000, 900, 800, 700, 600, 500, 400, 300, 200, 100, 50, 10, or less.

[0074] In some embodiments, the machine learning algorithm may comprise reinforcement learning algorithms. The reinforcement learning algorithm may be used for optimizing Markov decision processes (i.e., mathematical models used for studying a wide range of optimization problems where future behavior cannot be accurately predicted from past behavior alone, but rather also depends on random chance or probability). One example of reinforcement learning may be Q-learning. Reinforcement learning algorithms may differ from supervised learning algorithms in that correct training data input / output pairs are never presented, nor are sub-optimal actions explicitly corrected. The reinforcement learning algorithms may be implemented with a focus on real-time performance through finding a balance between exploration of possible outcomes (e.g., correct compound identification) based on updated input data and exploitation of past training.

[0075] In some embodiments, training data resides in a cloud-based database that is accessible from local and / or remote computer systems on which the machine learning-based sensor signal processing algorithms are running. The cloud-based database and associated software may be used for archiving electronic data, sharing electronic data, and analyzing electronic data. In some embodiments, training data generated locally may be uploaded to a cloud-based database, from which it may be accessed and used to train other machine learning-based detection systems at the same site or a different site.

[0076] The trained algorithm may accept a plurality of input variables and produce one or more output variables based on the plurality of input variables. The input variables may comprise one or more datasets indicative of a glycosylation feature. For example, the input variables may comprise a carbohydrate binding protein pattern, glycan structures, glycan features, clinical outcomes, diagnosis, or any combination thereof. The trained algorithm may be trained with a plurality of independent training samples. Each of the independent training samples may comprise a carbohydrate binding protein pattern and glycan structures or glycan features and diagnosis. The trained algorithm may be trained with at least about 5, at least about 10, at least about 15, at least about 20, at least about 25, at least about 30, at least about 35, at least about 40, at least about 45, at least about 50, at least about 100, at least about 150, at least about 200, at least about 250, at least about 300, at least about 350, at least about 400, at least about 450, at least about 500, at least about 1,000, at least about 1,500, at least about 2,000, at least about 2,500, at least about 3,000, at least about 3,500, at least about 4,000, at least about 4,500, at least about 5,000, at least about, 5,500, at least about 6,000, at least about 6,500, at least about 7,000, at least about 7,500, at least about 8,000, at least about 8,500, at least about 9,000, at least about 9,500, at least about 10,000, or more independent training samples.

[0077] The trained algorithm may be adjusted or tuned to improve one or more of the performance, accuracy, PPV, NPV, sensitivity, specificity, or AUC of associating the glycosylation feature. The trained algorithm may be adjusted or tuned by adjusting parameters of the trained algorithm (e.g., a set of cutoff values used to associate a glycosylation feature as described elsewhere herein, or weights of a neural network). The trained algorithm may be adjusted or tuned continuously during the training process or after the training process has completed.

[0078] After the trained algorithm is initially trained, a subset of the inputs may be identified as most influential or most important to be included for making high-quality predictions. For example, a subset of the data may be identified as most influential or most important to be included for making high-quality associations of carbohydrate binding protein patterns and glycan features or glycan features and diagnosis. The data or a subset thereof may be ranked based on classification metrics indicative of each parameter’s influence or importance toward making high-quality associations. Such metrics may be used to reduce, in some embodiments significantly, the number of input variables (e.g., predictor variables) that may be used to train the trained algorithm to a desired performance level (e.g., based on a desired minimum accuracy, PPV, NPV, sensitivity, specificity, AUC, or a combination thereof). For example, if training the trained algorithm with a plurality comprising several dozen or hundreds of input variables in the trained algorithm results in an accuracy of classification of more than 99%, then training the trained algorithm instead with only a selected subset of no more than about 5, no more than about 10, no more than about 15, no more than about 20, no more than about 25, no more than about 30, no more than about 35, no more than about 40, no more than about 45, no more than about 50, or no more than about 100 such most influential or most important input variables among the plurality can yield decreased but still acceptable accuracy of classification (e.g., at least about 50%, at least about 55%, at least about 60%, at least about 65%, at least about 70%, at least about 75%, at least about 80%, at least about 81%, at least about 82%, at least about 83%, at least about 84%, at least about 85%, at least about 86%, at least about 87%, at least about 88%, at least about 89%, at least about 90%, at least about 91%, at least about 92%, at least about 93%, at least about 94%, at least about 95%, at least about 96%, at least about 97%, at least about 98%, or at least about 99%). The subset may be selected by rank-ordering the entire plurality of input variables and selecting a predetermined number (e.g., no more than about 5, no more than about 10, no more than about 15, no more than about 20, no more than about 25, no more than about 30, no more than about 35, no more than about 40, no more than about 45, no more than about 50, or no more than about 100) of input variables with the best association metrics.

[0079] Systems and methods as described herein may use more than one trained algorithm to determine an output. Systems and methods may comprise 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, or more trained algorithms. A trained algorithm of the plurality of trained algorithms may be trained on a particular type of data. Alternatively, a trained algorithm may be trained on more than one type of data. The inputs of one trained algorithm may comprise the outputs of one or more other trained algorithms. Additionally, a trained algorithm may receive as its input the output of one or more trained algorithms.

[0080] In some embodiments, the likelihood may be expressed as a probability. In some embodiments, the likelihood may be expressed as a pseudo-probability. In some embodiments, the likelihood may be expressed as a ratio or product of one or more probabilities or pseudoprobabilities. In some embodiments, the likelihood may be expressed as a sum or difference of one or more probabilities. In some embodiments, the likelihood may be expressed as an odds ratio. In some embodiments, the likelihood may be expressed as the logarithm of an odds ratio.

[0081] EXAMPLES

[0082] A cancer-oriented interpretable neural network

[0083] To model responses to RSi drugs, we focused on the set of 718 genes assessed by current clinical cancer gene panels, including one or more of FoundationOne CDx (40), Tempos xT (41), and Project GENIE (42). The genomic alteration status of these genes in a tumor sample, including the presence / absence of mutation and copy number aberration (CNA), was used as input (Fig. 10, Panel A). Models were trained using drug response data for genomically characterized tumor cell lines harmonized from the Cancer Therapeutics Response Portal (CTRP) (43,44) and the Genomics of Drug Sensitivity in Cancer (GDSC) (45,46) databases. These databases included the measured responses to many RSi drugs targeting DNA replication or DNA damage response. In another recent study, Olivieri and colleagues employed genome-wide CRISPR chemogenetic screens to investigate DNA damage response against genotoxic agents (47). Therefore, we initially focused on six RSi agents characterized by Olivieri (cisplatin, gemcitabine, camptothecin, etoposide, olaparib, CD437) (Fig. 2, Panel A) which had also been examined in CTRP or GDSC. Instead of associating genetic alterations with drug responses via classical “black-box” machine learning, we implemented visible neural network models (VNN) in which the layers of artificial neurons are designed to propagate the effects of individual gene mutations over the NeST map of cancer protein assemblies (36). In constructing the VNN, each protein assembly encoded by genes on the clinical panel (131 assemblies, Fig. 8) was assigned a bank of artificial neurons to represent the in silico activity of that system (Fig 10, Panel A; Methods). The use of multiple neurons allowed protein assemblies to be multifunctional, with the ability to adopt a range of values along several dimensions. For the neurons assigned to an assembly, weighted input connections were permitted from the neurons assigned to each of its subassemblies in the preceding layers of the NeST hierarchy. This design enabled genetic information to flow from the input gene alterations to small protein complexes (e.g., “CDK holoenzyme complex”) then larger scale assemblies (e.g., “checkpoint-regulated DNA repair”), and finally to the root assembly, representing the whole cell (Fig. 10, Panel B). Connection weights were learned during training so that the integrated activity of the root neurons represented the predicted drug response of a tumor sample given its genetic alteration profile. In the multi-drug configuration, this architecture was extended so that the activity of the root was provided as input to an additional (final) neuronal processing layer that was optimized separately for each drug response (Fig. 10, Panel C; Methods).

[0084] Training and performance analysis of the models

[0085] We first trained single-drug VNN models for predicting the response to each of the six RSi agents (Fig. 2, Panel A). Training was conducted by minimizing the mean squared error between the predicted and observed drug responses using standard back-propagation techniques (Methods). The accuracy of drug response prediction was assessed for each model using nested cross- validation, in which 64% of cell lines were randomly selected for model training, 16% for model validation and hyperparameter tuning, and 20% as held-out cell lines not yet seen during training or validation (this entire procedure was repeated over five folds, Methods). This assessment produced predictive odds ratios (OR) in the range of 2.2 to 3.2 across the six RSi-drug models (Fig. 2B, greyscale-orange points). These models generally demonstrated performance that was comparable to, or better than, matched black-box neural networks (Fig. 2, Panel B, greyscale-blue points; ORs of 1.8 to 3.3; P = 0.04). Their performance was significantly better than that of two general pan-drug models (DrugCell, DeepCDR) which had not been specifically developed for RSi drug prediction (30,48) (Fig. 2, Panel B, greyscale-cyan and greyscale-purple points; with P- values of 2.5x 10-8 and 1.2x 10-10 respectively).

[0086] We also used multi-task learning to train a unified multi-drug VNN with six outputs, predicting the response to each of the six RSi drugs (Fig. 10, Panel C; Methods). This unified model yielded ORs of 3.0 to 4.9 (Fig. 2, Panel B, greyscale-red points), significantly outperforming both the single-task models (P = 4.3x 10-5) and the matched black-box models (P = 2.1 x 10-7). To further test the generality of this multi-drug model, we evaluated its accuracy in predicting cellline responses to additional RSi agents not yet seen in this study (predictions based on the average of model outputs, Methods). The new agents included RS-inducing chemotherapies such as doxorubicin, 5-fluorouracil, methotrexate, and bleomycin-a2 as well as drugs targeting DNA damage response factors such as ceralasertib (ATR inhibitor) and MK-1775 (WEE1 inhibitor). The predictive ORs were significant for all new RSi drugs (mean OR 3.2, Fig. 2, Panel C greyscale- red points, Fig. 9), and they were significantly higher than ORs obtained when predicting responses to a broad panel of 33 agents not associated with replication stress (mean OR 1.8, P = 3.1 x10-4; Fig. 2, Panel C greyscale-blue points). Furthermore, the performance of the multi-drug model decreased considerably when, during training, some of the RSi drug responses were substituted with unrelated drug responses (Fig. 11, Panel A). These results were consistent with the expectation that RSi drugs share common response pathways, a finding that was also reflected by varying degrees of similarity in their drug response profiles across tumor cell lines (Fig. 11, Panel B). Molecular assemblies important for RSi drug response prediction

[0087] We next scored all protein assemblies to rank their importance in prediction of RSi drug responses (Supplementary Table S3). For this purpose, we computed a system importance (SI) score, which measures the dependence of the drug response prediction on genetic alterations within an assembly, captured by changes in the in silico activity of that assembly’s neurons (Methods). We created an importance profde for each RSi agent using the single-drug models, then mapped these to the NeST protein assembly hierarchy (Fig. 3, Panel A); SI profdes were also computed for the multi-drug model, with qualitatively consistent results (Fig. 11, Panel C; Supplementary Table S3).

[0088] Due to the progressive integration of genetic information, we noted that importance scores for all models tended to increase with assembly size and depth in the hierarchy. To thus highlight distinct protein assemblies important to RSi drugs, we focused our examination on small-to- medium scale assemblies (124 assemblies with fewer than 100 genes), identifying 41 that were high scoring in the single- or multi-drug RSi models (Fig. 3 Panels A and B; Fig. 11, Panel C; Methods). These “RSi assemblies” recovered proteins associated with known drug indications or mechanisms of action, where applicable (camptothecin: TOPI; etoposide: TOP2A; olaparib: BRCA1, BRCA2, PARP1; the CD437 target POLA1 was absent from clinical gene panels). For example, in predicting etoposide response we identified “Checkpoint-regulated DNA repair” and “Gl / S phase” assemblies, which contained the etoposide target, topoisomerase 2a (TOP2A) (Supplementary Table S3). Turning from positive to negative controls, we found that the various RSi drug models yielded assembly importance scores that were distinct from those of non-RSi drugs but similar to one another (Figs 3C). Moreover, these RSi assemblies were specifically important for the prediction of the cellular response to RSi drugs, as opposed to non-RS drugs (Fig. 3, Panel D).

[0089] Validation of specific protein assemblies using genome-wide functional assays

[0090] For the assemblies identified as important to the RSi drug models, we investigated whether engineered genetic perturbations impacting each of these assemblies were able to predictably affect the drug response. For this purpose, we turned to the recent work of Olivieri and colleagues (47), who had conducted genome-wide drug sensitivity screens in the presence of each of 27 genotoxic agents, including the 6 RSi drugs used to train the models in our study (Fig. 12, Panel A). For each drug, genes were scored by the degree to which loss-of-function by CRISPR / Cas9 modulates drug sensitivity in either direction (i.e. absolute z-score, capturing both increases and decreases in sensitivity). Using these scores, we found that 24 of the 41 RSi assemblies were significantly enriched for genes affecting sensitivity to one or more RSi drugs, a rate higher than for non-RSi assemblies (59% versus 45%, P < 0.001; Fig. 12, Panel B). Such enrichment was particularly strong in the drug sensitivity screens for etoposide, camptothecin, and CD437 (Fig. 12, Panel C).

[0091] We then explored a second genome-wide loss-of-function screen, directly measuring immunofluorescent readouts of DNA damage and replication restart after an RS challenge (49). This screen had measured the effect of siRNA gene knockdowns on phosphorylation of histone variant H2AX (yH2AX), a marker of stalled replication forks, and on incorporation of EdU in DNA, a marker of active genome replication (Fig. 4, Panel A). The ratio of these two readouts, the “replication restart score” (RRS), was used to score all gene knockdowns for their importance in RS response. Genes that affected replication restart in either direction (absolute z-score) were enriched in 21 of the 41 RSi assemblies, a rate significantly higher than for non-RSi assemblies (51% versus 30%, P < 0.001; Fig. 4, Panel B). In addition, we found that the converse enrichment relationship was also true, in that NeST assemblies enriched in genes affecting drug sensitivity or replication restart were significantly more important to RSi drug models than to non-RSi models (Fig 4, Panels C and D; Fig. 12, Panels D-J; with P-values of 1 x 10-9 and 3 * 10-6 respectively). Summarizing the functional screening results, the 41 RSi assemblies included 30 with support from either the drug sensitivity or replication restart readouts (Fig. 4, Panel E). Among these, 15 assemblies had support from both, including assemblies involved in mismatch and excision repair, homologous recombination, Fanconi anemia repair, checkpoint-regulated DNA repair, locomotion, ubiquitin / proteasomal assemblies, and the RTK-JAK-STAT pathway (Fig. 4, Panel F; Supplementary Table S4). The set of 30 assemblies and the subset of 15 assemblies were equally important for predicting RSi drug response, surpassing non-RSi drug response (Fig. 12, Panel K). Prediction of clinical responses to cisplatin

[0092] Among adult solid tumors cataloged by The Cancer Genome Atlas (TCGA), five cancer subtypes commonly receive cisplatin therapy, with substantial numbers of associated samples (>30 cisplatin-treated subjects per subtype, Fig. 5, Panel A). Among these subtypes, we noted that patients with cervical and lung squamous carcinomas (CESC and LUSC) currently show cisplatin benefit, with long progression-free survival times (PFS, Fig. 5, Panel B) that are significantly improved over patients not receiving cisplatin (Fig. 5, Panels C and D; contrast to other types in Fig. 13, Panels A-C). Despite this benefit, approximately 35% of cervical and lung tumors continue to progress after treatment (Fig. 5, Panel B). Accordingly, we investigated whether these variable outcomes could be predicted by the cisplatin VNN (Fig. 5, Panel E; Methods). Indeed, we found that cisplatin-treated patients with CESC or LUSC, who were predicted as sensitive to treatment, had significantly better progression-free survival outcomes than those who were predicted as resistant (CESC Hazard Ratio = 2.2 at P = 0.02, LUSC HR = 3.4 at P = 0.03; Fig. 5, Panels F-G). This performance was significantly better than baseline random forest or elastic net models (Fig.

[0093] 5, Panel H; with P-values of 9.1 x 10-5 and 1.0x 10-2 for comparison with random forest and elastic net, respectively). Among the remaining cancer cohorts which failed to show cisplatin benefit — head-and-neck squamous cell carcinoma (HNSC), lung adenocarcinoma (LU AD), and ovarian carcinoma (OV) — we observed that the vast majority of these patients were predicted as resistant by the cisplatin VNN model, as expected (75%, Fig. 13, Panel D; compared to 25% for CESC and LUSC, Fig. 13, Panel E).

[0094] We next investigated which protein assemblies were responsible for the clinical predictions. Starting from the 29 assemblies found to be important to cisplatin response during training in cell lines (Fig. 6, Panel A), we computed the in silico activity of each of these assemblies (per patient) as the first principal component of its neuron values (Methods). This activity was then investigated, separately for each assembly, as a quantitative marker for the prediction of cisplatin response in cervical cancer patients. We found that these 29 assembly activities were generally predictive of cisplatin treatment outcomes and tended to outperform those drawn from other protein assemblies in the NeST map (Fig. 6B; Supplementary Table S5). This observation also applied to the 41 RSi assemblies, as well as to the sets of assemblies that were validated in functional screens. Notably, the 15 assemblies validated by both screening modes exhibited the highest predictive performance (Fig. 13, Panel F). They also surpassed the predictive power of single-gene biomarkers based on presence / absence of coding alterations in individual genes (Fig.

[0095] 6, Panel B). Some of the most predictive assemblies included roles in the regulation of -catenin transcriptional activity, DNA repair, epigenetic modification, and RTK-JAK-STAT signaling (Fig. 6, Panel B), most of which had also been validated in the earlier drug sensitivity or replication restart screens (Supplementary Table S4). For example, the top assembly with predictive power in patients was NeST: 126 (Regulation of P- catenin transcriptional activity, Fig. 4, Panel F and Fig. 6, Panel C), capturing activation of gene transcription by - catenin via multiple mechanisms, including P-catenin coactivation (EP300, CREBBP) (50,51), negative regulation of P-catenin activity (HDAC2) (52), and regulation of P-catenin ubiquitination (TP53) (53). The high in silico activity of this assembly was associated with significantly worse progression-free survival outcomes (Fig. 6, Panel D), reflecting that genetic alteration of this complex predicts cisplatin resistance. Notably, single proteins in this assembly with frequent genetic alterations in cancer patients, such as TP53, were not individually predictive of drug response (Fig. 6, Panel E; TP53 mutation frequency 14% in TCGA cervical cancer). BARD1 and TOP2A, two DNA repair genes that do not encode members of NeST: 126, exhibited relatively high predictive accuracy but had limited predictive value due to the rare occurrence of alterations in these genes in cervical cancer patients, making them inapplicable to a large portion of the population (Fig. 6, Panel E). The assembly as a whole, however, had both high alteration frequency and relatively high predictive accuracy (Fig. 6, Panel E; 42% alteration frequency, C-index: 0.62). Beyond the cisplatin VNN, we also evaluated the performance of the other RSi drug models on prediction of cisplatin survival outcomes, including the single-drug and multi-drug models. While other single-drug VNNs were less accurate than the cisplatin VNN, their C-index generally outperformed the non-RSi VNNs (Fig. 13, Panel G). The multi-drug VNN was also able to stratify cisplatin-treated patients into resistant and sensitive groups with comparable accuracy to the cisplatin VNN (Fig. 13, Panel G).

[0096] The RTK-JAK-STAT assembly as a predictive marker of cisplatin response

[0097] The RTK-JAK-STAT assembly (NeST: 89) provides an excellent illustration of how predictive accuracy is driven by the integration of multiple rare alterations in a protein assembly (Fig. 7, Panel A). It consists of dense interactions of upstream growth factors (EGF) and receptor tyrosine kinases (RTKs, including EGFR and ERBB2 / 3 / 4) with downstream signaling messengers and effectors (e.g. JAK1 / 2, STAT3 / 5A) (Fig. 7, Panel B). Genetic alterations in this assembly had been initially identified as important for predicting cisplatin and gemcitabine responses in cell lines (Fig. 3, Panel A). The impact of such alterations was subsequently validated in both the drug sensitivity and replication restart screens (Fig. 4, Panels C and F; Fig. 12, Panel I; Supplementary Table S4), and high activity of this assembly was shown to be predictive of drug resistance in cisplatin-treated cervical cancer patients (Fig. 6, Panel B and Fig. 7, Panel C, HR = 2.8, P = 0.004). Notably, none of the individual proteins in this complex were frequently genetically altered in cervical cancer, with all frequencies less than 10% (Fig. 7, Panel D). As an integrator of these rare alterations, however, the complex as a whole showed a relatively high frequency of alteration (44.3% in TCGA cervical cancer) and predictive accuracy (C-index: 0.58) (Fig. 7, Panel D). Perturbation of this complex (in silico activity > 0) was associated with drug resistance, driven predominantly by alterations spread over eight genes. These included copy number amplifications in ERBB2, JAK2, and EGFR, copy number deletions in ERBB4, CBL, and IRS1, and point mutations in CBLB and PDGFRA (Fig. 7, Panel E, several of these genes had multiple alteration types). In general, the presence of any of these alterations was able to signal high assembly activity, and thus cisplatin resistance, following approximate OR-type Boolean logic (Fig. 7, Panel F).

[0098] Discussion

[0099] Here we have advanced a set of interpretable deep learning models aimed at understanding genetic mechanisms of susceptibility and resistance to drugs that induce replication stress. Instead of associating single gene alterations with drug responses directly, the strategy is to project these alterations on a map of protein complexes and larger molecular assemblies associated with cancer. This approach is prompted and supported by the concept that cancer is a network-based disease arising from the action of hallmark cancer pathways (11 ,54). According to this concept, a particular driving mutation may occur rarely in a tumor population, but such rare events can sometimes be better understood and modeled by their impacts on common subcellular components. In this regard, our model identifies 41 protein assemblies in which genetic alterations modulate RSi drug responses, of which many could be corroborated by systematic gene loss-offunction assays (Fig. 4 and Fig. 12) and / or predictive power in clinical cohorts (Figs. 5-7).

[0100] Predictive assemblies identified by the models

[0101] It is widely recognized that cancer cells acquire resistance to the damage caused by cisplatin and other RSi agents through multiple mechanisms, with some of the key factors being the activation of DNA damage response pathways as well as rewiring of anti-apoptotic and prosurvival signaling cascades (24,55). In agreement with these previous studies, our models highlight the significant roles of alterations in DNA repair-related assemblies, including mismatch and excision repair (NeST:30) as well as homologous recombination and Fanconi Anemia repair (NeST:79) (Fig. 3, Panel A and Supplementary Table S3). Such assemblies also capture the previously observed interactions of multiple DNA polymerases, including pol 8, £, and q (POLDI, POLE, POLH) (56-58), with DNA repair proteins in regulating cisplatin resistance. Beyond DNA repair, an established component of cisplatin resistance is the dysregulation of cell growth and survival signaling, including the insulin receptor-PI3K-AKT-FOXO (INSR-FOXO, NeST:52) (59,60) and RAS-MAPK (NeST:73) (61) cascades, aspects that our models also capture (Fig. 3, Panel A and Supplementary Table S3).

[0102] In addition to these well-characterized response pathways, a growing body of evidence has begun to implicate aberrant epigenetic regulation cisplatin resistance, but the mechanistic details have remained unclear (62,63). Here, all RSi-drug models (including the specific agent models and the multi-drug model) identify an assembly of epigenetic regulators in which alterations to multiple genes converge to modulate drug resistance in both cell lines and patients (Epigenetic modification, NeST:34; cell lines: Fig. 3, Panel A and Fig. 11, Panel C; patients: Fig. 6, Panel B). This assembly integrates genetic alterations across DNA methyltransferases (DNMT1, DNMT3A, DNMT3B), histone methyltransferases (EZH2, PRDM1), histone demethylase (KDM5A), histone acetyltransferase (CREBBP), histone deacetylases (HDAC1-3), chromatin regulators (SMARCA1, CHD4, MTA1) and transcription factors (RUNX1, ZBTB2, ZNF217, BRD4, TP53). Among these, overexpression of the histone methyltransferase EZH2 has been shown to promote, and its inhibitor to effectively reverse, cisplatin resistance (64-66). One possibility is that this and other epigenetic factors can affect cisplatin response via interactions with transcription factors in this assembly, such as RUNX1 and the zinc-finger proteins ZBTB2 and ZNF217. The impact of alterations to other proteins in NeST:34 on cisplatin response is less well-studied. Beyond this particular assembly, our model also highlights the importance of assemblies engaged in protein transport (NeST: 12), locomotion (NeST: 14), and regulation of immune responses (NeST: 18). These predictive assemblies present promising avenues for further research and development.

[0103] The significance of integrating across many genetic alterations

[0104] Historically, drug response prediction has been pursued predominantly through the search for single-gene markers. Here, a key contribution of our work is to demonstrate how individual genetic alterations may be integrated into a unified quantitative assessment of drug resistance, calling attention to both well-known and understudied effects (Fig. 1, Panel A). For example, the predictive NeST:89 assembly (RTK-JAK-STAT signaling, Fig. 7) integrates and quantitatively weighs the effects of well-documented genetic markers of cisplatin resistance in cervical cancer, such as EGFR, JAK2, STAT3 and STAT5 (67-70). It also integrates additional predictive features that have been less well-appreciated, including copy number deletion of CBL and ERBB4, as well as mutations of PDGFRA. Negative regulators like ubiquitin ligases (CBL, CBLB) may counterbalance the activation of EGFR and other RTKs, influencing the sensitivity to RSi agents like cisplatin. Unlike other oncogenic ERBB family proteins in this assembly, ERBB4 can form tumor-suppressing homodimers and negatively regulate STAT5A (71). This unique aspect of the ERBB4 paralog may explain why, in our models, ERBB4 deletions are predictive of cisplatin resistance in cervical cancer patients (Fig. 7, Panels E and F). More generally, the extensive interactions in this assembly interconnecting RTKs to JAK-STAT factors (Fig. 7, Panel A) suggest that alterations of RTKs may trigger downstream JAK-STAT signaling to modulate the cellular response to cisplatin and potentially other RSi agents. Notably, the integration of all of these effects yields an integrated marker of cisplatin resistance with both high alteration frequency and predictive power (Fig. 7, Panel D).

[0105] Comparison to previous ML models

[0106] Our work extends previous RSi drug response models along several lines. A first group of models, representing the majority of ML drug response modeling efforts (72), have focused on mRNA expression features for prediction. For example, Jin et al. used expression features very effectively in their state-of-the-art pan-drug-response model, which was also interpretable (73). Here the VNN models are complementary, as they focus on the integration of rare and common genetic alterations, including somatic mutations and copy number aberrations. A second group of models have sought to identify gene signatures associated with responses to single RSi drugs (74- 78), such as Sui et al. who focused on predicting cisplatin resistance in non-small cell lung cancer (77). Conversely, at the opposite end of the spectrum are models trained across very large numbers of drugs, including DeepCDR and DrugCell (30,48). While such studies do not focus on RSi agents in particular, they nonetheless include representative RSi agents among hundreds of others, and some of these models also provide mechanistic pathway interpretation. Likely due to their generality, these models appear to sacrifice precision in RSi drug response prediction (e.g., see comparisons in Fig. 2, Panel B). In this respect, the multi-task implementation presented here appears to strike a useful balance between single RSi-drug models and pan-drug models, in that it can capture both drug-specific features alongside general mechanisms underlying multiple RSi responses (Fig. 10, Panel C). Systematic validation of assemblies via complementary functional assays

[0107] Our exploration of the predictive protein assemblies invoked two distinct types of genomewide functional assay, with complementary insights. The first sought to confirm assemblies in which loss of gene function affects tumor cell fitness during a drug treatment (47). This assay provides a direct test of the prediction that genetic alterations in an assembly affect a drug response, and it scales easily to multiple drugs using a pooled gene knockout library. On the other hand, drug sensitivity provides less information about the specific functions of an assembly, as it integrates the outputs of replication stress pathways with numerous other cellular responses such as growth signaling, apoptosis, immune response, drug export and so on. In contrast to this first screen, the second genome-wide assay sought to identify assemblies that relate specifically to mechanisms of DNA replication and replication restart (49). This second screen offers richer phenotypic readouts than cell fitness, using dual markers for replication damage and restart, respectively, and thus is a more direct probe of RS responses. On the other hand, not all genes underlying a replication phenotype necessarily affect the ultimate drug response. Furthermore, the direct replication phenotypes were screened in an arrayed format, one gene at a time, and thus were only available for a single RSi agent in our study. Regardless, 24 assemblies important to the RSi drug models could be supported by drug sensitivity screens and 21 by replication restart screens, covering 30 assemblies total (Fig. 4, Panel E).

[0108] Methods

[0109] Preparation of therapeutics response data

[0110] Drug response data were retrieved from the Genomics of Drug Sensitivity in Cancer database (GDSC, with separate datasets GDSC1 and GDSC2) and the Cancer Therapeutics Response Portal (CTRP, with separate datasets CTRP1 and CTRP2) in July of 2022 (43-46). Together, these repositories covered a total of 692,859 cell line-drug pairs, comprising 1244 cell lines and 888 drugs with some pairs missing. The numbers of cell lines with response data informing each RSi drug were as follows: cisplatin (947), gemcitabine (1210), camptothecin (799), etoposide (1181), olaparib (1213), and CD437 (825). Data from both repositories were harmonized as follows. Drug information: Each molecule’s published name, synonym, or SMILES string was queried using PubChemPy, and the corresponding associated InChiKey was extracted and stored. Duplicate drugs (within or between repositories) were then matched with one another using InChiKeys, and PubChemPy was used to extract isomeric SMILES strings. Some compounds with no matches were manually annotated. Cell viability data: For CTRP, the average percent viability fdes, which have been normalized to vehicle control, were used. For GDSC1, data were normalized to ‘cells-only’ controls on a per-plate basis. For GDSC2, data were normalized to DMSO control wells on a per-plate basis. Data were then averaged across all replicates for each dataset separately. For drug response measurement, we used Area Under the dose-response Curve (AUC) where AUC = 0 corresponds to complete cell killing and AUC = 1 corresponds to no cell killing; AUC > 1 represents a growth advantage conferred by the drug. The harmonized AUCs calculated in this study were in agreement with AUCs reported by the original consortia (Pearson correlations of 0.92, 0.83, 0.91, and 0.91 for CTRP1, CTRP2, GDSC1, and GDSC2, respectively).

[0111] Selection of gene features

[0112] A set of 718 clinical genes was assembled from the union of those in which mutations and / or copy number aberrations are assessed by one or more of the following clinical panels: FoundationOne CDx (40), Tempus xT (41), PALOMA-3 trial (80), or Project GENIE (42). To compile genotypes for all cell lines, we extracted non-synonymous coding mutations and copy number alterations for the clinical panel genes from the Cancer Cell Line Encyclopedia (CCLE, release 22Q1) (81). Gene mutations were marked as either present (“1”) or absent (“0”), with mutations fdtered for the following types: missense, nonsense and non-stop mutations, frame-shift insertions and deletions, splice site and region variations, and in-frame insertions and deletions. Gene copy number deletions or amplifications were marked separately, also using binary (1 / 0) indications. Together, mutations, copy number deletions, and copy number amplifications served as features for each of the clinical panel genes.

[0113] Model architecture

[0114] We queried the NeST hierarchy of cancer protein assemblies (36) to identify those that contained clinical panel genes. We define a hierarchy of protein assemblies as a multi-layered treelike structure where individual genes recursively cluster together to form assemblies that are connected to other assemblies in parent-child relationships all the way up to the root of the structure which represents a cancer cell (Fig. 9). In our models, protein assemblies were filtered to require >5 clinical panel genes or >1 child assembly, producing a final hierarchy consisting of 131 assemblies distributed over 7 layers. An assembly in NeST can have both single proteins and other assemblies as its children (Fig. 10, Panel A). We trained six single-drug visible neural networks (VNNs) (29) and one multi-drug VNN (Fig. 2, Panel A), where the VNN architecture followed the connections of the 718 genes and 131 assemblies in NeST (Fig. 10). Gene alterations (mutations, copy number deletions, and copy number amplifications) form the input feature layer, which connect together to form the gene layer (Fig. 10, Panel A). For any gene g, we denote the input features as a vector It and the output as g , where i e [1,718], li e [0, 1 ]3and gi e R. Hence, a gene-layer equation is given by: gi = B atchN orm(T anh(Linear(l i )))

[0115] BatchNorm indicates Batch Normalization (82), Tanh indicates a hyperbolic tangent function and Linear indicates a weighted linear transformation.

[0116] The remaining layers of the model represent the 131 NeST assemblies, where each assembly is represented by n neurons, and every parent-child connection follows the edges in the hierarchical map. An assembly-gene pair is connected through n X 1 connections and an assemblyassembly pair through n X n connections. The number of neurons is a hyperparameter. Dropout (83) with probability = 0.3 was added to layers five through eight after hyperparameter optimization. Given an assembly Aj connected to k child assemblies and m genes, its input is denoted by a vector Ij of dimension n * (n * k + m) and output by a vector Aj of dimension n , where Aj e Rnand j 6 [1,131], Thus, the equation for an assembly is given by:

[0117] Aj = BatchNorm (Tanh (Linear (Dropout(Ij))))

[0118] The state of an assembly, which we refer to as the ‘in silico activity’, is defined as a function of the states of its k child assemblies and m genes. To reduce in silico activity to a single dimension, we compute the first principal component resulting from principal component analysis (PCA) (84) of the set of neuron values for each assembly. The final output for a single-drug model is a linear transformation of the output of the root, resulting in a real number in the range [0,1], For the multi drug model, the output of the root node is connected to an additional layer of neurons, one for each of d drugs, resulting in an output vector OcRd. Hence, the output-layer equation is given by:

[0119] 0 = T anh(Linear(A root ))

[0120] The objective function (Loss) of the VNN aggregates the mean squared error (MSE) across every assembly in the hierarchy. The loss function for a single-drug model can be defined as:

[0121] The parameter a was set to 0.3 whereas was tuned by hyper-parameter optimization. Linear denotes the linear function used for transforming the vector / to a scalar. Note that for the singledrug models, 0 is already a scalar. W denotes the weights of the neural network. AdamW (arXiv 1711.05101) was used to optimize weights. For the multi-drug model, the final loss is the sum of loss for each drug, given by:

[0122] NB: Assembly states and weights remain the same because they are shared across all drugs.

[0123] Model training

[0124] All models were trained using a five-fold nested cross-validation procedure. For each fold setting, 64% of cell lines were split as a training set, 16% as a validation set (used in hyperparameter tuning), and 20% as a test set, ensuring that cell line replicate measurements (e.g., from different GDSC or CTRP datasets) were not split between test and training sets. Hyperparameter optimization was performed using Optuna (85). All VNN models were implemented in PyTorch and trained using five GPU servers containing four Nvidia Tesla VlOOs each with 5120 CUDA cores and 32GB GDDR6 RAM.

[0125] Alternate models for performance comparison

[0126] We assessed three models for benchmarking the performance of VNN models. We constructed a black-box artificial neural network (ANN) (35) using the Python scikit-learn library (86), which was allotted the same number of neurons and layers as the VNN model. The ANN was trained and optimized using the exact same procedure as the VNN models. We also evaluated two previously published models, DeepCDR (48) and DrugCell (30). Both models were re-trained against the exact mutation calls and drug responses used for the VNN models using five-fold cross- validation. We assessed the performance of these models on held-out cell lines, as compared to the held-out drug-cell line pairs used originally in these studies.

[0127] Predicting drug response using the multi-drug model

[0128] To predict cellular response to each drug (6 discovery and 39 test drugs), first, we divided the cell lines into five held-out sets, one for each fold of the multi-drug model, such that no cell line was used in training the model. The model predicted six responses per cell line, one for each RSi drug (O, Fig. 10, Panel C). Since each fold predicted response to a distinct set of cell lines, the responses from all the folds were simply aggregated to calculate one odds ratio per discovery drug and six odds ratio values per test drug. The odds ratio and its confidence interval were reported for the discovery drugs (Fig. 2, Panel B). The mean and standard error of the six odds ratios were used to compare the predictions across the test drugs (Fig. 2, Panel C).

[0129] System importance (SI) score

[0130] To determine the importance scores of protein assemblies for drug response prediction, we adopted a variation of “Relative Local Improvement in Predictive Power” previously reported by Ma et al. (29). For each of the five pre-trained models corresponding to each drug, we used Ridge Regression, an L2-norm regularized linear regression method, trained on assembly neuron values (A, Fig. 10, Panel A), to model the predicted drug response of an assembly (system). System importance was calculated using the Spearman correlation between the prediction of Ridge Regression and VNN drug response for the single-drug models and the in silico activity of the root node for the multi-drug model. Throughout the paper, the average scores of system importance were reported across all five models. A higher SI score indicated an assembly for which the neurons had a strong contribution to VNN predictions and could therefore be considered important; conversely, a low SI score indicated an assembly for which the neurons were weak predictors of VNN predictions and could therefore be considered of low importance. Assemblies with SI scores of > 0.5 were referred to as important assemblies (Fig. 3, Panel A).

[0131] Enrichment of important assemblies in functional screens

[0132] The function “gost” from the R package “g:Profiler” (87) was used to identify functional enrichment of genes in the NeST assemblies. An assembly-to-genes annotation file (.gmt) (Supplementary Table SI) was created using the function “CreatePathwayCollection” in the R package “pathwayPCA”. Genes were first rank-ordered by the degree (absolute z score) to which their loss-of-function modulates the response to an RSi drug in the CRISPR / Cas9 screens or the repair of replication fork (47,49). The ranks in each of the chemogenetic screens or the RNAi screen were then loaded into “gost” as input to find the over-representation of functions from the NeST. The Benj ami ni -Hochberg procedure was utilized to adjust p-values for multiple comparisons. Over-representation with FDR < 0.1 was considered enrichment in the CRISPR / Cas9 screens. (Fig. 12). Note that the models identify assemblies important for sensitivity or resistance, by integrating the effect of both gain and loss of function genetic events. However, the models currently only have binary annotations for mutations, without distinguishing gain or loss of function. Given the integrated bidirectional complexity of mutation functions and response outputs, we evaluated CRISPR knockouts or RNAi knockdowns as signless modifiers for systematic validations. A more detailed evaluation can be performed for a specific assembly by carefully examining individual patient mutations.

[0133] Preparation of cisplatin clinical data

[0134] Genotypes (mutations, copy number variations) for patients treated with cisplatin were queried from The Cancer Genome Atlas (TCGA, PanCancer Atlas) via cBioportal (cbioportal.org / datasets). Patients treated with multiple RSi drugs (cisplatin and etoposide; cisplatin and gemcitabine) were excluded, resulting in 368 cisplatin-treated patients with genomic profiles and clinical information. The patient population where cisplatin is the standard-of-care was selected for a focused model evaluation. We further selected cisplatin treated cancer types with at least 30 patients. This resulted in five types, including 106 patients with cervical squamous cell carcinoma (CESC), 45 patients with lung squamous carcinoma (LUSC), 92 patients with head and neck squamous cell carcinoma (HNSC), 58 patients with lung adenocarcinoma (LU AD) and 67 patients with ovarian serous cystadenocarcinoma (OV) (Fig. 5).

[0135] Model evaluation for cisplatin-treated patients

[0136] We evaluated the prediction performance of the cisplatin VNN model using the concordance index (C-index) and hazard ratio (HR) (88), which were calculated using “Lifelines”, a Python survival analysis library. C-index quantifies the fraction of concordant pairs in patients, comparing predicted AUC with actual survival time, out of all possible pairs. Those pairs were discarded if the earlier time was censored. For a given test set, its risk scores with event time and event status were used to calculate the C-index. A higher C-index value indicates a better time-to- event prediction model, with a C-index of 1.0 indicating perfect prediction and a C-index of 0.5 indicating random prediction. C-index less than 0.5 indicates an anti-concordance relation. Hazard ratio (HR) is a statistical measure that compares the relative hazard rates between predicted resistant and sensitive patients, using the Cox proportional hazards regression method (89). A hazard ratio greater than 1 indicates a higher risk in the predicted resistant group compared to the sensitive group (Figs. 5-7 and Fig. 13).

[0137] Data availability. The datasets used in this study are all publicly available: GDSC1 and GDSC2: cancerrxgene.org / ; CTRPL portals.broadinstitute.org / ctrp.vl / ; CTRP2: portals.broadinstitute.org / ctrp.v2.1 / ; DepMap 22Q2: doi.org / 10.6084 / m9. figshare.19700056.v2; cBioportal (cbioportal.org / datasets). The pre-trained models are available on GitHub in their respective repositories.

[0138] Code availability. The source codes are available at GitHub: VNN github.com / idekerlab / nest_vnn; multidrug ; VNN github.com / idekerlab / mutlitask_vnn.

[0139] Supplementary Tables. The supplementary tables are available at: doi.org / 10.1158 / 2159- 8290.CD-23-0641.

[0140] REFERENCES

[0141] 1. Burgers PMJ, Kunkel TA. Eukaryotic DNA Replication Fork. Annu Rev Biochem. 2017;86:417-38.

[0142] 2. Fragkos M, Ganier O, Coulombe P, Mechali M. DNA replication origin activation in space and time. Nat Rev Mol Cell Biol. 2015;16:360-74.

[0143] 3. Zeman MK, Cimprich KA. Causes and consequences of replication stress. Nat Cell Biol. 2014;16:2-9.

[0144] 4. Saxena S, Zou L. Hallmarks of DNA replication stress. Mol Cell. 2022;82:2298-314.

[0145] 5. Byun TS, Pacek M, Yee M-C, Walter JC, Cimprich KA. Functional uncoupling of MCM helicase and DNA polymerase activities activates the ATR-dependent checkpoint. Genes Dev. 2005;19: 1040-52.

[0146] 6. Zellweger R, Dalcher D, Mutreja K, Berti M, Schmid J A, Herrador R, et al. Rad51 -mediated replication fork reversal is a global response to genotoxic treatments in human cells. J Cell Biol. 2015;208:563-79.

[0147] 7. Saldivar JC, Cortez D, Cimprich KA. The essential kinase ATR: ensuring faithful duplication of a challenging genome. Nat Rev Mol Cell Biol. 2017;18:622-36.

[0148] 8. Ge XQ, Blow JJ. Chkl inhibits replication factory activation but allows dormant origin firing in existing factories. J Cell Biol. 2010;191: 1285-97.

[0149] 9. Toledo LI, Altmeyer M, Rask M-B, Lukas C, Larsen DH, Povlsen LK, et al. ATR Prohibits Replication Catastrophe by Preventing Global Exhaustion of RPA. Cell. 2013;155: 1088-103.

[0150] 10. Macheret M, Halazonetis TD. DNA replication stress as a hallmark of cancer. Annu Rev Pathol. 2015;10:425-48.

[0151] 11. Hanahan D, Weinberg RA. Hallmarks of cancer: the next generation. Cell. 2011;144:646-74.

[0152] 12. Ubhi T, Brown GW. Exploiting DNA Replication Stress for Cancer Treatment. Cancer Res. 2019;79: 1730-9.

[0153] 13. Dobbelstein M, Sorensen CS. Exploiting replicative stress to treat cancer. Nat Rev Drug Discov. 2015;14:405-23.

[0154] 14. Zhang J, Dai Q, Park D, Deng X. Targeting DNA Replication Stress for Cancer Therapy. Genes. 2016;7:51.

[0155] 15. Vesela E, Chroma K, Turi Z, Mistrik M. Common Chemical Inductors of Replication Stress: Focus on Cell-Based Studies. Biomolecules. 2017; 7(1): 19. 16. Zhang J, Chan DW, Lin S-Y. Exploiting DNA Replication Stress as a Therapeutic Strategy for Breast Cancer. Biomedicines. 2022; 10(11):2775.

[0156] 17. Shah MA, Schwartz GK. Cell cycle-mediated drug resistance: an emerging concept in cancer therapy. Clin Cancer Res. 2001;7:2168-81.

[0157] 18. Beretta GL, Gatti L, Perego P, Zaffaroni N. Camptothecin resistance in cancer: insights into the molecular mechanisms of a DNA-damaging drug. Curr Med Chem. 2013;20:1541-65.

[0158] 19. Dias MP, Moser SC, Ganesan S, Jonkers J. Understanding and overcoming resistance to PARP inhibitors in cancer therapy. Nat Rev Clin Oncol. 2021;18:773-91.

[0159] 20. Cao X, Hou J, An Q, Assaraf YG, Wang X. Towards the overcoming of anticancer drug resistance mediated by p53 mutations. Drug Resist Updat. 2020;49: 100671.

[0160] 21. Li L-Y, Guan Y-D, Chen X-S, Yang J-M, Cheng Y. DNA Repair Pathways in Cancer Therapy and Resistance. Front Pharmacol. 2020; 11 :629266.

[0161] 22. Jia Y, Xie J. Promising molecular mechanisms responsible for gemcitabine resistance in cancer. Genes Dis. 2015;2:299-306.

[0162] 23. Galluzzi L, Senovilla L, Vitale I, Michels J, Martins I, Kepp O, et al. Molecular mechanisms of cisplatin resistance. Oncogene. 2012;31:1869-83.

[0163] 24. Bhattacharjee R, Dey T, Kumar L, Kar S, Sarkar R, Ghorai M, et al. Cellular landscaping of cisplatin resistance in cervical cancer. Biomed Pharmacother. 2022;153: 113345.

[0164] 25. Nasrpour Navaei Z, Khalili-Tanha G, Sadra Zangouei A, Reza Abbaszadegan M, Moghbeli M. PI3K / AKT signaling pathway as a critical regulator of Cisplatin response in tumor cells. Oncol Res. Computers, Materials and Continua (Tech Science Press); 2021;29:235-50.

[0165] 26. Kiss RC, Xia F, Acklin S. Targeting DNA Damage Response and Repair to Enhance Therapeutic Index in Cisplatin-Based Cancer Treatment. International Journal of Molecular Sciences. 2021; 22(15):8199

[0166] 27. Li Y, Hostallero DE, Emad A. Interpretable deep learning architectures for improving drug response prediction performance: myth or reality? Bioinformatics, Volume 39, Issue 6, June 2023, btad390

[0167] 28. Deng L, Cai Y, Zhang W, Yang W, Gao B, Liu H. Pathway-Guided Deep Neural Network toward Interpretable and Predictive Modeling of Drug Sensitivity. J Chem Inf Model. 2020;60:4497-505. 29. Ma J, Yu MK, Fong S, Ono K, Sage E, Demchak B, et al. Using deep learning to model the hierarchical structure and function of a cell. Nat Methods. 2018;15:290-8.

[0168] 30. Kuenzi BM, Park J, Fong SH, Sanchez KS, Lee J, Kreisberg JF, et al. Predicting Drug Response and Synergy Using a Deep Learning Model of Human Cancer Cells. Cancer Cell. 2020;38:672-84. e6.

[0169] 31. Elmarakeby HA, Hwang J, Arafeh R, Crowdis J, Gang S, Liu D, et al. Biologically informed deep neural network for prostate cancer discovery. Nature. 2021;598:348-52.

[0170] 32. Chen H-IH, Chiu Y-C, Zhang T, Zhang S, Huang Y, Chen Y. GSAE: an autoencoder with embedded gene-set nodes for genomics functional characterization. BMC Syst Biol. 2018;12: 142.

[0171] 33. Hao J, Kim Y, Kim T-K, Kang M. PASNet: pathway-associated sparse deep neural network for prognosis prediction from high-throughput data. BMC Bioinformatics. 2018; 19:510.

[0172] 34. Huang X, Huang K, Johnson T, Radovich M, Zhang J, Ma J, et al. ParsVNN: parsimony visible neural networks for uncovering cancer-specific and drug-sensitive genes and pathways. NAR Genomics and Bioinformatics. 2021;3:lqab097.

[0173] 35. Hinton GE. Connectionist learning procedures. Artificial intelligence. 1989;40: 185-234.

[0174] 36. Zheng F, Kelly MR, Ramms DJ, Heintschel ML, Tao K, Tutuncuoglu B, et al. Interpretation of cancer mutations using a multiscale map of protein systems. Science. 2021;374:eabf3067.

[0175] 37. Caruana R. Multitask Learning. Machine Learning. 1997;28:41-75.

[0176] 38. Douglass EF Jr, Allaway RJ, Szalai B, Wang W, Tian T, Fernandez-Torras A, et al. A community challenge for a pancancer drug mechanism of action inference from perturbational profile data. Cell Rep Med. 2022;3: 100492.

[0177] 39. Yuan H, Paskov I, Paskov H, Gonzalez AJ, Leslie CS. Multitask learning improves prediction of cancer drug sensitivity. Sci Rep. 2016;6:31619.

[0178] 40. Frampton GM, Fichtenholtz A, Otto GA, Wang K, Downing SR, He J, et al. Development and validation of a clinical cancer genomic profiling test based on massively parallel DNA sequencing. Nat Biotechnol. 2013;31: 1023-31.

[0179] 41. Beaubier N, Tell R, Lau D, Parsons JR, Bush S, Perera J, et al. Clinical validation of the tempus xT next-generation targeted oncology sequencing assay. Oncotarget. 2019;10:2384-96.

[0180] 42. Smyth LM, Zhou Q, Nguyen B, Yu C, Lepisto EM, Amedos M, et al. Characteristics and Outcome of AKT1E17K-Mutant Breast Cancer Defined through AACR Project GENIE, a Clinicogenomic Registry. Cancer Discov. American Association for Cancer Research; 2020;10:526-35.

[0181] 43. Basu A, Bodycombe NE, Cheah JH, Price EV, Liu K, Schaefer GI, et al. An interactive resource to identify cancer genetic and lineage dependencies targeted by small molecules. Cell. 2013;154: 1151-61.

[0182] 44. Seashore-Ludlow B, Rees MG, Cheah JH, Cokol M, Price EV, Coletti ME, et al. Harnessing Connectivity in a Large-Scale Small-Molecule Sensitivity Dataset. Cancer Discov. American Association for Cancer Research; 2015;5:1210-23.

[0183] 45. Garnett MJ, Edelman EJ, Heidom SJ, Greenman CD, Dastur A, Lau KW, et al. Systematic identification of genomic markers of drug sensitivity in cancer cells. Nature. 2012;483:570-5.

[0184] 46. Iorio F, Knijnenburg TA, Vis DJ, Bignell GR, Menden MP, Schubert M, et al. A Landscape of Pharmacogenomic Interactions in Cancer. Cell. 2016;166:740-54.

[0185] 47. Olivieri M, Cho T, Alvarez-Quilon A, Li K, Schellenberg MJ, Zimmermann M, et al. A Genetic Map of the Response to DNA Damage in Human Cells. Cell. 2020; 182:481- 96.e21.

[0186] 48. Liu Q, Hu Z, Jiang R, Zhou M. DeepCDR: a hybrid graph convolutional network for predicting cancer drug response. Bioinformatics. 2020;36:i911-8.

[0187] 49. Kavanaugh G, Ye F, Mohni KN, Luzwick JW, Glick G, Cortez D. A whole genome RNAi screen identifies replication stress response genes. DNA Repair . 2015;35:55-62.

[0188] 50. Takemaru KI, Moon RT. The transcriptional coactivator CBP interacts with beta-catenin to activate gene expression. J Cell Biol. 2000;149:249-54.

[0189] 51. Hecht A, Vleminckx K, Stemmier MP, van Roy F, Kemler R. The p300 / CBP acetyltransferases function as transcriptional coactivators of beta-catenin in vertebrates. EMBO J. 2000;19: 1839-50.

[0190] 52. Shen Y, Xu L, Ning Z, Liu L, Lin J, Chen H, et al. ARHGAP4 regulates the cell migration and invasion of pancreatic cancer by the HDAC2 / p-catenin signaling pathway. Carcinogenesis. 2019;40: 1405-14.

[0191] 53. Sadot E, Geiger B, Oren M, Ben-Ze’ev A. Down-regulation of beta-catenin by activated p53. Mol Cell Biol. 2001;21 :6768-81.

[0192] 54. Vogelstein B, Kinzler KW. Cancer genes and the pathways they control. Nat Med. 2004;10:789-99. 55. Lugones Y, Loren P, Salazar LA. Cisplatin Resistance: Genetic and Epigenetic Factors Involved. Biomolecules. 2022; 12: 1365.

[0193] 56. Fuchs J, Cheblal A, Gasser SM. Underappreciated Roles of DNA Polymerase 8 in Replication Stress Survival. Trends Genet. 2021;37:476-87.

[0194] 57. Rocha CRR, Silva MM, Quinet A, Cabral-Neto JB, Menck CFM. DNA repair pathways and cisplatin resistance: an intimate relationship. Clinics . 2018;73:e478s.

[0195] 58. Tian W, Ji Z, Wang J, Meng J, Bi R, Ren Y, et al. Characterization of hotspot exonuclease domain mutations in the DNA polymerase c gene in endometrial cancer. Front Oncol. 2022;12: 1018034.

[0196] 59. Park J, Ko YS, Yoon J, Kim MA, Park J-W, Kim WH, et al. The forkhead transcription factor FOXO1 mediates cisplatin resistance in gastric cancer cells by activating phosphoinositide 3-kinase / Akt pathway. Gastric Cancer. 2014;17:423-30.

[0197] 60. Beretta GL, Como C, Zaffaroni N, Perego P. Role of FoxO Proteins in Cellular Response to Antitumor Agents. Cancers. 2019;l l:90.

[0198] 61. Achkar IW, Abdulrahman N, Al-Sulaiti H, Joseph JM, Uddin S, Mraiche F. Cisplatin based therapy: the role of the mitogen activated protein kinase signaling pathway. J Transl Med. 2018;16:96.

[0199] 62. Wang N, Ma T, Yu B. Targeting epigenetic regulators to overcome drug resistance in cancers. Signal Transduct Target Ther. 2023;8:69.

[0200] 63. Yang C, Zhang J, Ma Y, Wu C, Cui W, Wang L. Histone methyltransferase and drug resistance in cancers. J Exp Clin Cancer Res. 2020;39: 173.

[0201] 64. Sun J, Cai X, Yung MM, Zhou W, Li J, Zhang Y, et al. miR-137 mediates the functional link between c-Myc and EZH2 that regulates cisplatin resistance in ovarian cancer. Oncogene. 2019;38:564-80.

[0202] 65. Sun S, Zhao S, Yang Q, Wang W, Cai E, Wen Y, et al. Enhancer of zeste homolog 2 promotes cisplatin resistance by reducing cellular platinum accumulation. Cancer Sci. 2018;109: 1853-64.

[0203] 66. Qiu Z, Zhu W, Meng H, Tong L, Li X, Luo P, et al. CDYL promotes the chemoresistance of small cell lung cancer by regulating H3K27 trimethylation at the CDKN1C promoter.

[0204] Theranostics. 2019;9:4717-29. 67. Hugo de Almeida V, Guimaraes IDS, Almendra LR, Rondon AMR, Tilli TM, de Melo AC, et al. Positive crosstalk between EGFR and the TF-PAR2 pathway mediates resistance to cisplatin and poor survival in cervical cancer. Oncotarget. 2018;9:30594-609.

[0205] 68. Morgan EL, Macdonald A. JAK2 Inhibition Impairs Proliferation and Sensitises Cervical Cancer Cells to Cisplatin-Induced Cell Death. Cancers. 2019; 11: 1934.

[0206] 69. Gutierrez-Hoya A, Soto-Cruz I. Role of the JAK / STAT Pathway in Cervical Cancer: Its Relationship with HPV E6 / E7 Oncoproteins. Cells. 2020;9:2297.

[0207] 70. Huang L-L, Rao W. SiRNA interfering STAT3 enhances DDP sensitivity in cervical cancer cells. Eur Rev Med Pharmacol Sci. 2018;22:4098-106.

[0208] 71. Lucas LM, Dwivedi V, Senfeld JI, Cullum RL, Mill CP, Piazza JT, et al. The Yin and Yang of ERBB4: Tumor Suppressor and Oncoprotein. Pharmacol Rev. 2022;74: 18-47.

[0209] 72. Partin A, Brettin TS, Zhu Y, Narykov O, Clyde A, Overbeek J, et al. Deep learning methods for drug response prediction in cancer: Predominant and emerging trends. Front Med. 2023;10: 1086097.

[0210] 73. Jin I, Nam H. HiDRA: Hierarchical Network for Drug Response Prediction with Attention. J Chem Inf Model. 2021;61:3858-67.

[0211] 74. Mucaki EJ, Zhao JZL, Lizotte DJ, Rogan PK. Predicting responses to platin chemotherapy agents with biochemically-inspired machine learning. Signal Transduct Target Ther. 2019;4: l.

[0212] 75. Gao Y, Lyu Q, Luo P, Li M, Zhou R, Zhang J, et al. Applications of Machine Learning to Predict Cisplatin Resistance in Lung Cancer. Int J Gen Med. 2021 ; 14: 5911-25.

[0213] 76. Scarborough JA, Eschrich SA, Torres-Roca J, Dhawan A, Scott JG. Exploiting convergent phenotypes to derive a pan-cancer cisplatin response gene expression signature. NPJ Precis Oncol. 2023;7:38.

[0214] 77. Sui Q, Chen Z, Hu Z, Huang Y, Liang J, Bi G, et al. Cisplatin resistance-related multi-omics differences and the establishment of machine learning models. J Transl Med. 2022;20: 171.

[0215] 78. Berthelet J, Foroutan M, Bhuva DD, Whitfield HJ, El-Saafin F, Cursons J, et al. Computational Screening of Anti-Cancer Drugs Identifies a New BRCA Independent Gene Expression Signature to Predict Breast Cancer Sensitivity to Cisplatin. Cancers. 2022; 14(10):2404.

[0216] 79. Ngoi NYL, Pham MM, Tan DSP, Yap TA. Targeting the replication stress response through synthetic lethal strategies in cancer medicine. Trends Cancer Res. 2021;7:930-57. 80. Lira ME, Xie T, Deng S, Kinong J, Gao J, Zhu Z, et al. Abstract 2749: Liquid biopsy testing allows highly-sensitive detection of plasma cfDNA mutations in 87 breast cancer-related genes. Cancer Res 1 July 2017; 77 (13 Supplement): 2749

[0217] 81. Barretina J, Caponigro G, Stransky N, Venkatesan K, Margolin AA, Kim S, et al. The Cancer Cell Line Encyclopedia enables predictive modelling of anticancer drug sensitivity. Nature 2012 Mar 28;483(7391):603-7.

[0218] 82. Ioffe S, Szegedy C. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. In: Bach F, Blei D, editors. Proceedings of the 32nd International Conference on Machine Learning. Lille, France: PMLR; 07—09 Jul 2015. page 448-56.

[0219] 83. Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: A Simple Way to Prevent Neural Networks from Overfitting. J Mach Learn Res. 2014;15: 1929-58.

[0220] 84. Jolliffe IT. Principal Component Analysis. Springer New York, NY; 2013.

[0221] 85. Akiba T, Sano S, Yanase T, Ohta T, Koyama M. Optuna: A Next-generation Hyperparameter Optimization Framework. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. New York, NY, USA: Association for Computing Machinery; 2019. page 2623-31.

[0222] 86. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, et al. Scikit-leam: Machine learning in Python, the Journal of machine Learning research. JMLR. org;2011;12:2825-30.

[0223] 87. Raudvere U, Kolberg L, Kuzmin I, Arak T, Adler P, Peterson H, et al. g:Profiler: a web server for functional enrichment analysis and conversions of gene lists (2019 update), Nucleic Acids Research, Volume 47, Issue Wl, 02 July 2019, Pages W191-W198.

[0224] 88. Harrell FE Jr, Califf RM, Pryor DB, Lee KL, Rosati RA. Evaluating the yield of medical tests. JAMA. 1982;247:2543-6.

[0225] 89. Cox DR. Regression models and life-tables. J R Stat Soc. Wiley; 1972;34: 187-202.

[0226] 90. Singhal, A., et al. The Hallmarks of Predictive Oncology. Cancer Discovery (2025).

[0227] 91. Park, S., Silva, E., Singhal, A. et al. A deep learning model of tumor cell architecture elucidates response and resistance to CDK4 / 6 inhibitors. Nat Cancer 5, 996-1009 (2024).

Claims

CLAIMSWhat is claimed is:

1. A method, comprising: modeling a replication stress-inducing (RSi) drug response in a subject, by at least: detecting genetic alterations in a tumor sample of the subject, and incorporating the detected genetic alterations with information of cancer protein complexes, to predict the sensitivity or resistance to a RSi drug.

2. The method of claim 1, wherein the method implements a visible neural network (VNN) model to use a knowledge map of biological components and functions combined with genetic alterations associated with sensitivity or resistance to RSi drugs.

3. The method of any of claims 1 to 2, wherein the VNN predicts biomedical outcomes from genetic feature sets, and maps internal state changes in molecular mechanisms and pathways.

4. The method of any of claims 1 to 3, wherein multi-task learning is used to boost the efficacy of drug response and mechanism of action predictions, combined with knowledge bases of cancer protein complexes and pathways, such as NeST, applied to determine a response to RSi drugs.

5. The method of any of claims 1 to 4, wherein both single-drug and multi-task (multi-drug) models are evaluated.

6. The method of any of claims 1 to 5, further comprising training the model with artificial intelligence optimization procedures.

7. The method of any of claims 1 to 6, wherein the model is effective to highlight important protein assemblies or pathways that govern drug response.

8. The method of any of claims 1 to 7, further comprising: inputting genetic alterations of clinically acceptable genes by at least: taking in three binary input features per gene; describing the presence or absence of point mutation, insertion, deletion, copy number amplification, or copy number deletion; incorporating the three binary input features into a first layer of NeST-VNN to represent gene-level alteration features; and integrating the gene-level alteration features with respective protein assemblies into a second layer of the NeST-VNN, wherein the respective protein assemblies are represented by a bank of artificial neurons.

9. The method of claim 8, wherein the bank of artificial neurons are connected with larger assemblies of neurons, allowing for the flow of genetic information from small focal assemblies to effects on larger-scale assemblies and super assemblies.

10. A biologically interpretable deep learning model that combines a plurality of protein assemblies and genetic alterations to predict drug sensitivity or resistance responses to replication stress-inducing (RSi) agents in chemotherapy-treated patients.

11. The model of claim 10, wherein the model associates single gene alterations with drug responses with a map of protein complexes and larger molecular assemblies associated with cancer.

12. The model of any of claims 10 to 11, wherein the model comprises 41 protein assemblies in which genetic alterations modulate RSi drug responses as described herein.

13. A method of treating a subject in need comprising administering to the subject an effective amount of a drug identified by at least: detecting genetic alterations in a tumor sample of the subject, and incorporating the detected genetic alterations with information of cancer protein complexes, to predict the sensitivity or resistance to a RSi drug.

14. The method of claim 13, wherein the method implements a visible neural network (VNN) model to use a knowledge map of biological components and functions combined with genetic alterations associated with sensitivity or resistance to RSi drugs.

15. The method of any of claims 13 to 14, further comprising: inputting genetic alterations of clinically acceptable genes by at least: taking in three binary input features per gene; describing the presence or absence of point mutation, insertion, deletion, copy number amplification, or copy number deletion; incorporating the three binary input features into a first layer of NeST-VNN to represent gene-level alteration features; and integrating the gene-level alteration features with respective protein assemblies into a second layer of the NeST-VNN, wherein the respective protein assemblies are represented by a bank of artificial neurons.

Images