Mechanistic deep learning framework for non-toxic combination therapy design

Abstract

Methods and systems for efficiently predicting drug combination therapy candidates. A method includes: receiving an indication of a first drug; analyzing, by a trained mechanistic model, a drug profile associated with the first drug and at least one additional drug profile associated with one or more second drugs, to generate a plurality of joint profile features associated with combinations of the first drag and each respective second drug; generating, by the trained mechanistic model, a respective predicted drag combination outcome for each of the one or more drag combinations based on the plurality of joint profile features, the respective predicted drag combination outcome including a respective toxicity score and a respective potency score; and determining, by the trained mechanistic model, drug combination candidates of the one or more drug combinations based on the respective toxicity score and the respective potency score for the one or more drag combinations.

Description

Patent Application30275 / 70687 / PCMECHANISTIC DEEP LEARNING FRAMEWORK FOR NON-TOXIC COMBINATION THERAPY DESIGNCROSS REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of U.S. Provisional Patent Application No. 63 / 701,840, entitled “MECHANISTIC DEEP LEARNING FRAMEWORK FOR NON-TOXIC COMBINATION THERAPY DESIGN,” filed October 1, 2024. U.S. Provisional Patent Application No. 63 / 701,840 is hereby expressly incorporated by reference herein in its entirety.GOVERNMENT LICENSE RIGHTS

[0002] This invention was made with government support under AH50826, and GM137795 awarded by the National Institutes of Health. The government has certain rights in the invention.FIELD OF THE DISCLOSURE

[0003] The present disclosure relates generally to techniques for multidrug therapy discovery and, more particularly, to techniques for predicting both potency and side effects for drug combinations.BACKGROUND

[0004] The background description provided herein is for the purpose of generally presenting the context of the disclosure. Work of the presently named inventors, to the extent it is described in this background section, as well as aspects of the description that may not otherwise qualify as prior art at the time of filing, are neither expressly nor impliedly admitted as prior art against the present disclosure.

[0005] Many drugs can operate in combination to synergistically defend against pathogens, cancers, and more. Unfortunately, with thousands of drugs available for use in combination, it is difficult to identify drug combinations that are synergistic for potency for any specific application. As such, machine learning approaches to predict synergistic drug combinations have recently been developed.

[0006] Such machine learning approaches vary in how the interactions between drugs are studied and quantified. However, any effective drug combination must also present minimal side effects to successfully address the problem without causing further harm, and thesePatent Application30275 / 70687 / PC machine learning approaches are not equipped to efficiently predict the side effects of drug combination therapies. Typically, the side effects, or toxicity, of a drug combination is assessed after determining a drug combination is synergistic for potency. As a result, the conventional approaches are time consuming and inefficient, thereby imparting an unnecessarily large computational load, as well as high costs, and also slowing the development of drug combination therapies.

[0007] There is a need for techniques that can determine safe and effective drug combinations while reducing time to development and required resources for determining such.SUMMARY OF THE INVENTION

[0008] The present techniques include methods and systems for efficiently predicting drug combination therapy candidates. The method may comprise: (1) receiving, via one or more processors and at a trained mechanistic model trained based on one or more drug profiles, an indication of a first drug; (2) analyzing, via the one or more processors and by the trained mechanistic model, a drug profile of the one or more drug profiles associated with the first drug and at least one additional drug profile of the one or more drug profiles, the at least one additional drug profile associated with one or more second drugs, to generate a plurality of joint profile features associated with one or more respective combinations of the first drug and each respective second drug; (3) generating, via the one or more processors and by the trained mechanistic model, a respective predicted drug combination outcome for each of the one or more respective combinations based on the plurality of joint profile features, the respective predicted drug combination outcome including a respective toxicity score and a respective potency score; and (4) determining, via the one or more processors and by the trained mechanistic model, one or more preferable drug combination candidates of the one or more respective combinations based on the respective toxicity score and the respective potency score for the each of the one or more respective combinations.

[0009] In an example, a system for efficiently predicting drug combination therapy candidates comprises: one or more processors; a trained mechanistic model trained based one or more drug profiles; and a non-transitory computer readable medium including computer executable instructions that, when executed by the one or more processors, causePatent Application30275 / 70687 / PC the computing system to: (1 ) receive, via the one or more processors and at the trained mechanistic model, an indication of a first drug; (2) analyze, via the one or more processors and by the trained mechanistic model, a drug profile of the one or more drug profiles associated with the first drug and at least one additional drug profile of the one or more drug profiles, the at least one additional drug profile associated with one or more second drugs, to generate a plurality of joint profile features associated with one or more respective combinations of the first drug and each respective second drug; (3) generate, via the one or more processors and by the trained mechanistic model, a respective predicted drug combination outcome for each of the one or more respective combinations based on the plurality of joint profile features, the respective predicted drug combination outcome including a respective toxicity score and a respective potency score; and (4) determine, via the one or more processors and by the trained mechanistic model, one or more preferable drug combination candidates of the one or more respective combinations based on the respective toxicity score and the respective potency score for the each of the one or more respective combinations.

[0010] In another example, a non-transitory computer-readable medium having stored thereon instructions that, when executed by one or more processors of a computing system, cause the computing system to: (1) receive, via the one or more processors and at a trained mechanistic model trained based on one or more drug profiles, an indication of a first drug; (2) analyze, via the one or more processors and by the trained mechanistic model, a drug profile of the one or more drug profiles associated with the first drug and at least one additional drug profile of the one or more drug profiles, the at least one additional drug profile associated with one or more second drugs, to generate a plurality of joint profile features associated with one or more respective combinations of the first drug and each respective second drug; (3) generate, via the one or more processors and by the trained mechanistic model, a respective predicted drug combination outcome for each of the one or more respective combinations based on the plurality of joint profile features, the respective predicted drug combination outcome including a respective toxicity score and a respective potency score; and (4) determine, via the one or more processors and by the trained mechanistic model, one or more drug combination candidates of the one or more respectivePatent Application30275 / 70687 / PC combinations based on the respective toxicity score and the respective potency score for the each of the one or more respective combinations.BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The patent or application file contains at least one drawing executed in color. Copies of this patent or patent application publication with color drawing(s) will be provided by the Office upon request and payment of the necessary fee.

[0012] The figures described below depict various aspects of the system and methods disclosed herein. It should be understood that each figure depicts an example of aspects of the present systems and methods.

[0013] FIG. 1 is an example block diagram of a computing system configured to implement methods for predicting efficacious drug combinations, as described herein.

[0014] FIG. 2A is an example block diagram of a methodology for processing input data with mechanistic biochemical network models for subsequently identifying efficacious drug combinations, as described herein.

[0015] FIG. 2B is an example block diagram of data integration and synthesis for predicting efficacious drug combinations, as described herein.

[0016] FIG. 2C is an example block diagram of a machine learning model architecture for predicting efficacious drug combinations, as described herein.

[0017] FIG. 2D is an example block diagram for training of a machine learning model, as described herein.

[0018] FIG. 2E is an example block diagram for implementation of trained machine learning models, as described herein.

[0019] FIG. 3A is an example block diagram of output processing for a machine learning model for predicting efficacious drug combinations and for interpreting the model for identifying underlying biochemical mechanisms, as described herein.

[0020] FIG. 3B is an example block diagram for visualizing importance and directionality of combination interactions in example bacteria for various subsystems, as described herein.Patent Application30275 / 70687 / PC

[0021] FIG. 4A is an example block diagram of adverse events data processing for predicting efficacious drug combinations, as described herein.

[0022] FIG. 4B is an example graphical analysis of predicted toxicity scores for an example bacterium, as described herein.

[0023] FIG. 4C is an example graphical analysis of predicted toxicity scores for an example bacterium, as described herein.

[0024] FIG. 5A is an example graphical analysis of predicted combination toxicity scores and combination interaction potency scores of pairwise drug combinations for an example bacterium, as described herein.

[0025] FIG. 5B is an example graphical analysis of predicted combination toxicity scores and combination potency interaction scores of pairwise drug combinations for an example bacterium as described herein.

[0026] FIG. 6A is an example block diagram of experimental validation stages for predicting efficacious drug combinations, as described herein.

[0027] FIG. 6B is an example graphical analysis of individual and combination drug therapies for two example bacteria, as described herein.

[0028] FIG. 6C is an example graphical analysis of individual and combination drug therapies and predicted toxicity for two example bacteria, as described herein.

[0029] FIG. 7 illustrates a flowchart depicting an exemplary method for predicting efficacious drug combinations, implemented in the system of FIG. 1.DETAILED DESCRIPTION

[0030] Although the following text sets forth a detailed description of numerous different embodiments, it should be understood that the legal scope of the invention is defined by the words of the claims set forth at the end of this patent. The detailed description is to be construed as exemplary only and does not describe every possible embodiment, as describing every possible embodiment would be impractical, if not impossible. One could implement numerous alternate embodiments, using either current technology or technologyPatent Application30275 / 70687 / PC developed after the filing date of this patent, which would still fall within the scope of the claims.

[0031] In the field of pharmacology, the challenge of identifying synergistic drug combinations that are both effective against diseases and present minimal side effects is a significant hurdle. The disclosed embodiments address this challenge by leveraging a trained mechanistic model, which, depending on the implementation, includes one or more trained machine learning models to predict effective drug combination therapy candidates with minimal toxicity. Some embodiments include analyzing drug profiles to generate joint profile features for drug combinations and predicting toxicity and potency scores for the drug combinations based on the joint profile features. The conventional techniques include identifying drug combination therapies based upon the potency of the drug combinations, and thereafter determining if the drug combination therapy is a suitable candidate by determining the toxicity of the combination. In such conventional techniques, the potency and toxicity of a drug combination were compared against each other, however, the disclosed embodiments, specifically the trained mechanistic model, eliminate the need for such analysis by providing insight, or an explanation, as to why a drug combination is toxic. In some implementations, one or more portions of the trained mechanistic model correspond to respective subsystems of a metabolic network of an organism, thereby creating a model architecture that reflects known structure of cells and providing insight into the mechanistic interactions between subsystems that explain why and how a drug combination is toxic. Moreover, the disclosed embodiments, can predict synergistic drug combinations that are both effective and produce minimal side effects. Further, the disclosed techniques provide both high quality predictions and insight into the underlying mechanism of a given drug combination. Finally, the disclosed method can identify mechanisms to reverse the toxicity of a drug of interest by combining it with a one or more second drugs.

[0032] Generally, the present techniques improve the identification of synergistic drug combinations, through processing and analyzing large quantities of drug interaction data and mechanism of action data. Additionally, a first machine learning model is trained on drug combination toxicity data, enabling the prediction of low toxicity drug combination outcomes with improved accuracy. Depending on the implementation, the drug combination toxicity data may be obtained from adverse events reporting system databases. In furtherPatent Application30275 / 70687 / PC implementations, the drug combination toxicity data may be obtained and / or retrieved from other sources in addition to and / or in place of the adverse events reporting system database(s). For example, the drug combination toxicity data may be obtained from and / or derived from in vitro experiments, in vivo experiments, in silico simulation data on organ exposure, and / or any other such sources.

[0033] The second machine learning model represents a plurality of subsystems of a metabolic network for an organism (e.g., a human), integrating constraints such as multi- omics data (e.g., genome sequences, protein sequences, chemogenomic data, transcriptomics data, metabolomics data, etc.) and growth rates (e.g., bacterial growth rate, cell growth rate, biomarker growth rate, etc.), which allows for analysis of drug interactions at a metabolic level. By incorporating models such as those described herein, the present techniques can not only predict the effectiveness of drug combinations, but can also assess their potential side effects, thereby addressing a significant gap in existing machine learning approaches. Moreover, this multifaceted approach ensures that the predicted drug combinations are not only potent against the target conditions but also minimize adverse effects, thereby increasing their safety for patients and decreasing the time to development.

[0034] In summary, the disclosed embodiments provide significant advancements in the prediction of synergistic drug combinations and a solution to the challenges of identifying effective and safe drug therapies.

[0035] Turning first to FIG. 1, a system 100 configured to implement methods for predicting efficacious drug combinations includes a computing device 102 and an adverse events reporting database 104. The computing device 102 includes one or more processors 110, a network interface controller (NIC) 120, a memory 130, and a user interface module 160 including one or more display s / screens 162 and one or more suitable types of input / output (VO) devices 164.

[0036] The computing device 102 may be an individual server, a group (e.g., cluster) of multiple servers, or another suitable type of computing device or system (e.g., a collection of computing resources). For example, the computing device 102 may be a personal computer, a server, a mobile computing device, a smart phone, a tablet, a laptop, etc. Additionally, the computing device 102 may be the property of a customer, a company, an organization, etc.Patent Application30275 / 70687 / PC

[0037] The computing device 102 may include one or more processors 1 10 and a memory 130. The processors 110 may include any suitable number of processors and / or processor types, such as CPUs and one or more graphics processing units (GPUs). Generally, the processors 110 are configured to execute software instructions stored in a memory (e.g., the memory 130). For example, one or more CPUs of the one or more processors 110 of the computing device 102 may be configured to execute software instructions in the memory 130 for implementing the method 700 of FIG. 7. As another example, one or more GPUs of the one or more processors 110 of the computing device 102 may be configured to train one or more machine learning models (e.g., the drug combination outcome model 142). The memory 130 may include one or more persistent memories (e.g., a hard drive / solid state memory) and may store one or more sets of computer executable instructions / modules, including a mechanistic model 140, a drug combination analysis module 150, and a validation module 152, as described in more detail below. In some implementations, the mechanistic model 140 includes a drug combination outcome model 142 and a model 144, as described in more detail below.

[0038] The network interface controller (NIC) 120 includes at least one wireless communication interface which includes hardware, firmware, and / or software that is generally configured to communicate with other devices and / or over a network using one or more wireless communication protocols. For example, the NIC 120 may be configured to transmit and receive data using a Bluetooth protocol, a Wi-Fi® (IEEE 802.11 standard) protocol, a near-field communication (NFC) protocol, a cellular protocol (such as global system for mobile communications (GSM), code-division multiple access (CDMA), longterm evolution (LTE), worldwide interoperability for microwave access (WiMAX), etc.), a peer-to-peer wireless protocol, a short-range wireless protocol, and / or other suitable wireless communication protocols. Additionally, although not shown in FIG. 1, it is understood that, in some implementations, NIC 120 may include one or more wired communication interfaces which may be utilized by the computing device 102 to communicatively connect to the adverse event reporting database 104, a network of computing devices, and / or other devices via one or more wired communication protocols and / or wired data protocols. In some embodiments, the NIC 120 may include one or more suitable NICs, such as wired / wireless controllers (e.g., Ethernet controllers), and facilitate bidirectional / multiplexedPatent Application30275 / 70687 / PC networking over a network between the computing device 102 and other components of the environment 100 (c.g., the adverse event reporting database 104, another computing device, a remote computing device, etc.). In some embodiments, the NIC 120 may include one or more transceivers to support various different wireless communication protocols; however, for ease of reading (and not limitation purposes) herein, the NIC 120 may be referred to herein using the singular tense.

[0039] The user interface module 160 may include one or more suitable types of displays / screens (e.g., displays / screens 162) and / or one or more suitable types of input / output (I / O) devices 164. The displays / screens 162 may use any suitable display technology (e.g., LED, OLED, LCD, etc.), and in some embodiments may be integrated with the I / O devices as a touchscreen display. The I / O devices 164 may include one or more suitable types of user input devices, such as keyboards, mice, touch pads, touch screen displays, microphones, and / or any suitable types of remote and / or local user input devices. Further, the I / O devices 164 may include one or suitable types of output devices, such as displays, speakers, and the like. The user interface module 160 may enable a user of the computing device 102 to interact with graphical user interfaces (GUIs) provided by computing device 102 (e.g., via the displays / screens 162 and the I / O devices 164). The user interface module 160 may include one or more local interfaces, and / or may include one or more remote interfaces that receive instructions from the computing device 102 and are communicatively connected to the computing device 102 via one or more wireless / wired connections (e.g., via the NIC 120) and / or over a network (e.g., that are provided by an application, web browser, or other software executing on a computing device of a user). For ease of reading (and not limitation) purposes, user interface module 160 may be referred to herein using the singular tense.

[0040] Returning to the memory 130, the mechanistic model 140 (e.g., including the drug combination outcome model 142 and the model 144), the drug combination analysis module 150, and the validation module 152 may, generally, include machine-readable instructions corresponding to, for example, the operations represented by the flowcharts of this disclosure (e.g., the flowchart 700 of FIG. 7).Patent Application30275 / 70687 / PC

[0041] The memory 130 may include instructions for receiving properties, identifiers, or indications of drugs and obtaining respective drug profiles for the received drug indications. The mechanistic model 140 may be trained to analyze the drug profiles and generate a plurality of joint profile features for various combinations of the drugs and their respective drug profiles. Additionally, the mechanistic model 140 may be trained to generate predicted drug combination outcomes for the various drug combinations, including respective drug combination toxicity scores and respective drug combination potency scores, and the memory 130 may include instructions for determining one or more drug combination candidates (e.g., efficacious drug combinations) based on the toxicity scores and potency scores. In some embodiments, the memory 130 may include instructions for determining, by the mechanistic model 140, a suitable drug combination candidate of one or more drug combination candidates for one or more of: an individual organism, a disease, a tissue type, and / or a cell type. In some embodiments, the memory 130 may include instructions for determining, by the mechanistic model 140, one or more second drugs that reduce toxicity of a first drug for one or more of: an individual organism, a disease, a tissue type, and / or a cell type. In some embodiments, the drug candidates can be existing chemical compounds or new chemical entities, metabolites, peptides, proteins, or biologies.

[0042] The model 144 may represent a plurality of subsystems of a metabolic network for an organism (e.g., a human) and may include one or more integrated constraints, such as, multi-omics data (e.g., genome sequences, protein sequences, chemogenomic data, transcriptomics data, metabolomics data, etc.), media conditions (e.g., growth medium, nutrients, growth factors, temperature, etc.), metabolites, metabolic inhibitors, microbiome composition, exchange rates (e.g., exchange rate across a biological membrane, drug absorption and distribution, binding affinity, etc.), growth rates (e.g., bacterial growth rate, cell growth rate, biomarker growth rate, etc.), drug dosing, time between treatments, etc. In some embodiments, the model 144 may generate one or more respective sets of joint profile features, each corresponding to a respective subsystem represented by the model 144. Moreover, the mechanistic model 140 may generate the plurality of joint profile features for a drug combination using the model 144. In some embodiments, the model 144 may be a genome scale metabolic model (GEM). Additionally and / or alternatively, the model 144 (e.g., the GEM) may include one or more machine learning algorithms and / or models. InPatent Application30275 / 70687 / PC some embodiments, the model 144 may include other networks, such as signaling networks, gene regulatory networks, protein-protein interaction networks. In further embodiments, the model 144 may include other techniques, such as non-machine learning focused techniques.

[0043] The drug combination outcome model 142, and more generally the mechanistic model 140, may be trained on a plurality of drug profiles and drug combination toxicity data (e.g., data from the adverse events reporting database 104). In further implementations, the drug combination toxicity data may be obtained and / or retrieved from other sources in addition to and / or in place of the adverse events reporting system database(s). For example, the drug combination toxicity data may be obtained from and / or derived from in vitro experiments, in vivo experiments, in silico simulation data on organ exposure, and / or any other such sources. As such, it will be understood that, although the disclosure herein may refer to an adverse events reporting database 104, other sources may be envisioned in addition to and / or in place of the adverse events reporting database 104, except where such an interpretation would be contradictory.

[0044] In some embodiments, the drug combination outcome model 142 may be a trained artificial neural network. The mechanistic model 140 may process the plurality of joint profile features for a drug combination generated by the model 144 using the drug combination outcome model 142 to generate the predicted drug combination outcomes for the drug combinations. Moreover, the drug combination outcome model 142 may be trained to generate the drug combination toxicity scores and the drug combination potency scores. In some embodiments, the mechanistic model 140 may be trained to generate predicted toxicity scores for a drug combination based on a plurality of joint profile features for the drug combination. In some embodiments, the mechanistic model 140 may be trained to generate predicted potency scores for a drug combination based on a plurality of joint profile features for the drug combination. Moreover, the mechanistic model 140 may include two distinct drug combination outcome algorithms: a drug combination toxicity algorithm and a drug combination potency algorithm.

[0045] For example, the drug combination toxicity algorithm and the drug combination potency algorithm may be called separately (e.g., by one or more API) to generate a predicted toxicity score and / or generate a predicted potency score. In some embodiments, aPatent Application30275 / 70687 / PC predicted toxicity score generated by the drug combination toxicity algorithm and a predicted potency score generated by the drug combination potency algorithm may be aggregated into a single predicted drug combination outcome score. In some embodiments, a single predicted score may be generated by the drug combination toxicity algorithm or the drug combination potency algorithm. For example, a predicted toxicity score generated by the drug combination toxicity algorithm may exceed a toxicity threshold, and generating a predicted potency score may be unnecessary. As another example, a predicted potency score generated by the drug combination potency algorithm may not exceed a potency threshold, and generating a predicted toxicity score may be unnecessary. Although the drug combination toxicity algorithm and the drug combination potency algorithm are described with respect to two separate algorithms, it would be understood that a single algorithm (e.g., the drug combination outcome model 142) may also be envisioned. As shown in FIG. 3A and FIG. 3B, the mechanistic model 140 can be directly interpreted using the weights of the model to identify features of the model that are strongly associated with either toxicity or potency. Further, such interpretability may allow toxicity and potency to be attributed to particular subsystems of the metabolic network represented by the mechanistic model 140 (e.g., subsystems 1-n represented by the model 200c of FIG. 2C).

[0046] In some embodiments, the memory 130 may include instructions for receiving an identification or indication of a first drug and a second drug, and generating, by the mechanistic model 140, a drug combination toxicity score and drug combination potency score for the first drug and the second drug. Moreover, the mechanistic model 140 may be trained to analyze two provided drugs, and determine an efficacy (e.g., based on the drug combination toxicity and potency scores) of the combination of the two provided drugs. In further embodiments, the memory 130 may include instructions for receiving an indication of a drug, and generating, by the mechanistic model 140, drug combination toxicity scores and drug combination potency scores for a combination of the drug and one or more respective drugs of a plurality of drugs (e.g., a plurality of drugs the mechanistic model 140 has been trained on / for). Moreover, the mechanistic model 140 may be trained to analyze a provided drug against a plurality of other known drugs, and determine a plurality of potential drug combination candidates for the provided drug. In yet further embodiments, the memory 130 may include instructions for receiving an indication of a medical conditionPatent Application30275 / 70687 / PC(e.g., a bacterial infection, a disease, a cancer, and the like) and determining, by the mechanistic model 140, a plurality of potential drug combination candidates for the medical condition. In still yet further embodiments, the memory 130 may include instructions for receiving an indication of a patient profile (e.g., patient demographic data, genome data, medical history data, etc.), a medical condition of the patient, properties of a pathogen, properties of a cell line, or properties of an animal, and the memory 130 may include instructions for screening a plurality of potential drug combination candidates determined by the mechanistic model 140 based on the received indications.

[0047] The drug combination analysis module 150 may include instructions for determining drug combination candidates, or efficacious drug combinations, based on the drug combinations outcomes generated by the drug combination outcome model 142. In some embodiments, the drug combination analysis module 150 may include toxicity score thresholds and potency score thresholds, and the drug combination analysis module 150 may include instructions for comparing the toxicity scores and potency scores generated by the drug combination outcome model 142 for each drug combination to such thresholds. In some embodiments, the drug combination analysis module 150 may include instructions for computing a Spearman rank correlation (e.g., a non-parametric measure of rank correlation), or another statistical correlation metric, between experimental and predicted drug-interaction scores (e.g., a measure of whether a drug combination is synergistic, antagonistic, additive, etc., based on some combination of the toxicity and potency of a drug combination) when testing / validating the drug combination outcome model 142 on new data.

[0048] The validation module 152 may include instructions for determining the efficacy of a predicted drug combination based upon a comparison of experimental validation data for a drug combination and predicted drug combination outcomes generated by the drug combination outcome model 142. Cell viability data for individual drug and combination drug therapies may be determined using various experimental validation techniques. For example, one such technique may include cell plating and incubation, drug treatment, incubation of cells post-treatment, luciferin treatment, and subsequently taking a luminescence reading to determine cell viability for individual drug therapies and for combination drug therapies. The validation module 152 may include instructions for using the cell viability data for individual and combination drug therapies to compute a Bliss scorePatent Application30275 / 70687 / PC for each drug combination, the Bliss score used to evaluate interactions between drug combinations and representative of the experimentally determined safety of a drug combination (e.g., a lower Bliss score meaning the drug combination is less safe). The Bliss score for each drug combination may then be compared to the drug combination outcomes (e.g., a predicted toxicity score, a predicted potency score, a predicted interaction score, and / or some combination thereof) generated by the drug combination outcome model 142. Based on the comparison, drug combination candidates may be experimentally validated.

[0049] The adverse events reporting database 104 may store data / information for individual drugs and drug combinations and respective associated side effects. In some embodiments, the data stored on the adverse events reporting database 104 may be used to calculate or determine (e.g., using the drug combination outcome model 142) toxicity scores for a drug combination or an individual drug. For example, the data on drug combinations and their associated side effects may be used to generate a proportional reporting ratio (PRR) that quantifies the chances of occurrence of side effects due to a particular treatment in comparison to random chance. In some embodiments, the drug combination outcome model 142 may be trained on the PRR for a plurality of drugs and drug combinations. In exemplary embodiments, the adverse events reporting database 104 may store data for drug combinations of two drugs. In some embodiments, the adverse events reporting database 104 may store corresponding toxicity scores of the individual drugs and drug combinations. Regardless, it will be understood that the mechanistic model 140 may be trained on such data to generate predicted toxicity scores, predicted potency scores, and / or predicted interaction scores, for drug combinations of more than two drugs.

[0050] In some embodiments, at least a portion of the data stored in the adverse events reporting database 104 may be obtained / collected from a public or private adverse events reporting database, such as the food and drug administration (FDA) adverse events reporting system (FAERS), the world health organization vigilant database (e.g., VigiBase), an insurance claims database, a health records database, etc. In some embodiments, the data stored on the adverse event reporting database 104 may be obtained using an application programming interface (API). The API may employ a database query to directly request and retrieve data, utilize a file transfer protocol (FTP) for bulk downloading of data files, make a web service call to interact with web-based services for data extraction, etc. As anotherPatent Application 30275 / 70687 / PC example, the data stored on the adverse event reporting database 104 may be obtained by scraping web pages, subscribing to data feeds, using cloud storage access protocols, or other mechanisms that can be employed to gather the necessary data from such public adverse events reporting databases.

[0051] FIG. 2A illustrates an example block diagram of metabolic network modeling using flux balance analysis 200a for predicting efficacious drug combinations. At a high level, FIG. 2A depicts a methodology for processing input data with mechanistic biochemical network models for subsequently identifying drug combinations. The flux balance analysis 200a includes developing a model system 202, a mathematical representation 204, problem formulation 206, and a flux solution 208.

[0052] The model system 202 defines all metabolic reactions and metabolites for a metabolic network or a subsystem of a metabolic network. Moreover, the model system 202 may, in some embodiments, instead be representative of a model subsystem of a larger metabolic system or other cellular network system for an organism. The model system 202 shows a simplified example that includes three metabolites (e.g., metabolite A, metabolite B, and metabolite C) with reaction vl, reaction v4, reversible reactions v2 and v3, and three exchange fluxes bl, b2, and b3. The mathematical representation 204 corresponds to mass balance equations, or a set of differential equations, accounting for the reactions and transport mechanisms (e.g., reactions vl -v4 and exchange fluxes bl -b3) of the model system 202, or a model subsystem. The mathematical representation 204 includes the set of differential equations, or mass balance equations, in vector form (e.g., left side) and in matrix form (e.g., right side). For example, in the exemplary embodiment of FIG. 2A, the mathematical representation 204 may be

[0053] The problem formulation 206 includes a steady state solution of the set of differential equations from the mathematical representation 204 using various integratedPatent Application30275 / 70687 / PC constraints (e.g., exchange rates, growth rates, omics data, regulatory, environmental, thermodynamic, mass-balance, energy-balance, metabolic inhibitors, microbiomc composition, metabolites, drug dosing, time of treatments, etc.) such that S = b, where S is the matrix form of the mathematical representation 204 and v is a vector of fluxes corresponding to the reactions and exchange fluxes of the model system 202, and where b is the rate of change of metabolites, which can be set to zero to imply steady state. Using the steady state solution of the mathematical representation 204 an objective function can be solved to obtain the flux solution 208 which simulates reaction fluxes.

[0054] A genome-scale metabolic model (GEM) (e.g., the model 144 of FIG. 1) is a computational representation of an organism’s metabolic network (e.g., a computational representation of the model system 202) or subsystems of metabolic network, and, similar to the mathematical representation 204, is based on an annotated array of reactions and metabolites. Moreover, GEMs facilitate the determination of an objective function (e.g., such as a maximum and / or maximized growth) by integrating different constraints (e.g., problem formulation 206; multi-omics data) to simulate reaction fluxes(e.g., flux solution 208).

[0055] FIG. 2B illustrates an example block diagram of data integration 200b for predicting efficacious drug combinations. Data integration 200b includes constraints 210, a computational representation of a metabolic network 212 (e.g., the model 144 of FIG. 1), reaction fluxes 214, and joint profile feature 216. Constraints 210 include multi-omics data, such as genome sequences, chemogenomic data, transcriptomics data, metabolomics data, protein sequences, etc. for various drugs and media conditions (e.g., growth medium, nutrients, growth factors, temperature, pH, etc.).As part of data integration 200b, the system integrates constraints 210 to the computational representation of the metabolic network 212 to simulate reaction fluxes and solve an objective function to obtain the reaction fluxes 214. Further as pail of the data integration 200b, the system generates the joint profile features 216 based on the reaction fluxes 214 and processes the joint profile features 216 for input to a machine learning model (e.g., the drug combination outcome model 142 of FIG. 1).

[0056] FIG. 2C illustrates an example block diagram of a machine learning (ME) model 200c for predicting efficacious drug combinations. The ME model 200c (e.g., the drugPatent Application30275 / 70687 / PC combination outcome model 142 of FIG. 1) corresponds to various subsystems 1 to n (e.g., model system 202 of FIG. 2A) that form a metabolic network. The ML model 200c may include an input layer 220, one or more hidden layers (e.g., first hidden layer 222 and second hidden layer 224), and an output layer 226. In some embodiments, each of the input layer 220, the first hidden layer 222, and the second hidden layer 224, may include one or more sections, each corresponding to a subsystem of subsystems 1 to n. Additionally, the ML model 200c may include a concatenation layer 228a for converting the multidimensional output of the final hidden layer (e.g., hidden layer 224) into a output 226a. Furthermore, the nodes and the layers may be partially interconnected between subsystems 1 to n to represent cross-interactions between subsystems.

[0057] Joint profile features 229 (e.g., generated by the model 144 of FIG. 1; the joint profile features 216 of FIG. 2B) corresponding to each of subsystems 1 to n are fed in (e.g., by the computing device 102 of FIG. 1), to the input layer 220, as separate inputs to their corresponding subsystem. For example, the dimensionality of the input layer, or number of nodes in the input layer, corresponds to the dimensionality of joint profile features 229 of each subsystem. These inputs, or joint profile features, are fed in, by the input layer, to the first hidden layer 222 for processing, and subsequently, the outputs of the first hidden layer 222 are fed in to the second hidden layer 224. For example, the first hidden layer 222 and the second hidden layer 224 may process data by computing a weighted sum of the inputs from the previous layer, each node in a hidden layer corresponding to a set of weights used to weight each input element from the previous layer (e.g., a weight for each node of the previous layer) and a bias term that is added to each term of the sum, the weighted sum of each node is then passed to an activation function for the node, such as a Sigmoid function, a Rectified Linear Unit (ReLU), a Softmax function, etc. The output of the activation function for each node of a hidden layer is subsequently passed to next layer in the ML model 200c. Additionally, it should be understood that the ML model 200c may include any suitable number of hidden layers (e.g., one or more additional hidden layers after the second hidden layer 224). The concatenation layer 228a processes the outputs of each of subsystems 1 to n, from the final hidden layer, to combine the outputs into an output layer 226a, or score 226a, for a drug combination (e.g., a prediction of a drug combination outcome; a toxicity scorePatent Application30275 / 70687 / PC and / or a potency score). In some embodiments, the ML model 200c may be a random forest model, a support vector machine, an artificial neural network, or another type of ML model.

[0058] In some embodiments, underlying mechanisms of a metabolic network may be identified / determined using the ML model 200c. For example, the ML model 200c corresponds to various subsystems 1 to n, as mentioned above, and a toxicity score and / or potency score for a drug combination may be attributed to a particular subsystem. Each subsystem of the ML model 200c may be representative of a cellular network system of a greater metabolic network, and depending on the subsystem to which a toxicity score and / or potency score are attributed, underlying mechanisms associated with the toxicity score and / or potency score may be determined based on the weights of the ML model 200c. For example, the underlying mechanisms may include, biochemical pathways, molecular factors, etc.

[0059] FIG. 2D illustrates an example block diagram for training 200d of a machine learning (ML) model 230. The ML model 230 (e.g., the ML model 200c of FIG. 2C, and / or the drug combination outcome model 142 of FIG. 1) may include an input layer 232, one or more hidden layers 234, and an output layer 236. Additionally, in some embodiments, the ML model 230 may include a concatenation layer (not depicted) for generating linear or non-linear outputs. A training dataset 240 (e.g., drug combination toxicity data stored in the adverse event reporting database 104 of FIG. 1) includes training data 242 (e.g., proportional reporting ratio for adverse events for the drug combination toxicity data) and training data labels 244 (e.g., toxicity scores for the drug combination toxicity data). Training data on drug combination toxicity may also be generated from high throughput cell line screening, mining of health records, or mining of insurance claims. In some embodiments, the training data may also be generated in silica using generative models or genome-scale network modeling methods. The training data 242 may be vectorized (e.g., via an embedding layer / model) into a training vector 246 for input to the ML model 230. The ML model 230 may process the training vector 246 (e.g., the training vector 246 may be fed in to the input layer 232, processed by the hidden layers 234, and output by the output layer 236) to generate model output 250, and a training application 260 may compare the model output 250 to training data labels 244, or labelled training data 244, to generate an error value via, for example, an error function. Based on the error value for the model output 250, thePatent Application30275 / 70687 / PC training application 260 may adjust weights and / or bias terms of the ML model 230 (e.g., the weights and bias term for each node of the one or more hidden layers 234). For example, the training application 260 may adjust one or more weights and / or bias terms of one or more of the hidden layers 234 (e.g., line 262). In some embodiments, the ML model 230 may be or include an artificial neural network (ANN).

[0060] FIG. 2E illustrates an example block diagram for a method 200e of implementing a trained machine learning (ML) model 270 and the trained ML model 200c of FIG. 2C (e.g., the drug combination outcome model 142 of FIG. 1). In some embodiments, the method 200e may include generation of joint profile features 280 (e.g., similar to joint profile feature 216 processing of FIG. 2A). The trained ML model 270 may be trained on a plurality of drug-protein interactions to predict / generate drug-protein interactions scores for individual drugs and / or combinations of drugs. A first drug 272a and a second drug 272b corresponding to respective simplified molecular input line entry system (SMILES) notation(s) 274 (e.g., SMILES notation 274a and SMILES notation 274b) may be converted (e.g., via an encoding mechanism / conversion) to molecular access system (MACCS) keys 276 or any suitable encoding notation via any suitable encoding mechanism. While SMILES notation(s) 274 are explicitly discussed with regal'd to FIG. 2E, it will be understood that other alphanumeric representations may be used in addition to and / or in place of the SMILES notation(s) 274. For example, the drugs 272 may correspond to an international chemical identifier (InChi), InChi key, and / or any other such relevant methodology. In some such implementations, the corresponding alphanumeric representation may include stereoisomeric information and may be unique to each drug. In some such implementations, a particular drug may have multiple SMILES notations, particularly in cases where the drug has stereoisomeric information available. As such, InChi and InChi keys may provide preferable alternatives to SMILES notations. Depending on the implementation, InChi keys may be hashed values and have fixed lengths irrespective of a corresponding drug, while InChi notation may have variable lengths (e.g., based on the drug).

[0061] SMILES notation may use short ASCII (American Standard Code for Information Interchange) strings to define a chemical structure, whereas the MACCS keys are binary representations defining the presence of substructures of a chemical structure. For example,Patent Application30275 / 70687 / PC the encoding mechanism / conversion may be a binary encoding mechanism or a more complex encoding such as an encoding mechanism based on graph networks. The system may generate joint profile features 280 by inputting the MACCS keys 276 for the first drug 272a and the second drug 272b to the trained ML model 270 to generate: a drug-protein interaction profile 282a for the first drug 272a and a drug-protein interaction profile 282b for the second drug 272b. The ML model 270 may then determine a sigma score 284 (e.g., denoting similarity between the drug-protein interaction profiles 282a and 282b and / or a combined effect of the drugs in the combination) and / or a delta score 286 (e.g., denoting difference between the drug-protein interaction profile 282a and the drug-protein interaction profile 282b). The ML model 270 then concatenates the sigma score 284 and the delta score 286 and inputs the concatenation to the trained ML model 200c to generate a proportional reporting ratio (PRR) score indicative of the chances of occurrence of a side effect, or adverse event, from a combination of the first drug 272a and the second drug 272b and / or indicative of a toxicity of the combination. In some embodiments, similar to the trained ML model 200c, the trained ML model 270 may be a trained artificial neural network.

[0062] In some implementations, the ML model 270 may receive, in additional to and / or alternatively to the MACCS keys 276, other encoding mechanisms representative of the drugs 272. For example, the ML model 270 may receive Morgan fingerprints, Pubchem structural features, and / or any other such encoding mechanism outputs. As such, it will be understood that, although FIG. 2E refers to MACCS keys 276, that other encoding mechanisms are similarly envisioned.

[0063] In some implementations, rather than using only an encoded representation of the drug(s) 272 (e.g., using SMILES notation 274 and / or MACCS keys 276), the ML model 270 may additionally or alternatively use other techniques for profiling the drug(s) 272. For example, the ML model 270 may receive drug -protein interactions from literature sources, drug-protein predictions from computation models (not shown), reaction fluxes from mechanistic models (not shown) (e.g., genome-scale ecosystem models (GEMs)), omics data, gene-expression profiles, etc. In further implementations, the system may represent the drug(s) via a numerical representation rather than text to improve the ability for the ML model 270 to understand and / or process the input.Patent Application30275 / 70687 / PC

[0064] FIG. 3 A illustrates an example block diagram of output processing 300a for a machine learning (ML) model for predicting efficacious drug combinations. Output processing 300a includes comparing the performance and hidden features of the ML model for: (1) processing the joint profile features using all of the subsystems, subsystems 1 to n, and (2) processing the joint profile features using subsystems 1 to n-1. As depicted in FIG. 2C, for all subsystems contributing, the output / score 226a is generated by the subsystems of the ML model feeding into a concatenation layer 228a. For subsystems 1 to n-1 contributing, the output I score 226b is similarly generated by the concatenation layer 228b. The outputs 226a and 226b are respectively used to generate performance metrics 302a and 302b.

[0065] The output processing 300a includes determining a performance metric of the ML model 304, a directionality of the outputs 306, and subsystem knockoff impact 308.Moreover, the performance metric 302a and the performance metric 302b are compared (e.g., by computing their ratio) to determine the performance of the ML model 304, the score 226a and the score 226b (e.g., toxicity scores for a particular drug combination) are compared to determine the directionality of the outputs 306, and hidden feature(s) 308a and hidden feature(s) 308b are compared (e.g., by taking the square of the difference of 308a and 308b) to determine the subsystem knockoff impact 308 that measures the difference or impact due to the exclusion of subsystem n-1 and provides for an analysis of the internal representations or learned patterns of the ML model. By comparing the outputs of the ML model as such, with and without subsystem n, one or more subsystems with a significant influence on the predictions of the ML model can be identified. In some embodiments, the ML model (e.g., the drug combination outcome model 142 of FIG. 1, the ML model 200c of FIG. 2C, and / or the ML model 230 of FIG. 2D) may be a random forest model, a support vector machine, an artificial neural network, or another type of ML model.

[0066] In further implementations, the ML model may generate the hidden features (e.g., hidden feature 308a and / or hidden feature 308b) by performing a weight analysis. As described herein. In some such implementations, the machine learning model may perform a pathway weight analysis 310 by examining weights assigned to individual pathways for predicting potency and toxicity. In further implementations, the machine learning model may perform a combination weight analysis 312 by assessing contributions of pathways toPatent Application30275 / 70687 / PC the top and bottom combinations (e.g., the top / bottom 20 potent combinations, the top / bottom 10 toxic combinations, the top / bottom 5 potent and toxic combinations, etc.).

[0067] FIG. 3B illustrates an example diagram for visualizing importance and directionality of combination interactions 300b in example bacteria for various subsystems. The importance and directionality of combination interactions 300b is visualized as a bubble plot, more specifically, the bubble plot visualizes the importance and directionality of different subsystems for combination interactions 300b (e.g., combination toxicity and potency) for two example bacteria, Escherichia coli (E. coli) and Mycobacterium tuberculosis (M. tuberculosis). Each bubble corresponds to a subsystem (e.g., alternate carbon metabolism, cell envelope biosynthesis, etc.) within the metabolic network of an organism and their respective interaction with a drug combination. The size of each bubble corresponds to the relative importance of the subsystem in the context of drug combination interactions and / or toxicity (e.g., based on the subsystem knockoff impact 308 of FIG. 3A), with larger bubbles corresponding to greater importance. The color of each bubble corresponds to the directionality of the interaction / toxicity ranging from negative interactions to more positive interactions (e.g., based on the directionality of the outputs 306 of FIG. 3A). The exemplary bubble plot of FIG. 3B includes rows for: combination toxicity in humans, combination interaction in M. tuberculosis, and combination interactions in E. coli. It should be understood that the example bacteria M. tuberculosis and E. coli are provided for exemplary use only. Moreover, the disclosed embodiments are not limited to these two example bacterium and may be applied to other bacteria, viruses, fungi, cancer cells, illnesses, animal cells, etc.

[0068] FIG. 4A illustrates an example block diagram of adverse events data processing 400a for predicting efficacious drug combinations. Data processing 400a includes adverse event reporting system (AERS) database 402 (e.g., database corresponding to the Food and Drug Administration adverse event reporting system, or FDA AERS; or the adverse event reporting database 104 of FIG. 1), a contingency table 404, and proportional reporting ratio (PRR) score 406. The AERS database 402 stores data / information related to drugs and associated harmful effects, or potentially harmful effects, (e.g., drug side effects data). The data from the AERS database 402 may be processed as depicted in the contingency table 404 to generate PRR scores 406 for various drugs and drug combinations. The X-axis of thePatent Application30275 / 70687 / PC contingency table 404 represents combination drug treatment or exposure, and the Y-axis represents outcomes (c.g., a harmful effect from exposure to the drug combinations). The PRR scores 406 is a complex rational expression that may be computed as PRR = [A / (A+B)] / [C / (C+D)], with A representing the number of cases where an outcome was seen when exposed to a drug or drug combination, B representing the number of cases where an outcome was not seen when exposed to the drug, C representing the number of cases where an outcome was seen when not exposed to the drug, and D representing the number of cases when an outcome was not seen when not exposed to the drug, obtained based on random sampling of the whole study population. The PRR score 406 of a drug quantifies the chances of occurrence of side effects due to a particular treatment in comparison to random chance. More specifically, the PRR is computed as the ratio of the fraction of the number of cases where outcome was seen when exposed to a drug and the fraction of the number of cases where outcome was seen when not exposed to the drug, and generally, the PRR score 406 measures relative toxicity. The PRR scores 406 may be used to train the machine learning models, or artificial neural networks (e.g., the drug combination outcome model 142 of FIG. 1), described herein.

[0069] FIG. 4B illustrates an example graphical analysis 400b of predicted toxicity scores for an example bacterium. Similarly, FIG. 4C illustrates an example graphical analysis 400c of predicted toxicity scores for an example bacterium (e.g., predicted toxicity scores generated: by the drug combination outcome model 142 of FIG. 1, the ML model 200c of FIG. 2C, and / or using the method 200e of FIG. 2E). The example bacterium of FIG. 4B is Escherichia coli (E. coh). and the example bacterium of FIG. 4C is Mycobacterium tuberculosis (M. tuberculosis'). The predicted toxicity scores (e.g., the PRR scores 406 of FIG. 4A, or a toxicity score based on the PRR scores 406) are plotted against toxicity scores included in the AERS database 402 of FIG. 4A. N represents the number of interactions in a dataset, and R represents the Spearman rank correlation.

[0070] FIG. 5A illustrates an example graphical analysis 500a of predicted combination toxicity scores and combination interaction scores of pairwise drug combinations for an example bacterium. Similarly, FIG. 5B illustrates an example graphical analysis 500b of predicted combination toxicity scores and combination interaction scores of pairwise drug combinations for an example bacterium (e.g., predicted toxicity scores and combinationPatent Application30275 / 70687 / PC interaction scores generated: by the drug combination outcome model 142 of FIG. 1 , the ML model 200c of FIG. 2C, and / or using the method 200c of FIG. 2E). The drug combination interaction scores are indicative of whether a drug combination is synergistic, antagonistic, additive, etc. The example bacterium of FIG. 5A is Escherichia coli (E. coli). and the example bacterium of FIG. 5B is Mycobacterium tuberculosis (M. tuberculosis). The predicted combination toxicity scores for each example bacterium are plotted against the predicted combination interaction, or potency, scores for each example bacterium. Each point plotted on the graphs of FIG. 5A and FIG. 5B represents a specific pairwise drug combination (e.g., Amoxicillin and Vancomycin). FIG. 5A and FIG. 5B are graphical representations of the drug combination landscape for the respective example bacterium, E. coli and M. tuberculosis. Drug combinations 502, combinations 502a and combination 502b of FIG. 5A and FIG. 5B respectively, are chosen for experimental testing for their safety in vitro. Drug combinations 504, combinations 504a and combination 504b of FIG. 5A and FIG. 5B respectively, are other safe and effective predicted drug combinations (e.g., Amoxicillin and Clarithromycin). It should be understood that the drug combinations 502 and the drug combinations 504 are exemplary, and the ML models of the disclosed embodiments may analyze any known / available drug combinations and generate respective outcome scores.

[0071] FIG. 6A illustrates an example block diagram of experimental validation stages 600a for predicting efficacious and safe drug combinations. The validation stages 600a includes cell plating and incubation 602, drug treatment 604, incubation of cells posttreatment 606, luciferin treatment 608, and luminescence reading 610. For example, a cell viability assay (e.g., CellTiter-Glo Cell Viability Assay) is used for performing combination cytotoxicity experiments(e.g., cell plating and incubation 602, drug treatment 604, incubation of cells post-treatment 606, luciferin treatment 608) in mammalian cells (e.g., mammalian HEK293 and HEPG2 cells), and cell viability is measured using luminescence (e.g., luminescence reading 610). The validation stages 600a may be used to determine the efficacy of a particular drug combination candidate that has been identified using the disclosed techniques, moreover, the validation stages 600a may be used to: verify outputs of the ML models of the disclosed embodiments, validate a drug combination candidate for additional testing, etc. Other methods for validation include use of animal models, such asPatent Application30275 / 70687 / PC mouse models, and determining biomarkers of toxicity for specific tissues. Other validation modes may also include cell line, tissue, or organoid models in culture and toxicity measurements such as molecular read outs or imaging readouts.

[0072] FIG. 6B illustrates an example graphical analysis 600b of individual and combination drug therapies for two example bacteria. The example bacteria are Escherichia coli (E. coll) and Mycobacterium tuberculosis (M. tuberculosis). Information about cell viability in the presence of predicted drug combination treatments chosen for experimental testing (e.g., drug combinations 502a and 502b of FIG. 5A and FIG. 5B respectively) and the associated individual drug combination treatments are used to compute the Bliss scores 612 (e.g., a lower value for a Bliss toxicity score corresponding to a less safe drug combination). Other statistical methods for quantifying drug interactions, such as a Loewe model or fractional inhibitory concentration (FIC) models and highest single agent (HSA) models, may also be used. The combination toxicity scores (CTS) 614 are predicted toxicity scores (e.g., predicted toxicity scores generated: by the drug combination outcome model 142 of FIG. 1, the ML model 200c of FIG. 2C, and / or using the method 200e of FIG. 2E) for each drug combination used to compute the Bliss scores 612. FIG. 6C illustrates an example graphical analysis 600c of individual and combination drug therapies and predicted toxicity for the two example bacteria. Moreover, FIG. 6C depicts the CTS 614 plotted against the Bliss scores 612 for each drug combination.

[0073] FIG. 7 illustrates an example computer-implemented method 700 for predicting efficacious drug combinations, implemented by a computing system (e.g., system 100). It will be understood that the method 700 may be performed by other systems and / or components thereof (e.g., other systems configured to perform similar operations, and / or components thereof).

[0074] At block 702, an indication or identification of a first drug is received at a trained mechanistic model. In some embodiments, the trained mechanistic model (e.g., the mechanistic model 140 of FIG. 1) trained based on one or more drug profiles. In some embodiments, the trained mechanistic model includes a trained machine learning (ML) model and / or a trained artificial neural network (ANN), such as the drug combination outcome model 142 of FIG. 1. Additionally, the method 700 may include training the MLPatent Application30275 / 70687 / PC model (e.g., an ANN) based on drug combination toxicity data (e.g., data from the adverse events reporting database 104 of FIG. 1). In some embodiments, the trained mechanistic model may include a genome scale metabolic model (GEM) representing a plurality of subsystems of a metabolic network for an organism (e.g., a human) and including one or more integrated constraints. For example, the one or more integrated constraints may include at least one of: multi-omics data, media conditions, metabolites, exchange rates, growth rates, drug dosing, time of treatments, metabolic inhibitors, microbiome composition, or the like. In some embodiments, the method 700 may include determining one or more underlying mechanisms associated with the respective toxicity scores and the respective potency scores for the one or more respective combinations, by the trained mechanistic model, wherein the one or more underlying mechanisms include one or more of: biochemical pathways, and / or molecular factors. In some embodiments, the method 700 may include determining, by the trained mechanistic model, a suitable drug combination candidate of the one or more drug combination candidates for one or more of: an individual organism, a disease, a tissue type, or a cell type.

[0075] For example, an indication of a drug may be a computational representation of the drug, the chemical structure of the drug, etc., such as a simplified molecular' input line entry system (SMIEES) notation, InChi notation, and / or InChi keys notation for the chemical structure of the drug. In some embodiments, the indication of the drug may be the name of the drug and a computational representation of the drug may be obtained. The mechanistic model may convert representation of the drug (e.g., SMIEES notation for the drug), via an encoding mechanism (e.g., a binary encoder), to a format more suitable for processing by the mechanistic model, such as molecular access system (MACCS) keys, Morgan fingerprints, Pubchem structural features, another binary chemical structure representation, etc.

[0076] At block 704, a drug profile associated with the first drug is analyzed to generate a plurality of joint profile features associated with one or more respective combinations of the first drug and one or more second drugs. More specifically and in some embodiments, a drug profile associated with the first drug and at least one additional drug profile associated with one or more second drugs are analyzed by the trained mechanistic model to generate a plurality of joint profile features associated with one or more respective combinations of thePatent Application30275 / 70687 / PC first drug and each respective second drug. As mentioned above, the mechanistic model may be trained on one or more drug profiles which may include the drug profile associated with the first drug and the at least one drug profile associated with the one or more second drugs. Moreover, the method 700 may include analyzing, by the trained mechanistic model, a drug profile of the one or more drug profiles associated with the first drug and at least one additional drug profile of the one or more drug profiles, the at least one additional drug profile associated with one or more second drugs, to generate a plurality of joint profile features associated with one or more respective combinations of the first drug and each respective second drug.

[0077] In some embodiments, the method 700 may include generating, by the GEM, the plurality of joint profile features corresponding to the one or more respective combinations of the first drug and the one or more second drugs. Additionally, the plurality of joint profile features may include a two or more sets of joint profile features, each set corresponding to a respective subsystem of the plurality of subsystems represented by the GEM.

[0078] At block 706, a predicted drug combination outcome is generated for each of the drug combinations, the predicted drug combination outcomes each including a respective toxicity score and a respective potency score. In some embodiments, the respective predicted drug combination outcome for each of the one or more respective drug combinations are generated by the trained mechanistic model based on the plurality of joint profile features. In some embodiments, the method 700 may include processing, by the trained ML model (e.g., an ANN), the plurality of joint profile features to generate the respective predicted drug combination outcome for the each of the one or more respective combinations.

[0079] At block 708, one or more drug combination candidates of the one or more drug combinations are determined based on the respective toxicity score and the respective potency score for each combination. In some embodiments, the one or more drug combination candidates of the one or more respective drug combinations are determined by the trained mechanistic model based on the respective toxicity score and the respective potency score for the each of the one or more respective drug combinations.Patent Application30275 / 70687 / PC

[0080] In some embodiments, the one or more second drugs may be a single second drug and the indication of the first drug is a first indication, and the method 700 may include receiving, at the trained mechanistic model, a second indication of the single second drug, and generating an output based on the one or more drug combination candidates and indicative of whether a combination of the first drug and the single second drug is safe for a patient to take. In some embodiments, the method 700 may include determining, by the trained mechanistic model, one or more second drugs that reduce the toxicity of the first drug for one or more of: an individual organism, a disease, a tissue type, or a cell type.

[0081] In some implementations, the trained mechanistic model may additionally or alternatively receive information specific to an organ or tissue, such as omics data, geneexpression profiles, etc. Depending on the implementation, the trained mechanistic model may utilize such information to predict side effects of drug combinations specific to particular organs and / or tissues. For example, the trained mechanistic model may predict organ- specific and / or tissue-specific toxicity scores. Depending on the implementation, the method 700 may include using such predicted toxicity scores to (i) identify specific organs impacted by a combination, (ii) determine which combinations are less toxic to a particular organ or tissue, and / or (iii) generate one or more molecular-level insights regarding why certain combinations may have a particular toxicity level. In some such implementations, the trained mechanistic model determines toxicity of a particular combination relative to a current and / or selected treatment or standard of care.

[0082] In some embodiments, the method 700 may include generating an output based on the one or more drug combination candidates and indicative of which combinations of the one or more respective combinations are safe for a patient to take. In some embodiments, the output may be a list including the each of the one or more drug combination candidates and a respective toxicity score and a respective potency score for the each of the one or more drug combination candidates. For example, the list may be a graphical component of a graphical user interface, such as a chart, a table, a listing of the combination candidates, a graph, or some combination thereof.

[0083] In further implementations, the output may be or include one or more insights associated with why particular drug combination(s) vary in potency, toxicity, etc. based onPatent Application30275 / 70687 / PC association with various pathways, subsystems, and / or mechanisms (a nucleotide salvage pathway, for example). As such, the trained mechanistic model may generate an output indicative of the contribution and / or quantitative impact of each mechanism on a predicted potency and / or toxicity score of a particular drug combination.

[0084] The following list of examples reflects a variety of the embodiments explicitly contemplated by the present disclosure:

[0085] Example 1. A computer- implemented method for predicting drug combination therapy candidates, the method comprising: receiving, via one or more processors and at a trained mechanistic model trained based on one or more drug profiles, an indication of a first drug; analyzing, via the one or more processors and by the trained mechanistic model, a drug profile of the one or more drug profiles associated with the first drug and at least one additional drug profile of the one or more drug profiles, the at least one additional drug profile associated with one or more second drugs, to generate a plurality of joint profile features associated with one or more respective combinations of the first drug and each respective second drug; generating, via the one or more processors and by the trained mechanistic model, a respective predicted drug combination outcome for each of the one or more respective combinations based on the plurality of joint profile features, the respective predicted drug combination outcome including a respective toxicity score and a respective potency score; and determining, via the one or more processors and by the trained mechanistic model, one or more drug combination candidates of the one or more respective combinations based on the respective toxicity score and the respective potency score for the each of the one or more respective combinations.

[0086] Example 2. The computer-implemented method of example 1, wherein the trained mechanistic model includes an artificial neural network (ANN).

[0087] Example 3. The computer-implemented method of example 2, further comprising: processing, by the ANN, the plurality of joint profile features to generate the respective predicted drug combination outcome for the each of the one or more respective combinations.

[0088] Example 4. The computer-implemented method of example 1, further comprising: generating, by the trained mechanistic model, an indication of an impact on the respectivePatent Application30275 / 70687 / PC toxicity score or the respective potency score of at least one of: (i) a mechanism associated with the one or more drug combination candidates, (ii) a pathway associated with the one or more drug combination candidates, or (iii) a subsystem associated with the one or more drug combination candidates.

[0089] Example 5. The computer- implemented method of example 1, wherein the trained mechanistic model includes a genome scale metabolic model (GEM) representing a plurality of subsystems of a metabolic network for an organism and including one or more integrated constraints.

[0090] Example 6. The computer-implemented method of example 5, further comprising: generating, by the GEM, the plurality of joint profile features corresponding to the one or more respective combinations.

[0091] Example 7. The computer- implemented method of example 5, wherein the plurality of joint profile features includes a plurality of sets of joint profile features each corresponding to a respective subsystem of the plurality of subsystems represented by the GEM.

[0092] Example 8. The computer-implemented method of example 5, wherein the one or more integrated constraints include at least one of: multi-omics data, media conditions, metabolites, exchange rates, growth rates, drug dosing, or time of treatments.

[0093] Example 9. The computer-implemented method of example 1, wherein the one or more second drugs is a single second drug and the indication of the first drug is a first indication, the method further comprising: receiving, via the one or more processors and at the trained mechanistic model, a second indication of the single second drug; and generating, via the one or more processors, an output based on the one or more drug combination candidates and indicative of whether a combination of the first drug and the single second drug is safe for a patient to take.

[0094] Example 10. The computer-implemented method of example 1, further comprising: generating, via the one or more processors, an output based on the one or more drug combination candidates and indicative of which combinations of the one or more respective combinations are safe for a patient to take.Patent Application30275 / 70687 / PC

[0095] Example 1 1 . The computer-implemented method of example 10, wherein the output is a list including the each of the one or more drug combination candidates and a respective toxicity score and a respective potency score for the each of the one or more drug combination candidates.

[0096] Example 12. The computer-implemented method of example 1, further comprising: determining, via the one or more processors and by the trained mechanistic model, one or more underlying mechanisms associated with the respective toxicity scores and the respective potency scores for the one or more respective combinations, wherein the one or more underlying mechanisms include at least one of: a biochemical pathway, or a molecular factor.

[0097] Example 13. The computer-implemented method of example 1, further comprising: determining, via the one or more processors and by the trained mechanistic model, a drug combination candidate of the one or more drug combination candidates for one or more of: an individual organism, a disease, a tissue type, or a cell type.

[0098] Example 14. The computer-implemented method of example 1, further comprising: determining, via the one or more processors and by the trained mechanistic model, one or more second drugs that reduce toxicity of the first drug for one or more of: an individual organism, a disease, a tissue type, or a cell type.

[0099] Example 15. A computing system for predicting drug combination therapy candidates, the system comprising: one or more processors; a trained mechanistic model trained based one or more drug profiles; and a non-transitory computer readable medium including computer executable instructions that, when executed by the one or more processors, cause the computing system to: receive, via the one or more processors and at the trained mechanistic model, an indication of a first drug; analyze, via the one or more processors and by the trained mechanistic model, a drug profile of the one or more drug profiles associated with the first drug and at least one additional drug profile of the one or more drug profiles, the at least one additional drug profile associated with one or more second drugs, to generate a plurality of joint profile features associated with one or more respective combinations of the first drug and each respective second drug; generate, via the one or more processors and by the trained mechanistic model, a respective predicted drugPatent Application30275 / 70687 / PC combination outcome for each of the one or more respective combinations based on the plurality of joint profile features, the respective predicted drug combination outcome including a respective toxicity score and a respective potency score; and determine, via the one or more processors and by the trained mechanistic model, one or more drug combination candidates of the one or more respective combinations based on the respective toxicity score and the respective potency score for the each of the one or more respective combinations.

[0100] Example 16. The computing system of example 15, wherein the trained mechanistic model includes an artificial neural network (ANN).

[0101] Example 17. The computing system of example 16, further comprising instructions that, when executed by the one or more processors, cause the computing system to: train, via the one or more processors, the ANN based on drug combination toxicity data.

[0102] Example 18. The computing system of example 16, further comprising instructions that, when executed by the one or more processors, cause the computing system to: process, by the ANN, the plurality of joint profile features to generate the respective predicted drug combination outcome for the each of the one or more respective combinations.

[0103] Example 19. The computing system of example 15, wherein the trained mechanistic model includes a genome scale metabolic model (GEM) representing a plurality of subsystems of a metabolic network for an organism and including one or more integrated constraints.

[0104] Example 20. The computing system of example 19, further comprising instructions that, when executed by the one or more processors, cause the computing system to: generate, by the GEM, the plurality of joint profile features corresponding to the one or more respective combinations.

[0105] Example 21. The computing system of example 19, wherein the plurality of joint profile features includes a plurality of sets of joint profile features each corresponding to a respective subsystem of the plurality of subsystems represented by the GEM.

[0106] Example 22. The computing system of example 15, wherein: the one or more second drugs is a single second drug and the indication of the first drug is a first indication,Patent Application30275 / 70687 / PC and the computing system further comprising instructions that, when executed by the one or more processors, cause the computing system to: receive, via the one or more processors and at the trained mechanistic model, a second indication of the single second drug; and generate, via the one or more processors, an output based on the one or more drug combination candidates and indicative of whether a combination of the first drug and the single second drug is safe for a patient to take.

[0107] Example 23. A non-transitory computer- readable medium having stored thereon instructions that, when executed by one or more processors of a computing system, cause the computing system to: receive, via the one or more processors and at a trained mechanistic model trained based on one or more drug profiles, an indication of a first drug; analyze, via the one or more processors and by the trained mechanistic model, a drug profile of the one or more drug profiles associated with the first drug and at least one additional drug profile of the one or more drug profiles, the at least one additional drug profile associated with one or more second drugs, to generate a plurality of joint profile features associated with one or more respective combinations of the first drug and each respective second drug; generate, via the one or more processors and by the trained mechanistic model, a respective predicted drug combination outcome for each of the one or more respective combinations based on the plurality of joint profile features, the respective predicted drug combination outcome including a respective toxicity score and a respective potency score; and determine, via the one or more processors and by the trained mechanistic model, one or more drug combination candidates of the one or more respective combinations based on the respective toxicity score and the respective potency score for the each of the one or more respective combinations.

[0108] Throughout this specification, plural instances may implement components, operations, or structures described as a single instance. Although individual operations of one or more methods are illustrated and described as separate operations, one or more of the individual operations may be performed concurrently, and nothing requires that the operations be performed in the order illustrated. Structures and functionality presented as separate components in example configurations may be implemented as a combined structure or component. Similarly, structures and functionality presented as a single component may be implemented as separate components. These and other variations,Patent Application30275 / 70687 / PC modifications, additions, and improvements fall within the scope of the subject matter herein.

[0109] Additionally, certain embodiments are described herein as including logic or a number of routines, subroutines, applications, or instructions. These may constitute either software (e.g., code embodied on a machine-readable medium or in a transmission signal) or hardware. In hardware, the routines, etc., are tangible units capable of performing certain operations and may be configured or arranged in a certain manner. In example embodiments, one or more computer systems (e.g., a standalone, client or server computer system) such as system 100 of FIG. 1 or one or more hardware modules of a computer system (e.g., a processor or a group of processors) such as system 100 may be configured by software (e.g., an application or application portion) as a hardware module that operates to perform certain operations as described herein. Similarly, the computer systems or hardware modules of the computer systems may be configured to execute stored instructions on a memory as described herein to perform any such operations as described herein.

[0110] In various embodiments, a hardware module may be implemented mechanically or electronically. For example, a hardware module may comprise dedicated circuitry or logic that is permanently configured (e.g., as a special-purpose processor, such as a field programmable gate array (FPGA) or an application- specific integrated circuit (ASIC)) to perform certain operations. A hardware module may also comprise programmable logic or circuitry (e.g., as encompassed within a general-purpose processor or other programmable processor) that is temporarily configured by soil ware to perform certain operations. It will be appreciated that the decision to implement a hardware module mechanically, in dedicated and permanently configured circuitry, or in temporarily configured circuitry (e.g., configured by software) may be driven by cost and time considerations.

[0111] Accordingly, the term "hardware module" should be understood to encompass a tangible entity, be that an entity that is physically constructed, permanently configured (e.g., hardwired), or temporarily configured (e.g., programmed) to operate in a certain manner or to perform certain operations described herein. Considering embodiments in which hardware modules are temporarily configured (e.g., programmed), each of the hardware modules need not be configured or instantiated at any one instance in time. For example, where thePatent Application30275 / 70687 / PC hardware modules comprise a general-purpose processor configured using software, the general-purpose processor may be configured as respective different hardware modules at different times. Software may accordingly configure a processor, for example, to constitute a particular hardware module at one instance of time and to constitute a different hardware module at a different instance of time.

[0112] Hardware modules can provide information to, and receive information from, other hardware modules. Accordingly, the described hardware modules may be regarded as being communicatively coupled. Where multiple of such hardware modules exist contemporaneously, communications may be achieved through signal transmission (e.g., over appropriate circuits and buses) that connects the hardware modules. In embodiments in which multiple hardware modules are configured or instantiated at different times, communications between such hardware modules may be achieved, for example, through the storage and retrieval of information in memory structures to which the multiple hardware modules have access. For example, one hardware module may perform an operation and store the output of that operation in a memory device to which it is communicatively coupled. A further hardware module may then, at a later time, access the memory device to retrieve and process the stored output. Hardware modules may also initiate communications with input or output devices, and can operate on a resource (e.g., a collection of information).

[0113] The various operations of the example methods described herein may be performed, at least partially, by one or more processors that are temporarily configured (e.g., by software) or that are permanently configured to perform the relevant operations. Whether temporarily or permanently configured, such processors may constitute processor- implemented modules that operate to perform one or more operations or functions. The modules referred to herein may, in some example embodiments, comprise processor- implemented modules.

[0114] Similarly, the methods or routines described herein may be at least partially processor-implemented. For example, at least some of the operations of a method may be performed by one or more processors or by processor-implemented hardware modules. The performance of certain of the operations may be distributed among the one or morePatent Application30275 / 70687 / PC processors, not only residing within a single machine (having different processing abilities), but also deployed across a number of machines. In some example embodiments, the processors may be located in a single location (e.g., deployed in the field, in an office environment, or as part of a server farm), while in other embodiments the processors may be distributed across a number of locations.

[0115] Unless specifically stated otherwise, discussions herein using words such as "processing," "computing," "calculating," "determining," "presenting," "displaying," or the like may refer to actions or processes on a GPU thread that manipulates or transforms data represented as physical (e.g., electronic, magnetic, or optical) quantities within one or more memories (e.g., volatile memory, non-volatile memory, or a combination thereof), registers, or other machine components that receive, store, transmit, or display information.

[0116] As used herein any reference to "one embodiment" or "an embodiment" means that a particular element, feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. The appearances of the phrase "in one embodiment" in various places in the specification are not necessarily all referring to the same embodiment.

[0117] Some embodiments may be described using the expression "coupled" and "connected" along with their derivatives. For example, some embodiments may be described using the term "coupled" to indicate that two or more elements are in direct physical or electrical contact. The term "coupled," however, may also mean that two or more elements are not in direct contact with each other, but yet still co-operate or interact with each other. The embodiments are not limited in this context.

[0118] As used herein, the terms "comprises," "comprising," "includes," "including," "has," "having" or any other variation thereof, are intended to cover a non-exclusive inclusion. For example, a process, method, article, or apparatus that comprises a list of elements is not necessarily limited to only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Further, unless expressly stated to the contrary, "or" refers to an inclusive or and not to an exclusive or. For example, a condition A or B is satisfied by any one of the following: A is true (or present)Patent Application30275 / 70687 / PC and B is false (or not present), A is false (or not present) and B is true (or present), and both A and B arc true (or present).

[0119] In addition, use of the "a" or "an" are employed to describe elements and components of the embodiments herein. This is done merely for convenience and to give a general sense of the description. This description, and the claims that follow, should be read to include one or at least one and the singular’ also includes the plural unless it is obvious that it is meant otherwise.

[0120] This detailed description is to be construed as an example only and does not describe every possible embodiment, as describing every possible embodiment would be impractical, if not impossible. One could implement numerous alternate embodiments, using either current technology or technology developed after the filing date of this application.

Claims

Patent Application30275 / 70687 / PCWHAT IS CLAIMED:1) A computer-implemented method for predicting drug combination therapy candidates, the method comprising: receiving, via one or more processors and at a trained mechanistic model trained based on one or more drug profiles, an indication of a first drug; analyzing, via the one or more processors and by the trained mechanistic model, a drug profile of the one or more drug profiles associated with the first drug and at least one additional drug profile of the one or more drug profiles, the at least one additional drug profile associated with one or more second drugs, to generate a plurality of joint profile features associated with one or more respective combinations of the first drug and each respective second drug; generating, via the one or more processors and by the trained mechanistic model, a respective predicted drug combination outcome for each of the one or more respective combinations based on the plurality of joint profile features, the respective predicted drug combination outcome including a respective toxicity score and a respective potency score; and determining, via the one or more processors and by the trained mechanistic model, one or more drug combination candidates of the one or more respective combinations based on the respective toxicity score and the respective potency score for the each of the one or more respective combinations.2) The computer-implemented method of claim 1 , wherein the trained mechanistic model includes an artificial neural network (ANN).3) The computer-implemented method of claim 2, further comprising: processing, by the ANN, the plurality of joint profile features to generate the respective predicted drug combination outcome for the each of the one or more respective combinations.Patent Application30275 / 70687 / PC4) The computer-implemented method of any one of claims 1-3, further comprising: generating, by the trained mechanistic model, an indication of an impact on the respective toxicity score or the respective potency score of at least one of: (i) a mechanism associated with the one or more drug combination candidates, (ii) a pathway associated with the one or more drug combination candidates, or (iii) a subsystem associated with the one or more drug combination candidates.5) The computer-implemented method of any one of claims 1-4, wherein the trained mechanistic model includes a genome scale metabolic model (GEM) representing a plurality of subsystems of a metabolic network for an organism and including one or more integrated constraints.6) The computer-implemented method of claim 5, further comprising: generating, by the GEM, the plurality of joint profile features corresponding to the one or more respective combinations.7) The computer-implemented method of claim 5, wherein the plurality of joint profile features includes a plurality of sets of joint profile features each corresponding to a respective subsystem of the plurality of subsystems represented by the GEM.8) The computer-implemented method of claim 5, wherein the one or more integrated constraints include at least one of: multi-omics data, media conditions, metabolites, exchange rates, growth rates, drug dosing, or time of treatments.9) The computer-implemented method of any one of claims 1-8, wherein the one or more second drugs is a single second drug and the indication of the first drug is a first indication, the method further comprising:Patent Application 30275 / 70687 / PC receiving, via the one or more processors and at the trained mechanistic model, a second indication of the single second drug; and generating, via the one or more processors, an output based on the one or more drug combination candidates and indicative of whether a combination of the first drug and the single second drug is safe for a patient to take.10) The computer-implemented method of any one of claims 1-9, further comprising: generating, via the one or more processors, an output based on the one or more drug combination candidates and indicative of which combinations of the one or more respective combinations are safe for a patient to take.11) The computer-implemented method of claim 10, wherein the output is a list including the each of the one or more drug combination candidates and a respective toxicity score and a respective potency score for the each of the one or more drug combination candidates.12) The computer-implemented method of any one of claims 1-11, further comprising: determining, via the one or more processors and by the trained mechanistic model, one or more underlying mechanisms associated with the respective toxicity scores and the respective potency scores for the one or more respective combinations, wherein the one or more underlying mechanisms include at least one of: a biochemical pathway, or a molecular factor.13) The computer-implemented method of any one of claims 1-12, further comprising: determining, via the one or more processors and by the trained mechanistic model, a drug combination candidate of the one or more drug combination candidates for one or more of: an individual organism, a disease, a tissue type, or a cell type.14) The computer-implemented method of any one of claims 1-13, further comprising:Patent Application 30275 / 70687 / PC determining, via the one or more processors and by the trained mechanistic model, one or more second drugs that reduce toxicity of the first drug for one or more of: an individual organism, a disease, a tissue type, or a cell type.15) A computing system for predicting drug combination therapy candidates, the system comprising: one or more processors; a trained mechanistic model trained based one or more drug profiles; and a non-transitory computer readable medium including computer executable instructions that, when executed by the one or more processors, cause the computing system to implement the method of any one of claims 1-14.

Images