Methods, systems, and apparatus for classifying molecular drugs for drug optimization

Abstract

Methods, systems, and apparatuses are described for classifying one or more candidate molecular drugs for one or more desired target conditions for downstream drug optimization processes. A candidate molecular drug dataset of one or more molecular drugs is received, the candidate molecular drug dataset including data representing a set of physicochemical values for each candidate molecular drug and data associated with each candidate molecular drug for a dosing regimen and a target of a set of pharmacokinetic (PK) values, and wherein each candidate molecular drug in the candidate molecular drug dataset is classified as a large molecular drug. For each candidate molecular drug in the candidate molecular drug dataset, the set of corresponding physicochemical values and the set of PK values for the dosing regimen and the target are applied to a machine learning (ML) model configured for classifying whether each candidate molecular drug satisfies one or more desired conditions associated with a pharmacological or pharmacodynamic effect. The ML model is trained using a ML algorithm to train model parameters defining the ML model for classifying whether each candidate molecular drug satisfies the one or more desired conditions based on a training dataset, the training dataset including data representing a plurality of molecular drugs, each molecular drug represented by a set of corresponding physicochemical values and a set of PK values associated with the dosing regimen and the target, and annotated for corresponding pharmacological or pharmacodynamic effects, responses, and / or values related to the desired conditions. A set of candidate molecular drugs satisfying the desired conditions are output for downstream drug optimization processes.

Description

Technical Field

[0001] This specification relates to methods, systems, and apparatus for classifying whether one or more molecular drugs meet one or more conditions associated with pharmacological / pharmacodynamic effects prior to downstream computer simulation, in vitro and / or in vivo analysis and drug optimization, wherein each molecular drug is represented by: a set of physicochemical values; and a set of pharmacokinetic (PK) values ​​associated with the drug administration regimen and target. Background Technology

[0002] Monoclonal antibodies (mAbs) have revolutionized healthcare as therapeutic molecular drugs, becoming the best-selling form of medicine in recent years. Recently, the pharmaceutical industry has shown great interest in elucidating the pharmacokinetic (PK) principles of these antibodies, as these principles provide crucial information about drug efficacy and dosing regimens or strategies. During drug development, molecular drug candidates are typically studied iteratively to obtain key pharmacokinetic (PK) characteristic curves and endpoints or parameters associated with absorption, distribution, metabolism, and excretion (ADME) or release ADME (LADME). The pharmacokinetics of molecular drugs (e.g., mAbs) have been extensively studied, with additional focus on the distribution and elimination properties of the drug. Neonatal Fc receptor (FcRn) binding and molecular size have been shown to significantly influence the systemic PK of these antibodies. However, properties such as antibody charge and affinity for nonspecific interactions with charged cellular components remain among the few properties that are not yet fully understood. The quantitative relationship between the physicochemical properties of antibodies (such as molecular weight, size, charge, and FcRn binding affinity) and their PK response remains largely unknown, but it is still crucial for establishing a preliminary understanding of the PK of drugs and the expected pharmacological / pharmacodynamic effects (such as target binding or target occupancy).

[0003] During molecular drug discovery / optimization activities, candidate molecular drugs with unacceptable PK properties against targets are often abandoned (or reengineered). Due to budget constraints and / or ethical concerns, it is impossible to conduct in vivo testing, in vitro wet laboratory analysis, or high-throughput screening for a large number of potential candidate molecular drugs that meet certain desired physicochemical properties when targeting corresponding targets with certain desired physicochemical properties. There are many pairing combinations between molecular drugs and targets, and analyzing all possible molecular drug candidates with targets one by one is impractical. Such exhaustive analysis would place enormous strain on currently available computational and laboratory resources. This could lead to significant delays in the rollout of promising new molecular drugs (e.g., mAbs) and corresponding molecular drug therapies, and increase associated costs.

[0004] Therefore, there is an urgent need for a molecular drug discovery mechanism or framework that can minimize or reduce the computational and laboratory resources required to generate candidate molecular drugs (e.g., a candidate list of candidate molecular drugs) before performing further downstream drug optimization processes (e.g., in vitro and / or in vivo evaluations), as well as minimize or reduce the associated costs and delays of these candidate molecular drugs with the most promising PK properties and corresponding expected pharmacological / pharmacodynamic effects (e.g., target binding or target occupancy values, efficacy and / or toxicity effects / endpoints). Summary of the Invention

[0005] According to the first aspect, a computer-implemented method is provided for classifying one or more candidate molecular drugs for drug optimization based on one or more desired conditions associated with pharmacological or pharmacodynamic effects, the method comprising:

[0006] The system receives a candidate drug inference dataset representing one or more candidate drug molecules, wherein each candidate drug molecule in the dataset is represented by a set of physicochemical values ​​of the candidate drug molecule and a set of pharmacokinetic (PK) values ​​associated with each candidate drug molecule, having a dosing regimen and a target, and wherein each candidate drug molecule in the dataset is classified as a macromolecular drug; for each candidate drug molecule in the candidate drug inference dataset, the corresponding set of physicochemical values ​​and the set of PK values ​​for the dosing regimen and target are applied to a machine learning (ML) model, which is configured to determine whether each candidate drug molecule satisfies one or more of the pharmacological or pharmacodynamic effects associated with a pharmacological or pharmacodynamic effect. The model is trained using an ML algorithm to define its parameters. The ML model is used to classify each candidate molecular drug based on a training dataset to determine whether it satisfies one or more expected conditions associated with pharmacological or pharmacodynamic effects. The training dataset includes data representing multiple molecular drugs, represented by multiple sets of known and / or calculated physicochemical values ​​and multiple sets of pharmacokinetic (PK) values ​​associated with the dosing regimen, and labeled based on the one or more expected conditions associated with pharmacological or pharmacodynamic effects. The output represents a set of candidate molecular drugs classified according to these expected conditions associated with pharmacological or pharmacodynamic effects for use in downstream drug optimization processes.

[0007] The computer-implemented method as described in the first aspect, wherein each candidate molecular drug is an antibody. Alternatively, each candidate molecular drug is a monoclonal antibody. Alternatively, the monoclonal antibody has a size greater than or equal to 50 kDa. Alternatively, the monoclonal antibody is classified as a macromolecular drug. Alternatively, the target is an antigen.

[0008] The computer-implemented method as described in the first aspect, wherein the set of physicochemical values ​​for each molecular drug corresponds to one or more physicochemical parameters or properties derived from the group consisting of: molecular weight (MW); size or Stokes radius (ae); charge (Q); hydrophobicity; binding affinity; isoelectric point (pI); glycosylation; drug-target binding constant (KD); and any other type of physicochemical parameter or property associated with each molecular drug. Alternatively, the set of physicochemical properties for each molecular drug includes at least MW, ae, Q, and KD.

[0009] The computer-implemented method as described in the first aspect, wherein the set of PK values ​​for each molecule drug is associated with PK characteristic curves, endpoints, parameters and / or properties corresponding to ADME for absorption, distribution, metabolism and excretion or LADME for release, absorption, distribution, metabolism and excretion.

[0010] The computer-implemented method as described in the first aspect, wherein these PK characteristic curves, endpoints, parameters, and / or attributes comprise one or more of the group consisting of: dose (D); dosing interval (τ); maximum serum concentration (C). max ); Reaching C max Shortest time; minimum plasma concentration (C min Or C min,ss ); mean plasma concentration (C av Or C av,ss ); steady-state concentration (C) ss ); Distributed volume (V) d Absorption half-life; absorption rate constant; elimination half-life; time to reach minimum plasma concentration ( ); ), total solubles ( ), membrane target inhibition ( Elimination rate constant; area under the concentration-time curve (AUC) after a single dose or at steady state. ss ); clearance rate (C); and any other type of PK characteristic curve, endpoint, parameter, and / or attribute associated with ADME or LADME. Alternatively, this set of PK characteristic curves, endpoints, parameters, and / or attributes for each molecular drug shall include at least C. min AUC ss C max and C ss Alternatively, the PK characteristic curves, endpoints, parameters, and / or properties for each molecular drug in this group should at least include C. min AUC ss C max C ss , and .

[0011] The computer-implemented method as described in the first aspect, wherein the ML algorithm comprises one or more ML algorithms or combinations thereof from the group consisting of: neural network-based ML algorithms; decision tree-based classification ML algorithms; random forest decision tree ML algorithms; adaptive boosting-based ML algorithms; gradient boosting-based ML algorithms; bagging-based ML algorithms; ensemble-based ML algorithms; any other supervised ML algorithm suitable for training model parameters of an ML model for outputting a classification of whether one or more molecular drugs satisfy one or more desired conditions associated with pharmacological or pharmacodynamic effects, given input data representing a set of physicochemical values ​​and a set of PK values ​​corresponding to the one or more molecular drugs for targets and dosing regimens.

[0012] The computer-implemented method as described in the first aspect further includes iteratively training model parameters defining the ML model using the ML algorithm associated with the ML model. This training includes: receiving a training dataset associated with multiple molecular drugs, the training dataset comprising multiple training instances, each training instance associated with each molecular drug and including data representing a set of known and / or expected physicochemical values ​​and a set of corresponding known or calculated pharmacokinetic values ​​associated with the dosing regimen and the target, and each training instance being labeled or marked based on whether each molecular drug and target satisfies expected conditions associated with pharmacological or pharmacodynamic effects; iteratively applying the training dataset to the ML algorithm to train the model parameters of the ML model until a stopping criterion is met; and outputting model parameters representing the trained ML model.

[0013] The computer-implemented method as described in the first aspect, wherein the multiple sets of physicochemical values ​​corresponding to these molecular drugs in the training dataset are based on synthetically generated molecular drugs produced within a desired range of physicochemical values, wherein, for each synthetically generated molecular drug, a physiologically based pharmacokinetic (PBPK) model is used to calculate a corresponding PK value for each synthetically generated molecular drug, which is associated with the dosing regimen and the target and with pharmacological or pharmacodynamic effects, responses, and / or values, the PBPK model being configured to: simulate the effect of the synthetically generated molecular drug on the target in the subject's cellular system; and, for each synthetically generated molecule, output a set of corresponding PK values ​​from selected PK endpoints and characteristic curves, along with corresponding pharmacological or pharmacodynamic effects, responses, and / or values ​​associated with these desired conditions.

[0014] The computer-implemented method as described in the first aspect, wherein the multiple sets of physicochemical values ​​corresponding to each molecular drug in the training dataset, the PK values ​​associated with the dosing regimen and the target and associated with pharmacological or pharmacodynamic effects, responses and / or values, are based on known or measured physicochemical values ​​of real-world molecular drugs, and the PK values ​​associated with the dosing regimen and the target and associated with pharmacological or pharmacodynamic effects, responses and / or values.

[0015] The computer-implemented method as described in the first aspect, wherein the multiple sets of physicochemical values ​​corresponding to these molecular drugs in the training dataset are based on real-world molecular drugs that meet the expected range of physicochemical values, wherein, for each real-world molecular drug, a PBPK model is used to calculate a corresponding PK value for each real-world molecular drug, these PK values ​​being associated with the target and the dosing regimen and these pharmacological or pharmacodynamic effects, responses and / or values, the PBPK model being configured to: simulate the effect of the real-world molecular drug on the target in the physiological or cellular system of a subject; and, for each real-world molecular drug, output a set of corresponding PK values ​​and corresponding pharmacological or pharmacodynamic effects, responses and / or values ​​from selected PK endpoints and characteristic curves.

[0016] The computer-implemented method as described in the first aspect, wherein the multiple sets of physicochemical values ​​corresponding to these molecular drugs in the candidate molecular drug inference dataset are based on synthetically generated molecular drugs generated within a desired range of physicochemical values, wherein, for each synthetically generated molecular drug, a PBPK model is used to calculate a corresponding PK value corresponding to each synthetically generated molecular drug, these PK values ​​being associated with the dosing regimen and these corresponding pharmacological or pharmacodynamic effects, responses, and / or values, the PBPK model being configured to: simulate the effect of the synthetically generated molecular drug on the target in the physiological or cellular system of a subject; and, for each synthetically generated molecule, output a set of corresponding PK values ​​from selected PK endpoints and characteristic curves, along with corresponding pharmacological or pharmacodynamic effects, responses, and / or values.

[0017] The computer-implemented method as described in the first aspect, wherein the multiple sets of physicochemical values ​​corresponding to each molecular drug in the candidate molecular drug inference dataset, the PK values ​​associated with the dosing regimen and the target, and these pharmacological or pharmacodynamic effects, responses, and / or values, are based on known or measured physicochemical values ​​of real-world molecular drugs, and the PK values ​​associated with the dosing regimen and the target, and the pharmacological or pharmacodynamic effects, responses, and / or values.

[0018] The computer-implemented method as described in the first aspect, wherein the multiple sets of physicochemical values ​​corresponding to these molecular drugs in the candidate molecular drug inference dataset are based on real-world molecular drugs that satisfy a range of expected physicochemical values ​​in a set of physicochemical parameters, wherein, for each real-world molecular drug, the PBPK model is used to calculate a corresponding PK value for each real-world molecular drug, these PK values ​​being associated with the target and the dosing regimen and these pharmacological or pharmacodynamic effects, responses and / or values, the PBPK model being configured to: simulate the effect of the real-world molecular drug on the target in the cellular system of a subject; and, for each real-world molecular drug, output a set of corresponding PK values ​​and corresponding pharmacological or pharmacodynamic effects, responses and / or values ​​from selected PK endpoints and characteristic curves.

[0019] The computer-implemented method as described in the first aspect further includes: receiving a set of physicochemical values ​​corresponding to each molecular drug; receiving a set of one or more sets of physicochemical values ​​corresponding to the target; receiving a set of dosing regimens; and generating one or more unique molecular drug-target data pairs for each desired dosing regimen by pairing the set of desired physicochemical values ​​of the molecular drug with each of the set of one or more sets of desired physicochemical values ​​of the target.

[0020] For each dosing regimen, the expected physicochemical values ​​of the drug and the target in each drug-target pairing are applied to the PBPK model. For each dosing regimen and each drug-target pairing, the PBPK model is used to calculate a set of PK values ​​for each drug-target pairing and the corresponding pharmacological or pharmacodynamic effects, responses, and / or values ​​associated with these expected conditions. Training data instances or candidate drug inference data instances are generated for each drug-target pairing and the corresponding dosing regimen. These training data instances include... The training data instance includes data representing at least the set of physicochemical properties and the calculated PK value for each molecular drug, and is labeled based on the corresponding pharmacological or pharmacodynamic effects, responses, and / or values ​​calculated for each molecular drug paired with a target and the corresponding dosing regimen, as well as these expected conditions; the candidate molecular drug inference data instance includes data representing at least the set of physicochemical properties and the calculated PK value for each molecular drug paired with a target and the corresponding dosing regimen; these training data instances or candidate molecular drug inference data instances are stored for training or inference / classification, respectively.

[0021] The computer-implemented method as described in the first aspect, wherein outputting a set of candidate molecular drugs that satisfy these desired conditions associated with pharmacological or pharmacodynamic effects for downstream drug optimization further includes outputting a candidate list of candidate molecular drugs—each candidate molecular drug represented by a set of corresponding physicochemical values ​​and a set of PK values ​​for the dosing regimen and the target—for downstream processing, including in vitro wet laboratory analysis or high-throughput screening.

[0022] The computer-implemented method as described in the first aspect, wherein outputting a set of candidate molecular drugs that satisfy these desired conditions associated with pharmacological or pharmacodynamic effects for downstream drug optimization further includes outputting a candidate list of candidate molecular drugs—each candidate molecular drug represented by a set of corresponding physicochemical values ​​and a set of PK values ​​for the dosing regimen and the target—for downstream processing, which includes in vivo testing.

[0023] The computer-implemented method as described in the first aspect, wherein when one or more candidate molecular drugs in the candidate list of the candidate molecular drug are represented only by a set of physicochemical values ​​and a set of PK values, then one or more real-world molecular drugs that satisfy the set of corresponding physicochemical values ​​and the set of PK values ​​of the one or more candidate molecular drugs are generated before performing in vitro wet laboratory analysis, high-throughput screening, or in vivo testing.

[0024] The computer-implemented method as described in the first aspect, wherein generating one or more real-world molecular drugs further comprises: for each corresponding candidate molecular drug, calculating the structural and chemical properties required to realize or synthesize a real-world molecular drug that satisfies the corresponding set of physicochemical values, the set of PK values ​​and associated pharmacological or pharmacodynamic effects, reactions and / or values.

[0025] The computer-implemented method as described in the first aspect, wherein outputting the set of candidate molecular drugs that meet these desired conditions for downstream drug optimization further includes outputting data representing any of the following: a candidate list of candidate molecular drugs that meet these desired conditions, having PK values ​​for the dosing regimen and target, and associated pharmacological or pharmacodynamic effects, responses, and / or values, for review by one or more user or expert systems to further evaluate the output candidate list of candidate molecular drugs; or a list of candidate molecular drugs categorized by whether they meet these desired conditions, having PK values ​​for the dosing regimen and target, and associated pharmacological or pharmacodynamic effects, responses, and / or values, for review by one or more user or expert systems to further evaluate the output list of candidate molecular drugs.

[0026] According to a second aspect, an apparatus is provided, including a processor, a memory unit, and a communication interface, wherein the processor is connected to the memory unit and the communication interface, and wherein the processor and the memory are configured to implement the computer-implemented method of the first aspect and any one or more features thereof.

[0027] According to a third aspect, a computer-readable medium is provided, including data or instruction code that, when executed on a processor, causes the processor to implement the computer-implemented method of the first aspect and any one or more features thereof.

[0028] According to a fourth aspect, a computational system is provided, comprising: a molecule and target generation unit configured to generate one or more molecules represented by multiple sets of corresponding physicochemical values ​​and one or more targets represented by desired physicochemical values; a physiology-based PK unit configured to generate a set of PK values ​​and / or associated pharmacological or pharmacodynamic effects, responses, and / or values ​​for each molecular drug paired with a target as training data instances or candidate molecular drug inference data instances; and an ML module / unit configured to train an ML model using a training dataset containing multiple training data instances for specific drug administration. The scheme classifies whether one or more molecular drugs meet desired conditions associated with pharmacological or pharmacodynamic effects; or is used for reasoning using a trained ML model to classify whether one or more input candidate molecular drug reasoning data instances meet desired target conditions for a specific dosing regimen; and an output module / unit configured to output at least one candidate list of molecular drugs that meet these desired conditions associated with pharmacological or pharmacodynamic effects—each molecular drug represented by a set of physicochemical values ​​and a set of PK values—for generating real-world molecules with corresponding properties for in vivo assays, in vitro wet laboratory analysis, and / or high-throughput screening laboratory analysis.

[0029] The computational system as described in the fourth aspect, wherein the molecule and target generation unit, the physiology-based PK unit, the ML module / unit, and the output module / unit are configured to implement the computer-implemented method of the first aspect and any one or more of its features.

[0030] According to a fifth aspect, a non-transitory tangible computer-readable medium is provided, comprising data or instruction code for classifying one or more candidate molecular drugs for drug optimization in response to one or more desired conditions associated with pharmacological or pharmacodynamic effects, the data or instruction code, when executed on one or more processors, causing at least one of these processors to perform at least the following operations:

[0031] The system receives a candidate drug inference dataset representing one or more candidate drug molecules, wherein each candidate drug molecule in the dataset is represented by a set of physicochemical values ​​of the candidate drug molecule and a set of pharmacokinetic (PK) values ​​associated with each candidate drug molecule, including a dosing regimen and a target, and wherein each candidate drug molecule in the dataset is classified as a macromolecular drug. For each candidate drug molecule in the candidate drug inference dataset, the corresponding set of physicochemical values ​​and the set of PK values ​​associated with the dosing regimen and target are applied to an ML model, the ML model being configured to determine whether each candidate drug molecule satisfies a pharmacological or pharmacodynamic effect. The ML model is trained using an ML algorithm to define its parameters. The ML model is used to classify each candidate molecular drug based on a training dataset to determine whether it satisfies one or more of the desired conditions. This training dataset includes data representing multiple molecular drugs, represented by multiple sets of known and / or calculated physicochemical values ​​and multiple sets of pharmacokinetic (PK) values ​​associated with the dosing regimen and the target, and labeled with corresponding pharmacological or pharmacodynamic effects, responses, and / or values ​​related to these desired conditions. The output represents a set of candidate molecular drugs that satisfy these desired conditions for downstream drug optimization.

[0032] In various embodiments, a computer program instruction optionally stored on a non-transitory computer-readable medium, which, when executed by one or more processors of a data processing apparatus, causes the data processing apparatus to perform operations including one or more aspects of the embodiments described above and / or below (including one or more aspects of the appended claims).

[0033] In various embodiments, an apparatus is disclosed comprising a computer-readable storage medium and one or more processors having program instructions contained therein, the one or more processors being configured to execute the program instructions to cause the apparatus to perform operations including one or more aspects of the embodiments described above and / or below (including one or more aspects of the appended claims). The apparatus may include one or more processors or dedicated computing hardware. Attached Figure Description

[0034] To make the present invention more readily understood, embodiments of the invention will now be described by way of example only with reference to the accompanying drawings, in which:

[0035] Figure 1a illustrates an example ML model classification process according to some embodiments of the present invention;

[0036] Figure 1b illustrates an example ML model training process according to some embodiments of the present invention;

[0037] Figure 1c illustrates an example preliminary screening process for candidate molecules according to some embodiments of the present invention;

[0038] Figure 1d illustrates an example of generating a training dataset or candidate molecular drug inference dataset using a PBPK model according to some embodiments of the present invention;

[0039] Figure 2a illustrates an example PBPK / ML system according to some embodiments of the present invention;

[0040] Figure 2b shows an example decision tree of a trained tree-based ML model according to some embodiments of the present invention;

[0041] Figure 2c shows a set of example results of the tree-based ML model classification of Figures 2a and 2b according to some embodiments of the present invention;

[0042] Figure 2d shows another set of example results obtained by classifying synthetic real-world molecular drugs based on the tree-structured ML model of Figures 2a to 2c according to some embodiments of the present invention;

[0043] Figure 2e shows another set of example results obtained by classifying selected real-world molecular drugs based on the tree-structured ML model of Figures 2a to 2c according to some embodiments of the present invention;

[0044] Figure 2f shows several additional sets of example results for multi-label classification when a tree-based ML model is trained to classify three categories according to some embodiments of the present invention.

[0045] Figure 2g illustrates another PBPK / ML system 250 according to some embodiments of the present invention;

[0046] Figure 3 An example PBPK component 300 for implementing the mPBPK model 301 according to some embodiments of the present invention is shown;

[0047] Figure 4 This is a schematic diagram of a system / apparatus used to perform the methods described herein.

[0048] Throughout the accompanying drawings, common reference numerals are used to indicate similar features. Detailed Implementation

[0049] The various exemplary embodiments described herein relate to methods, apparatus, and systems for automatically, efficiently, and reliably performing computer simulation classification, prediction, or estimation of whether candidate molecular drugs (e.g., but not limited to, antibodies and / or monoclonal antibodies (mAbs)) satisfy various defined or desired conditions associated with pharmacological and / or pharmacodynamic effects (e.g., target occupancy conditions associated with target binding response or target occupancy values), wherein each candidate molecular drug (e.g., mAb) is represented by a set of physicochemical values, a set of PK values ​​associated with the drug administration regimen and the target (e.g., antigen), which are related to the PK characteristics of each molecular drug interacting with the target within a physiological or cellular system.

[0050] Pharmacological and / or pharmacodynamic effects include, but are not limited to, target binding response and / or target occupancy and / or any other multiple efficacy or toxicity-related endpoints (e.g., cytokine levels, biomarker concentrations, etc.). For example, desired conditions may be based on, but are not limited to, target binding or target occupancy conditions (e.g., a target occupancy percentage (TO%) value ≤ an upper TO% threshold and / or a lower TO% threshold <= a TO% value), where target occupancy can be considered an initial biomarker of the pharmacological efficacy of a given molecular drug / target combination and can be independent of disease / clinical symptoms. Other desired conditions are also applicable and may be based on, but are not limited to, multiple efficacy or toxicity-related endpoints (e.g., but not limited to, cytokine levels, biomarker concentrations, etc.), and / or any other suitable pharmacological and / or pharmacodynamic effects. Each of the target occupancy, efficacy, or toxicity-related endpoints (e.g., but not limited to, cytokine levels, biomarker concentrations, etc.) may be used to form one or more desired conditions. The advantage of using target occupancy to form one or more desired conditions is that target occupancy is a biomarker that can be used for a variety of clinical conditions, and the associated efficacy-toxicity endpoints may be highly dependent on the disease of interest.

[0051] This set of physicochemical values ​​for each molecular drug may correspond to one or more physicochemical parameters or properties from the group consisting of: molecular weight (MW), size or Stokes radius (ae), charge (Q), hydrophobicity, FcRn binding affinity, isoelectric point (pI), glycosylation, drug-target binding constant (KD), and / or any other type of physicochemical parameter or property associated with each molecular drug. In some embodiments, this set of physicochemical values ​​includes, but is not limited to, at least the following properties: MW, size or Stokes radius (ae), Q, and KD. This set of PK values ​​for each molecular drug is associated with PK characteristic curves, endpoints, parameters, and / or properties corresponding to ADME or LADME. These PK characteristic curves, endpoints, parameters, and / or properties include one or more from the group consisting of: dose (D) based on the desired dosing regimen, dosing interval (τ) based on the desired dosing regimen, and maximum serum concentration (C). max ), reaching C max Shortest time, minimum plasma concentration (C min,ss ), mean plasma concentration (C av,ss ), steady-state concentration (C) ss ), distributed volume or steady-state distributed volume (V) d or V ss Absorption half-life, absorption rate constant, elimination half-life; half-life ( ), time to reach minimum plasma concentration ( ), total solubles ( ), membrane target inhibition ( Elimination rate constant, area under the concentration-time curve after a single dose or at steady state (AUC) ss The clearance rate (C) (in mL / h / kg), and / or any other type of PK characteristic curve, endpoint, parameter, and / or attribute associated with ADME or LADME depending on application requirements. In some example embodiments, this set of PK characteristic curves, endpoints, parameters, and / or attributes for each molecular drug includes at least C. min AUC ss C max and C ss In some example embodiments, the set of PK characteristic curves, endpoints, parameters, and / or properties for each molecular drug include at least C. min AUC ss C max C ss , , , .

[0052] Computer simulations of classification, prediction, or estimation are performed by an appropriately trained ML model configured to classify, predict, or estimate whether the candidate molecular drugs satisfy various defined or desired conditions associated with pharmacological and / or pharmacodynamic effects (e.g., target binding responses or target occupancy). The ML model is trained using an ML algorithm and a training dataset comprising a variety of synthetic / virtual molecular drugs and / or real-world molecular drugs (e.g., mAbs). Each molecular drug in the training dataset is represented at least by data indicating that molecular drug, including a set of corresponding physicochemical values ​​and a set of PK values ​​for each molecular drug associated with drug dosing regimens and target interactions (e.g., antigens) (the set of PK values ​​being based on the PK characteristics of each molecular drug), and these molecular drugs are labeled according to how each molecular drug satisfies desired conditions associated with pharmacological and / or pharmacodynamic effects (e.g., target binding responses or target occupancy). The ML model is trained using a supervised training ML algorithm that computes model parameters defining the ML model against the training dataset. For example, ML algorithms may include one or more ML algorithms or combinations thereof that are derived from the group consisting of: neural network ML algorithms, decision tree-based classification ML algorithms, random forest decision tree ML algorithms, adaptive boosting ML algorithms, gradient boosting ML algorithms, bagging ML algorithms, ensemble ML algorithms, and / or any other supervised ML algorithms applicable to the model parameters used to train the ML model.

[0053] Once the model parameters of the ML model are defined and training is complete, the trained ML model is used to classify a candidate molecular drug inference dataset comprising a variety of synthetic / virtual molecular drugs and / or real-world molecular drugs (e.g., mAbs). Each molecular drug in the candidate molecular drug inference dataset is represented by data representing at least a set of corresponding physicochemical values ​​and a set of PK values ​​associated with drug dosing regimens and target interactions (e.g., antigens). The candidate molecular drug inference dataset is applied to and processed by the trained ML model to output data representing whether the candidate molecular drugs in the candidate molecular drug inference dataset meet desired conditions associated with pharmacological and / or pharmacodynamic effects (e.g., target binding response or target occupancy). For example, a candidate list of candidate molecular drugs can be obtained as output based on those candidate molecular drugs in the candidate molecular drug inference dataset that meet or conform to the desired conditions. The candidate list of candidate molecular drugs can be used in downstream drug optimization processes, including but not limited to, for example, in vivo testing, in vitro wet laboratory analysis, or high-throughput screening systems.

[0054] Training datasets and candidate molecular drug inference datasets can be generated by selecting data representing molecular drugs with desired physicochemical properties, data representing one or more targets of the same type with desired physicochemical properties, and dosing regimens, and calculating and / or measuring the PK characteristics corresponding to the molecular drug-target pairings. For example, data representing each selected molecular drug includes a set of corresponding physicochemical values ​​representing that molecular drug, wherein this set of physicochemical values ​​for the molecular drug is within a range of a set of physicochemical parameters set or selected by the user for the molecular drug being evaluated. Data representing each selected target includes a set of corresponding physicochemical values ​​representing that target, wherein this set of physicochemical values ​​for the target is within a range of a set of physicochemical parameters set or selected by the user for the target being evaluated. This can generate multiple molecular drugs and multiple targets.

[0055] Once generated, multiple unique drug-molecule-target pairings for each dosing regimen are derived from data representing the selected drug-molecule and one or more targets. If the PK characteristics of a particular drug-molecule-target pairing are unknown or unmeasured, a physiology-based pharmacokinetic (PBPK) model—configured to simulate the interaction of each drug-molecule with the target in a modeled physiological system according to the dosing regimen—can be applied to output the PK characteristics of that interaction. PBPK models represent drug-molecule-target interactions through a variety of mathematical relationships: cellular and tissue systems within a physiological or cellular system; and the interaction of each drug-molecule with the target within that physiological or cellular system according to the dosing regimen. PBPK models calculate or estimate a set of PK values ​​for the drug-molecule associated with the target and dosing regimen. This set of PK values ​​as output depends on the type of PBPK model and is based on a set of PK characteristic curves and / or endpoints. PBPK models can also calculate additional pharmacological and / or pharmacodynamic effects, responses, and values ​​associated with the effects of each drug-molecule-target pairing on the physiological or cellular system. The pharmacological and / or pharmacodynamic effects, responses, and values ​​of each drug-molecule paired with a target can be used to determine whether the corresponding drug-molecule meets the desired conditions. These effects, responses, and values ​​may include, but are not limited to, corresponding target binding responses and target binding / occupancy values, efficacy and / or toxicity-related endpoints. The pharmacological and / or pharmacodynamic effects, responses, and values ​​of each drug-molecule paired with a target can be calculated using a PBPK model, and / or calculated / estimated based on PK values ​​estimated from the PBPK model.

[0056] Alternatively, a training dataset can be generated comprising multiple training data instances, each containing data representing a drug-to-target pairing for each of a plurality of molecular drugs and each of a plurality of targets. Each training data instance is labeled or tagged based on whether the pharmacological and / or pharmacodynamic effects, responses, and values ​​of each corresponding drug-to-target pairing satisfy the expected conditions associated with the pharmacological and / or pharmacodynamic effects. The training data instances representing drug-to-target pairings include, for example: a set of physicochemical values ​​representing the molecular drug; a set of pharmacokinetic (PK) values ​​of the molecular drug associated with the target and drug dosing regimen; and annotations or labels associated with whether one or more corresponding pharmacological and / or pharmacodynamic effects, responses, and values ​​(e.g., target binding or occupancy) of each drug-to-target pairing satisfy the expected conditions.

[0057] Additionally, a candidate molecular drug inference dataset can be generated, comprising multiple candidate molecular drug data instances. Each candidate molecular drug data instance includes data representing a drug-target pairing for each of a plurality of molecular drugs and each of a plurality of targets. The candidate molecular drug data instances representing drug-target pairings include, for example, a set of physicochemical values ​​representing the molecular drug, and a set of p-values ​​of the molecular drug associated with the target and the drug dosing regimen.

[0058] Each of the training dataset and / or candidate molecular drug inference dataset can be populated with data representing real-world molecular drugs, each with a set of known physicochemical values ​​and a set of known or measured pharmacokinetic (PK) values ​​associated with the target and drug dosing regimen. The training dataset can be labeled based on whether the molecular drug-target pairing in each training data instance satisfies desired conditions associated with pharmacological and / or pharmacodynamic effects. The required set of physicochemical values, the set of known or measured PK values, and / or the corresponding pharmacological and / or pharmacodynamic effects, responses, and / or values ​​for each molecular drug can be data mined from various databases and / or publications associated with the evaluation of real-world molecular drugs to generate the training dataset or candidate molecular drug inference dataset.

[0059] Figure 1a illustrates an example ML model classification process 100 for classifying one or more candidate molecular drugs against one or more desired target conditions for drug optimization. The ML model classification process 100 includes at least the following steps:

[0060] In step 101, a candidate molecular drug inference dataset representing one or more candidate molecular drugs is received. Each candidate molecular drug in the candidate molecular drug dataset is represented by: a set of physicochemical values ​​of the candidate molecular drug; a set of PK values ​​associated with each candidate molecular drug and related to the dosing regimen and target; and a corresponding target binding or occupancy value for each candidate molecular drug.

[0061] In some examples, each candidate molecular drug is an antibody. In some embodiments, each candidate molecular drug is classified as a macromolecular drug. In some embodiments, each candidate molecular drug is a monoclonal antibody (mAb). In some embodiments, each candidate molecular drug is a mAb classified as a macromolecular drug. In examples, a large mAb can be a full-length mAb of conventional size with a MW ranging from, for example, but not limited to, 50 kDa to 150 kDa. In another example, a macromolecular drug or a large mAb can be considered a molecular drug or mAb with a size greater than or equal to 50 kDa. Candidate molecular drugs can also be classified as macromolecular drugs based on therapeutic efficacy or expected effect, wherein the MW of smaller-sized antibody / antibody fragments ranges from, for example, but not limited to, 3 kDa to 100 kDa, or 50 kDa to 150 kDa. In some embodiments, a candidate molecular drug corresponds to any size mAb or mAb fragment, as long as the corresponding PBPK model can use it to generate / estimate PK values ​​and / or pharmacological or pharmacodynamic effects, responses, and / or values ​​(e.g., target occupancy values), etc. The target is, for example, but not limited to, an antigen, or any other molecule to which the drug is designed to bind.

[0062] In step 102, for each candidate molecular drug in the candidate molecular drug inference dataset, data representing a set of corresponding physicochemical values ​​and a set of PK values ​​for dosing regimens and targets are applied to a machine learning (ML) model. This ML model is configured to classify whether each candidate molecular drug satisfies one or more desired conditions (e.g., target occupancy conditions, efficacy-toxicity conditions, etc.) associated with pharmacological and / or pharmacodynamic effects. For example, the ML model is trained using an ML algorithm to train model parameters that define the ML model for classifying whether each candidate molecular drug satisfies one or more desired conditions associated with pharmacological and / or pharmacodynamic effects based on a training dataset comprising data representing multiple molecular drugs represented by multiple sets of known and / or calculated physicochemical values ​​and multiple sets of PK values ​​associated with dosing regimens and targets, and labeled according to the corresponding pharmacological and / or pharmacodynamic effects, responses, or values ​​associated with the desired conditions.

[0063] In some embodiments, the ML algorithm may include one or more ML algorithms or combinations thereof that are derived from the group consisting of: neural network ML algorithms, decision tree-based classification ML algorithms, random forest decision tree ML algorithms, adaptive boosting ML algorithms, gradient boosting ML algorithms, bagging ML algorithms, ensemble ML algorithms, and / or any other supervised ML algorithm applicable to the model parameters for training the ML model accordingly.

[0064] In step 103, data representing a set of candidate molecular drugs categorized according to desired conditions associated with pharmacological and / or pharmacodynamic effects are output for downstream drug optimization processing. For example, a candidate list of molecular drugs can be output from an ML model, comprising data representing a set of candidate molecular drugs that meet or conform to desired conditions associated with pharmacological and / or pharmacodynamic effects for downstream drug optimization processing.

[0065] In some embodiments, outputting a set of candidate molecular drugs that meet desired conditions associated with pharmacological and / or pharmacodynamic effects for downstream drug optimization may include outputting a candidate list of candidate molecular drugs—each candidate molecular drug represented by a set of corresponding physicochemical values ​​and a set of PK values ​​for dosing regimens and targets—for downstream processing, including in vitro wet laboratory analysis or high-throughput screening. Alternatively or additionally, in some embodiments, outputting a set of candidate molecular drugs that meet desired conditions associated with pharmacological and / or pharmacodynamic effects for downstream drug optimization may further include outputting a candidate list of candidate molecular drugs—each candidate molecular drug represented by a set of corresponding physicochemical values ​​and a set of PK values ​​for dosing regimens and targets—for downstream processing, including in vivo testing.

[0066] If a classified list of one or more candidate molecular drugs, or a candidate list of one or more candidate molecular drugs, is represented solely by data representing a set of physicochemical values ​​and a set of PK values, and / or associated with pharmacological and / or pharmacodynamic effects, responses, or values ​​(e.g., target binding or occupancy values), then one or more real-world molecular drugs can be generated, calculated, and / or found prior to in vitro wet laboratory analysis, high-throughput screening, or in vivo testing. These real-world molecular drugs satisfy a corresponding set of physicochemical values ​​and a set of PK values ​​for the one or more candidate molecular drugs, as well as desired conditions associated with pharmacological and / or pharmacodynamic effects (e.g., target binding or occupancy values ​​satisfying target occupancy conditions). For example, the generation of one or more real-world molecular drugs can be based on: calculating, for each corresponding candidate molecular drug, the structural and chemical properties required to achieve or synthesize a real-world molecular drug that satisfies that set of corresponding physicochemical values, that set of PK values, and the associated pharmacological and / or pharmacodynamic effects (e.g., target binding or occupancy values). Therefore, a candidate list of real-world molecular drugs can be sent to downstream drug optimization processes, such as in vitro wet laboratory analysis, high-throughput screening, and / or in vivo testing.

[0067] ML models used to classify whether candidate molecular drugs satisfy one or more desired conditions associated with pharmacological and / or pharmacodynamic effects can be trained using supervised learning based on ML algorithms. Examples of ML algorithms or procedures that can be used to train the model parameters defining the ML model include, or can be based on, any ML algorithm or procedure that can train model parameters on, for example, a labeled molecular drug training dataset, representing a set of physicochemical values ​​and a set of PK values ​​associated with dosing regimens and targets, and labeled / labeled with corresponding pharmacological and / or pharmacodynamic responses / values ​​associated with one or more desired conditions associated with the corresponding pharmacological and / or pharmacodynamic effects (e.g., assigning labels to each molecular drug based on the satisfaction of one or more desired target occupancy conditions based on corresponding target binding or occupancy values). This can be used to classify whether one or more molecular drugs satisfy one or more desired conditions.

[0068] Examples of ML algorithms may include, or be based on (by way of example only but not limited to), one or more ML algorithms associated with random forests, decision tree learning, association rule learning, ML algorithms based on adaptive boosting (AdaBoost), ML algorithms based on gradient boosting, ML algorithms based on extreme gradient boosting (XGBoost), ML algorithms based on bagging, bootstrap aggregation, CoBoost, BrownBoost, data mining algorithms / methods, artificial neural networks (NNs), feedforward NNs, recursive NNs, deep NNs, deep learning, deep learning ANNs, convolutional NNs, support vector machines (SVMs), one or more combinations thereof or modifications thereof, and / or any other suitable ML algorithm depending on the application requirements. The selection of hyperparameters for training the resulting ML model may be based on performing a hyperparameter grid search across multiple hyperparameter ranges.

[0069] Figure 1b illustrates an example ML model training process 105 for training model parameters to define an ML model using an ML algorithm. This ML model is used to classify whether one or more molecular drugs meet or satisfy one or more expected conditions (e.g., one or more target occupancy conditions) associated with pharmacological or pharmacodynamic effects. Assume that an ML algorithm is chosen to classify one or more molecular drugs. The training dataset comprises multiple training data instances, each corresponding to a molecular drug-target pairing from a variety of molecular drugs and targets. This pairing can be generated based on the expected range of a set of expected physicochemical parameters of the molecular drug and target, where each molecular drug-target pairing is labeled with a corresponding pharmacological and / or pharmacodynamic response / value and one or more expected conditions. Assume that the training dataset corresponds to a specific dosing regimen. Therefore, if multiple dosing regimens are to be evaluated, multiple ML models corresponding to each dosing regimen need to be trained (e.g., an ensemble of ML models). The trained ML model, defined by the output model parameters, is then used by the ML modeling process 100 of Figure 1a. The ML model training process 105 includes iteratively training the model parameters that define the ML model using the ML algorithm associated with the ML model based on the following steps:

[0070] In step 106, a labeled training dataset associated with multiple molecular drugs is received. This training dataset includes multiple training instances, each associated with each molecular drug and including data representing a set of known and / or expected physicochemical values ​​for each molecular drug and a set of corresponding known or calculated pharmacokinetic (PK) values ​​associated with the dosing regimen and the target to which each molecular drug interacts. These training instances are labeled or tagged according to one or more expected conditions associated with pharmacological and / or pharmacodynamic effects. For example, each training instance is labeled by comparing the corresponding pharmacological and / or pharmacodynamic response / value of each molecular drug / target pair with one or more expected conditions associated with the pharmacological and / or pharmacodynamic effects. For example, a first label may be assigned to a molecular drug when the target binding or occupancy value corresponding to the molecular drug meets or satisfies the expected target occupancy condition, or a second label may be assigned to the molecular drug when the target binding or occupancy value corresponding to the molecular drug does not meet the expected target occupancy condition. Labels and / or tags may include two or more labels corresponding to two or more classification categories, which are based on the number of required target occupancy conditions and / or depend on application requirements.

[0071] In some examples, each training data instance of a molecular drug can be represented as a row in one or more lists or tables (e.g., a comma-separated file format (CSV)). If multiple tables / lists are used, each row can be associated to form each training data instance of the molecular drug. For example, the physicochemical values ​​or drug-target candidate properties of the molecular drug for each training data instance can be stored in separate columns in a first list or table, such as, but not limited to, target baseline, target half-life, and drug binding affinity (KD). This can be arranged in any order. Similarly, for each training data instance of a molecular drug, a set of corresponding PK values ​​and one or more pharmacological / pharmacodynamic values ​​(e.g., target occupancy % values) associated with the molecular drug of that training data instance can be stored in a row of another second list or table (e.g., stored in another CSV file), where separate columns represent different PK values ​​and one or more pharmacological / pharmacodynamic values ​​(e.g., target occupancy % values) from that set of PK values ​​and one or more pharmacological / pharmacodynamic values ​​(e.g., target occupancy % values). Each row in the second list or table may include additional columns for labeling based on whether one or more corresponding pharmacological / pharmacodynamic values ​​(e.g., target occupancy values) meet desired conditions associated with the corresponding pharmacological / pharmacodynamic effects. For example, the set of PK values ​​associated with a molecular drug is stored as a row in the second list / table, where each PK value of the PK attribute is stored in a separate column of that row, where multiple columns include, but are not limited to, for example: steady-state AUC; Cmin of the drug; soluble target inhibition; membrane target inhibition; minimum target occupancy in tight tissue, leaking tissue, and plasma; average target occupancy in tight tissue, leaking tissue, and plasma; maximum steady-state target occupancy in tight tissue, leaking tissue, and plasma; minimum steady-state target occupancy in tight tissue, leaking tissue, and plasma. Each pharmacological / pharmacodynamic value may also be stored in a separate column along with a corresponding column containing labels associated with whether the pharmacological / pharmacodynamic value meets desired conditions. Corresponding rows in the first list / table and the second list / table may be used or associated to form training data instances for molecular drugs.

[0072] In some embodiments, each training data instance for each molecular drug may be preprocessed and represented, but not limited to, a vector or embedding of data representing a set of known and / or expected physicochemical values ​​for each molecular drug, a set of corresponding known or calculated PK values ​​associated with the dosing regimen and the target to which each molecular drug interacts, and annotations or labels associated with corresponding pharmacological / pharmacodynamic values ​​(e.g., target binding or occupancy values) for each molecular drug.

[0073] In step 107, the labeled training dataset is iteratively applied to the ML algorithm to train the model parameters of the ML model until a stopping criterion is reached. For example, the stopping criterion may be based on one or more of a group of the following: reaching the maximum number of iterations or traversals of the training dataset; reaching an error threshold associated with the output classification.

[0074] When the ML algorithm is based on a decision tree or tree-based ML algorithm, the ML model represents a decision tree, which includes an acyclic tree structure comprising a root node, multiple decision nodes, and terminal nodes connected by labeled edges or branches. The root node includes a first rule or decision that splits into one or more decision nodes and / or terminal nodes, and each decision node includes an additional rule or decision that splits into one or more other decision nodes or terminal nodes via a labeled branch corresponding to that decision. Each terminal node outputs a classification accordingly. Stopping criteria may include, but are not limited to, predetermined constraints on the tree size, such as providing a minimum number of samples for node splits, a minimum number of samples for deploying terminal nodes, a maximum allowed tree depth (e.g., vertical depth), a maximum number of terminal nodes, a maximum number of features considered during splitting, and an impurity criterion.

[0075] In step 108, the output represents the model parameters of the trained ML model.

[0076] Figure 1c illustrates an example preliminary molecular drug screening process 110 for classifying a candidate molecular drug inference dataset using a trained ML model process 100. This dataset includes multiple molecular drugs that satisfy one or more desired conditions associated with pharmacological / pharmacodynamic effects. It is assumed that the ML model has been generated based on the ML model training process 105 of Figure 1b. The preliminary molecular drug screening process 110 includes the following steps:

[0077] In step 111, a candidate molecular drug inference dataset is received, which includes data representing multiple candidate molecular drugs. Each candidate molecular drug is represented by a set of corresponding physicochemical values ​​and a set of corresponding PK values ​​associated with the dosing regimen and target.

[0078] In step 112, each molecular drug in the candidate molecular drug inference dataset is processed to determine whether each molecular drug meets one or more desired conditions (e.g., one or more target occupancy conditions) to be included in the candidate list. For each candidate molecular drug, each candidate molecular drug in the candidate molecular drug inference dataset is selected for testing / evaluation based on the following steps:

[0079] In step 113, a set of corresponding physicochemical values ​​and a set of PK values ​​associated with the dosing regimen and the target are applied to a trained ML model (e.g., an ML model from ML model process 110 or an ML model output by process 105 of Figure 1a or Figure 1b), which is configured to classify whether the molecular drug satisfies one or more desired conditions (e.g., target occupancy conditions) specified when training the ML model.

[0080] In step 114, data representing classifications of whether a molecular drug meets one or more desired conditions (e.g., one or more target occupancy conditions) is received from the ML model.

[0081] In step 115, it is determined whether the classification indicates that the candidate molecular drug meets one or more desired conditions (e.g., target occupancy conditions) associated with the pharmacological / pharmacodynamic effects required for the preliminary screening of each candidate molecular drug. If the classification indicates that the molecular drug meets one or more desired conditions (e.g., target occupancy conditions) (e.g., Y), the process proceeds to step 116; otherwise (e.g., N), the process proceeds to step 117 to select the next molecular drug from the candidate molecular drug inference dataset. If there are no more molecular drugs in the candidate molecular drug inference dataset, the process proceeds to step 118; or if no molecular drug meets the target conditions, the preliminary screening process 110 is terminated.

[0082] In step 116, the molecular drug is added to the candidate list of candidate molecular drugs. Proceed to step 117.

[0083] In step 117, the next candidate molecular drug is selected from the candidate molecular drug reasoning dataset, and process 110 proceeds to step 113 to test the selected candidate molecular drug.

[0084] In step 118, a candidate list of candidate molecular drugs is output for use in downstream processing, such as in vitro wet laboratory analysis (including, for example, high-throughput screening laboratory analysis for efficacy and / or toxicity analysis), in vivo testing, and / or further drug optimization analysis or processes.

[0085] As previously mentioned, training datasets and candidate molecular drug inference datasets can be generated by selecting data representing molecular drugs with desired physicochemical properties, data representing one or more targets of the same type with desired physicochemical properties, and dosing regimens. However, the PK properties of molecular drug-target pairings may be unknown, unmeasured, or unavailable. Therefore, PK properties can be calculated by simulating the interaction of each molecular drug-target pairing in a physiological or cellular system. For example, PBPK models can be applied to estimate or calculate PK properties and / or one or more corresponding pharmacological / pharmacodynamic values ​​for use with one or more desired conditions for each molecular drug-target pairing with a specific dosing regimen.

[0086] Figure 1d illustrates PBPK process 120, which is used to calculate the PK properties required to generate a training dataset or candidate molecular drug inference dataset for use in Figures 1a through 1c and / or processes 100, 105, and 110 as described herein. PBPK process 120 includes the following steps:

[0087] In step 121, multiple molecular drugs are received, each represented by a set of physicochemical values. For example, a user can specify a set of physicochemical values ​​that the molecular drugs used for evaluation should meet within a desired range of a set of physicochemical properties or attributes, wherein multiple molecular drugs are generated, each represented by a unique set of physicochemical values ​​within a desired range for each physicochemical property of each molecular drug in the set of desired physicochemical properties / attributes.

[0088] In step 122, one or more sets of physicochemical values ​​corresponding to the target are received. For example, the user can specify a set of physicochemical values ​​that the target used for evaluation should meet within a desired range of a set of physicochemical properties or attributes, wherein multiple targets are generated, each target being represented by a unique set of physicochemical values ​​within the desired range of each physicochemical property of each target in the set of desired physicochemical properties / attributes.

[0089] In step 123, a set of dosing regimens is received. These can be specified by the user in an appropriate dosing regimen format.

[0090] In step 124, for each desired dosing regimen, one or more unique molecular drug-target data pairs are generated based on pairing a set of desired physicochemical values ​​for each molecular drug with each of a set of or more sets of desired physicochemical values ​​for the target.

[0091] In step 125, for each dosing regimen, a set of expected physicochemical values ​​for the drug and a set of expected physicochemical values ​​for the target are applied to the PBPK model in each drug-target data pairing.

[0092] In step 126, for each dosing regimen and each drug-to-target data pairing, a PBPK model is used to calculate a set of PK values ​​and corresponding pharmacological / pharmacodynamic values ​​(e.g., target binding or occupancy values) associated with each drug-to-target data pairing.

[0093] In step 127, for each drug-molecule paired with a target and its corresponding dosing regimen, add: a) a training data instance for the training dataset; or b) a candidate drug-molecule inference data instance for the candidate drug-molecule inference dataset. Each training data instance includes data representing at least a set of physicochemical properties for each drug-molecule, a set of calculated PK values ​​and pharmacological / pharmacodynamic values ​​calculated for each drug-molecule pair, and is labeled according to one or more desired conditions associated with the calculated pharmacological / pharmacodynamic values ​​(e.g., target binding or occupancy values). Each candidate drug-molecule data instance includes data representing at least a set of physicochemical properties for each drug-molecule and a set of calculated PK values ​​calculated for each drug-molecule paired with a target and its corresponding dosing regimen.

[0094] In step 128, training data instances are stored in the training dataset or candidate molecular drug inference data instances are stored in the candidate molecular drug inference dataset for use in training or inference / classification, respectively.

[0095] For example, data representing each selected molecular drug includes a set of corresponding physicochemical values ​​for that molecular drug, wherein these values ​​fall within a range of a set of physicochemical parameters set or selected by the user for the molecular drug being evaluated. Data representing each selected target includes a set of corresponding physicochemical values ​​for that target, wherein these values ​​fall within a range of a set of physicochemical parameters set or selected by the user for the target being evaluated. This can result in multiple molecular drugs and multiple targets.

[0096] Training datasets and candidate molecular drug inference datasets can be generated by selecting data representing molecular drugs with desired physicochemical properties, data representing one or more targets of the same type with desired physicochemical properties, and dosing regimens, and calculating and / or measuring the PK characteristics of the corresponding molecular drug-target pairing interactions. For example, data representing each selected molecular drug includes a set of corresponding physicochemical values ​​representing that molecular drug, wherein these values ​​are within a range of a set of physicochemical parameters set or selected by the user for the molecular drug being evaluated. Data representing each selected target includes a set of corresponding physicochemical values ​​representing that target, wherein these values ​​are within a range of a set of physicochemical parameters set or selected by the user for the target being evaluated. This can generate multiple molecular drugs and multiple targets.

[0097] Once generated, multiple unique drug-to-target pairings for each dosing regimen are derived from data representing the selected drug molecule and one or more targets. If the PK characteristics of a drug-to-target pairing are unknown or unmeasured, a physiology-based pharmacokinetic (PBPK) model—configured to simulate the interaction of each drug molecule with the target in a modeled physiological system according to the dosing regimen—can be applied to output the PK characteristics of that interaction. PBPK models represent drug-to-target interactions through a variety of mathematical relationships: cellular and tissue systems within a physiological or cellular system; and the interaction of each drug molecule with the target within that physiological or cellular system according to the dosing regimen. PBPK models calculate or estimate a set of PK values ​​for the drug molecule associated with the target and the dosing regimen. The output set of PK values ​​depends on the type of PBPK model and is based on a set of PK characteristic curves and / or endpoints. Furthermore, PBPK models can also be configured to calculate or estimate corresponding pharmacological / pharmacodynamic values ​​(e.g., target binding / occupancy values) as defined and calculated by the PBPK model. Pharmacological / pharmacodynamic values ​​can be used to label each molecular drug-target pairing in the training dataset against one or more desired conditions associated with the corresponding pharmacological / pharmacodynamic effects.

[0098] Alternatively, a training dataset can be generated comprising multiple training data instances, each containing data representing a drug-to-target pairing for each of a plurality of molecular drugs and each of a plurality of targets, and labeled according to corresponding calculated pharmacological / pharmacodynamic values ​​and desired conditions. The training data instances representing drug-to-target pairings include, for example: a set of physicochemical values ​​representing the molecular drug; a set of pharmacokinetic (PK) values ​​of the molecular drug associated with the target and drug dosing regimen; and corresponding labels for the pharmacological / pharmacodynamic values ​​(e.g., target binding or occupancy values) associated with one or more desired conditions related to the corresponding pharmacological / pharmacodynamic effects. Alternatively, a candidate molecular drug inference dataset can be generated comprising multiple candidate molecular drug data instances, each containing data representing a drug-to-target pairing for each of a plurality of molecular drugs and each of a plurality of targets. Each candidate drug data instance includes, for example, a set of physicochemical values ​​representing the candidate drug, and a set of PK values ​​associated with the target and drug administration regimen.

[0099] Each of processes 100, 105, 110, and 120 in Figures 1a to 1d can be implemented on a device including a processor, a memory unit, and a communication interface. The processor is connected to the memory unit and the communication interface. The processor and memory can be configured to implement processes 100, 105, 110, and 120 and / or each of other processes described herein as computer-implemented methods. Processes 100, 105, 110, 120 and / or combinations thereof, and modifications to these processes based on one or more other processes as described herein, can be stored on a computer-readable medium or a non-transitory computer-readable medium. The computer-readable medium may include data or instruction code that, when executed on a processor, causes the processor to implement processes 100, 105, 110, 120 and / or combinations thereof, and one or more modifications to these processes as described herein as one or more computer-implemented methods.

[0100] In some embodiments, the molecular drugs in the training dataset and / or candidate molecular drug inference dataset are monoclonal antibodies (mAbs) and / or mAb fragments. However, any type of antibody can be used. It may be preferred that the mAb be a macromolecular drug.

[0101] Although the following description of the PBPK and ML modeling system is based on processes 100, 105, 110 and 120 of Figures 1a to 1d and their ML models, this is only an example and the invention is not limited thereto. Those skilled in the art will understand that the principles described herein regarding the use of PBPK models and the generation of ML models can be applied to any type of molecular drug and target pairing, depending on the application requirements.

[0102] Figure 2a illustrates an example PBPK and ML model system 200, used to generate a training dataset and / or a candidate molecular drug inference dataset using a PBPK model; to train an ML model using the training dataset to classify whether candidate molecular drugs satisfy desired conditions associated with pharmacological and / or pharmacodynamic effects; and to classify the candidate molecular drug inference dataset accordingly using the trained ML model. In this example, the pharmacological and / or pharmacodynamic effects used are target binding or target occupancy, and the one or more desired conditions are one or more desired target occupancy conditions. In this example, the candidate molecular drug is a monoclonal antibody (mAb), the target is an antigen (Ag), and the desired target occupancy condition classifies whether the target occupancy value (TO%) of each candidate molecular drug is greater than or less than 90% (e.g., TO% ≥ 90%). Although the expected target occupancy condition is described as TO% ≥ 90%, this is merely an example, and the expected target occupancy condition is not limited to this. Those skilled in the art will understand that any type of target occupancy condition can be defined, such as a TO value greater than or equal to a first TO threshold, a TO value less than or equal to a second TO threshold, and / or any one or more rules associated with a TO value / threshold suitable for preliminary screening of candidate drugs for downstream drug optimization, etc. In this example, each candidate mAb is represented by a set of physicochemical properties, including but not limited to, for example, MW, antibody charge (Q), and binding constant (KD), and each target Ag is represented by a set of physicochemical properties, including but not limited to, for example, target half-life (…). ) and target baseline ( The PBPK and ML model system 200 combines PBPK model 202a and ML model 203b to infer optimal drug and target properties that achieve the desired target binding and occupancy. This can be used to determine the physicochemical properties of macromolecular drugs (e.g., mAbs) such as charge and drug-target binding. ADME characteristics and target attributes (such as target baseline) ),half life( The relationship between (and target receptor forms).

[0103] The PBPK and ML system 200 includes a molecular generation unit 201, a PBPK modeling unit 202, and an ML classification unit 203. The PBPK and ML system 200 is configured to perform model-based evaluation of target-binding responses of antibody candidates (candidate molecular drugs) by integrating a PBPK model 202a with a tree-based ML model 203a. In the PBPK and ML system 200, the molecular generation unit 201 generates multiple molecular drugs represented by physicochemical values ​​201a-1 to 201a-n, and multiple targets represented by target physicochemical values ​​201b-1 to 201b-m. In this case, the molecular drug is mAb, and the target is Ag. The molecular-target generation unit 201 is configured to generate multiple unique synthetic drug-target candidate pairings with different physicochemical properties. Once generated, each drug-target candidate pair is applied to the PBPK computation unit 202, where the PBPK model 202a is configured to generate PK values ​​(such as PK endpoints and characteristic curves 202b) and target binding / occupancy percentage values ​​202c for one or more scenarios of interest (e.g., one or more dosing regimens and / or targets 202d (e.g., one or more fixed doses, one or more fixed dosing schedules, fixed charges, etc.)). The ML classification unit 203 is configured to train one or more ML models 203b corresponding to one or more dosing regimens or scenarios 202d based on one or more classification-desired target occupancy conditions 203a. In this example, ML model 203b is an interpretable decision tree-based ML model trained by a decision tree-based ML algorithm against a desired target occupancy condition 203a, which is associated with classifying each candidate molecule based on whether the target binding / occupancy is greater than or less than a target occupancy threshold (e.g., TO% > 90% or TO% ≤ 90%; or TO% > 99% or TO% ≤ 99%). The decision tree-based ML model 203b is the classifier. The resulting interpretable decision tree-based ML model 203b is configured to classify data representing the physicochemical and pK values ​​of one or more molecule drugs against the desired target occupancy condition 203a for a specific dosing regimen / scenario 202d. Physicochemical properties of the drug-target, such as binding affinity, baseline concentration, target half-life, dose, dosing schedule, and antibody charge, are selected to help identify the combination of molecule drug physicochemical properties most likely to exhibit the desired target binding response / target occupancy.

[0104] In operation, the molecule-target generation unit 201 generates multiple synthetic candidate pairings of macromolecular drugs 201a-1 to 201a-n and their specific targets 201b-1 to 201b-m by modifying a set of physicochemical properties (such as dose, charge, target binding affinity, target baseline, and target half-life) to cover the corresponding selected or desired range. Each virtual candidate pairing 201a-1, 201b-1 represents a unique drug-target interaction. The PBPK calculation unit 202 uses the PBPK model 202a (e.g., Figure 3 The example mPBPK model is described to simulate and thus calculate the PK characteristic curve and PK endpoint 202b for each virtual candidate pair, as well as the pharmacological / pharmacodynamic effects, for calculating the target binding response, expressed as a percentage of target occupancy (TO%), for each corresponding selected scenario or dosing regimen 203d. The extent of the target binding response for each virtual candidate pair is assessed by calculating the percentage of target occupancy (TO%). The PBPK model 202a (or mPBPK model) can simulate and predict antibody PK characteristic curves, PK endpoints, and TO%, where TO% is calculated as the ratio of the bound target to the total target in plasma, tight tissue, and leaky tissue, as described in the following equation:

[0105] ,

[0106] in, It is the percentage of the target in the plasma (%). It is the concentration of the drug-target complex in plasma, and This is the concentration of the target in the plasma. Although the above equation describes a specific calculation of TO%, this is only an example, and the invention is not limited thereto. Those skilled in the art should understand that the PBPK model 202a can be further modified to calculate TO% using different parameters, values, and / or approximations, etc., where appropriate.

[0107] A supervised tree-based machine learning algorithm is used to train the tree-based ML model 203b to classify each virtual candidate (or drug-target interaction) as optimal or non-optimal based on the calculated TO% used to label each training data instance against the desired target occupancy condition (e.g., assigning a first label to the training data instance if the corresponding TO% > 90%; or assigning a second label to the training data instance if the corresponding TO% <= 90%). In this case, the classifier's target condition is whether the molecular drug associated with the target has a TO% > 90% or a TO% <= 90%. The tree-based ML model 203b helps infer the optimal combination of drug and target attributes that is more likely to provide the desired PK / TO endpoint and ultimately deliver the desired pharmacological effect.

[0108] As described above, by altering properties such as dose, charge, target binding affinity, target baseline, and target half-life, the molecule-target generation unit 201 can be used to generate numerous synthetic candidate pairings of macromolecular drugs and their specific targets. Each virtual synthetic candidate represents a unique drug-target interaction. The mPBPK model 202a is configured to simulate the PK characteristic curve and PK endpoint for each virtual synthetic candidate pairing. The degree of target binding response for each virtual candidate is evaluated by calculating the target occupancy percentage (TO%) based on the PK characteristic curve and PK endpoint generated by the mPBPK model simulation. When trained with supervised learning, the tree-based ML model 203b is used to classify each virtual candidate (or drug-target interaction) as optimal or non-optimal based on the TO% calculated as described herein. The ML model 203b helps infer the optimal combination of drug and target properties that are more likely to provide the desired PK / TO endpoint and ultimately the desired pharmacological effect.

[0109] Synthetic data is used to train the ML model 203a classifier for each of the different dose variations, dosing frequencies, antibody charge variations, and across different receptor forms (soluble or membrane-bound). This can be used to form an ensemble of ML models for classifying drug-target candidate pairings for one or more different dose variations, dosing frequencies, antibody charge variations, and across different receptor forms.

[0110] As an example, in the table below, the list of scenarios used to generate data for developing ML models is represented as a list of mPBPK model simulations for obtaining the PK endpoint of synthetic candidates (n = 10000):

[0111]

[0112] In the table above, the charge column represents the net surface charge on the antibody candidate; the dose column represents the administered antibody dose; the regimen column represents the dosing schedule, where IV refers to intravenous administration, Q1W refers to the dose administered every 1 week, Q2W refers to the dose administered every 2 weeks, and Q4W refers to the dose administered every 4 weeks; the criteria column represents the PK endpoint used for ML classification, where... Target occupancy (TO)% is calculated at the maximum drug concentration. The target occupancy (TO)% is calculated at the minimum drug concentration. Synthetic molecule pairings, along with PK endpoints (including target occupancy (TO%)), can be used to generate labeled training datasets for training ML models for each different scenario. These ML models can be combined to form ensemble ML models for classifying unknown candidate molecule-target pairings (e.g., synthetic molecule-target pairings or real-world molecule-target pairings).

[0113] Figure 2b illustrates an example trained tree-based ML model 210 based on an interpretable decision tree-based ML model 203b. The tree-based ML model 210 includes a root node 211 connected via edges to multiple decision nodes 212a to 212n and multiple terminal nodes 213a to 213m. These nodes are used by the trained tree-based ML model 210 to classify one or more molecular drugs associated with a target for dosing regimens. In the example, an ensemble of tree-based ML models can be used to classify one or more molecular drugs associated with a target for multiple dosing regimens, with each tree-based ML model in the ensemble associated with a specific dosing regimen among the multiple dosing regimens.

[0114] Referring to Figures 2a and 2b, the corresponding multiple sets of PK value groups 202b and target binding response / target occupancy value groups 202c for synthetic molecule-target pairings 201a-1 to 201a-n and 201b-1 to 201b-m, as well as various dosing regimens 202d, can be used to form a labeled training dataset for training ML models against different dose variations, dosing frequencies, antibody charge variations, and across different forms of receptors (soluble or membrane-bound). For example, the following example ML training dataset table shows a portion of the labeled training dataset, where only a subset of molecule-target pairings (e.g., only 5 pairs) are shown, with each row representing a different molecule-target pairing. In this example, only a subset of the physicochemical values ​​of molecule-target pairings are shown, and these physicochemical values ​​include, but are not limited to, target baselines (…). ), target half-life ( ), and the drug binding constant ( Similarly, each row was labeled / tagged (e.g., '1', '0') based on a comparison of the target occupancy value of the drug-target pairing with the target against two desired target occupancy conditions (e.g., TO% ≤ 90% and TO% > 90%). The label for each row was '1' when the target occupancy of the drug-target pairing was greater than 90%, and '0' when the target occupancy of the drug-target pairing was less than or equal to 90%.

[0115]

[0116] Therefore, labeled training datasets are used to train ML models for each different scenario, and these ML models can be combined to form an ensemble ML model for classifying unknown candidate molecule-target pairs (e.g., synthetic molecule-target pairs or real-world molecule-target pairs, etc.).

[0117] To generate a trained tree-based ML model 210 for a specific scenario (e.g., a specific dosing regimen), a tree-based ML algorithm is used to train the model parameters of the ML model using a labeled training dataset generated for that specific scenario. For example, when trained using 10,000 example synthetic drug-target candidate pairs, the mPBPK model 203a models the antibody as being administered via the intravenous (IV) route at a dose of 1 mg / kg every 2 weeks (Q2W) and binding only to the receptor in its soluble form. The decision tree-based models 203b / 210 target each input feature associated with a drug-target candidate ( Using an appropriate cutoff value, the drug-target candidate is classified into an optimal category (e.g., TO% > 90%) or a non-optimal category (e.g., TO% ≤ 90%) based on the TO% value observed at the minimum concentration in plasma under steady state. In this example, although the input features or attributes for drug-target candidate pairing include binding constant values ​​( or KD), target baseline ( ), and target half-life ( However, this is merely an example, and these need not be so limited. Those skilled in the art should understand that any set of input characteristics or physicochemical properties associated with a drug-target candidate can be considered, including but not limited to MW, antibody charge (Q), and binding constant (KD), target half-life (…). ) and target baseline ( (and / or any other suitable physicochemical properties depending on the application requirements.)

[0118] In this case, based on input features ( The decision tree-based model 203b / 210 was trained using 10,000 example synthetic drug-target candidate pairings, where antibodies were modeled as being administered via intravenous (IV) route at a dose of 1 mg / kg every 2 weeks (Q2W) and binding only to the receptor in its soluble form. The decision tree model 203b / 210 for binary classification was built using a labeled training dataset based on the 10,000 example synthetic drug-target candidate pairings. In this example, the decision tree algorithm used was based on, but not limited to, the sklearn package in Python, using default hyperparameters except for max_depth = 5, criterion = 'gini', min_samples_split = 3, splitter = 'best', and class_weight = 'balanced'. The resulting trained decision tree-based model 203b / 210 includes a set of rules derived from ML, generated by training the decision tree algorithm using a labeled training dataset based on 10,000 example synthetic drug-target candidate pairings labeled with one or more desired TO% conditions or other desired conditions associated with pharmacological / pharmacodynamic effects. In this case, the trained decision tree-based model 203b / 210 is configured to apply input features of the drug-target candidates (… These drug-target candidates are classified, and the following ML-derived rules generated using a decision tree algorithm are used to output a classification of either the best category (e.g., TO% > 90%) or a non-best category (e.g., TO% ≤ 90%):

[0119] If (KD <= 8.965) and ( <= 4.047) and ( If (KD <= 2.261) and (KD <= 8.503) and (KD <= 7.943), then the category is: Best (probability: 99.94%) | based on 3,605 samples;

[0120] If (KD > 8.965) and (KD > 8.988), then the category is: Non-optimal (probability: 100.0%) | based on 2,481 samples;

[0121] If (KD <= 8.965) and ( > 4.047) and ( <= 19.916) and ( > 9.114) and ( If <= 14.208), then the category is: Non-optimal (probability: 100.0%) | based on 1,152 samples;

[0122] If (KD <= 8.965) and ( > 4.047) and ( > 19.916) and ( <= 92.58) and ( If the probability is greater than 40.74, then the category is: Best (probability: 96.02%) | based on 251 samples;

[0123] If (KD <= 8.965) and ( > 4.047) and ( <= 19.916) and ( <= 9.114) and ( If <= 5.457), then the category is: Not optimal (probability: 97.14%) | based on 175 samples;

[0124] If (KD <= 8.965) and ( > 4.047) and ( > 19.916) and ( > 92.58) and ( If <= 137.423), then the category is: Non-optimal (probability: 100.0%) | based on 151 samples;

[0125] If (KD <= 8.965) and ( <= 4.047) and ( > 2.261) and ( If KD <= 7.626 and the value is greater than 2.182, then the category is: Best (probability: 100.0%) | Based on 136 samples

[0126] If (KD <= 8.965) and ( > 4.047) and ( > 19.916) and ( <= 92.58) and ( If <= 40.74), then the category is: Best (probability: 68.76%) | based on 80 samples;

[0127] If (KD <= 8.965) and ( > 4.047) and ( <= 19.916) and ( > 9.114) and ( If the value is > 14.208, then the category is: Not optimal (probability: 94.12%) | based on 68 samples;

[0128] If (KD <= 8.965) and ( <= 4.047) and (T0 > 2.261) and ( <= 2.182) and ( If the value is greater than 2.628, then the category is: Non-optimal (probability: 88.23%) | based on 68 samples;

[0129] If (KD <= 8.965) and ( > 4.047) and ( > 19.916) and ( > 92.58) and ( If the value is greater than 137.423, then the category is: Non-optimal (probability: 64.99%) | based on 60 samples;

[0130] If (KD <= 8.965) and ( <= 4.047) and ( > 2.261) and ( <= 2.182) and ( If <= 2.628), then the category is: Best (probability: 64.45%) | based on 45 samples;

[0131] If (KD <= 8.965) and ( > 4.047) and ( <= 19.916) and ( <= 9.114) and ( If the probability is > 5.457, then the category is: Best (probability: 90.7%) | based on 43 samples;

[0132] If (KD <= 8.965) and ( <= 4.047) and ( If (KD <= 2.261) and (KD <= 8.503) and (KD > 7.943), then the category is: Best (probability: 92.0%) | based on 25 samples;

[0133] If (KD <= 8.965) and ( <= 4.047) and ( <= 2.261) and (KD > 8.503) and ( If <= 0.28), then the category is: Best (probability: 100.0%) | based on 18 samples;

[0134] If (KD <= 8.965) and ( <= 4.047) and ( <= 2.261) and (KD > 8.503) and ( If the value is greater than 0.28, then the category is: Non-optimal (probability: 100.0%) | Based on 6 samples;

[0135] If (KD <= 8.965) and ( <= 4.047) and ( > 2.261) and ( If KD > 2.182 and KD > 7.626, then the category is: Best (probability: 50.01%) | based on 2 samples;

[0136] Although example rules for trained decision tree-based models 203b / 210 have been described above, these are merely examples, and trained decision tree-based models 203b / 210 are not limited thereto. Those skilled in the art should understand that the rules derived from the ML of trained decision tree-based models will vary depending on, but not limited to, the following: the type of molecule and its physicochemical properties, the target and its physicochemical properties, the PK value associated with the drug-molecule-target pair, the pharmacological / pharmacodynamic value associated with the drug-molecule-target pair, the expected conditions associated with the pharmacological / pharmacodynamic effects, the type of PBPK model, the dosing regimen and / or other scenarios, the training dataset, the number of training data instances, the model hyperparameters, the type of ML model algorithm, their combinations, modifications thereof, and / or variations according to application requirements.

[0137] Figure 2c shows a set of example figures 220 obtained using decision tree-based models 203b or 210 of Figures 2a and / or 2b (when trained with 10,000 synthetic drug-target candidate pairs), where the antibody is modeled as being administered via intravenous (IV) route at 1 mg / kg every 2 weeks (Q2W) and binding only to the receptor in its soluble form. Decision tree-based model 203b is designed for each input feature ( Using appropriate cutoff values, each virtual drug-target candidate is classified into an optimal category (large circles in Figures 222a, 224a, 226a) or a non-optimal category (small circles in Figures 222a, 224a, 226a) based on the TO% value observed at the minimum concentration in plasma under steady state. As shown in Figures 222a to 226a and 222b to 226b in Figure 2c, the decision tree-based model 203b indicates that for a class of soluble targets with a baseline value below 4 nM in plasma, when the binding constant ( When the concentration is below 9 nM, a drug administered at 1 mg / kg every two weeks should have a target binding / occupancy rate greater than 90%. In this example, the tree rule used for optimal class prediction was selected based on a Gini impurity close to 0 and an accuracy close to 100%. It is important to select the optimal rule with the highest accuracy, as this increases the likelihood of making true predictions on other unseen datasets. In this scenario, and (That is, combining constant values ​​and target baseline) is the most important physicochemical property for predicting over 90% target occupancy. Physicochemical parameters (i.e., target half-life) is shown to have the least impact on decision-making in this scenario. However, a class of soluble receptors with low half-lives and high baselines may be the least suitable combination of target attributes for achieving a target occupancy of greater than 90% in plasma. Figure 2b presents the binary classification trees of the trained decision tree-based model 203b for the above scenario. Each tree model was evaluated on a separate validation dataset based on the following evaluation metrics, such as classification metrics, confusion matrix, and area under the receiver operating curve (ROC).

[0138] The table below also shows the performance evaluation of tree-based ML models across various types of tree-based ML algorithms. In this example, these algorithms include decision trees, random forests, and gradient boosting ML classifier algorithms. In these examples, the decision trees are each trained for the same scenario (e.g., as indicated in Figure 2c above). The decision tree model for binary classification is built using the sklearn package in Python with default hyperparameters, except for max_depth = 5, criterion = 'gini', min_samples_split = 3, splitter = 'best', and class_weight = 'balanced'. Although a specific set of hyperparameters is described here, this is only an example and the invention is not limited thereto. Those skilled in the art should understand that different hyperparameters can be selected, set, and / or used to train tree-based ML models for binary and / or multi-class classification and / or the same or other types of scenarios and datasets, depending on the application requirements. Although a tree-based ML model has been described, this is merely an example and the invention is not limited thereto. Those skilled in the art should understand that any other suitable ML algorithm can be used to train a suitable ML model to perform binary and / or multi-class classification for the same scenario and / or dataset in various other scenarios and / or datasets, depending on the application requirements.

[0139] Training performance metrics (shown in the example training metric table below) include accuracy, precision, and F1 score, which are the average percentage of 5-fold cross-validation while training each algorithm. The standard deviation (SD) of each fold cross-validation is provided in curly braces. The example test metric table includes test metrics used to evaluate the model's performance on the test dataset, including accuracy, precision, and F1 score.

[0140]

[0141]

[0142] Figure 2d illustrates four figures, 230a, 230b, 230c, and 230d, which demonstrate the optimal drug-target properties for target engagement in different scenarios related to dosage and dosing regimens. In this example, a separate tree-based ML model (e.g., a decision tree) was trained for each scenario based on the tree-based ML model described with reference to any of Figures 1a through 2c. For synthetic data, separate tree-based ML models were trained and validated for each different scenario (including administration at different doses of 0.1 mg / kg, 1 mg / kg, and 10 mg / kg every 2 weeks (Q2W)), resulting in Figures 230a and 230b. Also for synthetic data, separate tree-based ML models were trained and validated for each different scenario based on administration at a fixed dose of 1 mg / kg once weekly (Q1W), once every two weeks (Q2W), and once every 4 weeks (Q4W), with the results shown in Figures 230c and 230d. In both cases, only the optimal property that satisfies the following condition is shown: target occupancy > 90% at the minimum plasma concentration in steady state.

[0143] These examples illustrate the differences in optimal rules and drug-target attribute values ​​under different doses and dosing regimens / plans. In Figures 230a and 230b, the corresponding trained decision tree-based ML models for each scenario provide drug-target binding constants at bolus and repeated IV doses of 0.1 mg / kg, 1 mg / kg, and 10 mg / kg. The optimal range for target baseline and target half-life values. The optimal rule for low-dose (0.1 mg / kg) scenarios indicates that... Values ​​below 1 nM, as shown in Figures 230a and 230b. For a class of soluble receptors with baseline values ​​below 2 nM, this low-dose administration of the drug can achieve 90% target occupancy. As shown in Figures 230a and 230b, optimal results were achieved when the dose was increased from 0.1 mg / kg to 1 mg / kg. The cutoff value increased to 9 nM, while the target baseline value increased to 4 nM. Similarly, as shown in Figures 230a and 230b, at a high dose of 10 mg / kg, The optimal cutoff value increased to 82 nM. At lower doses, drug clearance was faster, and target-mediated treatment was performed. Therefore, to achieve target occupancy greater than 90% at such low doses, drug-target affinity should be higher. The value should be lower). The best rule suggests that... Target baseline concentration and target half-life are relatively more important for classifying a given drug-target pairing based on target binding response. Target baseline concentration and target half-life are relevant target properties that determine drug-target interactions and desired target pharmacology. However, target half-life appears to be the least influential factor in the decision-making process for achieving target occupancy rates exceeding 90% at different doses.

[0144] Referring to Figures 230c and 230d, the effect of the dosing regimen on the optimal TO% was observed when the drug was administered at a repeated dose of 1 mg / kg once every 1 week (Q1W), once every 2 weeks (Q2W), and once every 4 weeks (Q4W). Similarly, a trained decision tree-based ML model was trained for these different regimen scenarios. The resulting tree-based ML rules indicate that as the dosing frequency increases, , and The optimal range also increases. The effect of dosing frequency on optimal drug and target properties can be observed in density maps 230c and 230d. For Q4W, Q2W, and Q1W dosing schedules, values ​​below 4.25, 9, and 15 can be observed, respectively. Values. As can be seen, for a given class of soluble receptors with baseline values ​​below 2 nM, 4 nM, and 8.3 nM, respectively, around 4 nM... The value will provide a target occupancy rate greater than 90%. As can be seen from Figure 230d, the target half-life has the least dependence on the dosing schedule / regimen scenario, and it is insufficient to determine the optimal category based on it alone.

[0145] Figure 2e shows two figures, 232a and 232b, used to validate the rules derived from the trained tree-based ML model 203b / 210 as described with reference to Figures 2b and 2c. In one study, the rules derived from the ML were validated using real-world compound / molecule data of clinically approved monoclonal antibodies that could be classified as macromolecules. The table below provides a list of literature-derived properties of approved monoclonal antibodies and their corresponding targets, as well as a list of recommended clinical doses and regimens used in this study:

[0146] Experimental data for approved monospecific antibodies.

[0147]

[0148] For simplicity, assumptions were made regarding the dosage to match the dosage and regimen to the 0.1 mg / kg, 1 mg / kg and 10 mg / kg used in the trained ML model 203b / 210.

[0149] In this study, the trained ML model 203b / 210 was able to predict fixed doses and regimens for intravenous administration. Given this limitation of the ML model 203b / 210, some necessary and reasonable assumptions were made to validate it against real-world data. For example, instead of directly using the recommended clinical doses presented in the table above, the recommended clinical doses for each antibody were matched (e.g., rounded) to the closest dose among the options of 0.1 mg / kg, 1 mg / kg, or 10 mg / kg. For example, the clinical dose of lerizobactam is 1.8 mg / kg, which means the closest dose to 0.1 mg / kg, 1.0 mg / kg, or 10 mg / kg is 1.0 mg / kg. Similarly, the clinical dose of balithimab is 20 mg / kg, which means the closest dose to 0.1 mg / kg, 1.0 mg / kg, or 10 mg / kg is 10 mg / kg. For each dosage / regimen scenario, a trained ML model 203b / 210 was used to predict the optimal dosage and recommended regimen for each antibody under intravenous administration. Values. As shown in Figures 232a and 232b, the optimal drug-target properties were validated for the following scenarios: 0.1 mg / kg bolus injection, 1 mg / kg Q4W, 1 mg / kg Q2W, and 10 mg / kg Q2W.

[0150] Real-world compounds / molecules of levozomab 233, anilumab 234, rituximab 235, abciximab 236, balithimab 237, adunab 238, and bevacizumab 239 are plotted in Figures 232a and 232b. Most antibodies, when administered at their respective doses and dosing regimens, exhibit experimental properties within the predicted optimal boundary of drug-target properties. As shown in Figure 232a, most real-world compounds / molecules 233, 234, 235, 237, 238, and 239 have targets falling within the predicted optimal boundary. Values, except for abciximab 236, where abciximab's target... The value fell outside the predicted optimal limit. For abciximab 236, this discrepancy may be due to the assumed dose being lower than the clinical dose (0.25 mg / kg bolus). For doses greater than 0.1 mg / kg, The optimal value will be higher, as can be seen from the dose dependence in Figure 2e. Furthermore, as shown in Figure 232b, the experimental observations for each of the control antibodies 233 to 239 show the optimal value for each scenario. The values ​​were well validated. Based on the examples (in which more compound / molecule data would help increase the confidence of the ML predictions), these preliminary real-world validations offer the advantage of validating the accuracy of the rules derived from ML.

[0151] Figure 2f shows the resulting pairwise scatter plots 240a, 240b, and 240c of the tree-based ML model 203b, which was trained to perform multi-label classification, specifically based on target occupancy percentage (TO%) < 50%, between 50% and 90%, and > 90%. In this case, as described herein, the tree-based ML model 203b was trained using synthetic data from drug administration at 1 mg / kg every two weeks (Q2W) to classify candidate molecular drugs based on whether they met one of several target occupancy criteria (e.g., TO% < 50%, TO% between 50% and 90%, TO% > 90%). In this case, ML model 203b was trained to classify candidate molecular drug data into three categories based on the TO% at the minimum concentration in plasma at steady state. The results in Figure 2f, Figures 240a, 240b, and 240c, illustrate the optimal regions for selecting drug-target binding constants for a given target with baseline and half-life to achieve low, intermediate, and high target occupancy. As observed, there are clear boundaries between the three categories of high target occupancy 242a, medium target occupancy 242b, and low target occupancy 242c, which can be used to distinguish the virtual drug-target attributes of each category. In this case, the trained tree-based ML model 203b has 99% accuracy, and 99%, 98%, and 99% category accuracy for low category 242c, medium category 242b, and high category 242a, respectively.

[0152] As can be seen in Figures 2a to 2f, decision tree-based classification models and / or other rule-based ML models offer several advantages, including fast training, efficient computational load during inference, and interpretability due to their rule-based interpretable nature; tree-based ML models can use recursive partitioning in a tree-like manner to identify the relationship between input drug-molecule-target pairs (e.g., drug-molecule-target attributes) and output target occupancy values ​​(TO%). This interpretable type of ML model depicts the process of obtaining predicted categories from given inputs, providing an advantage over other black-box machine learning algorithms (such as artificial neural networks).

[0153] Figure 2g illustrates another example of the PBPK / ML model system 200 based on Figure 2a, namely PBPK / ML model system 250, demonstrating further applications for drug optimization and downstream processes. For example, PBPK / ML model system 250 enables researchers to screen and optimize a large number of antibody candidates based on optimal ADME, PK, and target binding against specific antigens, reducing lead candidate attrition rates and shortening overall discovery time. PBPK / ML model systems 200 and 250 provide computer-simulated high-throughput screening, which helps significantly accelerate the hit-to-lead rate to identify which candidate drug-target pairings can be advanced to synthesis into real-world molecular drugs and subsequently downstream in vitro wet laboratory analysis and / or in vivo assays. This computer-simulated high-throughput screening and assessment of macromolecular drug properties allows for candidate evaluation at a very early stage for one or more pharmacological / pharmacodynamic effects (e.g., target occupancy and binding), significantly narrowing or reducing the candidate drug / target space within the drug discovery and development pipeline. This further significantly shortens the development cycle from candidate to real-world molecular medicine for treatment.

[0154] For example, some specific applications of the PBPK / ML modeling system 200 and / or 250 include one or more of the following:

[0155] a) Lead candidate generation, wherein the PBPK / ML model system 250 can be configured to determine optimal candidate properties for desired conditions associated with pharmacological / pharmacodynamic effects (e.g., target binding), which are required to advance from hit to lead candidate stage. This computer simulation high-throughput screening technique enables scientists to narrow down and generate a candidate list 251 of the most promising drug candidates based on a combination of structure-based properties, ADME, PK, and pharmacological / pharmacodynamic values ​​(e.g., TO). This candidate list 251 can be used to automatically or semi-automatically generate or synthesize real-world drug molecules with the properties of the drug molecules on the candidate list 251 for use in performing real-world in vitro wet laboratory analyses and / or in vivo assays, etc. The generated candidate list can significantly reduce the number of drug molecules that may be experimentally evaluated in the laboratory, further reducing development costs. For example, such a candidate list can be applied by the PBPK / ML model system 250 to medicinal chemistry, where prior knowledge of dose-driven antibody design is required. In order to prioritize candidate molecular drugs that have the greatest potential to achieve, for example, target binding and enter the lead phase at the lowest (or minimum) dose;

[0156] b) The PBPK / ML model system 250 can be configured or adapted to provide scientists with a tabular list 252 that lists all drug candidate attributes and specific target attribute combinations of candidates that achieve specific desired conditions associated with corresponding pharmacological / pharmacodynamic effects (e.g., target binding / occupancy greater than 90%). Users (e.g., scientists) can use this list as an initial reference table to focus on the most promising candidates.

[0157] c) The PBPK / ML model system 250 can be configured for research teams to perform in vitro ADME screening during the process from hit to lead or early lead optimization, and to visualize the optimal space of candidate properties to define the most critical combination of properties for screening.

[0158] d) The PBPK / ML model system 250 can be configured to support in vitro / in vivo correlation (IVIVC) work, where the current framework provides a platform that can incorporate more in vitro-derived properties (in addition to the binding constant KD), such as charge, FcRn binding, etc., and correlate these drug properties with corresponding pharmacological / pharmacodynamic effects (e.g., TO% and PK-based clearance (CL)). This attribute-based (human) clearance prediction is an important endpoint during IVIVC studies to inform recommendations for termination, study, and / or advancing the candidate pipeline to the next stage of the drug optimization pipeline / activity.

[0159] e) PBPK / ML model systems 200 or 250 can be further modified and / or configured to be integrated with alternative ML or molecular dynamics models 253, which predict the molecular properties of the antibody based on the structure or sequence of the antibody 254. This integration, performed in the generative molecule / target unit 201 of PBPK / ML model systems 200 and 250 to obtain molecular properties, can be used to replace and / or supplement the synthetic generation of these properties.

[0160] Figure 3 An example PBPK component 300 is shown for use with PBPK / ML model systems 200 and 250 and / or processes 100, 105, 110, 120, etc., as described with reference to Figures 1a to 2e. In this case, the PBPK component 300 uses a minimal PBPK (mPBPK) model 301, which is configured to account for the effects of antibody molecular weight, size, charge, binding to FcRn and different forms of specific receptors under different pH conditions. The mPBPK model 301 is configured to apply the two-well hypothesis to predict size-based clearance and exposure of full-length antibodies (e.g., 150 kDa) and antibody fragments (e.g., 50–100 kDa), and quantitatively correlate the effect of charge on PK parameters such as uptake rate, nonspecific binding affinity, and volume of distribution. The mPBPK model 301 is configured to predict target-mediated disposition of molecular drugs when compound-specific and target-specific properties are known. The combined effects of antibody molecular weight, size, charge, FcRn, and antigen have never before been incorporated into a single mPBPK model. By ultimately combining the physicochemical properties of multiple proteins and correlating them with observed PK, mPBPK model 301 provides a useful PBPK model for use in PBPK / ML model systems 200 and 230 and / or processes 100, 105, 110, 120, especially in the early stages of drug optimization development, where many of these properties can be optimized to improve the PK of the molecule and ultimately improve its efficacy.

[0161] Essentially, the mPBPK model 301 consists of plasma, lymph, compact tissue, and leaking tissue compartments. Each tissue compartment is further divided into vascular, endosomal, and interstitial spaces, as shown in the magnified right section 302. The plasma compartment is nested within the endosomal space. Monoclonal antibodies (A) are administered intravenously (IV) to the plasma compartment, where they bind to soluble antigens (T) in both the plasma compartment and the tissue vascular system to form antibody-antigen complexes (AT). mAb (A), soluble antigens (T), and complexes (AT) circulate to the nested endosomals within the plasma and tissue subcompartments. Free mAb (A) and mAb-antigen complexes (AT) enter the endosomal space via endocytosis and bind to FcRn (black rectangles) at an acidic pH. Unbound mAbs circulate in the endosomal space at a rate of Degradation occurs, while unbound antigens and bound mAb-antigen complexes are cleared from the body at a rate that allows for their elimination. Catabolism occurs. FcRn-bound mAbs and FcRn-bound mAb-antigen complexes are continuously recirculated from the endosomal space. Free mAbs (Y-shaped) also bind to membrane-binding receptors (rhomboids) and negatively charged membrane proteins (dashed lines) in the tissue vascular system, and are transported through paracellular pores based on their size and cleared systemically. Free mAbs and mAb-antigen complexes are transported back into the systemic circulation via the lymphatic compartment. Figure 3 The paper also illustrates the representation of nested endobodies within plasma compartment 303. The dynamics within the nested endobody space are analogous to those within the tissue endobody space. Figure 3 middle, It is the vascular reflection coefficient, and L, L1, and L2 are the lymphatic flow velocities in the lymphatic, compact tissue, and leaky tissue compartments, respectively.

[0162] Essentially, the structure of mPBPK model 301 remains the same across different species. In this example, the species tested with mPBPK model 301 is a mouse, but this is merely an example, and the invention is not limited thereto. Those skilled in the art will understand and adjust the mPBPK model parameters accordingly for other subjects (e.g., humans and / or other mammals). The equations for the mPBPK model are provided below as equations A1 to A51. The model parameters of mPBPK model 301 are broadly divided into two groups: physiologically based parameters and physicochemically or compound-based parameters. Additionally, some compound-specific parameters (such as molecular weight, size, and charge) can be correlated with the model parameters through derived and empirical equations. The following is a summary list of examples of physiologically based and kinetic parameters and their values ​​that can be used for the mouse example:

[0163]

[0164] Note: In this example, L = L1 + L2. It is calculated using transport time (8 min) and endosomal volume. For soluble receptors, For membrane-bound receptors, .

[0165] The corresponding example values ​​of physiological and kinetic parameters can be calculated, measured, estimated, and / or obtained from relevant literature studies relating to subjects of the species of interest (e.g., mice in this example) to be input into the mPBPK model 301. Physiological parameters (such as body weight, volume, flow rate, FcRn concentration, etc.) depend on the species of interest and can be adjusted for any species and / or subject (e.g., but not limited to, other animals and / or human subjects). Baseline concentrations of soluble and membrane-bound antigens can be obtained from literature studies.

[0166] Compound-specific parameters vary with the physicochemical properties of the drug. In the mPBPK model 301, the drug's specificity parameters are considered. Stokes radius of the drug ( ), net surface charge (Q) of the drug, fraction of available interstitial volume of the drug ( The association rate constant and dissociation rate constant of antibody-FcRn binding under slightly acidic pH (pH = 6) The association rate constant and dissociation rate constant of antibody-antigen binding at physiological pH (pH = 7.4) and slightly acidic pH (pH = 6) were calculated. as well as ), and the dissociation rate constant based on nonspecific interactions of charge ( The antibody-FcRn binding at acidic pH can be calculated, measured, and / or estimated, and / or obtained from published literature. Values ​​and antibody-antigen binding The size-based pore clearance was calculated using previously published derivation equations. Quantitative relationships between compound-specific parameters and PK parameters are provided in equations A49 to A51.

[0167] Equations A1 to A51 of the mPBPK model 301 are shown below:

[0168] Plasma section

[0169]

[0170]

[0171]

[0172] Plasma nested endosomal

[0173]

[0174]

[0175]

[0176]

[0177]

[0178]

[0179] Tight tissue (blood vessels)

[0180]

[0181]

[0182]

[0183]

[0184]

[0185]

[0186] Tightly packed tissue (internal body)

[0187]

[0188]

[0189]

[0190]

[0191]

[0192]

[0193] Tight tissue (interstitial tissue)

[0194]

[0195]

[0196] Leaking tissue (blood vessels)

[0197]

[0198]

[0199]

[0200]

[0201]

[0202]

[0203] Leaking tissue (internal body)

[0204]

[0205]

[0206]

[0207]

[0208]

[0209]

[0210] Leaking tissue (interstitium)

[0211]

[0212]

[0213] Lymphatic compartments

[0214]

[0215]

[0216] In the differential equations A1 to A39 describing the mPBPK model 301 above, A, T, ATC, FcRnA and FcRnATC are free antibody (A), free soluble antigen target (T), antibody-target complex (ATC), FcRn-antibody complex (FcRnA) and FcRn-antibody-target complex (FcRnATC), respectively. , , These are the endogenous concentration, vascular concentration, and interstitial concentration in the tightly packed tissue compartment and the leaking tissue compartment, respectively. It is the renal clearance rate. It is the internal recirculation / clearance rate. It is the rate of catabolism of the antibody-antigen complex, and This refers to the two-well clearance rate (equations A47 to A48). The effect of charge on antibody PK is reflected in the mPBPK model 301 through the following factors: the regulator of interstitial distribution volume (...). ), regulators of endocytosis rate ( ), and charge-dependent nonspecific equilibrium dissociation constant ( The subscripts p, 1, 2, L, v, is, and e represent plasma, tight tissue, leaking tissue, lymph, blood vessels, interstitium, and internal body compartments, respectively.

[0217] Calculation of two-pore mediated clearance rate

[0218]

[0219]

[0220]

[0221]

[0222]

[0223]

[0224]

[0225] These refer to the lymph flow rates through the macropores and micropores, respectively. It is the Stokes radius of the protein. These are the vascular reflectance coefficients for macropores and micropores, respectively. Molecular weight (MW) is expressed in g / mol or Daltons. In the above text, i = 1 represents compact tissue, and i = 2 represents leaky tissue.

[0226] Dual-pore removal rate via large and small pores ( ）

[0227]

[0228]

[0229] This can be calculated or estimated accordingly, or found in a literature review. In this example, The calculations were performed using derivation equations from the following literature study by Z. Li and DK Shah: "Two-pore physiologically based pharmacokinetic model with de novoderived parameters for predicting plasma PK of different size protein therapeutics," *Journal of Pharmacokinetics and Pharmacodynamics*, pp. 305-218, 2019. In this case, i = 1 represents tight tissue, and i = 2 represents leaky tissue. In this example, the relevant parameters for the two-pore transport model for the species of interest (e.g., mice) used for equations A40 to A48 are shown below:

[0230]

[0231] The example values ​​of the corresponding physiological and kinetic parameters can be calculated, measured, estimated, and / or found from relevant literature studies related to subjects of the species of interest (e.g., mice in this example) to be input into the mPBPK model 301.

[0232] Quantitative equations that correlate antibody properties (e.g., size (ae), charge (Q)) with PK parameters:

[0233]

[0234] Where Θ is the sieving coefficient of the renal filtration membrane.

[0235]

[0236]

[0237] In this example, mPBPK model 301 is configured to simulate and capture the effect of antibody size on plasma and tissue PK using equations A1 through A51. mPBPK model 301 is configured to predict the PK of large molecules (e.g., full-length IgG (150 kDa) and single-arm IgG (100 kDa)) in plasma and tissues. For larger antibodies (>100 kDa), size-based renal clearance is negligible. The larger size of these antibodies limits their movement through the size-selective glomerular filtration membrane and other pores in the kidneys. mPBPK model 301 can demonstrate plasma and tissue PK for smaller IgG fragments (50 kDa) without an intact Fc region. Additionally, mPBPK model 301 has been validated for the PK of FcRn-unbound antibody fragments ranging in size from 50 kDa to 100 kDa. mPBPK model 301 is configured to quantitatively correlate the effect of charge (Q) on nonspecific binding affinity (equation A50) and distribution volume fraction (equation A51) using empirical equations.

[0238] As can be seen, the mPBPK model 301 receives parameters based on physiology (e.g., organ volume, tissue lymphatic flow, blood flow), antibody-specific properties (e.g., MW, size, charge, binding affinity), and target properties (e.g., baseline expression, half-life). Based on equations A1 to A51, the mPBPK model simulates and predicts antibody PK characteristic curves, PK endpoints, and target occupancy percentages (e.g., TO%). The various PK values ​​calculated include, but are not limited to, the area under the curve at steady state (AUC). Minimum plasma concentration ( ), maximum plasma concentration ( ), time to reach minimum plasma concentration ( ), total solubles ( ) and membrane target inhibition ( The mPBPK model 301 is configured to calculate various PK endpoints and / or pharmacological / pharmacodynamic effects and / or values. For example, using equations A53 to A57 below, the mPBPK model 301 can output a set of PK parameters / values, including but not limited to, for example, C. min AUC ss C max C ss , Furthermore, the mPBPK model 301 can use equation A52 to calculate pharmacological / pharmacodynamic effects (including, but not limited to, target occupancy percentage (TO%)), which is based on the ratio of bound targets to total targets in plasma, tight tissue, and leaking tissue, where TO% is calculated for each scenario at the minimum and maximum drug concentrations in plasma. In this example, the mPBPK model 301 uses equations A52 to A57 to calculate the target occupancy percentage (TO%) and AUC. ss C min C max , The estimated value is calculated based on the following equation:

[0239]

[0240]

[0241]

[0242]

[0243]

[0244]

[0245] in, It is the target occupancy (%) calculated at time (t) and site (k) (e.g., plasma, dense tissue, leaky tissue). It is the concentration of the drug-target complex, and It is the concentration of the target. This is the AUC exposure of antibodies in plasma calculated under steady-state conditions. Target inhibition of soluble targets (T) and membrane-bound targets (Tm) is calculated as the difference between the maximum and minimum total target concentrations at site (k) (e.g., plasma, tight tissue, and leaky tissue). It is calculated within the mPBPK model 301 based on, for example, equations A1 to A57.

[0246] Despite the pharmacological effects of target occupancy (TO%) and PK value / endpoint (AUC) ss C min C max , The results are described as being calculated by mPBPK model 301, but this is only an example and mPBPK model 301 is not limited thereto. Those skilled in the art should understand that mPBPK model 301 can be configured to calculate any suitable pharmacological / pharmacodynamic effects / values ​​(e.g., TO%, toxicity or efficacy values / endpoints) and / or any suitable PK characteristic curves, endpoints, parameters and / or properties associated with ADME / LADME (as described herein), combinations thereof, modifications thereof and / or values ​​as required by the application.

[0247] Although this document describes an mPBPK model, it is only by way of example and the invention is not limited thereto. Those skilled in the art should understand that any other suitable PBPK model can be used to simulate the interaction between the molecular drug and the target in the physiological or cellular systems of subjects of the species of interest, and to calculate multiple sets of necessary PK values ​​corresponding to the PK characteristic curves / endpoints, as well as corresponding pharmacological / pharmacodynamic values ​​(e.g., target binding response / target occupancy) and / or values ​​as required by the application.

[0248] Although the PBPK and ML models and processes 100, 105, 110, 120, PBPK and ML systems 200 and 250, and PBPK model 301 described herein are described with reference to antibodies (especially large monoclonal antibodies, etc.), this is only an example and the invention is not limited thereto. Those skilled in the art should understand that the PBPK and ML models and processes 100, 105, 110, 120, PBPK and ML systems 200 and 250, and PBPK model 301 described herein can be applied to any molecule that can be described by at least a set of physicochemical properties and a set of PK values ​​associated with drug administration regimens and targets, such as, but not limited to, antibodies, CODV antibodies, enzymes, other proteins, and / or any other suitable protein forms, combinations thereof, modifications thereof, etc., and / or as required by the application.

[0249] Figure 4 This is a schematic diagram of a system / apparatus for performing the methods described herein. The system / apparatus shown is an example of a computing device. Those skilled in the art will understand that other types of computing devices / systems can be alternatively used to implement the methods described herein, such as distributed computing systems.

[0250] Device (or system) 400 includes one or more processors 402. The one or more processors control the operation of other components of the system / device 400. For example, the one or more processors 402 may include general-purpose processors. The one or more processors 402 may be single-core or multi-core devices. The one or more processors 402 may include a central processing unit (CPU) or a graphics processing unit (GPU). Alternatively, the one or more processors 402 may include dedicated processing hardware, such as a RISC processor or programmable hardware with embedded firmware. Multiple processors may be included.

[0251] The system / device includes working or volatile memory 404. One or more processors can access volatile memory 404 to process data and can control the storage of data in the memory. Volatile memory 404 may include any type of RAM, such as static RAM (SRAM), dynamic RAM (DRAM), or it may include flash memory, such as an SD card.

[0252] The system / device includes non-volatile memory 406. Non-volatile memory 406 stores an instruction set 408 for controlling the operation of processor 402 in the form of computer-readable instructions. Non-volatile memory 406 can be any type of memory, such as read-only memory (ROM), flash memory, or magnetically driven memory.

[0253] One or more processors 402 are configured to execute operation instructions 608 to cause the system / device to perform the actions described herein, as referred to in Figures 1a to 1b. Figure 3 Any of the methods or processes described. Operation instructions 408 may include code related to hardware components of system / device 400 (i.e., drivers), as well as code related to the basic operation of system / device 400. Generally, one or more processors 402 execute one or more instructions of operation instructions 408 that are permanently or semi-permanently stored in non-volatile memory 406, and use volatile memory 404 to temporarily store data generated during the execution of said operation instructions 408.

[0254] Implementations of the methods described herein can be achieved using digital electronic circuit systems, integrated circuit systems, specially designed ASICs (Application-Specific Integrated Circuits), computer hardware, firmware, software, and / or combinations thereof. These may include computer program products (e.g., software stored on, for example, a disk, optical disk, memory, or programmable logic device) comprising computer-readable instructions that, when read by a computer (e.g., regarding...),... Figure 4 When the computer described herein is executed, it causes the computer to perform one or more methods described herein.

[0255] Any system feature described in this paper can also be provided as a methodological feature, and vice versa. As used in this paper, means plus functions can be alternatively expressed by their corresponding structures. Specifically, methodological aspects can be applied to system aspects, and vice versa.

[0256] Furthermore, any, some, and / or all features of one aspect may be applied in any suitable combination to any, some, and / or all features of any other aspect. It should also be understood that specific combinations of the various features described and defined in any aspect of the invention may be implemented and / or provided and / or used independently.

[0257] Although several embodiments have been shown and described, those skilled in the art will understand that changes may be made to these embodiments without departing from the principles of this disclosure, the scope of which is defined in the claims.

Claims

1. A computer-implemented method for classifying one or more candidate molecular drugs for drug optimization based on one or more desired conditions associated with pharmacological or pharmacodynamic effects, the method comprising: Receive a candidate molecular drug inference dataset representing one or more candidate molecular drugs, wherein each candidate molecular drug in the candidate molecular drug dataset is represented by a set of physicochemical values ​​of the candidate molecular drug and a set of pharmacokinetic (PK) values ​​associated with each candidate molecular drug having a dosing regimen and a target, and wherein each candidate molecular drug in the candidate molecular drug dataset is classified as a macromolecular drug. For each candidate molecule in the candidate molecule inference dataset, the corresponding set of physicochemical values ​​and a set of pharmacokinetic (PK) values ​​for the dosing regimen and target are applied to a machine learning (ML) model. This ML model is configured to classify whether each candidate molecule satisfies one or more expected conditions associated with pharmacological or pharmacodynamic effects. The ML model is trained using an ML algorithm to train the model parameters defining the ML model. This ML model is used to classify whether each candidate molecule satisfies one or more expected conditions associated with pharmacological or pharmacodynamic effects based on a training dataset. The training dataset includes data representing multiple molecule drugs, represented by multiple sets of known and / or calculated physicochemical values ​​and multiple sets of PK values ​​associated with the dosing regimen, and labeled based on the one or more expected conditions associated with pharmacological or pharmacodynamic effects. The output represents data for a set of candidate molecular drugs classified according to these desired conditions associated with pharmacological or pharmacodynamic effects, for use in downstream drug optimization processes.

2. The computer-implemented method as described in claim 1, wherein, Each candidate molecular drug is an antibody.

3. The computer-implemented method as described in claim 1 or 2, wherein, Each candidate molecular drug is a monoclonal antibody.

4. The computer-implemented method as described in claim 3, wherein, This monoclonal antibody has a size greater than or equal to 50 kDa.

5. The computer-implemented method as described in claim 3 or 4, wherein, This monoclonal antibody is classified as a macromolecular drug.

6. The computer-implemented method as described in any of the preceding claims, wherein, The target is an antigen.

7. The computer-implemented method according to any one of claims 1 to 6, wherein, This set of physicochemical values ​​for each molecular drug corresponds to one or more physicochemical parameters or properties derived from a group consisting of the following: Molecular weight (MW); Size or Stokes radius ae; Charge Q; Hydrophobicity; Combined with affinity; isoelectric point pI; Glycosylation; Drug-target binding constant KD; and Any other type of physicochemical parameter or property associated with each molecular drug.

8. The computer-implemented method as described in claim 7, wherein, Each molecular drug’s physicochemical properties include at least MW, ae, Q, and KD.

9. The computer-implemented method according to any one of claims 1 to 8, wherein, The PK values ​​for each molecular drug are associated with PK characteristic curves, endpoints, parameters, and / or properties corresponding to ADME (absorption, distribution, metabolism, and excretion) or LADME (release, absorption, distribution, metabolism, and excretion).

10. The computer-implemented method according to any one of claims 1 to 6, wherein, These PK characteristic curves, endpoints, parameters, and / or attributes include one or more of the following groups: Dosage D; Dosing interval τ; Maximum serum concentration C max ; Reaching C max The shortest time; Minimum plasma concentration C min,ss ; Mean plasma concentration C av,ss ; steady-state concentration C ss ; Distributed volume V d ; Time to reach minimum plasma concentration ; Total soluble ; Membrane target inhibition ; Absorption half-life; Absorption rate constant; Eliminate half-life; Eliminate the rate constant; The area under the concentration-time curve (AUC) after a single dose or at steady state ss ; Sweep rate C; as well as Any other type of PK characteristic curve, endpoint, parameter, and / or attribute associated with ADME or LADME.

11. The computer-implemented method of claim 10, wherein, For each molecular drug, the PK characteristic curves, endpoints, parameters, and / or properties of this group should include at least C. min AUC ss C max C ss , and .

12. The computer-implemented method as described in any of the preceding claims, wherein, The ML algorithm includes one or more ML algorithms or combinations thereof that are derived from the group consisting of the following: Neural network-based ML algorithms; Classification ML algorithms based on decision trees; Random Forest Decision Tree ML Algorithm; ML algorithms based on adaptive boosting; Gradient boosting-based ML algorithms; ML algorithm based on bagging method; Based on ensemble ML algorithms; Any other supervised ML algorithm suitable for training model parameters of an ML model that, given input data representing a set of physicochemical values ​​and a set of PK values ​​for targets and dosing regimens corresponding to one or more molecular drugs, outputs a classification of whether one or more molecular drugs meet one or more desired conditions associated with pharmacological or pharmacodynamic effects.

13. The computer-implemented method as described in any of the preceding claims, further comprising iteratively training model parameters defining the ML model using the ML algorithm associated with the ML model, the training comprising: Receive a training dataset associated with multiple molecular drugs. The training dataset includes multiple training instances, each associated with each molecular drug and including data representing a set of known and / or expected physicochemical values ​​and a set of corresponding known or calculated PK values ​​associated with the dosing regimen and the target. Each training instance is labeled or marked based on whether each molecular drug and target meets the expected conditions associated with pharmacological or pharmacodynamic effects. Iteratively apply the training dataset to the ML algorithm to train the model parameters of the ML model until the stopping criterion is met; and The output represents the model parameters of the trained ML model.

14. The computer-implemented method as described in any of the preceding claims, wherein, The training dataset contains multiple sets of physicochemical values ​​corresponding to these molecular drugs, based on synthetically produced molecular drugs generated within the desired physicochemical value range. For each synthetically produced molecular drug, a physiologically based pharmacokinetic (PBPK) model is used to calculate the corresponding PK value for each synthetically produced molecular drug. These PK values ​​are associated with the dosing regimen and the target, and are also associated with pharmacological or pharmacodynamic effects, responses, and / or values. The PBPK model is configured to: simulate the effect of the synthetically produced molecular drug on the target in the subject's cellular system; and for each synthetically produced molecule, output a set of corresponding PK values ​​from selected PK endpoints and characteristic curves, along with the corresponding pharmacological or pharmacodynamic effects, responses, and / or values ​​associated with these desired conditions.

15. The computer-implemented method as described in any of the preceding claims, wherein, The training dataset contains multiple sets of physicochemical values ​​corresponding to each molecular drug, PK values ​​associated with the dosing regimen and the target, and related to pharmacological or pharmacodynamic effects, responses, and / or values, based on known or measured physicochemical values ​​of real-world molecular drugs, and PK values ​​associated with the dosing regimen and the target, and related to pharmacological or pharmacodynamic effects, responses, and / or values.

16. The computer-implemented method as described in any of the preceding claims, wherein, The training dataset contains multiple sets of physicochemical values ​​corresponding to these molecular drugs based on real-world molecular drugs that meet the expected range of physicochemical values. For each real-world molecular drug, a PBPK model is used to calculate the corresponding PK value for each real-world molecular drug. These PK values ​​are associated with the target and the dosing regimen, as well as the pharmacological or pharmacodynamic effects, responses, and / or values. The PBPK model is configured to: simulate the effect of the real-world molecular drug on the target in the physiological or cellular system of the subject; and for each real-world molecular drug, output a set of corresponding PK values ​​from selected PK endpoints and characteristic curves, along with the corresponding pharmacological or pharmacodynamic effects, responses, and / or values.

17. The computer-implemented method as described in any of the preceding claims, wherein, The multiple sets of physicochemical values ​​corresponding to these molecular drugs in the candidate molecular drug inference dataset are based on synthetically generated molecular drugs produced within the expected physicochemical value range. For each synthetically generated molecular drug, a PBPK model is used to calculate the corresponding PK value for each synthetically generated molecular drug. These PK values ​​are associated with the dosing regimen and the corresponding pharmacological or pharmacodynamic effects, responses, and / or values. The PBPK model is configured to: simulate the effect of the synthetically generated molecular drug on the target in the physiological or cellular system of the subject; and for each synthetically generated molecule, output a set of corresponding PK values ​​from selected PK endpoints and characteristic curves, along with the corresponding pharmacological or pharmacodynamic effects, responses, and / or values.

18. The computer-implemented method as described in any of the preceding claims, wherein, The candidate molecular drug inference dataset contains multiple sets of physicochemical values ​​corresponding to each molecular drug, PK values ​​associated with the dosing regimen and the target, and these pharmacological or pharmacodynamic effects, responses, and / or values, based on known or measured physicochemical values ​​of real-world molecular drugs, and PK values ​​associated with the dosing regimen and the target, and pharmacological or pharmacodynamic effects, responses, and / or values.

19. The computer-implemented method as described in any of the preceding claims, wherein, The multiple sets of physicochemical values ​​corresponding to these molecular drugs in the candidate molecular drug inference dataset are based on real-world molecular drugs that satisfy the expected range of physicochemical values ​​in a set of physicochemical parameters. For each real-world molecular drug, the PBPK model is used to calculate the corresponding PK value for each real-world molecular drug. These PK values ​​are associated with the target and the dosing regimen, as well as the pharmacological or pharmacodynamic effects, responses, and / or values. The PBPK model is configured to: simulate the effect of the real-world molecular drug on the target in the subject's cellular system; and for each real-world molecular drug, output a set of corresponding PK values ​​from selected PK endpoints and characteristic curves, along with the corresponding pharmacological or pharmacodynamic effects, responses, and / or values.

20. The computer-implemented method as described in any of the preceding claims, further comprising: Receive a set of physicochemical values ​​corresponding to each molecular drug; Receive one or more sets of physicochemical values ​​corresponding to the target; Receive a dosing regimen; For each desired dosing regimen, one or more unique molecular drug-target data pairs are generated by pairing the desired physicochemical values ​​of the molecular drug with each of the desired physicochemical values ​​of the target in one or more sets of values. For each dosing regimen, the expected physicochemical values ​​of the drug and the expected physicochemical values ​​of the target in the drug-target pairing data are applied to the PBPK model. For each dosing regimen and each drug-to-target data pairing, the PBPK model is used to calculate a set of PK values ​​for each drug-to-target data pairing and the corresponding pharmacological or pharmacodynamic effects, responses, and / or values ​​associated with these desired conditions. For each drug-molecule paired with a target and its corresponding dosing regimen, training data instances or candidate drug-molecule inference data instances are generated. The training data instances include data representing at least the set of physicochemical properties and the calculated pharmacokinetic (PK) values ​​for each drug-molecule. These training data instances are labeled based on the corresponding pharmacological or pharmacodynamic effects, responses, and / or values ​​calculated for each drug-molecule paired with a target and its corresponding dosing regimen, as well as these expected conditions. The candidate drug-molecule inference data instances include data representing at least the set of physicochemical properties for each drug-molecule and the calculated PK values ​​calculated for each drug-molecule paired with a target and its corresponding dosing regimen. Store these training data instances or candidate molecule drug inference data instances for use in training or inference / classification, respectively.

21. The computer-implemented method as described in any of the preceding claims, wherein, The output of the set of candidate molecular drugs that meet these desired conditions associated with pharmacological or pharmacodynamic effects for downstream drug optimization further includes outputting a candidate list of candidate molecular drugs—each candidate molecular drug represented by a set of corresponding physicochemical values ​​and a set of PK values ​​for the dosing regimen and the target—for downstream processing, including in vitro wet laboratory analysis or high-throughput screening.

22. The computer-implemented method according to any one of claims 1 to 20, wherein, The output of the set of candidate molecular drugs that meet these desired conditions associated with pharmacological or pharmacodynamic effects for downstream drug optimization further includes outputting a candidate list of candidate molecular drugs—each candidate molecular drug represented by a set of corresponding physicochemical values ​​and a set of PK values ​​for the dosing regimen and the target—for downstream processing, including in vivo testing.

23. The computer-implemented method as described in claim 21 or 22, wherein, When one or more candidate molecular drugs in the candidate list are represented by only one set of physicochemical values ​​and one set of PK values, then one or more real-world molecular drugs that meet the corresponding set of physicochemical values ​​and PK values ​​of the one or more candidate molecular drugs are generated before in vitro wet laboratory analysis, high-throughput screening or in vivo experiments.

24. The computer-implemented method of claim 23, wherein, Generating one or more real-world molecular drugs further includes, for each corresponding candidate molecular drug, calculating the structural and chemical properties required to realize or synthesize a real-world molecular drug that satisfies the corresponding set of physicochemical values, the set of PK values, and associated pharmacological or pharmacodynamic effects, reactions, and / or values.

25. The computer-implemented method according to any one of claims 1 to 20, wherein, The output of this set of candidate molecular drugs that meet these desired conditions is used for downstream drug optimization. Further output includes data representing any of the following: A candidate list of drug candidates that meet these desired criteria, including pharmacokinetic (PK) values ​​for the dosing regimen and target, as well as associated pharmacological or pharmacodynamic effects, responses, and / or values, is provided for review by one or more users or expert systems to further evaluate the output candidate list; or A list of candidate molecular drugs categorized according to whether they meet these expected criteria, with PK values ​​for the dosing regimen and target, as well as associated pharmacological or pharmacodynamic effects, responses, and / or values, is provided for review by one or more users or expert systems to further evaluate the output list of candidate molecular drugs.

26. An apparatus comprising a processor, a memory unit, and a communication interface, wherein, The processor is connected to the memory unit and the communication interface, wherein the processor and the memory are configured to implement a computer-implemented method according to any one of the preceding claims.

27. A computer-readable medium comprising data or instruction code that, when executed on a processor, causes the processor to perform the computer-implemented method as claimed in any one of claims 1 to 25.

28. A computing system, comprising: A molecule and target generation unit configured to generate one or more molecules represented by multiple sets of corresponding physicochemical values ​​and one or more targets represented by desired physicochemical values; Physiology-based pharmacokinetic PK unit, which is configured to generate a set of PK values ​​and / or associated pharmacological or pharmacodynamic effects, responses and / or values ​​for each molecular drug paired with a target as training data instances or candidate molecular drug inference data instances; The ML module is configured to train an ML model using a training dataset containing multiple training data instances to classify whether one or more molecular drugs meet desired conditions associated with pharmacological or pharmacodynamic effects for a specific dosing regimen; or to perform inference using the trained ML model to classify whether one or more input candidate molecular drug inference data instances meet desired target conditions for a specific dosing regimen. as well as An output module is configured to output a candidate list of at least one molecular drug that satisfies these desired conditions associated with pharmacological or pharmacodynamic effects—each molecular drug represented by a set of physicochemical values ​​and a set of PK values—to generate real-world molecules with corresponding properties for use in in vivo assays, in vitro wet laboratory analyses, and / or high-throughput screening laboratory analyses.

29. A non-transitory tangible computer-readable medium comprising data or instruction code for classifying one or more candidate molecular drugs for drug optimization in response to one or more desired conditions associated with pharmacological or pharmacodynamic effects, the data or instruction code, when executed on one or more processors, causing at least one of these processors to perform at least the following operations: Receive a candidate molecular drug inference dataset representing one or more candidate molecular drugs, wherein, Each candidate molecular drug in the candidate molecular drug dataset is represented by a set of physicochemical values ​​of the candidate molecular drug and a set of pharmacokinetic (PK) values ​​associated with each candidate molecular drug, including the dosing regimen and the target. For each candidate molecular drug in the candidate molecular drug inference dataset, the corresponding set of physicochemical values ​​and a set of pharmacokinetic (PK) values ​​associated with the dosing regimen and target are applied to a machine learning (ML) model. This ML model is configured to classify whether each candidate molecular drug satisfies one or more expected conditions related to pharmacological or pharmacodynamic effects. The ML model is trained using an ML algorithm to train the model parameters defining the ML model. The ML model is used to classify whether each candidate molecular drug satisfies the one or more expected conditions based on a training dataset. This training dataset includes data representing multiple molecular drugs, represented by multiple sets of known and / or calculated physicochemical values ​​and multiple sets of PK values ​​associated with the dosing regimen and target, and labeled with corresponding pharmacodynamic effects, responses, and / or values ​​related to these expected conditions. The output represents data for a set of candidate molecular drugs that meet these desired conditions, for use in downstream drug optimization processes.

Images