Predicting pharmacokinetic profiles of chemical compounds using machine learning

Abstract

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating a predicted pharmacokinetic (PK) profile of a chemical compound from the structure of the chemical compound using a multi-stage machine learning model. In one aspect, a system comprises processing a model input comprising the data characterizing the chemical structure of the chemical compound, using one or more PK parameter prediction machine learning models, to generate one or more predicted PK parameters, and processing a model input comprising the one or more predicted PK parameters generated by the one or more PK parameter prediction machine learning models, using a PK profile prediction machine learning model, to generate the predicted PK profile of the chemical compound.

Description

PREDICTING PHARMACOKINETIC PROFILES OF CHEMICAL COMPOUNDSUSING MACHINE LEARNINGTECHNICAL FIELD

[0001] This specification relates to predicting pharmacokinetic profiles of chemical compounds using machine learning.BACKGROUND

[0002] Machine learning models receive an input and generate an output, e.g., a predicted output, based on the received input. Some machine learning models are parametric models and generate the output based on the received input and on values of the parameters of the model.

[0003] A pharmacokinetic (PK) profile of a chemical compound can define a concentration versus time curve that can represent, e.g., the absorption, distribution, metabolism, and elimination (ADME) of the chemical compound from the body of a subject. PK profiles provide information about how chemical compounds behave in terms of concentration levels in various bodily compartments, such as the bloodstream, tissues, and organs. Obtaining accurate PK profiles for chemical compounds is essential in drug development and can influence dosing regimens, efficacy, and safety considerations.SUMMARY

[0004] This specification generally describes a system implemented as computer programs on one or more computers in one or more locations that can generate a predicted pharmacokinetic (PK) profile of a chemical compound from the structure of the chemical compound using a multi-stage machine learning model. In particular, the multi-stage machine learning model can include one or more PK parameter prediction models that process features characterizing the chemical structure of the chemical compound to generate one or more PK parameters and a PK profile prediction model to process a model input including the outputs of the one or more PK parameter prediction models to generate a predicted PK profile of the chemical compound that can be used in drug discovery.

[0005] In this specification, “PK parameters” are global summan' statistics characterizing properties of a chemical compound, e.g., characterizing one or more of absorption, distribution, metabolism, or elimination of the chemical compound in a subject. For example, a subject can be rat, dog, pig, human, or another species. A “PK” profile includes data thatdefines, for each time point in a sequence of time points, the concentration value of the chemical compound in a subject at the time point. In some cases, the PK profile for a particular chemical compound is different for different species.

[0006] According to a first aspect, there is a provided a method performed by one or more computers, the method comprising: processing data characterizing a chemical structure of a chemical compound using a multi-stage machine learning model to generate data defining a predicted pharmacokinetic (PK) profile of the chemical compound, comprising: processing a model input comprising the data characterizing the chemical structure of the chemical compound, using one or more PK parameter prediction machine learning models, to generate one or more predicted PK parameters of the chemical compound; and processing a model input comprising the one or more predicted PK parameters generated by the one or more PK parameter prediction machine learning models, using a PK profile prediction machine learning model, to generate the predicted PK profile of the chemical compound.

[0007] In some implementations, the one or more predicted PK parameters comprise one or more of: a clearance of the chemical compound, a volume of distribution of the chemical compound, a half-life of the chemical compound, a maximum concentration of the chemical compound, or a time to maximum concentration of the chemical compound.

[0008] In some implementations, the predicted PK profile of the chemical compound defines a concentration of the chemical compound in a subject over time after the chemical compound is administered to the subject.

[0009] In some implementations, the model input to the one or more PK profile prediction machine learning models further comprises features characterizing the chemical structure of the chemical compound.

[0010] In some implementations, the model input to the one or more PK profile prediction machine learning models comprises a plurality of features selected by performing operations comprising: identifying a collection of candidate features for inclusion in the model input to the one or more PK profile prediction machine learning models, wherein the collection of candidate features comprise: (i) features defining chemical compound PK parameters, and (ii) features charactenzing chemical compound structure; and selecting a proper subset of the collection of candidate features for inclusion in the model input to the one or more PK profile prediction machine learning models.

[0011] In some implementations, the features characterizing the chemical compound structure comprise molecular descriptor features and molecular fingerprint features.

[0012] In some implementations, selecting a proper subset of the collection of candidate features for inclusion in the model input to the one or more PK profile prediction machine learning models comprises selecting less than 50%, or less than 10%, or less than 5%, or less than 1% of the collection of candidate features for inclusion in the model input to the one or more PK profile prediction machine learning models.

[0013] In some implementations, selecting a proper subset of the collection of candidate features for inclusion in the model input to the one or more PK profile prediction machine learning models comprises applying a feature selection algorithm, to the collection of candidate features, in order to select a proper subset of the collection of candidate features that are predicted to optimize a performance of the one or more PK profile prediction machine learning models.

[0014] In some implementations, processing the model input comprising the one or more predicted PK parameters, using the PK profile prediction machine learning model, to generate the predicted PK profile of the chemical compound comprises: processing the model input comprising the one or more predicted PK parameters using each of a plurality’ of base machine learning models included in the PK profile prediction machine learning model; and generating the predicted PK profile of the chemical compound by combining a respective output of each of the plurality’ of base machine learning models included in the PK profile prediction machine learning model.|0015| In some implementations, generating the predicted PK profile of the chemical compound by combining the respective output of each of the plurality of base machine learning models included in the PK profile prediction machine learning model comprises summing the respective output of each of the plurality of base machine learning models included in the PK profile prediction machine learning model.

[0016] In some implementations, the plurality of base machine learning models included in the PK profile prediction machine learning model are associated with an ordering into a sequence; and each base machine learning model after a first base machine learning model in the sequence is configured, through training, to generate an output that corrects an error between: (i) a target PK profile, and (ii) a PK profile generated by combining the outputs of preceding base machine learning models in the sequence.

[0017] In some implementations, the PK profile prediction machine learning model has a set of PK profile prediction machine learning model parameters have been trained on a set of training examples by a machine learning training technique.

[0018] In some implementations, for each of one or more training examples from the set of training examples, the training example corresponds to a training chemical compound and comprises: (i) a training model input comprising one or more predicted PK parameters generated by the one or more PK parameter prediction machine learning models for the training chemical compound, and (ii) a target PK profde of the training chemical compound.

[0019] In some implementations, training the PK profde prediction machine learning model on the set of training examples comprises, for each training example, training the PK profde prediction machine learning model to reduce a discrepancy between: (i) the target PK profde of the training example, and (ii) a predicted PK profde generated by the PK profde prediction machine learning model by processing the training model input of the training example.

[0020] In some implementations, generating one or more predicted PK parameters of the chemical compound comprises processing the model input comprising the data characterizing the chemical structure of the chemical compound using each of a plurality of PK parameter prediction machine learning models to generate a respective predicted PK parameter as an output of each of the plurality of PK parameter prediction machine learning model.

[0021] In some implementations, each of the plurality of PK parameter prediction machine learning models has been configured through training by a machine learning training technique to generate a respective predicted PK parameter.

[0022] In some implementations, the data characterizing the chemical structure of the chemical compound comprises one or more molecular descriptors of the chemical compound.

[0023] In some implementations, the one or more molecular descriptors of the chemical compound define one or more of: a molecular weight of the chemical compound, a total number of hydrogen bond donors in the chemical compound, a total number of hydrogen bond acceptors in the chemical compound, a topological polar surface area of the chemical compound, a number of rotatable bonds in the chemical compound, a number of rings in the chemical compound, a molecular volume of the chemical compound, a dipole moment of the chemical compound, or a three-dimensional (3D) arrangement of atoms in the chemical compound.

[0024] In some implementations, the data characterizing the chemical structure of the chemical compound comprises a molecular fingerprint of the chemical compound.

[0025] In some implementations, the molecular fingerprint of the chemical compound defines a fixed-length binary representation that encodes a presence or absence of specific structural patterns or substructures within the chemical compound.

[0026] In some implementations, the data characterizing the chemical structure of the chemical compound is derived from a representation of the chemical structure of the chemical compound as a textual string.

[0027] In some implementations, the chemical structure of the chemical compound as a textual string comprises a Simplified Molecular Input Line Entry System (SMILES) string.

[0028] In some implementations, the method further comprises processing the predicted PK profile to determine an area under the predicted PK profile.

[0029] In some implementations, the method further comprises selecting the chemical compound for physical synthesis based at least in part on the predicted PK profile of the chemical compound.

[0030] In some implementations, the method further comprises physically synthesizing the chemical compound.

[0031] In some implementations, the method further comprises administering the physically synthesized chemical compound to a subject.

[0032] In some implementations, the method further comprises experimentally validating one or more properties of the physically synthesized chemical compound.

[0033] In some implementations, the method further comprises providing the data defining the predicted PK profile of the chemical compound, comprising one or more of: storing the data defining the predicted PK profile of the chemical compound in a memory; transmitting the data defining the predicted PK profile of the chemical compound over a data communications network; or displaying a representation of the predicted PK profile of the chemical compound on a display of a user device.

[0034] According to another aspect, there is provided a system comprising: one or more computers; and one or more storage devices communicatively coupled to the one or more computers, wherein the one or more storage devices store instructions that, when executed by the one or more computers, cause the one or more computers to perform operations of the methods described herein.

[0035] One or more non-transitory computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations of the methods described herein.

[0036] The subject matter described in this specification can be implemented in particular embodiments so as to realize one or more of the following advantages.

[0037] Drug discovery is a complex process that encompasses several distinct stages. Initially, in target identification and validation, potential biological targets for a specificdisease are identified. Next, in lead discovery and optimization many compounds directed to the specified target are synthesized, screened, and optimized with respect to their absorption, distribution, metabolism, elimination (ADME) characteristics and physiochemical (PC) properties. Finally, the selected compounds progress through preclinical development to assess their pharmacokinetics (PK), efficacy and toxicity7in animal models before potential regulatory approval and clinical testing. PK profile evaluation plays a critical role in understanding a compound’s safety and efficacy and determining its starting and efficacious dose for future clinical studies. Traditionally, PK profiles in humans are predicted by scaling PK profiles from pre-clinical species such as mice, rats, and non-human primates (NHP). However, traditional methods for obtaining PK profiles based on animal experimentation are expensive, time-consuming, non-scalable. and require sacrificing animals.

[0038] To address these issues, the system described in this specification provides a multistage machine learning model that can generate a predicted PK profile for a chemical compound by processing data characterizing a chemical structure of the chemical compound. That is. the machine learning model can translate from chemical structures to predicted PK profiles in-silico. without requiring physical experiments or animal testing for the chemical compound. The system can thus accelerate the drug development process while minimizing costs and reducing the need to sacrifice animals.

[0039] The system can generate a predicted PK profile for a chemical compound using a multi-stage machine learning model with two stages. In the first stage, the system processes data characterizing the chemical structure of the chemical compound using one or more PK parameter prediction machine learning models to generate one or more predicted PK parameters of the chemical compound, e.g., clearance, volume of distribution, half-life, etc. In the second stage, the system processes data including the predicted PK parameters using a PK profile prediction machine learning model to generate the predicted PK profile for the chemical compound. That is, the multi-stage machine learning model includes a first stage where one or more machine learning models translate chemical structure data to structure- related features, and a second stage where a machine learning model translates the features to full predicted PK profiles.

[0040] The multi-stage architecture of the machine learning model enables several technical advantages, which are described in more detail next.

[0041] First, by using a PK profile prediction machine learning model in the second stage, the system can avoid using a mechanistic PK model, e.g.. a one-compartment model, a two- compartment model, a compartmental model with enzyme kinetics, or a tissue compartmentmodel. Mechanistic PK models are explicit mathematical models that attempt to describe the underlying biological and physiological processes governing the behavior of drugs in the body. However, mechanistic PK models are underpinned by assumptions such as compartmentalization (e.g., that the body can be divided into distinct compartments or regions), mass balance, first-order kinetics, instantaneous mixing, and so forth. The assumptions underpinning mechanistic PK models are simplifications that are often inconsistent with real-world physiological processes and are typically difficult to validate.

[0042] In contrast, the PK profile prediction machine learning model does not rely on these assumptions, and rather learns to implicitly reason about underlying physical processes through information encoded in the parameter values of the PK profile prediction machine learning model through machine learning training. The PK profile prediction machine learning model can therefore generate PK profiles more accurately by avoiding errors and inconsistencies introduced by the assumptions underlying mechanistic PK models.

[0043] Second, the multi-stage architecture encodes an effective inductive bias by first predicting PK parameters and then predicting full PK profiles using the predicted PK parameters. More specifically, PK parameters are highly relevant to and predictive of PK profiles, and causing the machine learning model to perform the intermediate step of predicting PK parameters simplifies the complexity7of the prediction task performed by each stage of the machine learning model. In particular, performing the intermediate step of predicting PK parameters can reduce the amount of training data required for training, and can also enable the individual machine learning models implemented at the first and second stages of the multi-stage model to have less complex architectures than might otherwise be required. It will be appreciated that reducing the amount of training data and the complexity of machine learning model architectures further enables a reduction in consumption of computational resources, e.g., memory7and computing power, by the system.

[0044] Third, the multi-stage architecture enables feature selection and pruning to be implemented between the first and second stages of the model. In particular, the system can implement a feature selection algorithm to select a subset of features (e.g., from a full set of possible features including chemical structure features and PK parameter features) that are most predictive of PK profiles. The PK profile prediction machine learning model can then be configured to process only the selected subset of features, which can reduce the likelihood of overfitting and improve the robustness and prediction accuracy of the PK profile prediction machine learning model. Further, processing only the subset of features selected through a feature selection algorithm can reduce the amount of training data required fortraining the PK prediction machine learning model and allow the PK prediction machine learning model to have a less complex architecture, thus reducing consumption of computational resources.

[0045] The details of one or more embodiments of the subject matter of this specification are set forth in the accompanying drawings and the description below. Other features, aspects, and advantages of the subject matter will become apparent from the description, the drawings, and the claims.BRIEF DESCRIPTION OF THE DRAWINGS

[0046] FIG. 1 is a system diagram of an example pharmacokinetic (PK) profile prediction system that can process a chemical structure to generate a PK profile using a multi-stage machine learning model.

[0047] FIG. 2 is a block diagram of an example preprocessing protocol for the multi-stage machine learning model.

[0048] FIG. 3 depicts an example multi-stage machine learning model including a PK parameter prediction stage and a PK profile prediction stage.

[0049] FIG. 4 provides an overview of an example boosting model, which can be implemented by the first and second stage models of the multi-stage machine learning model of FIG. 3.

[0050] FIG. 5 illustrates the accuracy of predicted PK parameters by the example first stage model of FIG. 3.

[0051] FIG. 6 demonstrates the effectiveness of the multi-stage model in capturing nonlinearities with respect to two traditional PK approaches: a one-compartment model and noncompartmental analysis.

[0052] FIG. 7 illustrates the overall accuracy of the multi-stage model compared to the two traditional pharmacokinetic prediction approaches of FIG. 6.

[0053] FIG. 8 is a flow diagram of an example process for generating a predicted PK profile using a multi-stage machine learning model.

[0054] Like reference numbers and designations in the various drawings indicate like elements.DETAILED DESCRIPTION

[0055] FIG. 1 shows an example pharmacokinetic (PK) profile prediction system 100. The pharmacokinetic profile prediction system 100 is an example of a system implemented as computer programs on one or more computers in one or more locations in which the systems, components, and techniques described below are implemented.

[0056] The PK profile prediction system 100 can process data characterizing a chemical structure 110 of a chemical compound using an end-to-end machine learning approach to generate a PK profile 150 for the chemical compound. The chemical compound being analyzed by the system can be an organic or inorganic chemical compound, e.g., that is identified as a target in the identification and validation stage of drug development. The chemical structure 110 of the chemical compound can characterize the spatial arrangement or a graphical representation of the atoms in a chemical compound, such as a structural formula, ball-and- stick-model, perspective drawing, or space filling model. In some cases, data defining a three- dimensional (3D) arrangement of atoms in the chemical compound can be processed as the chemical structure 110.

[0057] The system can process the chemical structure 110 to generate one or more molecular representation features 120 of the chemical compound. As an example, the molecular representation features 120 can include descriptors of structure-based features that pertain to the chemical compound. As an example, the molecular representation features 120 can include quantitative estimation of drug-likeness (QED), topological polar surface area (TPSA), the number of hydrogen bond donors and acceptors, and the molecular weight of the chemical compound. As another example, the molecular representation features 120 can include a number of rotatable bonds in the chemical compound, a number of rings in the chemical compound, a molecular volume of the chemical compound, and a dipole moment of the chemical compound.

[0058] As another example, the molecular representation features 120 can be natural language based, such as a text string. In particular, the Simplified Molecular-Input Line-Entry (SMILES) representation can be used to flatten the chemical structure of a chemical compound into a single string. More specifically, the SMILES string can be represented as a molecular fingerprint, e.g., a fixed-length binary representation (bitvec of binary digits) that encodes a presence or absence of specific structural patterns or substructures within the chemical compound. An example SMILES representation will be covered in further detail in FIG. 2. As yet another example, the molecular representation features 120 can be graph-based. In particular, the graph can include a respective node for each atom and a respective bond for each edge.

[0059] The molecular representation features 120, including one or more of descriptor-based, natural-language-based, or graph-based features, can then be processed as input to the multistage machine learning model 140 to generate the PK profile 150 as described below.

[0060] In some cases, the molecular representation 120 can be further processed using a feature selection process 130 to identify the most relevant one or more molecular representation features 120 as inputs to the multi-stage model 140 and to mitigate overfitting. In particular, the system 100 can implement a feature selection algorithm to select a subset of molecular representation features 120, e.g., from a collection of candidate molecular representation features 120. For example, the system 100 can evaluate different combinations of the one or more descriptor-based, natural-language-based, and graph-based features molecular representation features 120, to identify the best-performing inputs to the model 140 as part of the feature selection process 130. An example feature selection process 130 will be covered in further detail in FIG. 2.

[0061] In the particular example depicted, the first stage of the multi-stage model 140 includes one or more PK parameter prediction model(s) 142 that can process the molecular representation features 120 to generate one or more predicted PK parameters as input to the second stage; and the second stage includes a PK profile prediction model 144 that processes inputs that can include one or more features selected from the predicted PK parameter predictions to generate a predicted PK profile 150. The PK parameter prediction model(s) 142 can include one or more models that are configurated to separately process the molecular representation features 120 in order to generate respective PK parameter predictions which can serve as inputs to the PK profile prediction model 144.|0062| In particular, the PK parameters can characterize absorption, distribution, metabolism, and elimination (ADME) properties of the chemical compound. As an example, the predicted PK parameters can include a clearance of the chemical compound, e.g., how fast a drug is eliminated or excreted from the body of the subject, a volume of distribution of the chemical compound, a half-life of the chemical compound, e.g., how long it takes the chemical compound to be reduced to half its value, a maximum concentration of the chemical compound, e.g., the maximum serum concentration that a chemical compound achieves in a specified area of the body of the subject, or a time to maximum concentration of the chemical compound, e.g., how long it takes for a drug to reach maximum concentration after administration and absorption. As another example, the predicted PK parameters can include an intrinsic clearance of the chemical compound, e.g., the rate at which the liver can metabolize the drug with unrestricted access, a lipophilicity of the chemical compound, e.g., the ability of the chemical compound to mix with an oily phase (as opposed to water), a solubility of the chemical compound, e.g., the degree to which a substance dissolves in a solvent to make a solution, a fraction unbound of the chemical compound, e.g., the portion that exerts a pharmacologiceffect, and an acidic pKa and a basic pKa of the chemical compound, e.g., an acid dissociation descriptor and a basic dissociation descriptor, respectively.

[0063] The system 100 can predict each of the PK parameters using respective PK parameter prediction model(s) 142. In some examples, the PK parameters in the model input for the PK profde prediction model 144 can be chosen by a PK scientist determining which PK parameters are most relevant to the specific chemical compound 110. In particular, a number of PK parameters can be determined, e.g., 5. 10. or 20 PK parameters and respective PK parameter prediction models 142 can be trained for each of the PK parameters.

[0064] In some examples, the system 100 can use the feature selection process 150 to determine the model input into the PK profile prediction model 144. More specifically, the system 100 can implement a feature selection algorithm to select one or more PK parameter features and optionally other features that are most predictive of PK profiles, e.g., from a collection of candidate features, using the feature selection process 150. As an example, the other features can include features related to the administration of the chemical compound 110 to a subject, e.g., the time of administration and the infusion duration, e.g., the time over which the drug is administered. An example feature selection process 150 will be covered in further detail in FIG. 2.

[0065] In particular, the system 100 can select different features for different subject types using the feature selection processes 130 and 150. For example, the system 100 can select a different set of features as input to the PK parameter prediction model(s) 142 and the PK profile prediction model 144 for different species (e.g., rat, dog, pig, human, etc ), since the PK profile 150 for a chemical structure 110 can be different for different species. In the case that the subject is a human, the system 100 can select a different set of features for different categories of person, e.g.. based on demographic criteria (e.g., age. gender, etc.), medical history (e.g., medical diagnoses), and so forth.

[0066] In some cases, the system 100 can maintain the selected features, e.g., as labels in a database that is indexed by subject ty pe. In this case, given a new subject, the system 100 can categorize that subject into a corresponding category’, and then generate a predicted PK profile 160 for that subject using the feature labels selected for that category of subject. For example, the system 100 can identify (i) the molecular representation features 120, (ii) the PK parameters, and (ii) any additional features corresponding with the feature labels for a particular subject and use the respective PK parameter prediction model(s) 142 to generate the PK parameters that can be processed using the PK profile prediction model 144 to generate the PK profile 150 for the subject.

[0067] In a particular example, an example multi-stage machine learning (ML) model 140 that includes a clearance prediction model, e.g., how fast a drug is eliminated or excreted from the body, and volume of distribution prediction model, e.g., the concentration over time of the chemical compound, as the PK parameter prediction models 142 will be covered in further detail in FIG. 3.

[0068] The PK profile prediction model 144 can be configured to process the one or more PK parameter model 142 outputs as well as additional features as a model input to generate the PK profile 150. In some examples, additional features characterizing the chemical compound structure, e.g., one or more features from the molecular representation 120 features or related PK experimental data, can also be processed by the PK profile prediction model 144.

[0069] More specifically, processing the predicted PK parameters enables the PK profile prediction model 144 to generate the PK profile 150 without any mechanistic assumptions, e.g., without any of the mechanistic assumptions inherent to the one-compartment or noncompartmental analysis models. In particular, these traditional approaches can ignore the impact of mechanistic processes, e.g., protein transport, kidney filtration, etc., that were not part of the initial mechanistic assumptions. In some cases, the two-stage approach can enable the multi-stage machine learning model 140 to capture non-linear trends that the other traditional methods based on mechanistic assumptions cannot, as will be described further in FIG. 6. Additionally, predicting the parameters from the molecular representation 120 features removes the need to collect more data, e.g., through additional animal testing.

[0001] Each of the one or more PK parameter prediction machine learning models and the PK profile prediction machine learning model can have any appropriate machine learning model architectures. For instance, each can be implemented, e.g., as a neural network, a decision tree, a random forest, a support vector machine, or as a combination thereof, and so forth. The PK parameter prediction machine learning models and the PK profile prediction machine learning model can each be implemented using a different machine learning model architecture or using the same machine learning model architecture. In particular, the architecture chosen for each model can depend on the amount of available training data. For example, if little training data is available, then a model with few training parameters can be chosen; or vice versa.

[0002] In some cases, the system can implement model 142, model 144, or both as a regression model, such as generalized linear models (based on simple or Poisson distributions), random forest, support vector regressor, or boosting model, e g., a light gradient-boosting machine or XGBoost. An example implementation of a multi-stage ML model 140 that uses an exampleboosting algorithm for both the one or more PK parameter prediction models 142 and the PK profile prediction model 144 will be covered in more detail with respect to FIG. 3.

[0003] In other cases, the models 142 and 144 can have any appropriate model architecture that enables the model to perform its described functions, e.g., for the PK parameter prediction model 142 to process one or more molecular representation features 120 to predict one or more PK parameters and for the PK profile prediction model 144 to process one or more PK parameters and potentially one or more molecular representation features 120 to predict a PK profile 150. As an example, the models 142 and 144 can each include any appropriate number of neural network layers (e.g., 1 layer, 5 layers, or 10 layers) of any appropriate type (e.g., fully connected layers, attention layers, convolutional layers, etc.) connected in any appropriate configuration (e.g., as a linear sequence of layers or as a directed graph of layers). In some cases, the models 142 and 144 can have different architectures. Furthermore, the one or more PK parameter prediction models 142 can have the same or different architectures.

[0004] The generated PK profile 150 defines the concentration vs. time profile of the chemical compound without extensive animal experiments or allometric scaling calculations. In particular, the PK profile 150 can be used to calculate the area under the PK profile curve (AUC), e.g., by integrating or performing a discrete sum to calculate the area under the curve, to quantify total systemic exposure to the chemical compound and to define the maximum concentration of the chemical compound in the subject as well as the half-life of the chemical compound.

[0005] In some examples, the data defining the predicted PK profile 150 of the chemical compound can be stored in a memory, transmitted over a data communications network, or both. Furthermore, in some examples, the data defining the predicted PK profile 150 can be displayed as a representation on a display of a user device.

[0006] As an example, the generated PK profiles for several compounds can be used for rank ordering of compounds, e.g., ascending or descending ordering with respect to one or more ADME properties, such as the area under curve (AUC). The rank ordering can then be used to select chemical compounds for physical synthesis. For example, chemical compounds can be deemed appropriate for synthesis and further drug development according to a deemed appropriate range of AUC. The PK profile can also be used to experimentally validate the properties of the physically synthesized chemical compounds, e.g., by comparing the generated PK profile 150 to the actual PK profile of the physically synthesized compound when applied to a subject. In some examples, the physically synthesized chemical compound can be administered to a subject. More specifically, if the PK profile of the chemical compounddefines a safe maximum concentration and half-life of the chemical compound in accordance with predefined thresholds, the drug can be cleared for administration to one or more subjects.

[0007] FIG. 2 is a block diagram that illustrates an example preprocessing protocol for a pharmacokinetic profile prediction system, e.g., the PK profile prediction system 100 of FIG. 1. In particular, FIG. 2 illustrates an example preprocessing protocol implemented on the chemical structure input for the chemical compound acetylsalicylic acid (aspirin).

[0008] In this case, the input chemical structure 110 for aspirin can be processed into a naturallanguage representation. More specifically, the system can process the input structure 110 to generate a SMILES representation 210, which flattens the chemical structure 110 into a single string.

[0009] The SMILES representation can be further processed to generate a molecular representation 120. In the particular example depicted, the molecular representation 120 can include both descriptor-based features 220 and molecular fingerprints 230. More specifically, the descriptor-based features for aspirin can include the topological polar surface area (TPSA), e.g., 63.6, the molecular weight, e.g., 180.4, the number of rotatable bonds in aspirin, e.g., 3, the number of rings in aspirin, e.g., 1, and the number of hydrogen donors or acceptors, e.g.. 1 and 3, respectively.

[0010] The fingerprint 230 is a fixed-length binary' representation that encodes a presence or absence of specific structural patterns or substructures within the chemical compound.|00111 The system can perform feature selection 130 to identify the most relevant features 220, 230 for the multi-stage machine learning model, e.g., the multi-stage machine learning model 140 of FIG. 1. In particular, the system can perform feature selection 130 to identify the selected features 240 as input to the PK parameter prediction model (s) 142, and the system can perform feature selection 150 to identify the most relevant features, e.g., including PK parameters generated using the PK parameter prediction model(s) 142, as input to the PK profile prediction model 144.

[0012] In some examples, the feature selection process 130, 150, or both can specify either a target number of selected features or a target distillation amount, e.g.. selecting less than 50%, or less than 10%, or less than 5%, or less than 1% of the collection of candidate features for inclusion in the model input.

[0013] In some examples, the system can implement forward or backwards variable selection as feature selection 130, 150, or both, in which features are added or dropped depending on a measurement of fit, e.g., R2. In other examples, the system can implement a variance-based method, such as variance inflation factor, to remove collinear features.

[0014] In some cases, the system can perform the feature selection 150 before the feature selection 130. In particular, the feature selection process 150 can be performed on PK and physiochemical property (PC) experimental results, e.g., from data gathered through previous experiments and molecular property data, and the descriptors 220 and fingerprints 230 to yield the selected features 260 as model inputs for the PK profile prediction model 144. In this case, the system can then use the selected features 260 to inform the configuration of and feature selection 130 for the PK parameter prediction models 142.

[0015] More specifically, as part of the feature selection process 130, the system can evaluate different combinations of the descriptor-based, natural-language-based, and graph-based features, including descriptors 220 and fingerprints 230 to identify the best-performing inputs to the PK parameter prediction model(s) 142, which provide the selected features 260 that sen e as inputs to the PK profile prediction model 144.

[0016] In the particular example depicted, the feature selection process 150 yields the clearance (CL) and volume of distribution (Vdss) of aspirin as the selected features 260 for the PK profile prediction model 144. Once selected, these features 260 can become the targets for prediction within the first stage of the multi-stage machine learning model, such that the PK parameter prediction model(s) 142 can generate clearance and volume of distribution predictions to be processed as model inputs by the PK profile prediction model 144.

[0017] In some cases, the selected features 260 can include one or more additional inputs to the PK profile prediction model 144. As an example, the selected features 260 can also include the TPS A and number of hydrogen donors. In this case, the PK profile prediction model 144 can process the predicted clearance and volume of distribution, e.g., as generated by the PK parameter prediction model(s) 142, along with the TPS A and the number of hydrogen donors to generate the PK profile.

[0018] For example, the system can perform feature selection 150 on a collection of candidate features that includes the volume of distribution, in-vivo clearance, pKa (acidic), pKa (basic), fraction unbound, lipophilicity, molecular weight, time of administration, and infusion duration for a chemical structure 110. In some cases, the system 100 can select a different set of features for subjects of different species, e.g., rat, dog. pig. human, etc. As an example, the system 100 can select the volume of distribution, in-vivo clearance, and time of administration as the selected features 260 for a rat subject. As another example, the system 100 can select volume of distribution, in-vivo clearance, acidic pKa, basic pKa, fraction unbound, lipophilicity, infusion duration, and time of administration as the selected features 260 for a human subject.

[0019] The system can then perform the feature selection 130 on a collection of candidate features that includes one or more descriptor 220 and fingerprint 230 features to identify the best-performing inputs to the PK parameter prediction model(s) 142, e.g., the selected features 240, that can be processed to generate the PK parameters selected as part of the features 260.

[0020] FIG. 3 depicts an example multi-stage machine learning model whose implementation is informed by the selected clearance and volume of distribution features chosen through feature selection in FIG. 2. In particular, the multi-stage machine learning model can include a first stage where one or more first machine learning models translates chemical structure data to predicted PK parameters, and a second stage where one or more second machine learning models translate the PK parameters to full predicted PK profiles. For example, the multi-stage machine learning model 300 can be used to predict the PK profile for a rat subject.

[0021] The multi-stage prediction model 300 can include two stages, e.g., a first stage 305 and a second stage 340. More specifically, the first stage 305 can involve the one or more PK parameter prediction model(s) 142 processing molecular representation features to generate predicted PK parameters 330 and the second stage 340 can involve the PK profile prediction model 144 processing the predicted PK parameters 330 to generate the PK profile 150.

[0022] In the particular example depicted, the predicted PK parameters 330 include a clearance (CL) prediction 332 and a volume of distribution (Vdss) prediction 334, which are generated by the corresponding CL model 315 and Vdss model 325. The parameter prediction models 142 do not necessarily have to process the same types of features. In this example, the CL model 315 processes molecular fingerprints 310 to generate a CL prediction 330 and the Vdss Model 325 processes molecular descriptors 320 to generate a Vdss prediction 335.

[0023] In some cases, an additional feature selection process, e.g., an additional process to the feature selection 130 of FIGS. 1 and 2. can be performed to determine the model inputs for each PK parameter prediction model 142 to determine which different molecular representation features result in higher fi del i ty predictions with respect to the training data set. In the particular example depicted, this additional feature selection process has resulted in the CL model 325 processing molecular fingerprints 310 to generate the CL prediction 332 and the Vdss model 325 processing molecular descriptors 320 to generate the Vdss prediction 334.

[0024] In the second stage 340, the PK profile prediction model 144 processes the predicted PK parameters 330, e.g., the CL prediction 332 and the Vdss prediction 334, to generate the PK profile 150. As an example, the PK profile prediction model 144 can implement the same algorithm as the PK parameter prediction models 142. In particular, both the first 305 andsecond stages 330 of the multi-stage ML model 300 can be implemented using a boosting model, which will be covered in more detail in FIG. 4.

[0025] FIG. 4 depicts an example boosting model. For example, one or more of the PK parameter prediction models 142 of FIG. 1, the PK profde prediction model 144 of FIG. 1, or both can be implemented using the depicted boosting model. In the particular example discussed below, the PK profile prediction model can be implemented as a boosting model 400 that includes one or more base machine learning models, e.g., model I 410, model 2 420, and model 4 440.

[0026] More specifically, the one or more base machine learning models 410, 420, and 440 can process the input to generate the predicted PK profile of the chemical compound by combining each respective output into an aggregated output 450. In the particular example depicted, the predicted PK profile is generated by summing each output of the one or more base machine learning models. In another example, the average of all the base model outputs can be used to generate the aggregated output 450.

[0027] In particular, each base machine learning model after a first base machine learning model can be associated with an ordering in a sequence and configured, through a training process, to generate an output that corrects an error between the target PK profile and the PK profile generated as the aggregated output 450 by combining the outputs of preceding base machine learning models in the sequence. In the particular example depicted, each of the base machine learning models can be implemented as a decision tree model. The first decision tree model can be a constant model 0 (not pictured) and a series of trees can be added that result in model 1 410 with output 415.

[0028] More specifically, the output of model 1 410 can have a certain residual error 410 with respect to how the decision rule implemented by the decision tree performs on the data, e.g., wi th respect to the target PK profile. The residual error on model 1 410 can then be predicted using model 2 to generate output 425, which can then be predicted using model 3 (not pictured) and onwards such that the residual of model k-1 can be input into model 4 440 to generate output 445.

[0029] The training process for boosting model 400 can include training the set of PK profile prediction machine learning model parameters using one or more training examples that include a target PK profile of the chemical compound and the PK profile generated by combining the outputs of preceding base machine learning models in the sequence. In particular, for each training example, the PK profile prediction machine learning model can be trained to reduce a discrepancy, e.g., a measure of distance, between the target PK profile ofthe training example and a predicted PK profile generated by the PK profile prediction learning model. For example, the PK profile prediction model can be trained using an optimization algorithm, such as gradient descent, to reduce a discrepancy between the target PK profile of a given training example and the predicted PK profile generated by the aggregated decision trees of the PK profile prediction model.

[0030] More specifically, the boosting model 400 demonstrates how the model can be implemented as a decision tree ensemble configured, through training, to generate an output that corrects an error between a target, e.g., a PK parameter or profile, and an output, e.g., a corresponding PK parameter or profile generated by combining the output of preceding decision tree models. In the particular example depicted, during training, the system can minimize prediction errors by adding sequential tree models to predict the residual errors of prior errors that can be optimized using gradient descent, e.g. to minimize the aggregated tree model loss.

[0031] In the particular example depicted, the gradient descent algorithm minimizes prediction errors by adding sequential tree models to minimize the aggregated boosting model 400 loss. In a particular example that implements XGBoost as the boosting model 400, the addition of new decision rule tree models with respect to the sequence of already existing base tree models can be evaluated by comparing the effectiveness of the added tree model, e.g., using a criterion to score the split represented by the decision rule implemented by the new decision tree against the same criterion without the new decision tree. In the case of XGBoost, a pruning technique uses these scores to keep new high-scoring decision rule trees in the ensemble and reject new low-scoring decision rule trees from the ensemble.

[0032] FIG. 5 illustrates the accuracy of the first stage of the multi-stage machine learning model, e.g., the one or more PK parameter prediction models 142 of model 300 in FIG. 3, in predicting PK parameters, e.g., for clearance (CL) and volume of distribution (Vdss). In particular, graphs 500 and 550 are scatterplots that depict the fidelity of the predicted distribution for CL and for VdSS. For example, the results of graphs 500 and 550 can result from the PK parameter prediction models 142 being implemented using XGBoost and trained as discussed in FIG. 4.

[0033] In the particular example depicted, the ground truth 510 experimental data is represented by an identify reference line that corresponds with perfect overlap between the predicted (measured) and observed experimental data. In this case, the CL model, e.g., the CL model 315 of FIG. 3. is can replicate the ground truth 510 clearance values with high accuracy and precision, as demonstrated by closeness to the identity reference line and the overlapbetween the ground truth reference line and the train 520 and test set 530 distribution predictions. Likewise, the Vdss model, e.g., the Vdss model 325 of FIG. 3. is able to replicate the ground truth 560 with high precision and accuracy, as demonstrated by closeness to the identity reference line between the ground truth reference line and the train 580 and test set 570 predictions. More specifically, both the CL predicted by the CL model and the Vdss predicted by the Vdss model maintain mean average percentage error (MAPE) less than 100% for both in-vivo Cl and Vdss.

[0034] FIG. 6 demonstrates how the multi-stage machine learning model is able to capture nonlinearities in the generated PK profile that other pharmacokinetic prediction approaches cannot account for given their mechanistic assumptions. In particular. FIG. 6 compares the PK profile prediction system 100 of FIG. 1 to the standard one-compartment model 612 and noncompartmental analysis model 614 for four compounds: compound A, B, C, and D. The results for compound A, B, C, and D are depicted respectively in graphs 610, 620, 630, and 640, respectively.

[0035] PK scientists and researchers can use one-compartment models to predict the efficacy and toxicity of chemical compounds over different body compartments, e.g., the bloodstream, tissues, and organs. One-compartment models assume mechanistic constraints with respect to the functioning of the defined body compartments that are often unknown in the very early phases of drug development, such as where the drug w ill go and how it will be eliminated. The results of the one-compartment model 612 for each of the four compounds, compound A, B, C, and D, are depicted in graphs 10, 620, 630, and 640.

[0036] PK scientists and researchers can also use noncompartmental analysis to predict the efficacy and toxicity of chemical compounds. In this case, noncompartmental analysis assumes that the chemical compound is distributed uniformly throughout the body and that the elimination of the compound is defined as a constant rate, which does not account for physiology. This modeling approach allows for rapid generation of PK profile estimates but does not account for any physiological mechanisms or biological processes that drive or alter the PK. In the particular example depicted, the example noncompartmental analysis model is a PBPK (physiologically-based pharmacokinetic) model that incorporates blood flow and tissue composition of organs. The results of the noncompartmental analysis model 614 for each of the four compounds, compound A, B, C, and D, are depicted in graphs 610, 620, 630, and 640.

[0037] For all compounds, as demonstrated by the graphs 610, 620, 630, and 640 for the four compounds, the ML approach 616 is better able to capture the nonlinear trends of the groundtruth PK data, as represented by the dots, than either the one-compartment 612 or PBPK model 614.

[0038] In particular, even though the one-compartment model 612 can capture the maximum concentration reasonably well, it fails to capture the lower concentration time points for multiple compounds which show bi-exponential behavior. In particular, the one-compartment model 612 results in PK profile curves with mono-exponential decay, which miss second order and higher dynamics, including the curvature of the target PK profile. On the other hand, the PBPK model 614 can capture the bi-exponential decay of the PK profile, since it considers distribution of drug to different tissues, but is unable to capture the maximum concentration in comparison to one-compartment model. It can be observed that the predictions generated by the proposed multi-stage machine learning model of this specification 612 can capture both mono-exponential and bi-exponential PK profiles in 610, 620, 630, and 640.

[0039] FIG. 7 illustrates the overall accuracy of the multi-stage machine learning model, compared to the one-compartmental and noncompartmental analy sis approaches. In particular, panel 700 demonstrates the accuracy of the multi-stage machine learning model in scatterplots 710, 720. and 730. e.g., the model illustrated in FIG. 3 predicting the PK profile with respect to the other traditional approaches. Additionally, table 750 demonstrates different performance metrics for the three approaches.

[0040] The scatterplots in panel 700 include the PK profile results for the multi-stage machine learning approach of this specification 710, the noncompartmental analysis model 720, and the one-compartment model 730, e g., the one-compartment and noncompartmental analysis models of FIG. 6, with respect to an identify reference line. The graph 710 illustrates how the ML results follow the identity reference line more uniformly, especially compared to the one- compartment model 730, which deviates from the identity line at lower concentrations.

[0041] Table 750 demonstrates performance metrics, such as the RMSE, MAPE, MAE, and R2, for the model. It can be observed that the MAPE and MAE for the pharmacokinetic profile prediction system of this specification are lower than that of traditional approaches, while the R2value of the pharmacokinetic profile prediction system of this specification is higher than that of traditional approaches.

[0042] FIG. 8 is a flow diagram of an example process for generating a predicted PK profile using a multi-stage machine learning model. For convenience, the process 800 will be described as being performed by a system of one or more computers located in one or more locations. For example, a training system, e.g., the pharmacokinetic (PK) profile predictionsystem 100 of FIG. 1, appropriately programmed in accordance with this specification, can perform the process 800.

[0043] In particular, the system can receive a chemical compound (step 810), and determine one or more PK features of the chemical compound (step 820). As an example, the PK parameters can include molecular representations, such as molecular fingerprints and descriptors, and PK and physiochemical (PC) features, such as a clearance of the chemical compound, a volume of distribution of the chemical compound, a half-life of the chemical compound, a maximum concentration of the chemical compound, or a time to maximum concentration of the chemical compound. In some cases, the features can be further distilled into a smaller set of features using feature selection.

[0044] The system can then process the features characterizing the chemical structure of the chemical compound using one or more PK parameter prediction models (step 830). In particular, the features can be processed using separate PK models to generate different data defining the PK parameter (step 840).

[0045] The one or more predicted PK parameters can then be processed using a PK profile prediction model (step 850). In some cases, the PK profile prediction model can additionally process one or more of the features identified in step 820 using the PK profile prediction model. As an example, the model inputs to the PK profile model can be selected through a feature selection process that identifies the most relevant inputs for the PK profile prediction model. More specifically, the PK profile prediction model can process the model inputs, including one or more predicted PK parameter outputs of each respective PK parameter prediction model, to generate a predicted PK profile of the chemical compound (step 860).

[0046] This specification uses the term “configured” in connection with systems and computer program components. For a system of one or more computers to be configured to perform particular operations or actions means that the system has installed on it software, firmware, hardware, or a combination of them that in operation cause the system to perform the operations or actions. For one or more computer programs to be configured to perform particular operations or actions means that the one or more programs include instructions that, when executed by data processing apparatus, cause the apparatus to perform the operations or actions.

[0047] Embodiments of the subject matter and the functional operations described in this specification can be implemented in digital electronic circuitry, in tangibly-embodied computer software or firmware, in computer hardware, including the structures disclosed in this specification and their structural equivalents, or in combinations of one or more of them.Embodiments of the subject matter described in this specification can be implemented as one or more computer programs, e.g., one or more modules of computer program instructions encoded on a tangible non-transitory storage medium for execution by, or to control the operation of, data processing apparatus. The computer storage medium can be a machine- readable storage device, a machine-readable storage substrate, a random or serial access memory device, or a combination of one or more of them. Alternatively or in addition, the program instructions can be encoded on an artificially-generated propagated signal, e.g.. a machine-generated electrical, optical, or electromagnetic signal, that is generated to encode information for transmission to suitable receiver apparatus for execution by a data processing apparatus.

[0048] The term “data processing apparatus" refers to data processing hardware and encompasses all kinds of apparatus, devices, and machines for processing data, including by way of example a programmable processor, a computer, or multiple processors or computers. The apparatus can also be, or further include, special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application-specific integrated circuit). The apparatus can optionally include, in addition to hardware, code that creates an execution environment for computer programs, e.g., code that constitutes processor firmware, a protocol stack, a database management system, an operating system, or a combination of one or more of them.|0049| A computer program, which may also be referred to or described as a program, software, a software application, an app, a module, a software module, a script, or code, can be written in any form of programming language, including compiled or interpreted languages, or declarative or procedural languages; and it can be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment. A program may, but need not, correspond to a file in a file system. A program can be stored in a portion of a file that holds other programs or data, e.g., one or more scripts stored in a markup language document, in a single file dedicated to the program in question, or in multiple coordinated files, e.g., files that store one or more modules, sub-programs, or portions of code. A computer program can be deployed to be executed on one computer or on multiple computers that are located at one site or distributed across multiple sites and interconnected by a data communication network.

[0050] In this specification the term “engine” is used broadly to refer to a software-based system, subsystem, or process that is programmed to perform one or more specific functions. Generally, an engine will be implemented as one or more software modules or components,installed on one or more computers in one or more locations. In some cases, one or more computers will be dedicated to a particular engine; in other cases, multiple engines can be installed and running on the same computer or computers.

[0051] The processes and logic flows described in this specification can be performed by one or more programmable computers executing one or more computer programs to perform functions by operating on input data and generating output. The processes and logic flows can also be performed by special purpose logic circuitry, e.g.. an FPGA or an ASIC, or by a combination of special purpose logic circuitry and one or more programmed computers.

[0052] Computers suitable for the execution of a computer program can be based on general or special purpose microprocessors or both, or any other kind of central processing unit. Generally, a central processing unit will receive instructions and data from a read-only memory or a random access memory or both. The essential elements of a computer are a central processing unit for performing or executing instructions and one or more memory devices for storing instructions and data. The central processing unit and the memory7can be supplemented by, or incorporated in, special purpose logic circuitry. Generally, a computer will also include, or be operatively coupled to receive data from or transfer data to. or both, one or more mass storage devices for storing data, e.g., magnetic, magneto-optical disks, or optical disks. However, a computer need not have such devices. Moreover, a computer can be embedded in another device, e.g., a mobile telephone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a Global Positioning System (GPS) receiver, or a portable storage device, e.g., a universal serial bus (USB) flash drive, to name just a few.

[0053] Computer-readable media suitable for storing computer program instructions and data include all forms of non-volatile memory7, media and memory devices, including by way of example semiconductor memory devices, e.g., EPROM, EEPROM, and flash memory devices; magnetic disks, e.g., internal hard disks or removable disks; magneto-optical disks; and CD-ROM and DVD-ROM disks.

[0054] To provide for interaction with a user, embodiments of the subject matter described in this specification can be implemented on a computer having a display device, e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor, for displaying information to the user and a keyboard and a pointing device, e.g., a mouse or a trackball, by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback, e.g., visual feedback, auditory feedback, or tactile feedback; and input from the user can be received in any form, including acoustic, speech, or tactile input. In addition, acomputer can interact with a user by sending documents to and receiving documents from a device that is used by the user; for example, by sending web pages to a web browser on a user’s device in response to requests received from the web browser. Also, a computer can interact with a user by sending text messages or other forms of message to a personal device, e.g., a smartphone that is running a messaging application, and receiving responsive messages from the user in return.

[0055] Data processing apparatus for implementing machine learning models can also include, for example, special-purpose hardware accelerator units for processing common and compute-intensive parts of machine learning training or production, e.g., inference, workloads.

[0056] Machine learning models can be implemented and deployed using a machine learning framework, e.g., a TensorFlow framework, or a Jax framework.

[0057] Embodiments of the subject matter described in this specification can be implemented in a computing system that includes a back-end component, e.g., as a data server, or that includes a middleware component, e.g., an application server, or that includes a front-end component, e.g.. a client computer having a graphical user interface, a web browser, or an app through which a user can interact with an implementation of the subject matter described in this specification, or any combination of one or more such back-end, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication, e.g., a communication network. Examples of communication networks include a local area network (LAN) and a wide area network (WAN), e.g., the Internet.

[0058] The computing system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. In some embodiments, a server transmits data, e.g., an HTML page, to a user device, e.g., for purposes of displaying data to and receiving user input from a user interacting with the device, which acts as a client. Data generated at the user device, e.g.. a result of the user interaction, can be received at the server from the device.

[0059] While this specification contains many specific implementation details, these should not be construed as limitations on the scope of any invention or on the scope of what may be claimed, but rather as descriptions of features that may be specific to particular embodiments of particular inventions. Certain features that are described in this specification in the contextof separate embodiments can also be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment can also be implemented in multiple embodiments separately or in any suitable subcombination. Moreover, although features may be described above as acting in certain combinations and even initially be claimed as such, one or more features from a claimed combination can in some cases be excised from the combination, and the claimed combination may be directed to a subcombination or variation of a subcombination.

[0060] Similarly, while operations are depicted in the drawings and recited in the claims in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations be performed, to achieve desirable results. In certain circumstances, multitasking and parallel processing may be advantageous. Moreover, the separation of various system modules and components in the embodiments described above should not be understood as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0061] Particular embodiments of the subject matter have been described. Other embodiments are within the scope of the following claims. For example, the actions recited in the claims can be performed in a different order and still achieve desirable results. As one example, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some cases, multitasking and parallel processing may be advantageous.

Claims

What is claimed is:

1. A method performed by one or more computers, the method comprising: processing data characterizing a chemical structure of a chemical compound using a multi-stage machine learning model to generate data defining a predicted pharmacokinetic (PK) profile of the chemical compound, comprising: processing a model input comprising the data characterizing the chemical structure of the chemical compound, using one or more PK parameter prediction machine learning models, to generate one or more predicted PK parameters of the chemical compound; and processing a model input comprising the one or more predicted PK parameters generated by the one or more PK parameter prediction machine learning models, using a PK profile prediction machine learning model, to generate the predicted PK profile of the chemical compound.

2. The method of claim 1, wherein the one or more predicted PK parameters comprise one or more of: a clearance of the chemical compound, a volume of distribution of the chemical compound, a half-life of the chemical compound, a maximum concentration of the chemical compound, or a time to maximum concentration of the chemical compound.

3. The method of any preceding claim, wherein the predicted PK profile of the chemical compound defines a concentration of the chemical compound in a subject over time after the chemical compound is administered to the subject.

4. The method of any preceding claim, wherein the model input to the PK profile prediction machine learning model further comprises features characterizing the chemical structure of the chemical compound.

5. The method of any preceding claim, wherein the model input to the PK profile prediction machine learning model comprises a plurality of features selected by performing operations comprising: identifying a collection of candidate features for inclusion in the model input to the PK profile prediction machine learning model, wherein the collection of candidate features comprise: (i) features defining chemical compound PK parameters, and (ii) features characterizing chemical compound structure; andselecting a proper subset of the collection of candidate features for inclusion in the model input to the PK profile prediction machine learning model.

6. The method of claim 5, wherein the features characterizing the chemical compound structure comprise molecular descriptor features and molecular fingerprint features.

7. The method of any one of claims 5-6, wherein selecting a proper subset of the collection of candidate features for inclusion in the model input to the PK profile prediction machine learning model comprises: selecting less than 50%, or less than 10%, or less than 5%, or less than 1% of the collection of candidate features for inclusion in the model input to the PK profile prediction machine learning model.

8. The method of any one of claims 6-7, wherein selecting a proper subset of the collection of candidate features for inclusion in the model input to the PK profile prediction machine learning model comprises: applying a feature selection algorithm, to the collection of candidate features, in order to select a proper subset of the collection of candidate features that are predicted to optimize a performance of the PK profile prediction machine learning model.

9. The method of any preceding claim, wherein processing the model input comprising the one or more predicted PK parameters, using the PK profile prediction machine learning model, to generate the predicted PK profile of the chemical compound comprises: processing the model input comprising the one or more predicted PK parameters using each of a plurality of base machine learning models included in the PK profile prediction machine learning model; and generating the predicted PK profile of the chemical compound by combining a respective output of each of the plurality of base machine learning models included in the PK profile prediction machine learning model.

10. The method of claim 9, wherein generating the predicted PK profile of the chemical compound by combining the respective output of each of the plurality of base machine learning models included in the PK profile prediction machine learning model comprises: summing the respective output of each of the plurality of base machine learning models included in the PK profile prediction machine learning model.

11. The method of any one of claims 9-10, wherein the plurality of base machine learning models included in the PK profile prediction machine learning model are associated with an ordering into a sequence; and each base machine learning model after a first base machine learning model in the sequence is configured, through training, to generate an output that corrects an error between: (i) a target PK profile, and (ii) a PK profile generated by combining the outputs of preceding base machine learning models in the sequence.

12. The method of any preceding claim, wherein the PK profile prediction machine learning model has a set of PK profile prediction machine learning model parameters have been trained on a set of training examples by a machine learning training technique.

13. The method of claim 12, wherein for each of one or more training examples from the set of training examples, the training example corresponds to a training chemical compound and comprises: (i) a training model input comprising one or more predicted PK parameters generated by the one or more PK parameter prediction machine learning models for the training chemical compound, and (ii) a target PK profile of the training chemical compound.

14. The method of claim 13, wherein training the PK profile prediction machine learning model on the set of training examples comprises, for each training example, training the PK profile prediction machine learning model to reduce a discrepancy between: (i) the target PK profile of the training example, and (ii) a predicted PK profile generated by the PK profile prediction machine learning model by processing the training model input of the training example.

15. The method of any preceding claim, wherein generating one or more predicted PK parameters of the chemical compound comprises: processing the model input comprising the data characterizing the chemical structure of the chemical compound using each of a plurality of PK parameter prediction machine learning models to generate a respective predicted PK parameter as an output of each of the plurality of PK parameter prediction machine learning model.

16. The method of claim 15, wherein each of the plurality of PK parameter prediction machine learning models has been configured through training by a machine learning training technique to generate a respective predicted PK parameter.

17. The method of any preceding claim, wherein the data characterizing the chemical structure of the chemical compound comprises one or more molecular descriptors of the chemical compound.

18. The method of claim 17, wherein the one or more molecular descriptors of the chemical compound define one or more of a molecular weight of the chemical compound, a total number of hydrogen bond donors in the chemical compound, a total number of hydrogen bond acceptors in the chemical compound, a topological polar surface area of the chemical compound, a number of rotatable bonds in the chemical compound, a number of rings in the chemical compound, a molecular volume of the chemical compound, a dipole moment of the chemical compound, or a three-dimensional (3D) arrangement of atoms in the chemical compound.

19. The method of any preceding claim, wherein the data characterizing the chemical structure of the chemical compound comprises a molecular fingerprint of the chemical compound.

20. The method of claim 19, wherein the molecular fingerprint of the chemical compound defines a fixed-length binary representation that encodes a presence or absence of specific structural patterns or substructures within the chemical compound.

21. The method of any preceding claim, wherein the data characterizing the chemical structure of the chemical compound is derived from a representation of the chemical structure of the chemical compound as a textual string.

22. The method of claim 21, wherein the chemical structure of the chemical compound as a textual string comprises a Simplified Molecular Input Line Entry System (SMILES) string.

23. The method of any preceding claim, further comprising processing the predicted PK profile to determine an area under the predicted PK profile.

24. The method of any preceding claim, further comprising selecting the chemical compound for physical synthesis based at least in part on the predicted PK profile of the chemical compound.

25. The method of claim 24, further comprising physically synthesizing the chemical compound.

26. The method of claim 25, further comprising administering the physically synthesized chemical compound to a subject.

27. The method of claim 25, further comprising experimentally validating one or more properties of the physically synthesized chemical compound.

28. The method of any preceding claim, further comprising providing the data defining the predicted PK profile of the chemical compound, comprising one or more of: storing the data defining the predicted PK profile of the chemical compound in a memory; transmitting the data defining the predicted PK profile of the chemical compound over a data communications network; or displaying a representation of the predicted PK profile of the chemical compound on a display of a user device.

29. A system comprising: one or more computers; and one or more storage devices communicatively coupled to the one or more computers, wherein the one or more storage devices store instructions that, when executed by the one or more computers, cause the one or more computers to perform operations of the respective method of any one of claims 1-24 and 28.

30. One or more non-transitory computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations of the respective method of any one of claims 1-24 and 28.

Images