Predicting pharmacokinetic profiles of compounds using machine learning

Abstract

A method, system, and apparatus comprising a computer program encoded on a computer storage medium for generating a predicted pharmacokinetic (PK) profile of a compound from the structure of a compound using a multi-stage machine learning model. In one embodiment, the system includes processing a model input, which includes data characterizing the chemical structure of a compound, using one or more PK parameter predictive machine learning models to generate one or more predicted PK parameters, and processing a model input, which includes one or more predicted PK parameters generated by the one or more PK parameter predictive machine learning models, using a PK profile predictive machine learning model to generate a predicted PK profile of the compound.

Description

TECHNICAL FIELD

[0001] This specification relates to the prediction of pharmacokinetic profiles of compounds using machine learning.

BACKGROUND ART

[0002] A machine learning model receives an input and generates an output, for example a predicted output, based on the received input. Some machine learning models are parametric models, which generate outputs based on the received inputs and the values of the parameters of the model.

[0003] The pharmacokinetic (PK) profile of a compound may define, for example, a concentration-versus-time curve that may represent the absorption, distribution, metabolism and excretion (ADME) of the compound from the body of a subject. The PK profile provides information on how a compound behaves in terms of concentration levels in various body compartments, such as blood flow, tissues and organs. Obtaining an accurate PK profile of a compound is essential for drug development, and may influence considerations of dosing regimens, efficacy and safety.

SUMMARY OF THE INVENTION

MEANS FOR SOLVING THE PROBLEMS

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

[0005] In this specification, “PK parameters” characterize global summary statistics that characterize the properties of a compound, such as its absorption, distribution, metabolism, or excretion in a subject. For example, the subject may be a rat, dog, pig, human, or other species. A “PK” profile includes data that defines the concentration value of the compound in the subject at each point in time in a series of time points. In some cases, the PK profile for a particular compound will differ for different species.

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

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

[0008] In some implementations, the predicted PK profile of a compound defines the time course of the compound in the subject after it has been administered to the subject.

[0009] In some implementations, the model input to one or more PK profile predictive machine learning models further includes features that characterize the chemical structure of a compound.

[0010] In some implementations, the model input to one or more PK profile predictive machine learning models includes a set of candidate features to be included in the model input to one or more PK profile predictive machine learning models, selected by performing an operation that includes identifying a set of candidate features including (i) features that define compound PK parameters and (ii) features that characterize the compound structure, and selecting an appropriate subset of the set of candidate features to be included in the model input to one or more PK profile predictive machine learning models.

[0011] In some implementations, the features that characterize the compound structure include molecular descriptor features and molecular fingerprint features.

[0012] In some implementations, selecting an appropriate subset of candidate features to include in the model input to one or more PK profile predictive machine learning models involves selecting less than 50%, or less than 10%, or less than 5%, or less than 1% of the set of candidate features to include in the model input to one or more PK profile predictive machine learning models.

[0013] In some implementations, selecting an appropriate subset of candidate features to include in the model input to one or more PK profile predictive machine learning models involves applying a feature selection algorithm to the set of candidate features to select an appropriate subset of candidate features that is predicted to optimize the performance of one or more PK profile predictive machine learning models.

[0014] In some implementations, using a PK profile prediction machine learning model to process a model input containing one or more predicted PK parameters to generate a predicted PK profile of a compound includes processing a model input containing one or more predicted PK parameters using each of several underlying machine learning models included in the PK profile prediction machine learning model, and generating a predicted PK profile of a compound by combining the respective outputs of each of the several underlying machine learning models included in the PK profile prediction machine learning model.

[0015] In some implementations, generating a predicted PK profile of a compound by combining the outputs of each of the multiple underlying machine learning models included in the PK profile prediction machine learning model involves summing the outputs of each of the multiple underlying machine learning models included in the PK profile prediction machine learning model.

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

[0018] In some implementations, for each of the one or more training examples from the set of training examples, the training example corresponds to a training compound and includes a training model input that contains (i) one or more predicted PK parameters generated by one or more PK parameter predictive machine learning models for the training compound, and (ii) a target PK profile of the training compound.

[0019] In some implementations, training a PK profile prediction machine learning model on a set of training examples involves training the PK profile prediction machine learning model to reduce the discrepancy between (i) the target PK profile of the training example and (ii) the training model input of the training example, thereby reducing the discrepancy between the PK profile prediction machine learning model and the predicted PK profile generated by the PK profile prediction machine learning model.

[0020] In some implementations, generating one or more predicted PK parameters for a compound involves using each of several PK parameter prediction machine learning models to process model inputs containing data characterizing the compound's chemical structure, and generating each predicted PK parameter as the output of each of the several PK parameter prediction machine learning models.

[0021] In some implementations, each of the multiple PK parameter prediction machine learning models is configured to generate its own predicted PK parameters through training using machine learning training techniques.

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

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

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

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

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

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

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

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

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

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

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

[0033] In some implementations, the method further includes providing data defining a predicted PK profile of a compound, which comprises one or more of storing the data defining the predicted PK profile of the compound in a memory, transmitting the data defining the predicted PK profile of the compound via a data communication network, or displaying a representation of the predicted PK profile of the compound on a display of a user device.

[0034] In another embodiment, a system is provided comprising one or more computers and one or more storage devices communicably coupled to one or more computers, wherein one or more storage devices, when executed by one or more computers, store instructions causing one or more computers to perform operations in the manner described herein.

[0035] One or more non-temporary computer storage media that, when executed by one or more computers, store instructions causing one or more computers to perform operations in the manner described herein.

[0036] The subject matter described herein may be implemented to achieve one or more of the following advantages in particular:

[0037] Drug discovery is a complex process encompassing several distinct stages. First, in target identification and validation, potential biological targets for specific diseases are identified. Next, in lead discovery and optimization, numerous compounds targeted at the specified targets are synthesized, screened, and optimized for their absorption, distribution, metabolism, and removal (ADME) characteristics and physicochemical (PC) properties. Finally, preclinical development is conducted to evaluate the pharmacokinetics (PK), efficacy, and toxicity of the selected compounds in animal models before potential regulatory approval and clinical trials. PK profile evaluation plays a crucial role in understanding the safety and efficacy of a compound and determining the initiation point and effective dose for future clinical trials. Traditionally, human PK profiles have been predicted by scaling PK profiles from preclinical species such as mice, rats, and non-human primates (NHPs). However, conventional methods for obtaining PK profiles based on animal experiments are expensive, time-consuming, unscalable, and require animal sacrifice.

[0038] To address these issues, the system described herein provides a multi-stage machine learning model capable of generating a predicted PK profile of a compound by processing data characterizing the compound's chemical structure. That is, the machine learning model can translate the chemical structure into a predicted PK profile in silico, without requiring physical experiments or animal testing of the compound. This system, therefore, can accelerate the drug development process while minimizing costs and reducing the need to sacrifice animals.

[0039] This system can generate a predicted PK profile of a compound using a multi-stage machine learning model with two stages. In the first stage, the system uses one or more PK parameter prediction machine learning models to process data characterizing the chemical structure of a compound and generate one or more predicted PK parameters of the compound, such as clearance, volume of distribution, half-life, etc. In the second stage, the system uses a PK profile prediction machine learning model to process data containing the predicted PK parameters and generate a predicted PK profile of the compound. In other words, the multi-stage machine learning model includes a first stage in which one or more machine learning models convert chemical structure data into structure-related features, and a second stage in which the machine learning models convert these features into a predicted complete PK profile.

[0040] The multi-stage architecture of machine learning models enables several technical advantages, which are detailed below.

[0041] Firstly, this system avoids the use of mechanical PK models, such as one-compartment models, two-compartment models, compartment models with enzyme dynamics, or tissue compartment models, by using a PK profile prediction machine learning model in the second stage. Mechanical PK models are explicit mathematical models that attempt to describe the underlying biological and physiological processes that govern the behavior of drugs in the body. However, mechanical PK models are supported by assumptions such as compartmentalization (e.g., the body can be divided into separate compartments or regions), mass balance, primary dynamics, and instantaneous mixing. The assumptions supporting mechanical PK models often contradict real-world physiological processes and are typically simplifications that are difficult to verify.

[0042] In contrast, PK profile prediction machine learning models do not rely on these assumptions; rather, they learn to implicitly infer about the underlying physical processes through information encoded in the parameter values ​​of the PK profile prediction machine learning model during machine learning training. PK profile prediction machine learning models can therefore generate PK profiles more accurately by avoiding errors and inconsistencies arising from the underlying assumptions of mechanical PK models.

[0043] Secondly, the multi-stage architecture encodes effective induction bias by first predicting the PK parameters and then using the predicted PK parameters to predict the complete PK profile. More specifically, the PK parameters are highly correlated with the PK profile, and the complexity of the prediction task performed by each stage of the machine learning model is simplified by having the machine learning model perform an intermediate step of predicting the PK profile. In particular, performing an intermediate step of predicting the PK parameters can reduce the amount of training data required for training, and the complexity of the architecture of the individual machine learning models implemented in the first and second stages of the multi-stage model can be reduced compared to what would otherwise be required. It will be understood that by reducing the amount of training data and the complexity of the machine learning model architecture, the system's consumption of computational resources, such as memory and computing power, can be further reduced.

[0044] Thirdly, the multi-stage architecture allows for the implementation of feature selection and pruning between the first and second stages of the model. In particular, the system can implement a feature selection algorithm to select a subset of the features that best predict the PK profile (for example, from a complete set of possible features including chemical structure features and PK parameter features). The PK profile prediction machine learning model can then be configured to process only the selected subset of features, which can reduce the possibility of overfitting the PK profile prediction machine learning model and improve its robustness and prediction accuracy. Furthermore, by processing only the subset of features selected through the feature selection algorithm, the amount of training data required to train the PK prediction machine learning model can be reduced, thereby reducing the complexity of the PK prediction machine learning model architecture and thus reducing the consumption of computational resources.

[0045] Details of one or more embodiments of the subject matter of this specification are shown in the accompanying drawings and the following description. Other features, aspects and advantages of the subject matter will become apparent from the description, drawings and claims. [Brief explanation of the drawing]

[0046] [Figure 1] This is a system diagram of an exemplary pharmacokinetic (PK) profile prediction system that can process chemical structures and generate PK profiles using a multi-stage machine learning model. [Figure 2] This is a block diagram of an exemplary preprocessing protocol for a multi-stage machine learning model. [Figure 3] An exemplary multi-stage machine learning model is shown, including PK parameter prediction and PK profile prediction stages. [Figure 4] Figure 3 shows an overview of an example of a boosting model that can be implemented by the first and second stage models of the multi-stage machine learning model. [Figure 5] Figure 3 shows the accuracy of the PK parameters predicted by the illustrative first-stage model. [Figure 6]Two conventional PK methods: the effectiveness of a multi-stage model in capturing nonlinearities in a single-segment model and a non-segment analysis is demonstrated. [Figure 7] Figure 6 shows the overall accuracy of the multi-stage model compared to two conventional pharmacokinetic prediction methods. [Figure 8] This is an illustrative flowchart of a process that uses a multi-stage machine learning model to generate a predicted PK profile. [Modes for carrying out the invention]

[0047] Similar reference numbers and designations in various drawings refer to the same elements.

[0048] Figure 1 shows an exemplary pharmacokinetic (PK) profile prediction system 100. The pharmacokinetic profile prediction system 100 is an example of a system implemented as a computer program on one or more computers in one or more locations, in which the systems, elements, and techniques described below are implemented.

[0049] The PK profile prediction system 100 can use an end-to-end machine learning approach to process data characterizing the chemical structure 110 of a compound and generate a PK profile 150 of that compound. The compounds analyzed by this system may be, for example, organic or inorganic compounds identified as targets in the identification and validation phases of drug development. The chemical structure 110 of a compound can characterize the spatial arrangement or graphical representation of atoms in the compound, such as a structural formula, ball-and-stick model, perspective drawing, or space-filling model. In some cases, data defining the three-dimensional (3D) arrangement of atoms in a compound can be processed as the chemical structure 110.

[0050] This system can process a chemical structure 110 to generate one or more molecular representation features 120 of a compound. For example, the molecular representation features 120 may include descriptors of structure-based features related to the compound. For example, the molecular representation features 120 may include a quantitative drug-likeness estimate (QED), topological polar surface area (TPSA), the number of hydrogen bond donors and acceptors, and the molecular weight of the compound. As another example, the molecular representation features 120 may include the number of rotatable bonds in the compound, the number of rings in the compound, the molar volume of the compound, and the dipole moment of the compound.

[0051] As another example, molecular representation features 120 can be based on natural language, such as strings. In particular, the chemical structure of a compound can be flattened into a single string using a Simple Molecular Input Line Input System (SMILES) representation. More specifically, a SMILES string can be represented as a molecular fingerprint, for example, a fixed-length binary representation (a bit vector of binary digits) that encodes the presence or absence of a particular structural pattern or substructure within the compound. An exemplary SMILES representation is described in more detail in Figure 2. As yet another example, molecular representation features 120 can be based on a graph. In particular, the graph may include each node for each atom and each bond for each edge.

[0052] A molecular representation feature 120, which includes one or more descriptors, natural language, or graph-based features, can then be processed as input to a multi-stage machine learning model 140 to generate the PK profile 150 described below.

[0053] In some cases, the molecular representations 120 can be further processed using a feature selection process 130 to identify one or more of the most relevant molecular representation features 120 as input to a multi-stage model 140, while also mitigating overfitting. In particular, the system 100 can execute a feature selection algorithm to select a subset of molecular representation features 120 from, for example, a set of candidate molecular representation features 120. For example, as part of the feature selection process 130, the system 100 can evaluate different combinations of molecular representation features 120 based on one or more descriptors, natural language, and graphs to identify the optimal force for the model 140. An exemplary feature selection process 130 is described in more detail in Figure 2.

[0054] In the specific example illustrated, the first stage of the multi-stage model 140 includes one or more PK parameter prediction models 142 that process 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 an input which may include one or more features selected from the predicted PK parameter predictions to generate a predicted PK profile 150. The PK parameter prediction models 142 may include one or more models configured to process the molecular representation features 120 separately to generate each PK parameter prediction that can serve as input to the PK profile prediction model 144.

[0055] In particular, PK parameters can characterize the absorption, distribution, metabolism, and elimination (ADME) properties of a compound. For example, predicted PK parameters may include the clearance of the compound, e.g., the rate at which the drug is eliminated or excreted from the subject's body; the volume of the compound's distribution; the half-life of the compound, e.g., the time required for the compound to decay to half its value; the maximum concentration of the compound, e.g., the maximum serum concentration that the compound reaches in a specified area of ​​the subject's body or the time to reach the maximum concentration of the compound, e.g., the time it takes for the drug to reach its maximum concentration after administration and absorption. As another example, predicted PK parameters may include the intrinsic clearance of the compound, e.g., the rate at which the liver can metabolize the drug with unrestricted access; the lipophilicity of the compound, e.g., the ability of the compound to mix with oily phases (as opposed to water); the solubility of the compound, e.g., the degree to which the substance dissolves in a solvent to form a solution; the unbound fraction of the compound, e.g., the proportion that exerts pharmacological effects; and the acidic and basic pKa of the compound, e.g., the acid dissociation descriptor and the basic dissociation descriptor, respectively.

[0056] System 100 can predict each of the PK parameters using each PK parameter prediction model 142. In some examples, the PK parameters in the model input of the PK profile prediction model 144 may be selected by a PK scientist to determine the PK parameters most relevant to a particular compound 110. In particular, a certain number of PK parameters, for example 5, 10, or 20 PK parameters, can be determined, and each PK parameter prediction model 142 can be trained for each of the PK parameters.

[0057] In some examples, system 100 can use a feature selection process 150 to determine the model input to the PK profile prediction model 144. More specifically, system 100 can use the feature selection process 150 to implement a feature selection algorithm that, for example, selects one or more PK parameter features that best predict the PK profile from a set of candidate features, and optionally selects other features. As an example, the other features may include features related to the administration of compound 110 to a subject, such as administration time and infusion duration, e.g., the time over which the drug is administered. An example of a feature selection process 150 is described in more detail in Figure 2.

[0058] In particular, system 100 can use feature selection processes 130 and 150 to select different features for different types of subjects. For example, since the PK profile 150 of the chemical structure 110 may differ for each different species, system 100 can select different sets of features as input to the PK parameter prediction model 142 and the PK profile prediction model 144 for different species (e.g., rats, dogs, pigs, humans, etc.). If the subject is human, system 100 can select different sets of features for different categories of people based on, for example, demographic criteria (e.g., age, sex, etc.), medical history (e.g., medical diagnosis), etc.

[0059] In some cases, system 100 can maintain the selected features as labels in a database indexed according to the type of object, for example. In this case, given a new object, system 100 can classify the object into a corresponding category and then use the selected feature labels for that category to generate a predicted PK profile 160 for the object. For example, system 100 can identify (i) molecular representation features 120, (ii) PK parameters, and (ii) any additional features corresponding to feature labels for a particular type of object, and use each PK parameter prediction model 142 to generate PK parameters that can be processed using the PK profile prediction model 144 to generate a PK profile 150 for the object.

[0060] In a specific example, an exemplary multi-stage machine learning (ML) model 140, which includes a clearance prediction model, such as a model predicting the rate and distribution of drug elimination or excretion from the body, and a PK parameter prediction model 142, such as the concentration of a compound over time, is described in more detail in Figure 3.

[0061] The PK profile prediction model 144 may be configured to process additional features along with one or more PK parameter model 142 outputs as model inputs to generate a PK profile 150. In some examples, additional features characterizing the compound structure, such as molecular representation 120 features or one or more features of the associated PK experimental data, may also be processed by the PK profile prediction model 144.

[0062] More specifically, by processing the predicted PK parameters, the PK profile prediction model 144 can generate a PK profile 150 without mechanical assumptions, such as those inherent in one-compartment or non-compartmental analysis models. In particular, these conventional methods can ignore the effects of mechanical processing, such as protein transport and renal filtration, which were not part of the initial mechanical assumptions. In some cases, the two-stage method allows the multi-stage machine learning model 140 to capture nonlinear trends that were not possible with other conventional methods based on mechanical assumptions, as further described in Figure 6. Furthermore, by predicting parameters from molecular representation 120 features, it becomes unnecessary to collect more data, for example, through additional animal studies.

[0063] Each of the one or more PK parameter prediction machine learning models and PK profile prediction machine learning models may have any suitable machine learning model architecture. For example, each may be implemented as, for example, a neural network, decision tree, random forest, support vector machine, or a combination thereof. The PK parameter prediction machine learning models and PK profile prediction machine learning models may be implemented using different machine learning model architectures or the same machine learning model architecture. In particular, the architecture selected for each model may depend on the amount of training data available. For example, if there is little training data available, a model with fewer training parameters may be selected, and vice versa.

[0064] In some cases, the system can implement Model 142, Model 144, or both as generalized linear models (based on simple or Poisson distributions), random forests, supported vector regressions, or boosting models, such as regression models like light gradient boosting machines or XG boosting. An exemplary implementation of a multistage ML model 140 using exemplary boosting algorithms for both one or more PK parameter prediction models 142 and PK profile prediction models 144 is described in more detail with reference to Figure 3.

[0065] In other cases, models 142 and 144 may have any suitable model architecture that enables the model to perform its described function, for example, in the case of PK parameter prediction model 142, to process one or more molecular representation features 120 to predict one or more PK parameters, and in the case of 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, models 142 and 144 may each include any suitable number of neural network layers (e.g., 1, 5, or 10 layers) of any suitable type (e.g., fully connected layers, attention layers, convolutional layers, etc.) connected in any suitable configuration (e.g., a linear sequence of layers or a directed graph of layers). In some cases, models 142 and 144 may have different architectures. Furthermore, one or more PK parameter prediction models 142 may have the same or different architectures.

[0066] The generated PK profile 150 defines the concentration-versus-time profile of the compound without extensive animal experiments or allometric scaling calculations. In particular, the area under the PK profile curve (AUC) can be calculated by using the PK profile 150 to quantify systemic exposure to the compound and define the maximum concentration of the compound and the compound's half-life within the subject, for example by performing integration, i.e., discrete summation.

[0067] In some examples, the data defining the predicted PK profile 150 of a compound may be stored in memory, transmitted over a data communication network, or both. Furthermore, in some examples, the data defining the predicted PK profile 150 may be displayed as a representation on the user device's display.

[0068] As an example, the PK profiles generated for several compounds can be used to rank the compounds, ordering them in ascending or descending order based on one or more ADME properties, such as area under the curve (AUC). The rank ranking can then be used to select compounds for physical synthesis. For example, a compound may be considered suitable for synthesis and further drug development according to a suitable AUC range. PK profiles can also be used to experimentally verify the properties of physically synthesized compounds, for example, by comparing them to the actual PK profile of a physically synthesized compound when the generated PK profile 150 is applied to the subject. In some examples, the physically synthesized compound can be administered to the subject. More specifically, if the PK profile of a compound defines a safe maximum concentration and half-life of the compound according to a predetermined threshold, the drug may be considered suitable for administration to one or more subjects.

[0069] Figure 2 is a block diagram illustrating an exemplary pretreatment protocol for a pharmacokinetic profile prediction system, such as the PK profile prediction system 100 in Figure 1. Specifically, Figure 2 shows an exemplary pretreatment protocol performed for the chemical structure input of the compound acetylsalicylic acid (aspirin).

[0070] In this case, the input chemical structure 110 of aspirin can be processed and converted into a natural language representation. More specifically, the system can process the input structure 110 to generate a SMILES representation 210 that flattens the chemical structure 110 into a single string.

[0071] The SMILES representation can be further processed to generate a molecular representation 120. In the particular example shown, the molecular representation 120 may include both descriptor-based features 220 and molecular fingerprints 230. More specifically, the descriptor-based features of aspirin may have a topological polar surface area (TPSA) of, for example, 63.6, a molecular weight of, for example, 180.4, a number of rotatable bonds in aspirin of, for example, 3, a number of rings in aspirin of, for example, 1, and a number of hydrogen donors or acceptors of, for example, 1 and 3, respectively.

[0072] Fingerprint 230 is a fixed-length binary representation that encodes the presence or absence of a specific structural pattern or substructure within a compound.

[0073] This system can perform feature selection 130 to identify the most relevant features 220 and 230 for a multi-stage machine learning model, such as the multi-stage machine learning model 140 in Figure 1. In particular, this system can perform feature selection 130 to identify the selected features 240 as input to the PK parameter prediction model 142, and this system can perform feature selection 150 to identify the most relevant features, including PK parameters generated using the PK parameter prediction model 142, as input to the PK profile prediction model 144.

[0074] In some examples, feature selection processes 130, 150, or both may specify a target number of selected features or a target distillation amount, for example, less than 50%, less than 10%, less than 5%, or less than 1% of the set of candidate features to be included in the model input.

[0075] In some examples, the system can perform forward or backward variable selection as feature selection 130, 150, or both, and the feature is a measurement of fit, e.g., R 2 Features are added or removed accordingly. In other examples, this system can remove collinear features by implementing variance-based methods, such as the variance inflation coefficient.

[0076] In some cases, the system can perform feature selection 150 before feature selection 130. In particular, feature selection processing 150 can be performed, for example, on PK and physicochemical property (PC) experimental results from data and molecular property data collected through previous experiments, as well as descriptors 220 and fingerprints 230, to generate 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 configure the PK parameter prediction model 142 and notify the system of feature selection 130.

[0077] More specifically, as part of the feature selection process 130, the system can evaluate different combinations of features based on descriptors 220 and fingerprints 230, natural language, and graphs to identify the optimal input to the PK parameter prediction model 142 that provides selected features 260 to be input to the PK profile prediction model 144.

[0078] In the specific example illustrated, the feature selection process 150 generates aspirin clearance (CL) and volume of distribution (Vdss) as selected features 260 for the PK profile prediction model 144. Once selected, these features 260 can become prediction targets within the first stage of a multi-stage machine learning model, so that the PK parameter prediction model 142 can generate the clearance and volume of distribution predictions, which are then processed as model inputs by the PK profile prediction model 144.

[0079] In some cases, the selected features 260 may include one or more additional inputs to the PK profile prediction model 144. For example, the selected features 260 may also include the number of TPSAs and hydrogen donors. In this case, the PK profile prediction model 144 can generate a PK profile by processing, for example, the predicted clearance and distribution volume generated by the PK parameter prediction model 142 together with the number of TPSAs and hydrogen donors.

[0080] For example, the system can perform feature selection 150 on a set of candidate features including the volume of distribution, in vivo clearance, pKa (acidic), pKa (basic), unbound fraction, lipophilicity, molecular weight, administration time, and infusion duration of the chemical structure 110. In some cases, the system 100 can select different sets of features for different subjects, such as rats, dogs, pigs, and humans. As an example, the system 100 can select the volume of distribution, in vivo clearance, and administration time as selected features 260 for rat subjects. As another example, the system 100 can select the volume of distribution, in vivo clearance, acidic pKa, basic pKa, unbound fraction, lipophilicity, infusion duration, and administration duration as selected features 260 for human subjects.

[0081] The system can then perform feature selection 130 on a set of candidate features including one or more descriptors 220 and fingerprint features to identify selected features 240 that can be processed to generate the best performance input to the PK parameter prediction model 142, for example, PK parameters selected as part of features 260.

[0082] Figure 3 shows an exemplary multi-stage machine learning model in which execution is indicated by the selected clearance and the volume of distributional features selected through feature selection in Figure 2. In particular, the multi-stage machine learning model may include a first stage in which one or more first machine learning models convert chemical structure data into predicted PK parameters, and a second stage in which one or more second machine learning models convert the PK parameters into a complete predicted PK profile. For example, the multi-stage machine learning model 300 can be used to predict the PK profile of a rat subject.

[0083] The multi-stage prediction model 300 may include two stages, for example, a first stage 305 and a second stage 340. More specifically, the first stage 305 may include one or more PK parameter prediction models 142 that process molecular representation features to generate predicted PK parameters 330, and the second stage 340 may include a PK profile prediction model 144 that processes the predicted PK parameters 330 to generate a PK profile 150.

[0084] In the specific example illustrated, the predicted PK parameters 330 include clearance (CL) predictions 332 and volume-of-distribution (Vdss) predictions 334 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 the molecular fingerprint 310 to generate the CL prediction 330, and the Vdss model 325 processes the molecular descriptor 320 to generate the Vdss prediction 335.

[0085] In some cases, additional feature selection processing may be performed, for example, additional processing on the feature selections 130 in Figures 1 and 2, to determine the model input to each PK parameter prediction model 142 and to determine which different molecular representation features yield higher fidelity predictions for the training data set. In the particular example illustrated, this additional feature selection processing results in a CL model 325 that processes the molecular fingerprint 310 to generate a CL prediction 332 and a Vdss model 325 that processes the molecular descriptor 320 to generate a Vdss prediction 334.

[0086] In the second stage 340, the PK profile prediction model 144 processes the predicted PK parameters 330, such as the CL prediction 332 and the Vdss prediction 334, to generate a PK profile 150. For example, the PK profile prediction model 144 can execute the same algorithm as the PK parameter prediction model 142. In particular, as will be described in more detail in Figure 4, both the first stage 305 and the second stage 330 of the multistage ML model 300 can be implemented using a boosting model.

[0087] Figure 4 shows an exemplary boosting model. For example, one or more of the PK parameter prediction model 142, the PK profile prediction model 144, or both of them in Figure 1 may be implemented using the boosting models shown. In the specific examples described below, the PK profile prediction model may be implemented as a boosting model 400 that includes one or more base machine learning models, e.g., model 1410, model 2420, and model 4440.

[0088] More specifically, one or more base machine learning models 410, 420, and 440 can process the input by combining their respective outputs to form an aggregated output 450, thereby generating a predicted PK profile of the compound. In the particular example illustrated, the predicted PK profile is generated by summing the individual outputs of one or more base machine learning models. In another example, the aggregated output 450 can be generated using the average of all base model outputs.

[0089] In particular, each subsequent basic machine learning model after the first basic machine learning model may be associated with ordering within the sequence and configured to produce an output that corrects the error between the target PK profile and the PK profile generated as an aggregated output 450 by combining the outputs of the preceding basic machine learning models in the sequence through the training process. In the specific example illustrated, each of the basic machine learning models may be implemented as a decision tree model. The first decision tree model may be stationary model 0 (not shown), and a series of trees are added to obtain model 1410 having output 415.

[0090] More specifically, the output of model 1410 may have a specific residual 410 with respect to the target PK profile, for example, regarding how the decision rules implemented in the decision tree are performed on the data. The residual of model 1410 can then be predicted using model 2 to produce output 425, which can then be predicted using models 3 (not shown) and later, so that the residual of model k-1 can be input into model 4440 to produce output 445.

[0091] The training process for the boosting model 400 may include training a set of PK profile prediction machine learning model parameters using one or more training examples, each including a target PK profile of a compound and a PK profile generated by combining the output of a preceding base machine learning model in the sequence. In particular, for each training example, the PK profile prediction machine learning model may be trained to reduce the discrepancy, e.g., the measure of distance, between the target PK profile of the training example and the 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 the discrepancy between the target PK profile of a given training example and the predicted PK profile generated by the aggregated decision tree of the PK profile prediction model.

[0092] More specifically, Boosting Model 400 demonstrates how the model can be implemented as a decision tree ensemble configured to produce an output that corrects the 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 outputs of preceding decision tree models, throughout training. In the specific example illustrated, the system can minimize prediction errors, for example, by minimizing the loss of the aggregate tree model, by adding successive tree models during training to predict residuals of previous errors that can be optimized using gradient descent.

[0093] In the specific example illustrated, the gradient descent algorithm minimizes prediction error by adding successive tree models to minimize the aggregate boosting model 400 loss. In the specific example implementing XGBoost as the boosting model 400, the addition of new decision rule tree models to a sequence of existing base tree models can be evaluated by comparing the effectiveness of the additional tree models using a scoring criterion for partitions represented by the decision rules implemented in the new decision trees against the same criteria as the new decision trees. In the case of XGBoost, the pruning technique uses these scores to retain new high-scoring decision rule trees in the ensemble and reject new low-scoring decision rule trees from the ensemble.

[0094] Figure 5 shows the accuracy of the PK parameter prediction model 142 for the first stage of a multistage machine learning model in predicting PK parameters such as clearance (CL) and volume of distribution (Vdss), for example, one or more models 300 in Figure 3. In particular, graphs 500 and 550 are scatter plots showing the fidelity of the predicted distributions of CL and VdSS. For example, the results in graphs 500 and 550 are obtained from the PK parameter prediction model 142 being implemented using XGBoost and trained as described in Figure 4.

[0095] In the specific example illustrated, the ground truth 510 experimental data is represented by an identity reference line corresponding to the perfect overlap between the predicted (measured) experimental data and the observed experimental data. In this case, the CL model, e.g., the CL model 315 in Figure 3, can replicate the ground truth 510 clearance values ​​with high accuracy and precision, as demonstrated by its proximity to the identity reference line and the overlap between the ground truth reference line and the training 520 and test set 530 distribution predictions. Similarly, the Vdss model, e.g., the Vdss model 325 in Figure 3, can replicate the ground truth 560 with high accuracy and precision, as demonstrated by its proximity to the identity reference line between the ground truth reference line and the training 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 a mean percentage error (MAPE) of less than 100% for both in vivo CL and Vdss.

[0096] Figure 6 demonstrates how a multi-stage machine learning model can capture the nonlinearity in the generated PK profile that other pharmacokinetic prediction methods cannot account for under their mechanical assumptions. In particular, Figure 6 compares the PK profile prediction system 100 from Figure 1 with a standard single-compartment model 612 and a non-compartmental analysis model 614 for four compounds, namely compounds A, B, C, and D. The results for compounds A, B, C, and D are shown in graphs 610, 620, 630, and 640, respectively.

[0097] PK scientists and researchers can use a single-compartment model to predict the efficacy and toxicity of compounds across different body compartments, such as blood flow, tissues, and organs. The single-compartment model assumes mechanical constraints on the function of defined body compartments, such as the destination and elimination of drugs, which are often unknown in the very early stages of drug development. The results of the single-compartment model 612 for four compounds, namely compounds A, B, C, and D, are shown in graphs 610, 620, 630, and 640.

[0098] PK scientists and researchers can also use non-compartmental analysis to predict the efficacy and toxicity of compounds. In this case, non-compartmental analysis assumes that the compound is uniformly distributed throughout the body and that the elimination of the compound is defined as a steady rate that does not take physiological factors into account. This modeling scheme allows for the rapid generation of PK profile estimates but does not take into account the physiological mechanisms or biological processes that drive or modify PK. In the specific example illustrated, the exemplary non-compartmental analysis model is a PBPK (physiology-based pharmacokinetics) model that incorporates organ blood flow and tissue composition. The results of the non-compartmental analysis model 614 for each of four compounds, namely compounds A, B, C, and D, are shown in graphs 610, 620, 630, and 640.

[0099] As demonstrated by graphs 610, 620, 630, and 640 for the four compounds, for all compounds, the ML method 616 can capture the nonlinear trends of the ground truth PK data represented by dots better than either the single-section 612 or the PBPK model 614.

[0100] In particular, even if the single-compartment model 612 can reasonably well capture the maximum concentration, it cannot capture the lower concentration time points of multiple compounds exhibiting dual exponential behavior. Specifically, the single-compartment model 612 yields a PK profile curve exhibiting single-exponential decay and lacks the second and higher-order dynamics, including the curvature of the target PK profile. On the other hand, the PBPK model 614 can capture the dual-exponential decay of the PK profile because it considers the distribution of drugs to different tissues, but it cannot capture the maximum concentration compared to the single-compartment model. It can be observed that the predictions generated by the multi-stage machine learning model 612 proposed herein can capture both single-exponential and dual-exponential PK profiles 610, 620, 630, and 640.

[0101] Figure 7 shows the overall accuracy of the multi-stage machine learning model compared to single-partition and non-partition analysis methods. In particular, Panel 700 shows the accuracy of the multi-stage machine learning model in the model shown in Figure 3, predicting the PK profile for other conventional methods, for example, as shown in scatter plots 710, 720, and 730. In addition, Table 750 shows different performance metrics for the three methods.

[0102] The scatter plot of panel 700 includes the PK profile results for the multi-stage machine learning methods, non-compartmental analysis model 720 and one-compartmental model 730, as described herein, with respect to the identity reference line, for example, the one-compartmental and non-compartmental analysis model in Figure 6. Graph 710 shows how the ML results follow the identity reference line compared to the one-compartmental model 730, which deviates from the identity line particularly at low concentrations.

[0103] Table 750 demonstrates the performance metrics of the model, including RMSE, MAPE, MAE, and R2. While the MAPE and MAE of the pharmacokinetic profile prediction system described herein are lower than those of conventional systems, the R2 value of the pharmacokinetic profile prediction system described herein may be higher than that of conventional systems.

[0104] Figure 8 is a flowchart illustrating an exemplary process for generating a predicted PK profile using a multi-stage machine learning model. For convenience, the process 800 is described as being performed by one or more computer systems located in one or more locations. For example, a training system, e.g., the pharmacokinetic (PK) profile prediction system 100 of Figure 1, appropriately programmed according to this specification, can perform the process 800.

[0105] In particular, this system can accept a compound (step 810) and determine one or more PK features of the compound (step 820). For example, PK parameters may include molecular fingerprints and descriptors and PK and physicochemical (PC) features, such as the clearance of the compound, the volume of distribution of the compound, the half-life of the compound, the maximum concentration of the compound, or the time to reach the maximum concentration of the compound. In some cases, these features can be further aggregated into smaller sets of features using feature selection.

[0106] The system can then process features that characterize the chemical structure of a compound using one or more PK parameter prediction models (step 830). In particular, these features can be processed using separate PK models to generate different data that define the PK parameters (step 840).

[0107] One or more predicted PK parameters may then be processed using a PK profile prediction model (step 850). In some cases, the PK profile prediction model may additionally process one or more of the features identified in step 820. For example, the model input to the PK profile model may be selected through a feature selection process that identifies the inputs most relevant to the PK profile prediction model. More specifically, the PK profile prediction model may process the model input, which includes one or more predicted PK parameter outputs from each PK parameter prediction model, to generate a predicted PK profile of the compound (step 860).

[0108] This specification uses the term “configured” in relation to systems and computer program elements. A system of one or more computers configured to perform a particular operation or action means that software, firmware hardware, or a combination thereof is installed on the system that causes the system to perform the operation or action while in operation. One or more computer programs being configured to perform a particular operation or action means that one or more programs, when executed by a data processing device, contain instructions that cause the device to perform the operation or action.

[0109] The subject matter and functional embodiments described herein may be implemented in digital electronic circuits, tangibly implemented computer software or firmware, computer hardware including structures disclosed herein and their structural equivalents, or one or more combinations thereof. Embodiments of the subject matter described herein may be implemented as one or more modules of computer program instructions encoded in a tangible non-temporary storage medium, for example, executed by a data processing device or controlling the operation of a data processing device. The computer storage medium may be a machine-readable storage device, a machine-readable storage board, a random or serial access memory device, or one or more combinations thereof. Alternatively or additionally, the program instructions may be encoded in artificially generated propagating signals, for example, mechanically generated electrical, optical or electromagnetic signals generated to encode information for transmission to a suitable receiving device, executed by a data processing device.

[0110] The term "data processing device" refers to data processing hardware and encompasses all kinds of devices, machines, and equipment for data processing, including, for example, a programmable processor, a computer, or multiple processors or computers. A device may be, or further include, a dedicated logic circuit, such as an FPGA (Field-Programmable Gate Array) or ASIC (Application-Specific Integrated Circuit). Optionally, in addition to hardware, a device may include code that generates an execution environment for computer programs, such as processor firmware, a protocol stack, a database management system, an operating system, or one or more combinations thereof.

[0111] Computer programs may also be called or described as programs, software, software applications, apps, modules, software modules, scripts, or code, and may be written in any form of programming language, including compiler or interpreter languages ​​or declarative or procedural languages, and may be deployed in any form, including standalone programs or modules, elements, subroutines, or other units suitable for use in a computer environment. A program may, but may not, correspond to a file in a file system. A program may be stored in part of a file that holds other programs or data, such as one or more scripts stored in a markup language document, a single file or multiple collaborating files dedicated to the program in question, such as a file that stores one or more modules, subprograms, or parts of code. Computer programs may be deployed to run on a single computer or on multiple computers located in one place or distributed across multiple locations and interconnected by a data communication network.

[0112] In this specification, the term “engine” is used broadly to refer to a software-based system, subsystem, or process programmed to perform one or more specific functions. Generally, an engine is implemented as one or more software modules or elements installed on one or more computers in one or more locations. In some cases, one or more computers are dedicated to a particular engine, while in other cases, multiple engines may be installed and operate on the same computer or on multiple computers.

[0113] The processes and logic flows described herein may be executed by one or more programmable computers that execute one or more computer programs to perform functions by acting on input data and producing outputs. The processes and logic flows may be executed by dedicated logic circuits, such as FPGAs or ASICs, or by a combination of dedicated logic circuits and one or more programmed computers.

[0114] A computer suitable for running computer programs may be based on a general-purpose or dedicated microprocessor, or both, or any other type of central processing unit. Generally, the central processing unit receives instructions and data from read-only memory, random-access memory, or both. Essential elements of a computer are a central processing unit that executes or carries out instructions, and one or more memory devices that store instructions and data. The central processing unit and memory may be reinforced by or incorporated into dedicated logic circuits. Generally, a computer may include one or more mass storage devices for storing data, such as magnetic disks, magneto-optical disks, or optical disks, or may be operablely coupled to receive data from or transfer data to them, or both. However, a computer does not have to have such devices. Furthermore, a computer may be embedded in another device, for example, a mobile phone, 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, such as a Universal Serial Bus (USB) flash drive.

[0115] Computer-readable media suitable for storing computer program instructions and data include, for example, semiconductor memory devices such as EPROM, EEPROM, and flash memory devices; magnetic disks such as internal hard disks or removable disks; magneto-optical disks; and all forms of non-volatile memory, media, and memory devices, including CD-ROMs and DVD-ROMs.

[0116] To enable user interaction, embodiments of the subject matter described herein may be implemented in a computer having a display device for displaying information to the user, such as a CRT (cathode ray tube) or LCD (liquid crystal display) monitor, and a pointing device to which the user can provide input to the computer, such as a keyboard and a mouse or trackball. Other types of devices may also be used to enable user interaction; for example, the feedback provided to the user may be any form of sensory feedback, such as visual feedback, auditory feedback, or tactile feedback, and input from the user may be received in any form, including acoustic, voice, or tactile input. In addition, the computer may interact with the user by sending and receiving documents to and from the device used by the user, for example, by sending a web page to a web browser on the user's device in response to a request received from a web browser. The computer may also interact with the user by sending text messages or other forms of messages to a personal device, such as a smartphone running a messaging application, and then receiving response messages from the user.

[0117] Data processing devices that implement machine learning models may include, for example, dedicated hardware accelerator units for processing the common and computationally intensive parts of machine learning training or production, such as inference, and workloads.

[0118] Machine learning models can be implemented and deployed using machine learning frameworks, such as the TensorFlow framework or the Jax framework.

[0119] Embodiments of the subject matter described herein may be implemented in a computing system that includes, for example, a backend element as a data server, or a middleware element, such as an application server, or a frontend element, such as a client computer having a graphical user interface, a web browser, or an application that allows a user to interact with the implementation of the subject matter described herein, or any combination of one or more such backend, middleware, or frontend elements. The elements of the system may be interconnected by any form or medium of digital data communication, such as a communication network. Examples of communication networks include local area networks (LANs) and wide area networks (WANs), such as the Internet.

[0120] A computing system may include a client and a server. The client and server are generally remote from each other and typically interact via a communication network. The relationship between the client and the server arises from computer programs running on each computer that have a client-server relationship with each other. In some embodiments, the server sends data, such as an HTML page, to a user device for the purpose of displaying data to a user interacting with a device acting as a client and receiving user input from the user. Data generated on the user device, such as the results of user interaction, may be received by the server from the device.

[0121] This specification includes many details of specific implementations, but these should not be interpreted as limiting the scope of any invention or claim, but rather as describing features specific to particular embodiments of a particular invention. Certain features described herein in relation to separate embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in relation to a single embodiment may be implemented separately or in any suitable partial combination in multiple embodiments. Furthermore, while features may be described above as operating in a particular combination and initially claimed as such, one or more features of the claimed combination may be excluded from that combination in some cases, and the claimed combination may relate to a partial combination or a variation of a partial combination.

[0122] Similarly, although the operations are illustrated and described in the claims in a specific order, it should not be understood that such operations must be performed in a specific illustrated or sequential order, or that all illustrated operations must be performed, in order to obtain the desired results. In certain situations, multitasking and parallel processing may be advantageous. Furthermore, the separation of various system modules and elements in the embodiments described above should not be understood as requiring such separation in all embodiments, and it should be understood that the program elements and systems described above can generally be integrated into a single software product or packaged into multiple software products.

[0123] Specific embodiments of the subject matter have been described. Other embodiments are also included in the scope of the following claims. For example, the actions described in the claims may still yield the desired results even if they are performed in a different order. As an example, the processes shown in the accompanying drawings do not necessarily have to be performed in the specific order or sequential order shown to obtain the desired results. In some cases, multitasking and parallel processing may be advantageous.

Claims

1. A method performed by one or more computers, Using a multi-stage machine learning model, the data characterizing the chemical structure of a compound is processed to generate data defining the predicted pharmacokinetic (PK) profile of the compound, Using one or more PK parameter predictive machine learning models, process the model input, which includes the data characterizing the chemical structure of the compound, to generate one or more predicted PK parameters of the compound. Using a PK profile prediction machine learning model, process the model input, which includes the one or more predicted PK parameters generated by the one or more PK parameter prediction machine learning models, to generate the predicted PK profile of the compound. Including generating A method that includes this.

2. The method according to claim 1, wherein the one or more predicted PK parameters include one or more of the clearance of the compound, the distribution volume of the compound, the half-life of the compound, the maximum concentration of the compound, or the time to reach the maximum concentration of the compound.

3. The method according to claim 1 or 2, wherein the predicted PK profile of the compound defines the time-dependent concentration of the compound in the subject after the compound has been administered to the subject.

4. The method according to any one of claims 1 to 3, wherein the model input to the PK profile prediction machine learning model further includes features that characterize the chemical structure of the compound.

5. The model input to the PK profile predictive machine learning model is: Identifying a set of candidate features to be included in the model input to the PK profile predictive machine learning model, the set of candidate features including (i) features that define the compound PK parameters, and (ii) features that characterize the compound structure. Selecting an appropriate subset of the set of candidate features to include in the model input to the PK profile predictive machine learning model. The method according to any one of claims 1 to 4, comprising a plurality of features selected by performing an operation including the operation.

6. The method according to claim 5, wherein the features characterizing the compound structure include molecular descriptor features and molecular fingerprint features.

7. The method according to claim 5 or 6, wherein selecting an appropriate subset of the set of candidate features to be included in the model input to the PK profile predictive machine learning model includes selecting less than 50%, less than 10%, less than 5%, or less than 1% of the set of candidate features to be included in the model input to the PK profile predictive machine learning model.

8. The method according to claim 6 or 7, wherein selecting an appropriate subset of the set of candidate features to include in the model input to the PK profile predictive machine learning model comprises applying a feature selection algorithm to the set of candidate features to select an appropriate subset of the set of candidate features that is predicted to optimize the performance of the PK profile predictive machine learning model.

9. Using the PK profile prediction machine learning model, processing the model input, which includes one or more predicted PK parameters, to generate the predicted PK profile of the compound is: Processing the model input, which includes one or more predicted PK parameters, using each of the multiple basic machine learning models included in the PK profile prediction machine learning model, The predicted PK profile of the compound is generated by combining the outputs of each of the multiple basic machine learning models included in the PK profile prediction machine learning model. The method according to any one of claims 1 to 8, including

10. The method according to claim 9, wherein generating the predicted PK profile of the compound by combining the respective outputs of each of the plurality of basic machine learning models included in the PK profile prediction machine learning model includes summing the respective outputs of each of the plurality of basic machine learning models included in the PK profile prediction machine learning model.

11. The plurality of basic machine learning models included in the PK profile predictive machine learning model are associated with ordering into a sequence, and The method according to claim 9 or 10, wherein each subsequent basic machine learning model in the sequence is configured, through training, to produce an output that corrects the error between (i) a target PK profile and (ii) a PK profile generated by combining the outputs of preceding basic machine learning models in the sequence.

12. The method according to any one of claims 1 to 11, wherein the PK profile prediction machine learning model has a set of PK profile prediction machine learning model parameters trained with a set of training examples by a machine learning training technique.

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

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

15. The method according to any one of claims 1 to 14, wherein generating one or more predicted PK parameters of the compound comprises using each of a plurality of PK parameter predictive machine learning models to process the model input, which includes the data characterizing the chemical structure of the compound, and generating each of the predicted PK parameters as the output of each of the plurality of PK parameter predictive machine learning models.

16. The method according to claim 15, wherein each of the plurality of PK parameter prediction machine learning models is configured to generate its respective predicted PK parameters through training by machine learning training techniques.

17. The method according to any one of claims 1 to 16, wherein the data characterizing the chemical structure of the compound includes one or more molecular descriptors of the compound.

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

19. The method according to any one of claims 1 to 18, wherein the data characterizing the chemical structure of the compound includes the molecular fingerprint of the compound.

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

21. The method according to any one of claims 1 to 20, wherein the data characterizing the chemical structure of the compound is derived from a representation of the chemical structure of the compound as a string.

22. The method according to claim 21, wherein the chemical structure of the compound as a string includes a Simple Molecular Input Line Input System (SMILES) string.

23. The method according to any one of claims 1 to 22, further comprising processing the predicted PK profile to determine the area below the predicted PK profile.

24. The method according to any one of claims 1 to 23, further comprising selecting the compound for physical synthesis based at least partially on the predicted PK profile of the compound.

25. The method according to claim 24, further comprising physically synthesizing the compound.

26. The method according to claim 25, further comprising administering the physically synthesized compound to the target.

27. The method according to claim 25, further comprising experimentally verifying one or more properties of the physically synthesized compound.

28. To provide the data that defines the predicted PK profile of the compound, The data defining the predicted PK profile of the compound is stored in memory. Transmitting the data defining the predicted PK profile of the compound via a data communication network, or Displaying the predicted PK profile representation of the compound on the user device's display. The method according to any one of claims 1 to 27, further comprising providing one or more of the following.

29. One or more computers, One or more storage devices communicably connected to one or more computers and A system comprising, wherein one or more storage devices, when executed by the one or more computers, store instructions causing the one or more computers to perform the operation of each of the methods described in any one of claims 1 to 24 or 28.

30. One or more non-temporary computer storage media that, when executed by one or more computers, store instructions causing one or more computers to perform the operation of each of the methods described in any one of claims 1 to 24 or 28.

Images