Synthon insertion for DNA-encoded library modeling

Abstract

Embodiments of the present disclosure relate to modeling DEL data using factorized molecular representations (e.g., hierarchical single and double synthon building blocks), which leverages the hierarchical structure inherent to these molecules. Using factorized molecular representations, machine learning models are trained to learn the underlying binding affinity of compounds to targets and one or more covariates (e.g., loading / repeat noise). This results in machine learning models producing improved predictions in the form of higher enrichment scores that have good correlation with compound-target binding affinity.

Description

Cross-references to related applications

[0001] This application claims the benefit and priority of U.S. Provisional Patent Application No. 63 / 540,425, filed September 26, 2023, the entire disclosure of which is incorporated herein by reference in its entirety for all purposes. Background Technology

[0002] Optically encoded DNA libraries (DELs) have demonstrated their effectiveness as a powerful method for efficient exploration in a vast chemical space. These small molecule libraries are synthesized combinatorially by combining multiple building blocks with compatible chemical properties. DNA barcodes covalently linked to the molecules specify the unique combination of building blocks for each molecule. These small molecule libraries are then used for screening experiments against target proteins, where multiple rounds of washing and elution are performed before identifying viable library molecules. While proven to be an efficient process for large-scale exploration of the chemical space, these screening experiments are often highly noisy, requiring computational methods with proper inductive bias to extract useful signals for downstream applications such as hit discovery and lead compound optimization. Therefore, improved methods are needed to process the output of DEL experiments to build improved machine learning models for exploring the chemical space. Summary of the Invention

[0003] This paper discloses a method, non-transitory computer-readable medium, and system for modeling DEL data by decomposing molecular representations into their hierarchical monosynontic and disynontic building blocks, leveraging the inherent hierarchical structure of these molecules. The disclosed method explicitly decomposes molecular representations in a motivated manner. Specifically, it learns individual synthon representations from their respective decomposed representations, including corresponding disynons, trisynons, and other combinations of synthons. This avoids the necessity of enumerating the entire molecular structure, which is typically an error-prone and tedious process.

[0004] Furthermore, covariate factors are integrated into the modeling to more effectively account for data noise. For example, the model trained in this paper considers various experimental biases, including two main sources of noise inherent in DEL data: pre-screening and repetition level bias. Since DEL molecules are synthesized using a split-merge method, the relative abundance of each library member in the final mixture is uncertain. While the library itself is sequenced to obtain a rough estimate of the molecular distribution, this counting data is also susceptible to potential synthesis and sequencing biases. Different experimental or sequencing noises are also expected between different replicates. Our model attempts to mitigate the effects of these factors to better model the observed counting data and learn useful potential enrichments of DEL molecules. The published machine learning model demonstrates robust performance compared to the counting baseline, enriches the correct pharmacophores, and provides valuable insights through its inherent interpretable structure, thus offering a powerful tool for DEL data analysis.

[0005] In summary, the machine learning models disclosed herein can be used for a variety of applications, including virtual compound screening, performing hit selection and analysis, and identifying common binding motifs. Virtual compound screening enables the identification of compounds from a library (e.g., a virtual library) that may bind to a target (e.g., a protein target). Performing hit selection enables the identification of compounds that may exhibit the desired activity. For example, a hit could be a compound that binds to a target (e.g., a protein target) and thus exhibits the desired effect by binding to the target. Predicting the binding affinity between a compound and a target can lead to the identification of compounds exhibiting the desired binding affinity. For example, binding affinity values ​​can be continuous values, thus indicating different types of binders (e.g., strong or weak binders). This makes it possible to identify and classify compounds that exhibit different binding affinities to a target. Identifying common binding motifs is useful for understanding the mechanisms between target binders. Understanding binding motifs is useful for developing additional novel small molecule compounds (e.g., during medicinal chemistry activities). In various embodiments, the predicted binding affinity is correlated with the activity of the compound. For example, a compound with a higher predicted binding affinity may be correlated with higher activity. In various implementations, the predicted binding affinity may not be directly related to the activity of the compound. For example, in some cases, a compound with a higher predicted binding affinity may exhibit lower activity compared to a second compound with a lower predicted binding affinity.

[0006] This paper discloses a method for performing molecular screening of compound-target binding, the method comprising: obtaining a plurality of synthons forming the compound; converting the plurality of synthons into a plurality of synthon representations; combining the plurality of synthon representations into a molecular embedding; and analyzing the molecular embedding using a machine learning model to generate at least one target enrichment prediction representing a measure of binding between the compound and the target. In various embodiments, the method is further characterized by performing probabilistic modeling using at least the target enrichment prediction by applying a probability density function that models the experimental target count. In various embodiments, the probability density function is represented by any one of a Poisson distribution, a binomial distribution, a gamma distribution, a binomial-Poisson distribution, or a negative binomial distribution. In various embodiments, the Poisson distribution is a zero-inflated Poisson distribution.

[0007] In various embodiments, analyzing the molecular embedding using the machine learning model also generates covariate predictions. In various embodiments, the method does not include the step of enumerating the compound from the plurality of synthons. In various embodiments, converting the plurality of synthons into a plurality of synthon representations includes generating one or more single synthon representations from the plurality of synthons. In various embodiments, generating one or more single synthon representations from the plurality of synthons includes analyzing the plurality of synthons using a learned representation model, optionally wherein the learned representation model is a multilayer perceptron. In various embodiments, converting the plurality of synthons into a plurality of synthon representations further includes generating one or more bisynthesis representations from the one or more single synthon representations. In various embodiments, generating one or more bisynthesis representations from the one or more single synthon representations includes analyzing the one or more single synthon representations using a learned representation model, optionally wherein the learned representation model is a multilayer perceptron. In various embodiments, converting the plurality of synthons into a plurality of synthon representations further includes generating one or more trisynthesis representations from the one or more bisynthesis representations. In various embodiments, generating one or more trisynthesis representations from the one or more bisynthesis representations includes analyzing the one or more bisynthesis representations using a learned representation model, optionally wherein the learned representation model is a multilayer perceptron.

[0008] In various embodiments, the plurality of synthesizer representations includes one or more single synthesizer representations. In various embodiments, the plurality of synthesizer representations includes one or more dual synthesizer representations. In various embodiments, the plurality of synthesizer representations includes one or more triple synthesizer representations. In various embodiments, the plurality of synthesizer representations includes one or more quad synthesizer representations. In various embodiments, the plurality of synthesizer representations includes one or more single synthesizer representations, one or more dual synthesizer representations, and one or more triple synthesizer representations. In various embodiments, the plurality of synthesizer representations includes three single synthesizer representations, three dual synthesizer representations, and one triple synthesizer representation.

[0009] In various embodiments, the machine learning model includes a neural network. In various embodiments, the neural network includes a feedforward artificial neural network. In various embodiments, the neural network includes a multilayer perceptron (MLP). In various embodiments, the machine learning model includes one or more parameters learned through supervised training techniques. In various embodiments, the method disclosed herein further includes determining a binding affinity value between the compound and the target using the target enrichment prediction. In various embodiments, the method disclosed herein further includes ranking the compound based on at least the target enrichment prediction.

[0010] In various embodiments, combining the multiple synthon representations into molecular embeddings includes implementing a multi-head attention mechanism across the multiple synthon representations. In various embodiments, implementing the multi-head attention mechanism includes using one or more learned attention weights of the multiple synthon representations. In various embodiments, the methods disclosed herein also include ranking the ability of the multiple synthons to bind to the target using the one or more learned attention weights. In various embodiments, the covariate prediction is derived from one or more covariates including either nonspecific binding or noise. In various embodiments, nonspecific binding includes one or more of binding to beads, binding to the matrix, streptavidin binding to beads, binding to biotin, binding to the gel, binding to the DEL container surface, or binding to a tag. In various embodiments, the noise includes one or more of loading bias, duplication bias, enrichment in other negative control panning, enrichment in other target panning, confounding, compound synthesis yield, reaction type, initial tag imbalance, initial loading population, experimental conditions, chemical reaction yield, byproducts and truncated products, library synthesis errors, DNA affinity for the target, sequencing depth, and sequencing noise.

[0011] In various embodiments, the covariate prediction is derived from loading noise. In various embodiments, the covariate prediction is derived from repetitive noise. In various embodiments, analyzing the molecular embedding using the machine learning model also generates a second covariate prediction. In various embodiments, the covariate prediction and the second covariate prediction are each independently selected from nonspecific binding or noise. In various embodiments, the covariate prediction is derived from loading noise, and the second covariate prediction is derived from repetitive noise. In various embodiments, converting the plurality of synthons into the representation of the plurality of synthons includes applying one or more trained representation models. In various embodiments, the machine learning model is trained using one or more training compounds with corresponding DNA-encoded library (DEL) outputs. In various embodiments, the corresponding DNA-encoded library (DEL) outputs of the training compounds include: experimental control counts determined by a first panning experiment; and experimental target counts determined by a second panning experiment. In various embodiments, for one of the training compounds, the machine learning model is trained by: generating target enrichment predictions and covariate predictions from molecular embeddings generated by combining multiple synthetic representations transformed from multiple synthetics forming the training compound; combining the target enrichment predictions and the covariate predictions to generate predicted target counts; and determining a loss value based on at least the predicted target counts and the experimental target counts, according to a loss function. In various embodiments, the machine learning model is trained based on the determined loss value. In various embodiments, the method disclosed herein further includes jointly training the machine learning model with one or more learned representation models based on the determined loss value. In various embodiments, the loss value is also determined based on the covariate predictions and the experimental control counts. In various embodiments, the loss function is any one of negative log-likelihood loss, binary cross-entropy loss, focal loss, arc loss, cosface loss, cosine-based loss, or a loss function based on a BEDROC metric.

[0012] In various implementations, combining the target enrichment prediction and the covariate prediction to generate the predicted target count includes applying a probability density function modeling the experimental target count. In various implementations, the probability density function is represented by any one of a Poisson distribution, binomial distribution, gamma distribution, binomial-Poisson distribution, or negative binomial distribution. In various implementations, the Poisson distribution is a zero-inflated Poisson distribution. In various implementations, the machine learning model is also trained to generate the predicted control count from the covariate prediction by applying the probability density function modeling the experimental control count. In various implementations, the probability density function modeling the experimental control count is represented by any one of a Poisson distribution, binomial distribution, gamma distribution, binomial-Poisson distribution, or negative binomial distribution. In various implementations, the Poisson distribution is a zero-inflated Poisson distribution. In various implementations, the binding metric is any one of binding affinity, DEL count, DEL reading, or DEL exponent.

[0013] In various embodiments, the molecular screening is a virtual molecular screening. In various embodiments, the compound is derived from a virtual library of compounds. In various embodiments, the target includes a protein target. In various embodiments, the protein target is a human carbonic anhydrase IX (CAIX) protein target, a horseradish peroxidase (HRP) protein target, a discoid domain receptor tyrosine kinase 1 (DDR1) protein target, or a mitogen-activated protein kinase 14 (MAPK14) protein target. In various embodiments, the methods disclosed herein further include: identifying common binding motifs in a subset of one or more compounds, wherein compounds in the subset have a predicted binding metric above a threshold binding value.

[0014] This document also discloses a method for generating molecular embeddings of a compound, the method comprising: obtaining a plurality of synthons forming the compound; converting the plurality of synthons into a plurality of synthon representations, wherein the conversion comprises: generating one or more single synthon representations by analyzing the plurality of synthons using a first learned representation model; generating one or more bisynthesis representations by analyzing the one or more single synthon representations using a second learned representation model; generating one or more trisynthesis representations by analyzing the one or more bisynthesis representations using a third learned representation model; and combining the plurality of synthon representations into a molecular embedding. In various embodiments, combining the plurality of synthon representations into a molecular embedding comprises implementing a multi-head attention mechanism across the plurality of synthon representations. In various embodiments, converting the plurality of synthons into a plurality of synthon representations further comprises generating one or more N-synthesis representations, wherein N is 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, or 20. In various embodiments, the first learned representation model comprises a multilayer perceptron. In various embodiments, the second learned representation model comprises a multilayer perceptron. In various embodiments, the representation model of the third learning includes a multilayer perceptron. In various embodiments, the plurality of synthesizer representations includes one or more single synthesizer representations, one or more dual synthesizer representations, one or more triple synthesizer representations, or one or more quad synthesizer representations. In various embodiments, the plurality of synthesizer representations includes one or more single synthesizer representations, one or more dual synthesizer representations, and one or more triple synthesizer representations. In various embodiments, the plurality of synthesizer representations includes three single synthesizer representations, three dual synthesizer representations, and one triple synthesizer representation.

[0015] This paper also discloses a method for predicting experimental counts of DNA-coding libraries (DELs), the method comprising: obtaining a molecular embedding of a compound generated from a plurality of synthetic representations of the compound; analyzing the molecular embedding using a machine learning model to generate (A) a target enrichment prediction representing a measure of binding between the compound and the target, and (B) one or more covariate predictions; and combining the target enrichment predictions and the one or more covariate predictions by applying a probability density function modeling the experimental target count to generate a predicted target count. In various embodiments, the probability density function is represented by any one of a Poisson distribution, a binomial distribution, a gamma distribution, a binomial-Poisson distribution, or a negative binomial distribution. In various embodiments, the Poisson distribution is a zero-expanded Poisson distribution. In various embodiments, the one or more covariate predictions are derived from one or more covariates including any one of nonspecific binding or noise. In various embodiments, nonspecific binding includes one or more of binding to beads, binding to a matrix, binding to streptavidin of beads, binding to biotin, binding to a gel, binding to the surface of a DEL container, or binding to a tag. In various embodiments, the noise includes one or more of the following: loading bias, duplication bias, enrichment in other negative control panning, enrichment in other target panning, confounding, compound synthesis yield, reaction type, initial tag imbalance, initial loading population, experimental conditions, chemical reaction yield, byproducts and truncated products, library synthesis errors, DNA affinity for the target, sequencing depth, and sequencing noise. In various embodiments, at least one of the one or more covariate predictions originates from loading noise. In various embodiments, at least one of the one or more covariate predictions originates from duplication noise. In various embodiments, a first covariate prediction originates from loading noise, and a second covariate prediction originates from duplication noise. In various embodiments, the one or more covariate predictions include two, three, four, five, six, seven, eight, nine, ten, eleven, twelve, thirteen, fourteen, fifteen, sixteen, seventeen, eighteen, nineteen, or twenty covariate predictions.

[0016] This paper also discloses a method for predicting experimental counts of DNA-encoded libraries (DELs), comprising: obtaining a target enrichment prediction representing a measure of binding between a compound and a target, and modeling the experimental target counts of the DEL by applying a probability density function, performing probabilistic modeling using at least the target enrichment prediction.

[0017] In various implementations, the probability modeling includes implementing any one of a Poisson distribution, a binomial distribution, a gamma distribution, a binomial-Poisson distribution, or a negative binomial distribution. In various implementations, the Poisson distribution is a zero-inflated Poisson distribution.

[0018] In various embodiments, the method further includes obtaining covariate predictions, and the covariate predictions are used in performing the probabilistic modeling. In various embodiments, the covariate predictions are derived from one or more covariates including either nonspecific binding or noise. In various embodiments, the nonspecific binding includes one or more of binding to beads, binding to the matrix, streptavidin binding to beads, binding to biotin, binding to the gel, binding to the DEL container surface, or binding to a tag. In various embodiments, the noise includes one or more of loading bias, duplication bias, enrichment in other negative control panning, enrichment in other target panning, confounding, compound synthesis yield, reaction type, initial tag imbalance, initial loading population, experimental conditions, chemical reaction yield, byproducts and truncated products, library synthesis errors, DNA affinity for the target, sequencing depth, and sequencing noise. In various embodiments, the covariate predictions are derived from loading noise. In various embodiments, the covariate predictions are derived from duplication noise.

[0019] In various embodiments, the target enrichment prediction is generated by a machine learning model trained using one or more training compounds with corresponding DNA-coding library (DEL) outputs. In various embodiments, the machine learning model includes a neural network. In various embodiments, the neural network includes a feedforward artificial neural network. In various embodiments, the neural network includes a multilayer perceptron (MLP). In various embodiments, the machine learning model includes one or more parameters learned through supervised training techniques.

[0020] In various embodiments, the machine learning model generates the target enrichment prediction by analyzing molecular embeddings to generate the target enrichment prediction, which at least represents a measure of binding between the compound and the target. In various embodiments, the corresponding DNA-coding library (DEL) output of the training compound includes: experimental control counts determined by a first panning experiment and experimental target counts determined by a second panning experiment. In various embodiments, where for one of the training compounds, the machine learning model is trained by generating target enrichment predictions and covariate predictions from molecular embeddings generated by combining multiple synthon representations transformed from multiple synthons forming the training compound, combining the target enrichment predictions and the covariate predictions to generate predicted target counts, and determining a loss value based on at least the predicted target counts and the experimental target counts according to a loss function. In various embodiments, the machine learning model is trained based on the determined loss value.

[0021] In various implementations, the machine learning model further includes jointly training the machine learning model with one or more learned representation models based on the determined loss value. In various implementations, the loss value is also determined based on the covariate prediction and the experimental control count. In various implementations, the loss function is any one of negative log-likelihood loss, binary cross-entropy loss, focal loss, arc loss, cosface loss, cosine-based loss, or a loss function based on the BEDROC metric. In various implementations, combining the target enrichment prediction and the covariate prediction to generate the predicted target count includes applying a probability density function modeling the experimental target count. In various implementations, the probability density function is represented by any one of a Poisson distribution, binomial distribution, gamma distribution, binomial-Poisson distribution, or negative binomial distribution. In various implementations, the Poisson distribution is a zero-inflated Poisson distribution.

[0022] In various implementations, the machine learning model is further trained by predicting the predicted control counts from the covariate by applying a probability density function that models the experimental control counts. In various implementations, the probability density function that models the experimental control counts is represented by any one of a Poisson distribution, a binomial distribution, a gamma distribution, a binomial-Poisson distribution, or a negative binomial distribution. In various implementations, the Poisson distribution is a zero-expanded Poisson distribution. In various implementations, the binding metric is any one of binding affinity, DEL count, DEL reading, or DEL exponent. In various implementations, the target includes a protein target.

[0023] In various implementations, the protein target is a human carbonic anhydrase IX (CAIX) protein target, a mitogen-activated protein kinase 14 (MAPK14) protein target, a discoid domain receptor tyrosine kinase 1 (DDR1) protein target, or a horseradish peroxidase (HRP) protein target.

[0024] This document further discloses a non-transitory computer-readable medium including instructions that, when executed by a processor, cause the processor to perform any of the methods disclosed herein. This document also discloses a system comprising: a processor; and a non-transitory computer-readable medium including instructions that, when executed by the processor, cause the processor to perform any of the methods disclosed herein. Attached Figure Description

[0025] These and other features, aspects, and advantages of the invention will be better understood from the following description and accompanying drawings. It should be noted that, where feasible, similar or identical reference numerals may be used in the drawings and may indicate similar or identical functions. For example, letters following a reference numeral, such as "DEL Experiment 115A," indicate that the text specifically refers to the element having that particular reference numeral. Reference numerals without subsequent letters in the text, such as "DEL Experiment 115," refer to any or all elements in the drawings bearing that reference numeral (e.g., "DEL Experiment 115" in the text refers to "DEL Experiment 115A" and / or "DEL Experiment 115B" in the drawings). As another example, "composite component 310" refers to any or all elements of "composite component 310A," "composite component 310B," and "composite component 310C."

[0026] Figure 1A An example system environment involving a synthetic submodeling system is described, based on an implementation scheme.

[0027] Figure 1B An example DNA-encoded library (DEL) panning experiment is described according to one implementation scheme.

[0028] Figure 2 A block diagram of the synthetic subsystem modeling system is depicted, based on one implementation scheme.

[0029] Figure 3A A flowchart for analyzing the decomposition of synthons to generate molecular embeddings is depicted, according to one implementation scheme.

[0030] Figure 3B A flowchart illustrating the implementation of a machine learning model for predicting DEL experiment counts is depicted, according to one implementation scheme.

[0031] Figure 4A An example flow for generating target enrichment predictions is described, based on an implementation.

[0032] Figure 4B An example flow for generating molecular representations is described, according to one implementation.

[0033] Figure 4C An example flow for predicting DEL experimental counts is described, based on an implementation scheme.

[0034] Figure 5A An example flowchart for training a machine learning model is depicted, based on one implementation scheme.

[0035] Figure 5B A sample flowchart for training a machine learning model is further described, according to one implementation scheme.

[0036] Figure 6 An example flow for training a machine learning model is described, based on an implementation scheme.

[0037] Figure 7A The following is shown for implementation Figure 1A-1B Example computing devices for the systems and methods described in 2, 3A-3B, 4A-4C, 5A-5B and 6.

[0038] Figure 7B The overall system environment for implementing the synthetic submodeling system is described, according to an implementation scheme.

[0039] Figure 7C It is used for implementation Figure 7B An example depiction of a distributed computing system environment.

[0040] Figure 8A A general paradigm for modeling DEL molecules using their combinatorial properties is described.

[0041] Figure 8B A schematic diagram depicting the machine learning architecture and data flow is provided.

[0042] Figure 9A The known pharmacophores of carbonic anhydrase IX (CA-IX) are shown, arranged in order of binding affinity.

[0043] Figure 9B The known pharmacophores of horseradish peroxidase (HRP) are shown, arranged in order of binding affinity.

[0044] Figure 10A-10D The predicted average enrichment of molecules under control and target conditions is shown by grouping synthons by the B or C position of CA-IX and HRP.

[0045] Figure 11A-11D The performance comparison shows the difference between models using decomposed molecular representations and models using whole molecular representations. Each model was trained using a different percentage of data retention.

[0046] Figure 12A and 12B It has been demonstrated that the zero probability of prediction is a good measure of the prediction noise of CA-IX and HRP.

[0047] Figure 13A and 13B The model attention distributions for the CA-IX and HRP datasets are depicted respectively.

[0048] Figure 14 Describing kinase inhibitors The distribution of chemical properties in the DNA-encoded library (KinDEL) dataset. These selected properties are typically used to assess the drug-likeness of molecules. The light blue areas conform to the Lipinski and Veber drug-likeness rules. QED: Quantitative estimate of drug-likeness; PSA: Polar surface area; HBA: Hydrogen bond acceptor; HBD: Hydrogen bond donor.

[0049] Figure 15 A 3D cube visualization of the KinDEL dataset is presented, where each axis corresponds to a different cycle in DEL. The points in the figure are the most enriched compounds (using a Poisson enrichment measure). Linear patterns can be interpreted as enriched bisynthetics, i.e., combinations of two synthons that typically bind to a protein target. Detailed Implementation

[0050] definition Unless otherwise stated, the terms used in the claims and specification are defined as follows.

[0051] The phrase "to obtain multiple synthons that form a compound" includes generating multiple synthons of a compound or obtaining multiple synthons of the compound, for example, from a third party that generates multiple synthons of the compound.

[0052] The phrase "syntheticon" refers to a molecular building block of a compound. In various embodiments, a syntheticon refers to a starting reagent in the synthesis of the compound. A compound may consist of multiple syntheticons. In various embodiments, a compound consists of two syntheticons. In various embodiments, a compound consists of three syntheticons. In various embodiments, a compound consists of four syntheticons. In various embodiments, a compound consists of five, six, seven, eight, nine, ten, eleven, twelve, thirteen, fourteen, fifteen, sixteen, seventeen, eighteen, nineteen, or twenty syntheticons. In various embodiments, a compound consists of more than twenty syntheticons.

[0053] The phrase "synthesizer representation" refers to the representation of a synthon, such as transforming a synthon into a representation space. First, synthons can be represented in a specific structural format, including any of the following formats: Simplified Molecular Input Line Input System (SMILES) strings, MDL Molfile (MDL MOL), Structural Data File (SDF), Protein Database (PDB), Molecular Specification File (xyz), International Union of Pure and Applied Chemistry (IUPAC) International Chemical Identifier (InChI), and Tripos Mol2 file (mol2). In various embodiments, synthons can be represented as codes, such as fingerprints or graphs of synthons. The representation of a synthon can be a transformation of a synthon in a specific structural format. In various embodiments, the representation of a synthon can be continuous or discrete. An example synthon representation can be an embedding of a synthon, which is a numerical representation of the synthon. In various embodiments, the embedding of a synthon is generated using one of a neural network, a graph neural network, a transformer, or a multilayer perceptron.

[0054] The phrase "target enrichment prediction" refers to an informative prediction of the binding metric between a compound and a target, learned by a machine learning model. In various embodiments, the target enrichment prediction is a value or score. Typically, the target enrichment prediction is informative (e.g., relevant) of the binding metric between the compound and the target and is a denoised prediction to account for covariate predictions (e.g., unaffected by covariates and other noise sources). In various embodiments, the target enrichment prediction is learned by attempting to predict experimental DEL counts, which include counts from noise sources and covariates.

[0055] The phrase "covariate prediction" refers to a prediction derived from a covariate learned by a machine learning model. In various implementations, the covariate prediction is a value or score. Example covariates may include noise sources (e.g., noise sources in a DEL experiment) and nonspecific bindings (e.g., binding to beads, binding to the matrix, streptavidin binding to beads, binding to biotin, binding to the gel, binding to the DEL container surface, binding to a tag, such as a DNA tag or protein tag). Example noise sources include biases (e.g., pre-screening count bias or duplication bias), enrichment in other negative control pannings, enrichment in other target pannings, confounding, compound synthesis yield, reaction type, initial tag imbalance, initial loading population, experimental conditions, chemical reaction yield, byproducts and truncated products, library synthesis errors, DNA affinity for the target, sequencing depth, and sequencing noise such as PCR bias. In a particular implementation, the covariate prediction is a prediction of nonspecific bindings (e.g., binding to the matrix). In a particular implementation, the covariate prediction is a prediction of pre-screening count bias. In a particular implementation, the covariate prediction is a prediction of loading bias.

[0056] The phrases "pre-screening count bias" and "loading bias" are used interchangeably and generally refer to the bias in the true signal caused by differences in the starting population during a DEL panning experiment. For example, some molecules may be present in different quantities compared to others (e.g., some molecules may be present in quantities 10-1000+ times higher than others). Differences in the starting population can lead to loading bias because the molecular series may have survived only by chance.

[0057] The phrase "repetition bias" refers to a deviation in the true signal caused by sequencing or experimental problems in technical experiments. Example problems may include poor protein constructs and / or inaccurate volume transfer within wells / repetitions. In various implementations, repetition bias may also be caused by different sequencing depths between repetitions. In summary, these problems can lead to significant signal attenuation that is not a true reflection of the DEL experiment.

[0058] The phrase "MAPK14" refers to mitogen-activated protein kinase 14. The phrase "DDR1" refers to discoid domain receptor tyrosine kinase 1. The phrase "CAIX" refers to carbonic anhydrase IX. The phrase "HRP" refers to horseradish peroxidase.

[0059] It should be noted that, as used in the specification and appended claims, the singular forms “a,” “an,” and “the” include plural referents unless the context clearly specifies otherwise.

[0060] System Environment Overview Figure 1A An example system environment involving the synthetic submodeling system 130 is described, according to one implementation. Specifically, Figure 1A DNA-encoded library (DEL) experiments 115A and 115B are described to generate DEL outputs (e.g., DEL outputs 120A and 120B) for use by a synthon modeling system 130 to train and deploy machine learning models. In certain embodiments, the machine learning model is useful for generating target enrichment predictions that can be correlated with a binding metric between a compound and a target, for example, for performing virtual compound screening or for selecting and analyzing hits.

[0061] DEL experiments may involve incubating a mixed collection of DNA-barcode-tagged compounds with an immobilized protein target in a process called panning. The mixture is then washed to remove non-binding compounds and elute the remaining binding compounds. In various embodiments, the remaining binding compounds may undergo one or more additional rounds of incubation, washing, and elution. For example, the remaining binding compounds may undergo two, three, four, five, six, seven, eight, nine, or ten additional rounds of incubation, washing, and elution. The remaining binding compounds are amplified and sequenced to identify the putative bindings. DEL provides quantitative readings for numerous (e.g., up to billions) compounds.

[0062] like Figure 1A As shown, two DEL experiments 115A and 115B can be performed. However, in various embodiments, fewer or additional DEL experiments can be performed. In various embodiments, Figure 1A The different DEL experiments 115A and 115B shown may refer to different repetitions of similar / identical experimental conditions. In various implementations, the example system environment involves at least three DEL experiments, at least four DEL experiments, at least five DEL experiments, at least six DEL experiments, at least seven DEL experiments, at least eight DEL experiments, at least nine DEL experiments, at least ten DEL experiments, at least fifteen DEL experiments, at least twenty DEL experiments, at least thirty DEL experiments, at least forty DEL experiments, at least fifty DEL experiments, at least sixty DEL experiments, at least seventy DEL experiments, at least eighty DEL experiments, at least ninety DEL experiments, or at least one hundred DEL experiments. The output of one or more DEL experiments (e.g., DEL outputs) can be provided to the synthetic sub-modeling system 130 to train and deploy machine learning models.

[0063] In various embodiments, DEL experiments involve screening small molecule compounds for a DEL library against a target. In some embodiments, DEL experiments involve screening multiple DEL libraries (e.g., in a single pool or across multiple pools). Typically, DEL experiments (e.g., DEL experiments 115A or 115B) involve constructing small molecule compounds using chemical building blocks (also called synthons). In various embodiments, small molecule compounds can be generated using two chemical building blocks, these are called disyntons. In various embodiments, small molecule compounds can be generated using three chemical building blocks, these are called trisyntons. In various embodiments, small molecule compounds can be generated using four or more, five or more, six or more, seven or more, eight or more, nine or more, ten or more, fifteen or more, twenty or more, thirty or more, forty or more, or fifty or more chemical building blocks. In various embodiments, the DNA-coding library (DEL) used for the DEL experiment can include at least 10 3 Each is a unique small molecule compound. In various embodiments, the DNA-coding library (DEL) used for DEL experiments may include at least 10 4 Each is a unique small molecule compound. In various embodiments, the DNA-coding library (DEL) used for DEL experiments may include at least 10 5 Each is a unique small molecule compound. In various embodiments, the DNA-coding library (DEL) used for DEL experiments may include at least 10 6 Each is a unique small molecule compound. In various embodiments, the DNA-coding library (DEL) used for DEL experiments may include at least 10 7 Each is a unique small molecule compound. In various embodiments, the DNA-coding library (DEL) used for DEL experiments may include at least 10 8 Each is a unique small molecule compound. In various embodiments, the DNA-coding library (DEL) used for DEL experiments may include at least 10 9 Each is a unique small molecule compound. In various embodiments, the DNA-coding library (DEL) used for DEL experiments may include at least 10 10 Each is a unique small molecule compound. In various embodiments, the DNA-coding library (DEL) used for DEL experiments may include at least 10 11 Each is a unique small molecule compound. In various embodiments, the DNA-coding library (DEL) used for DEL experiments may include at least 10 12 A unique small molecule compound.

[0064] Typically, the small molecule compounds in a DEL are composed of individual chemical building blocks, also referred to herein as synthons. In various embodiments, synthons can be individually tagged. In various embodiments, synthons can be individually tagged via adapters. Therefore, a small molecule compound can be tagged with multiple tags corresponding to the synthons constituting that small molecule compound. In various embodiments, the small molecule compound can be covalently linked to a unique tag. In various embodiments, the tag comprises a nucleic acid sequence. In various embodiments, the tag comprises a DNA nucleic acid sequence.

[0065] In various embodiments, for DEL experiments (e.g., DEL experiment 115A or 115B), a tagged small molecule compound is incubated with an immobilized target. In various embodiments, the target is a nucleic acid target, such as a DNA target or an RNA target. In various embodiments, the target is a protein target. In a particular embodiment, the protein target is immobilized on beads. The mixture is washed to remove small molecule compounds that are not bound to the target. Small molecule compounds bound to the target are eluted and may undergo one or more additional rounds of incubation, washing, and elution. The corresponding tag sequence of the remaining compound is amplified. In various embodiments, the tag sequence is amplified by one or more rounds of polymerase chain reaction (PCR) amplification. In various embodiments, the tag sequence is amplified using an isothermal amplification method, such as loop-mediated isothermal amplification (LAMP). The amplified sequence is sequenced to determine a quantitative reading of the putative number of small molecule compounds bound to the target. More detailed information on methods for constructing DNA-encoded libraries using small molecule compounds and for identifying putative binders of DEL targets is described in McCloskey et al.'s "Machine Learning on DNA-Encoded Libraries: A New Paradigm for Hit Finding." Med. Chem. 2020, 63, 16, 8857–8866, and Lim, K. et al., "Machine learning on DNA-encoded library count data using an uncertainty-aware probabilistic loss function." arXiv: 2108.12471, are each incorporated herein by reference in their entirety.

[0066] refer to Figure 1B The paper describes an example DNA-encoded library (DEL) panning experiment, according to one implementation scheme. DNA-encoded libraries (DELs) can be constructed by sequentially assembling molecular building blocks (i.e., synthons) into molecules labeled with unique DNA barcode identifiers. These are... Figure 1BThe library is displayed as a "linked small molecule" with a DNA barcode. Once synthesized, the library undergoes affinity testing against the target (e.g., the target protein target) through a series of selection experiments. For example, as... Figure 1B As shown, the target can be a protein fixed on the bead.

[0067] The experiment, also referred to in this paper as panning, involves assembling DEL molecules into an immobilized target solution (e.g., Figure 1B Step 1 is shown. Figure 1B Step 2, as shown, involves multiple rounds of washing of the resulting mixture. Due to washing, unbound and weakly bound molecules are removed. This procedure leaves behind DEL members that remain bound (e.g., bound to the target or other elements such as the matrix). Step 3 involves eluting the bound DEL molecules. The eluted DEL molecules are then amplified in step 4. It is worth noting that some DEL molecules may remain bound to the matrix (in...). Figure 1B These matrix binders (displayed as "matrix binders") are therefore not washed away during step 2. These matrix binders may represent covariates and / or noise and are not actually binders to the target. In contrast, the DEL molecules that actually bind to the target (in...) Figure 1B The protein complex (shown as "protein conjugate") was also obtained.

[0068] In step 5, the presence of DEL molecules is subsequently identified using next-generation DNA sequencing. The bioinformatics-processed results can include DNA readings and the corresponding molecules. Therefore, the relative abundance of identified DEL members (e.g., the number of DEL counts) is theoretically a plausible proxy for their binding affinity.

[0069] In various implementations, for DEL experiments (e.g., DEL experiment 115A or 115B), small molecule compounds are screened against a target using a solid medium containing the target. Here, the target is integrated into the solid medium, unlike panning systems that use targets immobilized on beads. For example, the screening may involve running a DEL using electrophoresis on a solid medium containing the target, such as a gel. The gel is then sliced ​​to obtain tags for labeling the small molecule compounds. The presence of the tags indicates that the small molecule compound is a putative binder of the target integrated into the gel. The tags are amplified (e.g., by PCR or isothermal amplification processes such as LAMP) and then sequenced. Further details of gel electrophoresis methods for identifying putative binders are described in International Patent Application No. PCT / US2020 / 022662, entitled "Methods and Systems for Processing or Analyzing Oligonucleotide Encoded Molecules," filed March 13, 2020, which is incorporated herein by reference in its entirety.

[0070] In various implementations, one or more DNA-encoded library experiments 115 are performed to model one or more covariates (e.g., off-target covariates or covariate predictions). Typically, covariates refer to experimental effects that influence the DEL output of a DEL experiment (e.g., DEL counts) and thus serve as confounding factors when determining the actual binding between the small molecule compound and the target. Example covariates may include noise sources (e.g., noise sources in the DEL experiment) and nonspecific bindings (e.g., binding to beads, binding to the matrix, streptavidin binding to beads, binding to biotin, binding to the gel, binding to the DEL container surface, binding to tags, such as DNA tags or protein tags). Example noise sources include biases (e.g., pre-screening count bias or replication bias), enrichment in other negative control pannings, enrichment in other target pannings, confounding, compound synthesis yield, reaction type, initial tag imbalance, initial loading population, experimental conditions, chemical reaction yields, byproducts and truncated products, library synthesis errors, DNA affinity for the target, sequencing depth, and sequencing noise such as PCR bias. In one specific implementation, the covariate is the pre-screening count bias. In another specific implementation, the covariate is the loading bias. In yet another specific implementation, the first covariate is the pre-screening count bias, and the second covariate is the loading bias. Therefore, different DEL experiments can be performed to model the pre-screening count bias and the loading bias.

[0071] For example, DEL experiment 115 can be designed to model the covariate of small molecule compound binding to beads. Here, if the small molecule compound binds to the beads instead of, or in addition to, an immobilized target on the beads, subsequent washing and elution steps may result in the small molecule compound being detected and identified as a putative binder, even if the small molecule compound does not specifically bind to the target. Therefore, DEL experiment 115 for modeling the covariate of nonspecific binding to beads can involve incubating the small molecule compound with the beads in the absence of an immobilized target on the beads. The mixture of small molecule compound and beads is washed to remove unbound compounds that are not bound to the beads. The small molecule compound bound to the beads is eluted, and the corresponding tag sequence is amplified (e.g., by PCR or isothermal amplification such as LAMP amplification). The amplified sequence is sequenced to determine a quantitative reading of the number of small molecule compounds bound to the beads. Thus, this quantitative reading can be a DEL output (e.g., DEL output 120) from the DEL experiment (e.g., DEL experiment 115), which is then provided to the synthon modeling system 130.

[0072] As another example, DEL experiment 115 can be designed to model the covariate of small molecule compound binding to streptavidin adapters on beads. Here, the streptavidin adapters on the beads are used to attach a target (e.g., a target protein) to the beads. If the small molecule compound binds to the streptavidin adapter instead of, or in addition to, the immobilized target on the beads, subsequent washing and elution steps may result in the small molecule compound being detected and identified as a putative binder, even if the small molecule compound does not specifically bind to the target. Therefore, DEL experiment 115 for modeling the covariate of nonspecific binding to beads can involve incubating the small molecule compound with the streptavidin adapters on the beads in the absence of an immobilized target on the beads. The mixture of the small molecule compound and the streptavidin adapters on the beads is washed to remove the non-binding compound. The small molecule compound bound to the streptavidin adapters on the beads is eluted, and the corresponding tag sequence is amplified (e.g., by PCR or isothermal amplification such as LAMP amplification). The amplified sequence is sequenced to determine a quantitative reading of the number of small molecule compounds that bind to the streptavidin linker on the beads. Thus, this quantitative reading can be a DEL output (e.g., DEL output 120) from a DEL experiment (e.g., DEL experiment 115) and then provided to the synthon modeling system 130.

[0073] As another example, DEL experiment 115 can be designed to model the covariate of small molecule compound binding to the gel, which occurs when implementing the nDexer method. Here, if the small molecule compound binds to the gel during electrophoresis and not, or only to, an immobilized target on beads, the subsequent washing and elution steps may result in the small molecule compound being detected and identified as a putative binder, even if the small molecule compound is not bound to the target. Therefore, DEL experiment 115 can involve incubating the small molecule compound with a control gel that does not contain the target. The small molecule compound bound or immobilized in the gel is eluted, and the corresponding tag sequence is amplified (e.g., by PCR or isothermal amplification such as LAMP). The amplified sequence is sequenced to determine a quantitative reading of the amount of small molecule compound bound or immobilized in the gel. Thus, this quantitative reading can be a DEL output (e.g., DEL output 120) from the DEL experiment (e.g., DEL experiment 115), which is then provided to the synthon modeling system 130.

[0074] In various embodiments, at least two DEL experiments 115 are performed to model one covariate. For example, a first DEL experiment is performed against a target, while a second DEL experiment is performed to model a covariate. In various embodiments, at least two DEL experiments 115 are performed to model at least two covariates. In various embodiments, at least three DEL experiments 115 are performed to model at least three covariates. In various embodiments, at least four DEL experiments 115 are performed to model at least four covariates. In various embodiments, at least five DEL experiments 115 are performed to model at least five covariates. In various embodiments, at least six DEL experiments 115 are performed to model at least six covariates. In various embodiments, at least seven DEL experiments 115 are performed to model at least seven covariates. In various embodiments, at least eight DEL experiments 115 are performed to model at least eight covariates. In various embodiments, at least nine DEL experiments 115 are performed to model at least nine covariates. In various embodiments, at least ten DEL experiments 115 are performed to model at least ten covariates. The DEL output from each DEL experiment can be provided to the synthesizer modeling system 130. In various embodiments, the DEL experiment 115 for modeling covariates can be performed multiple times. For example, technical repetitions of the DEL experiment 115 for modeling covariates can be performed. In a particular embodiment, at least three repetitions of the DEL experiment 115 for modeling covariates can be performed.

[0075] The DEL outputs from each DEL experiment (e.g., DEL output 120A and / or DEL output 120B) may include DEL readings of the small molecule compounds from the DEL experiment. In various embodiments, the DEL outputs may be DEL counts of the small molecule compounds from the DEL experiment. Therefore, small molecule compounds that are putative binders of the target will have higher DEL counts compared to small molecule compounds that are not putative binders of the target. As an example, the DEL count may be a unique molecular index (UMI) count determined by sequencing. As an example, the DEL count may be the number of counts observed in a specific index of a solid medium (e.g., a gel). In various embodiments, the DEL outputs may be DEL readings corresponding to small molecule compounds. For example, a DEL reading may be a sequence reading derived from a tag labeling the corresponding small molecule compound. In various embodiments, the DEL outputs may be DEL indexes. For example, a DEL index may refer to a slice number of a solid medium (e.g., a gel) that indicates how far a DEL member has moved within the solid medium.

[0076] Typically, the synthon modeling system 130 generates molecular embeddings from multiple synthons transformed from decomposed synthons and further trains and / or deploys machine learning models. These machine learning models are trained to learn the potential binding affinity of compounds to targets and one or more covariates (e.g., loading / repetition bias). This results in improved predictions from the machine learning models in the form of higher enrichment scores, which correlate well with compound-target binding affinity. Therefore, these machine learning models trained and / or deployed by the synthon modeling system 130 are useful for predicting expected target binding in virtual compound screening activities.

[0077] Figure 2 A block diagram of a synthetic submodeling system 130 is depicted according to one implementation. Figure 2 The various components of the synthetic submodeling system 130 are introduced, including examples such as the synthetic subrepresentation module 140, the model training module 150, the model deployment module 155, the DEL output analysis module 160, and the DEL data storage 170.

[0078] Referring to synthon representation module 140, which generates representations of synthons (e.g., derived from synthons of a compound or from training synthons of a training compound). In various embodiments, synthon representation module 140 generates representations of synthons by obtaining multiple decomposition synthons of a compound. Here, synthons of a compound can be represented as codes, such as fingerprints, graphs of synthons, or 3-D point clouds. Example fingerprints of synthons can be represented as Morgan fingerprints or subunits of Morgan fingerprints. Other example codes of synthons can be represented in specific structures, such as Simplified Molecular Input Line Input System (SMILES) strings, MDL Molfiles (MDL MOL), Structural Data Files (SDF), Protein Databases (PDB), Molecular Specification Files (xyz), International Union of Pure and Applied Chemistry (IUPAC) International Chemical Identifiers (InChI), and Tripos Mol2 file (mol2) formats. In various embodiments, synthon representation module 140 generates multiple synthon representations by transforming multiple synthons. In various implementations, the synthesizer representation module 140 applies one or more machine learning models (referred to herein as learned representation models) to transform multiple synthesizers into multiple synthesizer representations. In various implementations, the one or more learned representation models are neural networks, such as multilayer perceptrons (MLPs). Further details of the methods performed by the synthesizer representation module 140 are described herein.

[0079] A reference model training module 150 is provided, which trains a machine learning model using a training dataset. Typically, the model training module 150 trains the machine learning model to effectively denoise DEL experimental data to generate target enrichment predictions representing the binding between the compound and the target. In a particular embodiment, the model training module 150 trains the machine learning model to effectively denoise the DEL experimental data, taking into account one or both of loading bias and duplication bias to improve the target enrichment predictions. Therefore, the method disclosed herein relates to training a machine learning model to generate target enrichment predictions that are better correlated with binding measurements compared to previous work. Further details of the training process performed by the model training module 150 are described herein.

[0080] Reference model deployment module 155 deploys machine learning models to generate target enrichment predictions representing the binding between compounds and targets. Target enrichment predictions are useful for various applications, such as performing virtual compound screening, selecting and analyzing hits, and identifying common binding motifs on targets (e.g., protein targets). Further details of the process performed by model deployment module 155 are described herein.

[0081] Referring to the DEL output analysis module 160, which analyzes the output of one or more trained machine learning models. In various embodiments, the DEL output analysis module 160 converts the predictions from the machine learning model outputs into values ​​representing a binding metric between a compound and a target. As a specific example, the DEL output analysis module 160 can convert target enrichment predictions from the machine learning model outputs into binding affinity values. In various embodiments, the DEL output analysis module 160 ranks compounds based on at least their target enrichment predictions or based on the binding metric. In various embodiments, the DEL output analysis module 160 identifies candidate compounds that are likely target conjugates based on the target enrichment predictions from the machine learning model outputs. For example, candidate compounds may be compounds ranked highly based on their target enrichment predictions or based on their binding metric. Thus, candidate compounds can be synthesized, for example, as part of medicinal chemistry activities, and experimentally screened against the target to validate their binding and efficacy. In various embodiments, the DEL output analysis module 160 identifies common binding motifs in conjugates that may contribute to effective binding between the conjugate and the target. This enables the identification of valuable binding motifs that can be further integrated into the design of additional compounds to achieve the desired activity. More details about the process executed by the DEL output analysis module 160 are described in this article.

[0082] Example methods for generating target enrichment predictions As described herein, methods for generating target enrichment predictions involve training and / or deploying machine learning models to analyze molecular embeddings derived from decomposition synthons. The machine learning models are further trained to denoise the target enrichment predictions by considering the influence of one or more covariates. Thus, the machine learning models are able to generate target enrichment predictions that are better correlated with experimental binding affinity measurements. In various implementations, experimental binding affinity measurements encompass any known method for measuring the binding affinity of a compound to a biological target (e.g., DNA, RNA, and / or protein). Example experimental methods include, but are not limited to, fluorescence polarization, plasmon resonance / surface plasmon resonance (SPR), enzyme-linked immunosorbent assay (ELISA), isothermal titration calorimetry (ITC), radioligand binding assays, fluorescence resonance energy transfer (FRET) assays, and / or equilibrium dialysis.

[0083] Now for reference Figure 3A and Figure 3B Specifically, Figure 3A A flowchart illustrating the synthesis of synthons for analytical decomposition to generate molecular embeddings is depicted, according to one embodiment. Furthermore, Figure 3B A flowchart illustrating the implementation of a machine learning model for predicting DEL experiment counts is depicted, according to one implementation scheme. In various implementation schemes, Figure 3AThe steps in the flowchart shown can be executed by the synthesizer representation module 140.

[0084] Methods for generating molecular representations Figure 3A It begins with multiple synthons, such as synthons 310A, 310B, and 310C. These multiple synthons can be synthons of a compound. For example, the multiple synthons can be a subset of the synthons of a compound. As another example, the multiple synthons are all the synthons of a compound. The compound can be included as part of a virtual library of compounds for performing molecular screening (e.g., virtual molecular screening) against a target. In various embodiments, the target can be a protein target. In a particular embodiment, the target can be a human protein target. The protein target may be associated with a disease, therefore, virtual molecular screening is useful for identifying candidate compounds that can bind to the protein target and modulate its behavior in the disease. As a specific example, the protein target can be a human carbonic anhydrase IX (CAIX) protein target. As another specific example, the protein target can be a horseradish peroxidase (HRP) protein target. As another specific example, the protein target can be a discoid domain receptor tyrosine kinase 1 (DDR1) protein target. As another specific example, the protein target could be a mitogen-activated protein kinase 14 (MAPK14) protein target. However, as those skilled in the art will understand, other known target proteins can be used.

[0085] although Figure 3A Three synthons 310A, 310B, and 310C are explicitly described, but in various embodiments, there may be fewer or additional synthons. Therefore, the following description can be similarly applied to embodiments that may have fewer (e.g., one or two synthons) or additional synthons (e.g., four, five, six, seven, eight, nine, ten, eleven, twelve, thirteen, fourteen, fifteen, sixteen, seventeen, eighteen, nineteen, or twenty or more synthons). As discussed herein, by processing the compound at the synthon level (e.g., in the form of decomposed synthons), this process avoids the step of enumerating the compound from multiple synthons, which is typically a costly and error-prone process.

[0086] Typically, the plurality of synthesizers 310 are converted into a plurality of synthesizer representations. For example, the plurality of synthesizers 310 undergo one or more transformations to generate the plurality of synthesizer representations. In various embodiments, converting the plurality of synthesizers into a plurality of synthesizer representations includes using at least two transformations, at least three transformations, at least four transformations, at least five transformations, at least six transformations, at least seven transformations, at least eight transformations, at least nine transformations, at least ten transformations, at least eleven transformations, at least twelve transformations, at least thirteen transformations, at least fourteen transformations, at least fifteen transformations, at least sixteen transformations, at least seventeen transformations, at least eighteen transformations, at least nineteen transformations, or at least twenty transformations to transform the plurality of synthesizers. In a particular embodiment, converting the plurality of synthesizers into a plurality of synthesizer representations includes using two transformations to transform the plurality of synthesizers. In a particular embodiment, converting the plurality of synthesizers into a plurality of synthesizer representations includes using three transformations to transform the plurality of synthesizers. In a particular embodiment, converting the plurality of synthesizers into a plurality of synthesizer representations includes using four transformations to transform the plurality of synthesizers.

[0087] In various embodiments, one or more transformations involve applying a learned representation model. In various embodiments, each transformation involves applying a learned representation model. In various embodiments, the learned representation model for a first transformation is a different learned representation model from that used for another transformation. In various embodiments, each learned representation model for one transformation is different from another learned representation model used for another transformation. In various embodiments, the learned representation model is a neural network. In a particular embodiment, the learned representation model is a multilayer perceptron (MLP).

[0088] In various embodiments, converting the plurality of synthesizers into a plurality of synthesizer representations includes generating one or more single synthesizer representations from the plurality of synthesizers. In various embodiments, converting the plurality of synthesizers into a plurality of synthesizer representations includes generating one or more bi-synthetic representations. In various embodiments, converting the plurality of synthesizers into a plurality of synthesizer representations includes generating one or more tri-synthetic representations. In various embodiments, converting the plurality of synthesizers into a plurality of synthesizer representations includes generating one or more quad-synthetic representations. In various embodiments, generating one or more single synthesizer representations from the plurality of synthesizers includes analyzing the plurality of synthesizers using a learned representation model. In various embodiments, generating one or more bi-synthetic representations from the one or more single synthesizer representations includes analyzing the one or more single synthesizer representations using a learned representation model. In various embodiments, generating one or more tri-synthetic representations from the one or more bi-synthetic representations includes analyzing the one or more bi-synthetic representations using a learned representation model.

[0089] In various embodiments, the plurality of synthesizers is converted into a plurality of synthesizer representations, including one or more single synthesizer representations, one or more double synthesizer representations, and one or more triple synthesizer representations. In various embodiments, the plurality of synthesizer representations includes three single synthesizer representations, three double synthesizer representations, and one triple synthesizer representation.

[0090] return Figure 3A The plurality of synthons 310 undergo a hierarchical transformation process to generate a plurality of synthon representations. In doing so, the method utilizes the inherent hierarchical structure of the molecule (in its synthon form). Here, the plurality of synthon representations includes a monosynon representation 330A, a disynon representation 330B, and a trisynon representation 330C. Those skilled in the art will understand that in additional embodiments, there may be additional synthon representations (e.g., tetrasynon representations, pentagonal representations, etc.).

[0091] The plurality of synthons 310 undergo a first conversion 325A to generate a plurality of single synthon representations 330A. Here, a single synthon representation 330A may represent a synthon building block of the compound. In various embodiments, the number of single synthon representations 330A is equal to the number of synthons in the plurality of synthons 310. For example, if there are three synthons 310, there may be three corresponding single synthon representations 330A.

[0092] In various implementations, the first transition 325A involves an applied learned representation model, such as a multilayer perceptron. In various implementations, the first transition 325A from the plurality of synthesizers 310 to a single synthesizer representation 330A can be represented as:

[0093] in These refer to the synthesizers in the first, second, and third positions, respectively. It is a monosynthetic embedding.

[0094] The monosynthetic representation 330A is further transformed by a second conversion 325B to generate a disynthetic representation 330B. Here, the disynthetic representation 330B can represent a disynthetic compound (e.g., two synthons). In various embodiments, the disynthetic representation 330B includes one, two, three, four, or five representations. In various embodiments, the disynthetic representation 330B includes two representations. For example, given a synthon... The bisynthesis representation 330B may include a first bisynthesis representation. Second bisynthesis representation In various implementations, the dual-synthetic representation 330B includes three representations. For example, given a synthesizer... The bisynthesis representation 330B may include a first bisynthesis representation. The second disynthesis represents and the third disynthesis .

[0095] In various implementations, the second transformation 325B involves an applied learned representation model, such as a multilayer perceptron. In various implementations, the second transformation 325B from the plurality of monosynthetic representations 330A to the bisynthetic representation 330B can be represented as: = MLP

[0096] = MLP

[0097] = MLP

[0098] in These refer to the synthesizers in the first, second, and third positions, respectively, and This indicates a bisynthesis embedding.

[0099] The disynton representation 330B is further transformed via a third conversion 325C to generate the trisynton representation 330C. Here, the trisynton representation 330C represents a compound with three syntons (e.g., three syntons). In various embodiments, the trisynton representation 330A includes one, two, three, four, or five representations. In various embodiments, the trisynton representation 330C includes one representation. For example, given a synton... The trisynthesis representation of 330C can include the trisynthesis representation. In various implementations, the trisynthesis representation 330C includes more than one representation. For example, given a synthesizer... The ternary representation 330C may include a first ternary representation. Second and third synthesis representation .

[0100] In various implementations, the third transformation 325C involves an applied learned representation model, such as a multilayer perceptron. In various implementations, the third transformation 325C from the plurality of bisynthetic representations 330B to the trisynthetic representation 330C can be represented as: = MLP

[0101] in These refer to the synthesizers in the first, second, and third positions, respectively, and This indicates the embedding of a trisomy.

[0102] In some implementations, the third transformation 325C considers only a subset of all bisynthetic representations 330B. For example, the third transformation 325C can be represented as: = MLP

[0103] in These refer to the synthesizers in the first, second, and third positions, respectively, and This indicates the embedding of a trisomy.

[0104] although Figure 3A Not shown in the text, but one or more additional transformations can be performed to generate additional higher-order composite representations. As used herein, a higher-order composite representation refers to a composite representation that represents a larger number of composites. For example, a ternary composite representation would be a higher-order representation than a duocomposite representation.

[0105] These multiple synthons indicate that, in Figure 3A The composite representation includes a single synthetic representation 330A, a double synthetic representation 330B, and a triple synthetic representation 330C, which are combined to generate a molecular embedding (z) 340. In various embodiments, the composite synthetic representation includes applying a model to generate the molecular embedding (z) 340. In various embodiments, the composite synthetic representation includes applying a feedforward model, such as a feedforward neural network, to generate the molecular embedding (z) 340. In various embodiments, the composite synthetic representation includes applying a recurrent model, such as a recurrent neural network, to generate the molecular embedding (z) 340. For example, the composite synthetic representation may include inputting two single synthetic representations 330A into a recurrent model, and then further inputting a third single synthetic representation into the recurrent model.

[0106] In various implementations, the combinatorial synthon representation includes a polymeric synthon representation, and then a model is applied to the polymeric synthon representation to generate a molecular embedding (z) 340. For example, the combinatorial synthon representation includes a polymeric synthon representation and applies a multilayer perceptron to the polymeric synthon representation.

[0107] In various embodiments, combining the plurality of synthon representations into a molecular embedding includes implementing a multi-head attention mechanism across the plurality of synthon representations. In various embodiments, implementing the multi-head attention mechanism includes using one or more learned attention weights of the plurality of synthon representations. Here, the learned attention weights of the plurality of synthon representations may be useful for identifying which synthon representations play a role in the binding of the compound. For example, a synthon representation assigned a higher weight may be considered to contribute more to the binding of the compound to the target, while a synthon representation assigned a lower weight may be considered to contribute less to the binding of the compound to the target. In various embodiments, the one or more learned attention weights are used to rank the ability of the plurality of synthons (corresponding to synthon representations) to bind to the target.

[0108] In various implementations, the molecular embedding (z) 340 can be represented as:

[0109] Now for reference Figure 4B It describes an example process for generating molecular representations, according to one embodiment. Step 435 involves obtaining multiple synthons that form a compound.

[0110] Step 440 involves converting the plurality of synthesizers into a plurality of synthesizer representations. In various embodiments, step 440 involves performing a hierarchical transformation, wherein higher-order synthesizer representations are constructed hierarchically from lower-order synthesizer representations.

[0111] like Figure 4B As shown, step 440 may involve multiple sub-steps, such as steps 445, 450, and 455. Specifically, step 445 involves generating one or more single-composite representations by analyzing the plurality of composites using a first learned representation model. Step 450 involves generating one or more double-composite representations by analyzing the one or more single-composite representations using a second learned representation model. Then, step 455 involves generating one or more triple-composite representations by analyzing the one or more double-composite representations using a third learned representation model.

[0112] Step 460 involves combining the multiple synthon representations (e.g., monosynone, disynone, trisynone representations) to generate a molecular embedding.

[0113] Example methods for generating target enrichment predictions The methods disclosed in this paper also cover the use of, for example, molecular embeddings to generate target enrichment predictions. See below for further details. Figure 3B The molecular embedding (z) 340 can be provided as input to the machine learning model 345. Here, the machine learning model 345 is trained to generate target enrichment predictions. 350 represents the binding measure between the compound and the target. Specifically, the target enrichment prediction 350 is a prediction learned by the machine learning model 345. For example, the target enrichment prediction 350 represents a prediction of the binding between the compound and the target, which has been denoised (e.g., unaffected by covariates and other noise sources).

[0114] In various implementation schemes, machine learning model 345 also generates one or more covariate predictions. 355. Covariate prediction 355 refers to a learned prediction of the effect of one or more covariates (e.g., noise sources in a DEL experiment). For example, covariate prediction can be a learned prediction of the effect of one or more covariates, including nonspecific binding (e.g., determined from controls) and / or other target data (e.g., binding to beads, streptavidin binding to beads, biotin binding, gel binding, binding to the DEL container surface) or other noise sources (e.g., loading bias, replication bias, initial tag imbalance, experimental conditions, chemical reaction yields, byproducts and truncated products, library synthesis errors, DNA affinity for the target, sequencing depth, and sequencing noise such as PCR bias). In a particular embodiment, covariate prediction is derived from loading bias. In a particular embodiment, covariate prediction is derived from replication noise.

[0115] although Figure 3B Show predictions for only a single covariate 355, but in various implementations, the machine learning model 345 can be trained to predict two or more covariate predictions. 355, each covariate prediction represents a learned prediction of the effect of a particular covariate. In a specific implementation, machine learning model 345 is trained to predict two covariate predictions. 355. For example, the first covariate prediction is derived from loading bias, and the second covariate prediction is derived from duplication bias.

[0116] In various implementation schemes, Figure 3B All the steps related to the machine learning model 345 shown can be performed during the training and deployment of the machine learning model 345. However, in some implementations, such as Figure 3B As shown by the dashed line, Figure 3B Some of the steps shown may only be performed when training the machine learning model 345, and not during the deployment of the machine learning model 345. For example, the machine learning model 345 does not need to generate covariate predictions during deployment. Instead, it is generated only during the training of the machine learning model 345. In this case, covariate prediction 355, count modeling 358, and predicted target count 360 do not need to be generated. In various implementations, covariate prediction is generated during the deployment of the machine learning model 345. 355 is generated but discarded and not used. In this type of implementation, the predicted target count 360 is not generated.

[0117] like Figure 3B As shown, in various embodiments, target enrichment prediction 350 can be directly output by machine learning model 345. Typically, target enrichment prediction 350 represents a measure of binding between the compound and the target. Target enrichment prediction 350 can be used to calculate the binding affinity value of the compound-target complex. In various embodiments, target enrichment prediction 350 can be converted into a binding affinity value. In various embodiments, the binding affinity value is obtained by balancing the dissociation constant (…). ) measurement. In various implementation schemes, the affinity value is combined with the negative logarithm of the equilibrium dissociation constant ( ) measurement. In various implementations, the affinity value is combined with the balance inhibition constant ( ) measurement. In various implementation schemes, the affinity value is combined with the negative logarithm of the equilibrium inhibition constant ( Measurements are taken using the following methods: In various embodiments, the binding affinity is measured by the half-maximum inhibitory concentration (IC50). In various embodiments, the binding affinity is measured by the half-maximum effective concentration (EC50). In various embodiments, the binding affinity is measured by the equilibrium binding constant (EC50). ) measurement. In various implementations, the binding affinity value is measured by balancing the negative logarithm of the binding constant ( Measurement. In various embodiments, the binding affinity value is measured by an activation percentage value. In various embodiments, the binding affinity value is measured by an inhibition percentage value.

[0118] In various implementations, the target enrichment prediction 350 is converted into a binding affinity value according to a predetermined conversion relationship. This predetermined conversion relationship can be determined using DEL experimental data, for example, based on DEL outputs previously generated from DEL experiments (e.g., Figure 1A The DEL outputs 120A and 120B are shown. In various embodiments, the predetermined transformation relationship is a linear equation. Here, the target enrichment prediction 350 can be correlated with the binding affinity value. In various embodiments, the predetermined transformation relationship is any of a linear, exponential, logarithmic, nonlinear, or polynomial equation.

[0119] In various implementations, target enrichment prediction 350 can be used to rank compounds. For example, a first compound with a target enrichment prediction associated with a strong binding affinity to the target can be ranked higher than a second compound with a target enrichment prediction associated with a weak binding affinity to the target. Typically, in medicinal chemistry activities such as hit-to-lead compound optimization, binding affinity values ​​are often used to evaluate and select the next compound to synthesize. Therefore, target enrichment predictions associated with binding affinity values ​​can be used to rank compounds, thus directly guiding the design.

[0120] In various embodiments, the ranking of compounds uses target enrichment prediction 350 and probabilities obtained from a probability density function. In various embodiments, the probability density function is any of a Poisson distribution, binomial distribution, gamma distribution, binomial-Poisson distribution, or negative binomial distribution. In a particular embodiment, the probability density function is a Poisson distribution. In a particular embodiment, the Poisson distribution is a zero-expanded Poisson distribution. As further discussed herein, the probability density function can be a learned distribution used to model DEL counts (e.g., target counts or control counts). The probability obtained from the probability density function can be the predicted zero probability. .here, p These are parameters of the probability distribution. Therefore, the ranking of compounds can be determined based on a metric. Determined, it is represented as ,in This indicates a target enrichment prediction of 350.

[0121] In various embodiments, the ranking of compounds is used to identify conjugates and non-conjugates. In various embodiments, identifying conjugates involves ranking the top compounds in the list... Z One compound was identified as a conjugate. Not included in the preceding text. Z The compounds in the compound are considered non-binding. In various embodiments, the former... Z The term "compound" refers to any one of the following: the first 5 compounds, the first 10 compounds, the first 20 compounds, the first 30 compounds, the first 40 compounds, the first 50 compounds, the first 75 compounds, the first 100 compounds, the first 200 compounds, the first 300 compounds, the first 400 compounds, the first 500 compounds, the first 1000 compounds, or the first 5000 compounds.

[0122] In various embodiments, the compound identified as a target conjugate can be further analyzed to characterize the conjugate. In various embodiments, a conjugate can be defined as a compound having a predicted binding affinity above a threshold binding value. In one case, the conjugate is analyzed to identify common binding motifs within the conjugate that may contribute to effective binding between the conjugate and the target. In various embodiments, common binding motifs refer to those appearing at least...X The chemical groups in the compound. In various embodiments, X% It is at least 10% of the binder, at least 20% of the binder, at least 30% of the binder, at least 40% of the binder, at least 50% of the binder, at least 60% of the binder, at least 70% of the binder, at least 80% of the binder, at least 90% of the binder, or at least 95% of the binder. In various embodiments, X % is a 100% compound.

[0123] As a specific example, the target protein could be human carbonic anhydrase IX (CAIX) protein. However, as those skilled in the art will understand, other known target proteins can be used. Using the methods described herein, target enrichment predictions based on machine learning models can identify compounds that bind to the target protein. A common binding motif in many predicted compounds (e.g., conjugates) that bind to the target protein is a benzenesulfonamide group.

[0124] Now for reference Figure 4A It describes an example flow for generating target enrichment predictions, according to one implementation.

[0125] Step 410 involves obtaining multiple synthons that form the compound.

[0126] Step 415 involves converting the plurality of synthesizers into a plurality of synthesizer representations.

[0127] Step 420 involves combining the multiple synthon representations into a molecular embedding 420.

[0128] Step 425 involves using a machine learning model to analyze the molecular embedding to generate a target enrichment prediction that at least represents a measure of binding between the compound and the target. Typically, the machine learning model is trained to predict target enrichment predictions that are denoised to account for one or more covariate predictions (e.g., unaffected by covariates and other noise sources).

[0129] like Figure 4A As shown, steps 410, 415, 420, and 425 can be repeated for one or more additional compounds to generate target enrichment predictions for each additional compound. Therefore, compounds (e.g., conjugates) that bind to the target can be readily identified. Thus, virtual compound screening can be performed efficiently by repeating steps 410, 415, 420, and 425 in a large-scale or high-throughput manner.

[0130] Optionally, step 430 involves identifying common binding motifs in one or more compounds predicted to bind to the target.

[0131] Example methods for generating predicted target counts This paper also discloses methods for generating predicted target counts (e.g., for DEL). Typically, methods for generating predicted target counts involve implementing a machine learning model and one or more probability density functions to model the target counts. For example, a method for generating predicted target counts for DEL could involve analyzing molecular embeddings using a trained machine learning model trained to output target enrichment predictions and one or more covariate predictions. The target enrichment predictions and one or more covariate predictions are further analyzed, for example, using one or more probability density functions, to model at least the experimental target counts for DEL.

[0132] Refer again Figure 3B This is an example procedure for generating a predicted target count 360 for DEL. Here, the predicted target count 360 can be a predicted DEL output from a DEL panning experiment. For example, the predicted target count 360 can represent the DEL output of one or more DEL panning experiments, examples of which include predictions of DEL counts and / or average counts across multiple DEL panning experiments. Here, the predicted target count 360 is a prediction of DEL counts that includes various covariates, such as off-target bindings or noise (e.g., background, matrix, covariates). Therefore, for a given DEL panning experiment, Figure 3B The process can be used exist On the computer Predict the DEL count to be observed in the panning experiment.

[0133] Figure 3B Starting with molecular embedding (z) 340, as previously described. Machine learning model 345 analyzes molecular embedding (z) 340 and outputs target enrichment predictions. and one or more covariate predictions 355.

[0134] In various implementation schemes, target enrichment prediction )350 and covariate prediction 355 are combined to generate a predicted target count of 360. As an example, combined target enrichment predictions... )350 and covariate prediction 355 relates to performing count modeling 358. In various implementations, the count modeling 358 step includes implementing a probability density function trained to model the predicted target count 360. Therefore, in Figure 3B In the illustrated implementation, count modeling 358 will involve implementing a probability density function trained to model a predicted target count 360 for a DEL experiment, the predicted target count 360 including contributions from the binding between the compound and the target, as well as other covariates.

[0135] In various implementations, the probability density function is represented by any of a Poisson distribution, a binomial distribution, a gamma distribution, a binomial-Poisson distribution, or a negative binomial distribution. In a particular implementation, the probability density function is a Poisson distribution. In a particular implementation, the Poisson distribution is a zero-inflated Poisson distribution.

[0136] Typically, this probability density function includes one or more learnable parameters (e.g., learned and / or adjusted during training, as further described herein). For example, the probability density function may include parameters... This parameter enables the probability density function to more accurately model the predicted target count while taking into account covariates (e.g., noise / bias) in the DEL experiment.

[0137] In various implementation schemes, the predicted target count 360 is represented as:

[0138] in Indicates the target number j The target counts are repeated predictions, where "ZIPoisson" indicates a zero-inflated Poisson distribution. l It is normalized pre-screened count data (normalized to account for sequencing depth differences across experiments), and It is the learning parameter of the zero-inflated Poisson distribution, used to account for covariates (e.g., loading bias and / or repeating bias). This indicates that the covariate prediction is 355. This indicates a target enrichment prediction of 350.

[0139] although Figure 3B Not shown in the text, but additional count modeling steps can be performed by implementing one or more additional probability density functions to model additional DEL counts. For example, this can be achieved by implementing analytical covariate prediction. 355. An additional probability density function is generated to perform additional count modeling steps, specifically to generate the predicted control counts. As used herein, "predicted control counts" refers to the predicted DEL counts resulting from one or more covariate factors. Therefore, in such embodiments, implementing an additional probability density function enables modeling of the predicted number of DEL counts due to covariates. In various embodiments, these DEL counts from covariate effects can be discarded because they are not generated by the binding between the compound and the desired target.

[0140] In various implementation schemes, the predicted control count is expressed as:

[0141] in Represents the first covariatei The number of repeated predicted controls, where "ZIPoisson" indicates a zero-inflated Poisson distribution. l It is normalized pre-screened count data. These are the learned parameters of the zero-inflated Poisson distribution, and This indicates that the covariate prediction is 355.

[0142] In various implementations, the machine learning model 345 outputs multiple covariate predictions 355 (e.g., two or more, three or more, four or more, five or more, six or more, seven or more, eight or more, nine or more, ten or more, eleven or more, twelve or more, thirteen or more, fourteen or more, fifteen or more, sixteen or more, seventeen or more, eighteen or more, nineteen or more, or twenty or more covariate predictions). In such implementations, target enrichment predictions are performed. 350 combines count modeling 358 with multiple covariate predictions 355 to generate predicted target counts 360. In various implementations, additional probability density functions may be implemented for one or more of the multiple covariate predictions 355 to model the DEL counts resulting from any corresponding covariate effects.

[0143] As described herein, in various implementations, multiple probability density functions can be implemented to generate the predicted target count 360 and one or more additional DEL counts, such as one or more predicted control counts generated by covariates. In such implementations, each of the multiple probability density functions can independently be one of a Poisson distribution, a binomial distribution, a gamma distribution, a binomial-Poisson distribution, or a negative binomial distribution. In a particular implementation, each probability density function is a Poisson distribution. In a particular implementation, each Poisson distribution is a zero-inflated Poisson distribution. In various implementations, each of the multiple probability density functions includes one or more learnable parameters that are learned / adjusted during training. .

[0144] Figure 4C An example flow for predicting DEL experimental counts is described, based on an implementation scheme.

[0145] Step 485 involves obtaining a molecular intercalation of the compound. As described herein, this molecular intercalation can be generated from multiple synthetic representations of the compound. Such synthetic representations can be derived from the decomposition synthetics of the compound, thus eliminating the need for a compound enumeration step.

[0146] Step 490 involves using a machine learning model to analyze the molecular embedding to generate (A) a target enrichment prediction representing a measure of binding between the compound and the target, and (B) one or more covariate predictions.

[0147] Step 495 involves combining the target enrichment prediction and the one or more covariate predictions by applying a probability density function that models the experimental target count to generate a predicted target count. In various implementations, the probability density function is a Poisson distribution, such as a zero-inflated Poisson distribution.

[0148] Training machine learning models The embodiments disclosed herein describe the training of a machine learning model that analyzes molecular embeddings derived from factorized synthons. Typically, the machine learning model is trained to generate target enrichment predictions, which represent the learned binding strength between a compound and a target. Therefore, target enrichment predictions can be used to identify and / or rank potential bindings, for example, in virtual compound screening. In various embodiments, target enrichment predictions represent intermediate predictions from the machine learning model. For example, target enrichment predictions are learned by training the machine learning model to predict experimentally observed target counts and / or experimentally observed control counts generated by background / matrix / covariates.

[0149] Typically, the machine learning models described in this paper are trained using training synthons of training compounds with corresponding DNA-encoded library (DEL) outputs. A training synthon refers to a factorized synthon of training compounds. As used herein, a training compound is a compound with known corresponding experimental counts generated through one or more DEL panning experiments. Therefore, these experimental counts can represent the true values ​​used to train the machine learning model.

[0150] In various implementations, the training synthons of the training compounds have known corresponding experimental target counts from a DEL panning experiment. Experimental target counts can refer to signals in DEL data from the DEL experiment, including various noise sources (e.g., background, matrix, covariates). For example, a DEL experiment may include immobilizing protein targets on beads, exposing protein targets to DEL compounds, washing the mixture to remove unbound compounds, and eluting, amplifying, and sequencing the tag sequences. Therefore, the experimental target counts obtained from this DEL experiment can include data generated by various noise sources.

[0151] In various implementations, the training synthons of the training compounds have one or more known corresponding experimental control counts from a DEL panning experiment. Experimental control counts can refer to signals in the DEL data from the DEL experiment, which include only one or more noise sources (e.g., background, matrix, covariates). For example, the DEL experiment can simulate a covariate (e.g., nonspecific binding to beads). This involves incubating a small molecule compound with beads in the absence of a fixed target on the beads. The mixture is washed to remove non-bindings, followed by elution, sequence amplification, and sequencing. Therefore, the experimental control counts obtained from this DEL experiment include data generated by noise sources but exclude data generated by the actual binding of the compound to the target.

[0152] In various implementations, the training synthon of the training compound simultaneously has 1) one or more known corresponding experimental control counts from one or more additional DEL panning experiments and 2) known corresponding experimental target counts from the DEL panning experiments. Specifically, the corresponding DNA-coding library (DEL) output of the training compound includes: 1) experimental control counts generated by covariates determined by a first panning experiment; and 2) experimental target counts determined by a second panning experiment. In such implementations, both the experimental control counts and the experimental target counts can be used as reference ground truth values ​​for training the machine learning model. For example, a machine learning model can be trained to generate target enrichment predictions by attempting to predict the experimental control counts and experimental target counts observed for the training compound.

[0153] Typically, during training iterations involving training synthons of training compounds, methods for training machine learning models involve obtaining multiple training synthons that form training compounds, converting the multiple training synthons into multiple training synthon representations, and combining the multiple training synthon representations into molecular embeddings.

[0154] Here, the step of obtaining multiple training synthons that form the training compound can be performed in a similar or identical manner to that described for the synthons of the reference compound above (e.g., as in the reference). Figure 3A The steps of converting the plurality of training synthons into a plurality of training synthon representations can be performed in a similar or identical manner to that described in the referenced compounds above (e.g., as described in the reference). Figure 3A (As described in transformations 325A, 325B, and 325C). For example, the plurality of training synthesizers can be generated by a hierarchical transformation process, wherein higher-order synthesizer representations are generated from lower-order synthesizer representations. Furthermore, the step of combining the plurality of training synthesizer representations into molecular embeddings can be performed in a similar or identical manner to that described above with reference to the plurality of synthesizer representations (e.g., as...). Figure 3AThe monosynthetic designation described in the text is 330A, the disynthetic designation is 330B, and the trisynthetic designation is 330C.

[0155] Furthermore, during training iterations involving training synthons of training compounds, a machine learning model is implemented to analyze molecular embeddings to generate target enrichment predictions and one or more covariate predictions. Here, this step can be performed in a similar or identical manner to the molecular embeddings described above during the deployment of the reference machine learning model (e.g., as referenced). Figure 3B (As described in the section on molecular embedding (z)340).

[0156] Furthermore, the training iterations involving the training compounds also include combining target enrichment predictions and one or more covariate predictions to generate predicted target counts. In various embodiments, combining target enrichment predictions and covariate predictions to generate predicted target counts involves applying a probability density function that models the predicted target counts. In various embodiments, the probability density function is represented by any of a Poisson distribution, a binomial distribution, a gamma distribution, a binomial-Poisson distribution, or a gamma-Poisson distribution. In a particular embodiment, the probability density function is represented by a Poisson distribution. In various embodiments, the Poisson distribution is a zero-inflated Poisson distribution.

[0157] In various embodiments, training iterations involving the training compound further include analyzing one or more covariate predictions to generate one or more predicted control counts. In various embodiments, generating one or more predicted control counts includes applying a probability density function that models the experimental control counts. In various embodiments, the probability density function is represented by any of a Poisson distribution, a binomial distribution, a gamma distribution, a binomial-Poisson distribution, or a gamma-Poisson distribution. In a particular embodiment, the probability density function is represented by a Poisson distribution. In various embodiments, the Poisson distribution is a zero-expanded Poisson distribution.

[0158] In various embodiments, the training iterations involving the training compound further include analyzing two covariate predictions to generate two predicted control counts. For example, the first covariate prediction may consider a first covariate (e.g., loading bias), and the second covariate prediction may consider a second covariate (e.g., replication bias). In various embodiments, generating each predicted control count includes applying a probability density function that models the corresponding experimental control count. In various embodiments, the probability density function is represented by any of a Poisson distribution, a binomial distribution, a gamma distribution, a binomial-Poisson distribution, or a gamma-Poisson distribution. In a particular embodiment, the probability density function is represented by a Poisson distribution. In various embodiments, the Poisson distribution is a zero-inflated Poisson distribution.

[0159] Furthermore, the training iterations involving the training compounds also include determining a loss value based on at least the predicted target count and the experimental target count, according to a loss function. The loss value can then be used (e.g., backpropagation) to adjust at least the parameters of the machine learning model to improve its predictions. In various implementations, the loss value is calculated using the predicted target count and the experimental target count. For example, the closer the predicted target count is to the experimental target count, the smaller the loss value. Therefore, the machine learning model can be trained (e.g., the parameters of the machine learning model are tuned) to minimize the loss value.

[0160] In various implementations, the loss value is calculated using predicted control counts and experimental control counts. For example, the closer the predicted control count is to the experimental control count, the smaller the loss value. In various implementations, the loss value is calculated using the predicted control count and experimental control count of a first covariate and the predicted control count and experimental control count of a second covariate. In various implementations, the loss value is calculated using each of the predicted target count, the experimental target count, one or more predicted control counts, and one or more experimental control counts. In such implementations, the closer the predicted target count is to the experimental target count, and the closer each of the one or more predicted control counts is to its corresponding one or more experimental control counts, the smaller the loss value. In various implementations, the loss value is determined by calculating the root mean square error (RMSE). For example, the RMSE value can be calculated as the square root of the sum of 1) the difference between the predicted target count and the experimental target count and 2) the difference between one or more predicted control counts and one or more corresponding experimental control counts.

[0161] In various implementations, the loss value is determined based on a probability density function modeling the experimental target count and the experimental control count. In various implementations, the loss value is determined based on a first probability density function modeling the experimental target count and a second probability density function modeling the experimental control count.

[0162] In various implementations, the probability density function is represented by any of the Poisson, binomial, gamma, binomial-Poisson, or gamma-Poisson distributions. In a particular implementation, the probability density function is represented by a Poisson distribution. In various implementations, the Poisson distribution is a zero-inflated Poisson distribution. An example zero-inflated Poisson (ZIP) distribution is described and implemented according to equations (2) and (3) in the examples below (e.g., for calculating...). and In a particular implementation, the Poisson distribution is based on one or more parameters. Characterize. Example parameters and The Poisson distribution is described by equations (2) and (3) in the example below.

[0163] In various implementations, the loss function is any one of the following: negative log-likelihood loss, binary cross-entropy loss, focal loss, arc loss, cosface loss, cosine-based loss, or a loss function based on the BEDROC metric. In a specific implementation, the loss function is negative log-likelihood loss.

[0164] Now for reference Figure 5A It describes an example flowchart for training a machine learning model based on one implementation scheme. See also: Figure 5B The document further describes an example flowchart for training machine learning models based on one implementation scheme. Figure 5A and Figure 5B Together, they depicted a single training iteration of the training synthons of the training compounds. Therefore, Figure 5A and Figure 5B The flowchart shown can be executed multiple times in multiple iterations to train a machine learning model.

[0165] Figure 5A The example flowchart begins with multiple training synthons 510A, 510B, and 510C of the training compound. These multiple training synthons 510 can undergo a hierarchical transformation process to generate multiple synthon representations (e.g., as shown in the diagram). Figure 5A The single synthesizer representation shown is 530A, the dual synthesizer representation is 530B, and the triple synthesizer representation is 530C. Specifically, the plurality of training synthesizers 510 undergo transformation 325A to generate a single synthesizer representation 530A. The single synthesizer representation 530A undergoes transformation 325B to generate a dual synthesizer representation 530B. The dual synthesizer representation 530B undergoes transformation 325C to generate a triple synthesizer representation 530C. In various embodiments, although... Figure 5A The representation is not shown, but further transformations can be performed to generate higher-order synthon representations.

[0166] Each of transformations 325A, 325B, and 325C can be performed by a representation model. In various embodiments, each representation model is a machine learning model, such as a neural network. In a particular embodiment, each representation model is a multilayer perceptron.

[0167] The plurality of synthesizers represent (e.g., such as) Figure 5A The single synthon representation 530A, the double synthon representation 530B, and the triple synthon representation 530C shown are combined to generate a molecular embedding (z) 540. In various embodiments, combining the multiple synthon representations involves implementing a multi-head attention mechanism across the multiple synthon representations.

[0168] Next reference Figure 5BThe molecular embedding (z)540 is fed into the machine learning model 345. The machine learning model generates target enrichment predictions representing the binding between the training compound and the target (e.g., a protein target). 550. The machine learning model also generates covariate predictions. 555 (e.g., noise prediction). Here, target enrichment prediction 550 represents the learned enrichment value, which represents the binding between the training compound and the target, excluding noise sources (e.g., background, matrix, covariates). Covariate prediction 555 represents the learned value or score attributable to non-target binding sources and / or other noise sources (e.g., background, matrix, covariates).

[0169] Target enrichment prediction 550 and covariate prediction 555 are combined to generate a predicted target count 560. The predicted target count 560 represents a prediction of the DEL count from the DEL panning experiment, including various non-target binding sources and / or other noise sources (e.g., background, matrix, covariates). In various embodiments, combining target enrichment prediction 550 and covariate prediction 555 involves adding target enrichment prediction 540 and covariate prediction 555. In various embodiments, combining target enrichment prediction 540 and covariate prediction 555 involves performing a linear or nonlinear combination of target enrichment prediction 540 and covariate prediction 555. For example, in some embodiments, combining target enrichment prediction 540 and covariate prediction 555 may involve performing a weighted summation of target enrichment prediction 540 and covariate prediction 555, where the weights are previously learned (e.g., learned weights from a machine learning model, such as a neural network) or may be fixed weights determined according to a predetermined weighting scheme. In various embodiments, for example... Figure 5B The embodiment shown performs count modeling 562 to generate a predicted target count 560 by combining target enrichment prediction 550 and covariate prediction 555. Here, count modeling 562 involves implementing a probability density function to model the predicted target count 560. As described herein, the probability density function may be a Poisson distribution, optionally a zero-inflated Poisson distribution. The probability density function includes one or more learnable parameters (e.g., parameters...). ).

[0170] Given a predicted target count of 560, calculate the loss value. Here, the loss value can be calculated based on a combination of the predicted target count of 560 and the experimental target count of 570. For example... Figure 5B As shown, the loss value can be represented as the difference between the predicted target count 560 and the experimental target count 570. Here, the loss value can be backpropagated to train at least a machine learning model.

[0171] In a particular implementation, the experimental target count 570 is an observed dataset, such as a set of DEL counts. The predicted target count 560 can be represented as a distribution that maximizes the likelihood of that observed data. Here, this distribution is parameterized by the weights predicted by the model. To compute the loss value for a single training example (e.g., a single training molecule), the likelihood of each count observation under the predicted distribution is determined. Assuming that each observation data (e.g., a count observation) is independent, the loss value for a single training molecule can be a product of a single probability or likelihood (associated with each count observation of that molecule). In a particular implementation, the loss value is computed by taking the negative log-likelihood (NLL) as the loss.

[0172] like Figure 5B As further shown in the figure, covariate prediction It can be analyzed, for example, by performing count modeling 558 to generate a predicted control count 564. In various implementations, count modeling 558 involves implementing a probability density function to model the predicted control count 564. As described herein, the probability density function may be a Poisson distribution, optionally a zero-inflated Poisson distribution. The probability density function includes one or more learnable parameters (e.g., parameters). ).

[0173] Given a predictive control count of 564, calculate the loss value. Here, the loss value can be calculated based on a combination of the predictive control count of 564 and the experimental control count of 565. For example... Figure 5B As shown, the loss value can be represented as the difference between the predicted control count 564 and the experimental control count 565. Here, the loss value can be backpropagated to train at least a machine learning model.

[0174] although Figure 5B Two separate loss values ​​backpropagated are shown, but in various implementations, the two loss values ​​can be combined into a single loss value, which is backpropagated to train at least a machine learning model. In various implementations, the machine learning model outputs two or more covariate predictions 555, which model the two or more covariates. Therefore, count modeling 558 can be performed on the two or more covariate predictions 555 to generate two or more predictive control counts 564. The two or more predictive control counts 564 are combined with the corresponding two or more experimental control counts 565 for each of the two or more covariates to generate the loss value used for backpropagation.

[0175] The loss value is backpropagated to train at least machine learning model 345. The parameters of machine learning model 345 are adjusted based on the calculated loss value. Specifically, the parameters of machine learning model 345 are adjusted to minimize the calculated loss value. In various implementations, the backpropagated loss value is also used to train one or more additional machine learning models, including performing... Figure 5A The representation model shown for conversion 325A, 325B, or 325C, or for performing... Figure 5B The count model shown is a probability density function of 558 or 562. Specifically, Figure 5A and 5B The trainable elements shown are depicted with dashed lines. In various implementations, machine learning model 345 executes... Figure 5A The representation model of the conversion 325A, 325B, or 325C shown, and the method for performing... Figure 5B Each of the probability density functions in the counting modeling 558 or 562 shown is jointly trained. In other words, the parameters of the machine learning model 345, the parameters representing the model, and the parameters of the probability density function can be jointly tuned during training iterations, enabling the machine learning model to better predict experimentally observed target counts and experimentally observed control counts generated by background / matrix / covariates, thereby intrinsically learning target enrichment prediction.

[0176] Now for reference Figure 6 It describes an example process for training a machine learning model based on one implementation scheme.

[0177] Step 610 involves obtaining a plurality of training synthons that form training compounds. Here, the plurality of training synthons refer to factorized synthons of the training compounds.

[0178] Step 615 involves converting the plurality of training synthesizers into a plurality of training synthesizer representations. In various embodiments, the step of converting the plurality of training synthesizers may involve a hierarchical conversion process, for example, referring to... Figure 3A The described process generates the plurality of training synthesizer representations. In short, converting the plurality of training synthesizers into a plurality of training synthesizer representations may involve applying one or more representation models to hierarchically construct higher-order synthesizer representations.

[0179] Step 620 involves combining the plurality of training synthesizer representations (e.g., single synthesizer representations, dual synthesizer representations, triple synthesizer representations, etc.) into a molecular embedding.

[0180] Step 625 involves using a machine learning model to analyze molecular embeddings to generate target enrichment predictions and one or more covariate predictions. In a particular implementation, the machine learning model generates two covariate predictions (e.g., loading bias and duplication bias).

[0181] Step 630 involves combining target enrichment prediction and covariate prediction to generate predicted target counts. In various implementations, combining target enrichment prediction and covariate prediction involves performing count modeling by implementing a probability density function that models the predicted target counts (e.g., DEL counts).

[0182] Step 635 involves determining a loss value based on at least the predicted target count and the experimental target count. Here, the experimental target count is used as the true value.

[0183] Step 640 involves training a machine learning model based on a determined loss value. In various embodiments, step 640 also involves training one or more representation models and one or more models that model the predicted target counts using a probability density function. In various embodiments, the determined loss value is used to jointly train each of the machine learning model, the one or more representation models, and the models that model the predicted target counts using a probability density function. Thus, in training iterations, target enrichment predictions are learned by attempting to predict at least the experimental control counts (e.g., observed experimental control counts from a DEL experiment modeled on a specific covariate).

[0184] Machine learning model benchmarks In various implementations, the methods described herein are evaluated relative to known methods to determine the relative performance of the disclosed model. Examples of known methods include, but are not limited to, Random Forest (RF), XGBoost, k-Nearest Neighbors (kNN), and Deep Neural Networks (DNN) and / or Graph Isomorphic Networks (GIN). Evaluation metrics for model performance may include any known machine learning performance metrics (e.g., loss value, Spearman correlation between model predictions and experimental results, F1 score, accuracy, precision, and / or recall).

[0185] Systems and computing devices In various implementations, the methods described herein are performed on a computing device. Examples of computing devices may include personal computers, desktop computers, laptops, server computers, computing nodes within a cluster, message processors, handheld devices, multiprocessor systems, microprocessor-based or programmable consumer electronics, network PCs, minicomputers, mainframes, mobile phones, PDAs, tablets, pagers, routers, switches, etc.

[0186] Figure 7A The following is shown for implementation Figure 1A-1B Example computing devices for the systems and methods described in sections 2, 3A-3B, 4A-4C, 5A-5B, and 6. Furthermore, Figure 7B An implementation scheme describes the overall system environment for implementing the synthetic submodeling system. Figure 7C It is used for implementation Figure 7B An example depiction of a distributed computing system environment.

[0187] In some implementation schemes, Figure 7AThe computing device 700 shown includes at least one processor 702 coupled to a chipset 704. The chipset 704 includes a memory controller hub 720 and an input / output (I / O) controller hub 722. A memory 706 and a graphics adapter 712 are coupled to the memory controller hub 720, and a display 718 is coupled to the graphics adapter 712. A storage device 708, an input interface 714, and a network adapter 716 are coupled to the I / O controller hub 722. Other embodiments of the computing device 700 have different architectures.

[0188] Storage device 708 is a non-transitory computer-readable storage medium, such as a hard disk drive, optical disc read-only memory (CD-ROM), DVD, or solid-state storage device. Memory 706 stores instructions and data used by processor 702. Input interface 714 is a touchscreen interface, mouse, trackball, or other type of input interface, keyboard, or some combination thereof, and is used to input data into computing device 700. In some embodiments, computing device 700 may be configured to receive input (e.g., commands) from input interface 714 via gestures from a user. Graphics adapter 712 displays images and other information on display 718. Network adapter 716 couples computing device 700 to one or more computer networks.

[0189] Computing device 700 is adapted to execute computer program modules to provide the functions described herein. As used herein, the term "module" refers to computer program logic for providing specified functions. Therefore, modules can be implemented in hardware, firmware, and / or software. In one embodiment, the program module is stored on storage device 708, loaded into memory 706, and executed by processor 702.

[0190] The type of computing device 700 may differ from the embodiments described herein. For example, computing device 700 may lack some of the components described above, such as graphics adapter 712, input interface 714, and display 718. In some embodiments, computing device 700 may include processor 702 for executing instructions stored in memory 706.

[0191] In various implementation schemes, Figure 7B The different entities depicted can implement one or more computing devices to perform the methods described above, including methods for training and deploying one or more machine learning models. For example, the synthetic sub-modeling system 130, third-party entity 740A, and third-party entity 740B can each employ one or more computing devices. As another example, one or more subsystems of the synthetic sub-modeling system 130 (such as...) Figure 1B (As shown) The above method can be performed using one or more computing devices.

[0192] Methods for training and deploying one or more machine learning models can be implemented in hardware or software, or a combination of both. In one embodiment, a non-transitory machine-readable storage medium, such as the medium described above, is provided, comprising data storage material encoded with machine-readable data, capable of displaying any dataset of the machine learning models disclosed herein, along with execution and results, when used with a machine programmed to use instructions for the data.

[0193] The above-described method can be implemented in a computer program that executes on a programmable computer, which includes a processor, a data storage system (including volatile and non-volatile memory and / or storage elements), a graphics adapter, an input interface, a network adapter, at least one input device, and at least one output device. A display is coupled to the graphics adapter. Program code is applied to the input data to perform the above-described functions and generate output information. The output information is applied to one or more output devices in a known manner. The computer can be, for example, a conventionally designed personal computer, microcomputer, or workstation.

[0194] Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system. However, if desired, the program can be implemented in assembly language or machine language. In any case, the language can be a compiled language or an interpreted language. Each such computer program is preferably stored on a storage medium or device (e.g., ROM or disk) readable by a general-purpose or special-purpose programmable computer for configuring and operating the computer to execute the program described herein when the storage medium or device is read by the computer. The system can also be considered as implemented as a computer-readable storage medium configured with a computer program, wherein such a storage medium causes the computer to operate in a particular and predetermined manner to perform the functions described herein.

[0195] The signature pattern and its database can be provided in various media for ease of use. "Media" refers to an article of manufacture capable of recording and reproducing the signature pattern information of this invention. The database of this invention can be recorded on a computer-readable medium, for example, any medium that can be directly read and accessed by a computer. Such media include, but are not limited to, magnetic storage media such as floppy disks, hard disk storage media, and magnetic tape; optical storage media such as CD-ROMs; electrical storage media such as RAM and ROM; and hybrids of these categories such as magnetic / optical storage media. Those skilled in the art will readily understand how to use any currently known computer-readable medium to create an article of manufacture including records of information from this database. "Record" refers to the process of storing information on a computer-readable medium using any such method known in the art. Any convenient data storage structure can be selected based on the means used to access the stored information. Storage can be performed using various data processing programs and formats, such as word processing text files, database formats, etc.

[0196] System Environment Figure 7B According to one implementation scheme, the overall system environment for implementing the synthetic submodeling system is described. The overall system environment 725 includes the synthetic submodeling system 130 (as previously referenced). Figure 1A (as described) and one or more third-party entities 740A and 740B communicating with each other via network 730. Figure 7A One implementation of the overall system environment 725 is depicted. In other implementations, additional or fewer third-party entities 740 may be included that communicate with the synthesizer modeling system 130. Typically, the synthesizer modeling system 130 implements machine learning models to make predictions, such as predictions of compound binding, virtual screening, or hit selection and analysis. The third-party entities 740 communicate with the synthesizer modeling system 130 for purposes related to implementing the machine learning models or obtaining predictions or results from the machine learning models.

[0197] In various implementations, the methods described above, performed by the synthesizer modeling system 130, can be distributed between the synthesizer modeling system 130 and a third-party entity 740. For example, third-party entity 740A or 740B can generate training data and / or train a machine learning model. The synthesizer modeling system 130 can then deploy the machine learning model to generate predictions, such as predictions of compound binding, virtual screening, or hit selection and analysis.

[0198] Third-party entities In various implementations, third-party entity 740 represents a partner entity of synthetic sub-modeling system 130, operating upstream or downstream of synthetic sub-modeling system 130. As an example, third-party entity 740 operates upstream of synthetic sub-modeling system 130 and provides information to synthetic sub-modeling system 130 to enable the training of a machine learning model. In this case, synthetic sub-modeling system 130 receives data, such as DEL experimental data collected by third-party entity 740. For example, third-party entity 740 may have performed experiments on one or more DEL experiments (e.g., Figure 1A The analysis of DEL experiments 115A or 115B shown is performed, and the DEL experimental data from these experiments are provided to the synthesizer modeling system 130. Here, the third-party entity 740 can synthesize small molecule compounds of DEL, incubate the small molecule compounds of DEL with immobilized protein targets, elute the bound compounds, and amplify / sequencing DNA tags to identify the putative binders. Therefore, the third-party entity 740 can provide sequencing data to the synthesizer modeling system 130.

[0199] As another example, a third-party entity 740 operates downstream of the synthesizer modeling system 130. In this case, the synthesizer modeling system 130 can identify predicted conjugates through virtual screening and provide the third-party entity 740 with information related to the predicted conjugates. The third-party entity 740 can then use the information identifying the predicted conjugates for its own purposes. For example, the third-party entity 740 could be a drug developer. Therefore, the drug developer can synthesize the predicted conjugates for further research.

[0200] network This disclosure contemplates any suitable network 730 capable of enabling connectivity between the synthetic submodeling system 130 and the third-party entity 740. Network 730 may include any combination of local area networks (LANs) and / or wide area networks (WANs) using wired and / or wireless communication systems. In one embodiment, network 730 uses standard communication technologies and / or protocols. For example, network 730 includes communication links using technologies such as Ethernet, 802.11, WiMAX, 3G, 4G, CDMA, and Digital Subscriber Line (DSL). Examples of network protocols used for communication via network 730 include Multiprotocol Label Switching (MPLS), Transmission Control Protocol / Internet Protocol (TCP / IP), Hypertext Transfer Protocol (HTTP), Simple Mail Transfer Protocol (SMTP), and File Transfer Protocol (FTP). Data exchanged via network 730 may be represented using any suitable format, such as Hypertext Markup Language (HTML) or Extensible Markup Language (XML). In some embodiments, all or part of the communication links of network 730 may be encrypted using any suitable technology.

[0201] Application Programming Interface (API) In various implementations, the synthetic sub-modeling system 130 communicates with third-party entities 740A or 740B through one or more application programming interfaces (APIs) 735. API 735 can define data fields, calling protocols, and function exchanges between the computing system maintained by the third-party entity 740 and the synthetic sub-modeling system 130. API 735 can be implemented to define or control parameters of data received or provided by the third-party entity 740 and data received or provided by the synthetic sub-modeling system 130. For example, API 735 can be implemented to provide access only to information generated by one of the subsystems including the synthetic sub-modeling system 130. API 735 can support the implementation of permission restrictions and tracking mechanisms for information provided by the synthetic sub-modeling system 130 to the third-party entity 740. Such permission restrictions and tracking mechanisms supported by API 735 can be implemented using blockchain-based networks, secure ledgers, and information management keys. Examples of APIs include remote APIs, Web APIs, operating system APIs, or software application APIs.

[0202] APIs can be provided as libraries, comprising specifications for routines, data structures, object classes, and variables. In other cases, APIs can be provided as specifications for remote calls exposed to API consumers. API specifications can take various forms, including international standards such as POSIX, vendor documentation such as the Microsoft Windows API, or libraries of programming languages, such as the Standard Template Library in C++ or the Java API. In various implementations, the synthetic sub-modeling system 130 includes a set of custom APIs specifically developed for the synthetic sub-modeling system 130 or its subsystems.

[0203] Distributed computing environment In some implementations, the methods described above, including methods for training and implementing one or more machine learning models, are executed in a distributed computing system environment, wherein local and remote computer systems are linked via a network (via a hardwired data link, a wireless data link, or a combination of hardwired and wireless data links), and both perform the task. In some implementations, one or more processors used to implement the methods described above may be located in a single geographic location (e.g., in a home environment, an office environment, or a server farm). In various implementations, one or more processors used to implement the methods described above may be distributed across multiple geographic locations. In a distributed computing system environment, program modules may reside in local and remote memory storage devices.

[0204] Figure 7C It is used for implementation Figure 7BAn example of a distributed computing system environment is depicted above. The distributed computing system environment 750 may include a control server 760, such as computing devices 700, connected via a communication network to at least one distributed computing resource pool 770, an example of which is described above with reference to Figure 7. In various embodiments, additional distributed pools 770 may coexist with the control server 760 within the distributed computing system environment 750. Computing resources may be dedicated to exclusive use within a distributed pool 770 or shared with other pools within the distributed processing system and with other applications outside the distributed processing system. Furthermore, computing resources in the distributed pool 770 can be dynamically allocated, adding or removing computing devices 700 from the pool 710 as needed.

[0205] In various implementations, control server 760 is a software application that provides control and monitoring of computing devices 700 within the distributed pool 770. Control server 760 itself can be implemented on the computing device (e.g., see reference above). Figure 7A The computing device 700 is described. Communication between the control server 760 and the computing device 700 in the distributed pool 770 can be facilitated through an application programming interface (API) (e.g., a Web services API). In some implementations, the control server 760 provides users with management and computing resource management functions for controlling the distributed pool 770 (e.g., defining resource availability, submitting, monitoring, and controlling tasks performed by the computing device 700, controlling the timing of tasks to be completed, prioritizing tasks, or storing / transferring data generated by completed tasks).

[0206] In various implementations, control server 760 identifies computational tasks to be executed in distributed computing system environment 750. Computational tasks can be divided into multiple units of work that can be executed by different computing devices 700 in a distributed pool 770. By dividing and executing computational tasks among computing devices 700, computational tasks can be executed efficiently in parallel. This allows tasks to be completed with improved performance (e.g., faster, less resource consumption) compared to non-distributed computing system environments.

[0207] In various implementations, the computing devices 700 in the distributed pool 770 can be configured differently to ensure efficient performance of their respective tasks. For example, a first group of computing devices 700 may be dedicated to performing the collection and / or analysis of phenotypic data. A second group of computing devices 700 may be dedicated to performing the training of machine learning models. Given that more resources may be required when training machine learning models, the first group of computing devices 700 may have less random access memory (RAM) and / or processors than the second group of computing devices 700.

[0208] The computing devices 700 in the distributed pool 770 can execute their respective jobs in parallel, and upon completion, can store the results in persistent storage and / or transmit the results back to the control server 760. The control server 760 can compile the results, or, if necessary, reassign the results to the appropriate computing device 700 for further processing.

[0209] In some implementations, the distributed computing system environment 750 is implemented in a cloud computing environment. In this specification, "cloud computing" is defined as a model that enables on-demand network access to a set of shared, configurable computing resources. For example, the control server 760 and the computing devices 700 of the distributed pool 770 can communicate via the cloud. Therefore, in some implementations, the control server 760 and the computing devices 700 are located in geographically different locations. Cloud computing can be used to provide on-demand access to a shared set of configurable computing resources. The shared set of configurable computing resources can be rapidly configured through virtualization and released with low management effort or service provider interaction, and then scaled accordingly. The cloud computing model can consist of various features, such as on-demand self-service, broad network access, resource pooling, rapid elasticity, metered services, etc. The cloud computing model can also expose various service models, such as Software as a Service ("SaaS"), Platform as a Service ("PaaS"), and Infrastructure as a Service ("IaaS"). The cloud computing model can also be deployed using different deployment models, such as private cloud, community cloud, public cloud, hybrid cloud, etc. In this specification and claims, a "cloud computing environment" is an environment employing cloud computing.

[0210] Example Example 1: Example Model Architecture These examples illustrate the published model, which this paper refers to as the "factorization" model, that learns factorized synthon representations, constructs corresponding bi-synthetic and tri-synthetic representations from the factorized synthon representations, and generates target enrichment predictions.

[0211] Molecules as a composite of synthons The factorization model disclosed in this paper extensively utilizes the combinatorial properties of DEL molecules and creates a combination of representations using the individual building blocks of each molecule. Since the DEL selection data signal is highly correlated with its synthon composition, this hierarchical decomposition captures the nuances of noise in the data. Therefore, this paper describes a fully generative model that captures the underlying data generation process of DEL count data. First, this paper introduces the mathematical notation: set up Let { be the set of DEL molecules in the dataset, { } represents the synthetic subsets at the first, second, and third positions, respectively. Each molecule is represented as... ∈ The subscript indicates the identity of the synthesizer at a specific position. ∈ , ∈ , ∈ To simplify notation, the subscript of a specific composite element is omitted if it does not exist. For example, This indicates the synthesizer corresponding to the second position. molecules, This indicates the synthesizer corresponding to the first position. Second position synthesizer The process of combining molecules can be easily generalized to higher-order combinations of more than three synthons, but for ease of demonstration, we focus here on the setup of the ternary DEL molecule. The DEL molecule is used for selection experiments, where the molecule undergoes multiple rounds of washing to determine the strongest binding under each experimental condition. Two experimental conditions are presented: the target condition describes the data for selection against the protein target of interest, and the matrix condition describes the data in the absence of a protein target. The observed data is DNA read counts, expressed as... and , representing target and matrix readout counts, respectively. Here, ( ) represent the number of repetitions for the target and control counts, respectively. Furthermore, DEL data are typically calibrated using additional readouts from the library itself, which we denote as (This symbol is lowercase because there is usually only a single readout of the library.) This library readout is a noisy estimate of the relative abundance of each molecular member.

[0212] DEL data generation model The overall goal is to maximize the given input molecule The likelihood of the count data was observed at that time. The plate model of this paradigm is as follows: Figure 8A As shown. For simplicity, assume there is only one count repetition, but this can be easily extended to multiple observations. Let... For potential synthon embedding; in its most basic form, the target is decomposed into Equation 1: (1) To better remove noise from the contribution of actual molecule binding to the readout, the latent variable was explicitly defined as... The binding affinity of the captured molecules under target and matrix experimental conditions was analyzed. While many factors influence the final read count in DEL experiments, two prominent factors were selected for inclusion in the model. These two factors include the selection of pre-library readouts. and repetitive level noise, denoted as The latter explains the variance between different replicates of the same experiment, since the different coding DNA in the replicates may be related to PCR bias noise. The model can then be generated based on the decomposition of Equation 2, where... It is the set of model parameters for learning.

[0213] (2) Synthetome and molecule representations parameterized using neural networks refer to Figure 8B It depicts a schematic diagram of the factorization model architecture and the corresponding data flow. Let... Let be a collection of DEL molecules, in which These are the composite subsets at the first, second, and third positions, respectively.

[0214] In addition, each set also includes empty elements. This indicates that a synthon does not exist at that position (because not all molecules are tripons). To simplify notation, the subscript is omitted if a specific synthon does not exist. For example, This indicates the synthesizer corresponding to the second position. molecules, This indicates the synthesizer corresponding to the first position. Second position synthesizer Combined molecules.

[0215] set up For the output embedding of molecules. There exists some kind of transformation. It receives input molecules and maps them to Dimensional embedding. The simplest transformation. It can be in The fingerprint representation is based on a multilayer perceptron (MLP), that is, = = ).

[0216] However, complete molecules Careful enumeration is required, which is typically a costly process. Since DEL data are highly correlated within specific synthoid groups, individual synthoid information is preserved during the construction of molecular embeddings. Therefore, this paper proposes a model that eliminates the need for manual enumeration. First, the embeddings of each single synthoid are constructed as follows: = Next, The bisynthesis embedding is constructed as = MLP Trisomers (complete molecular embedding) are constructed as The molecular embedding of the polymer is then: (1) Here, molecular embedding Used to predict the intrinsic properties of molecules: , representing the binding affinity of the molecule to the control / matrix and the target, respectively. ∈ (0, 1), which is a measure of the probability of predicting noise / uncertainty.

[0217] set up These are the count data from the i / jth replicates for the control / matrix and target, respectively. To include pre-loading / selection and replicate bias, additional terms were introduced, including: [The text abruptly ends here, so the translation stops as well.] For normalized selection of count data ( = and set The weights for learning are used to account for replication bias in control and target experiments.

[0218] The count data is modeled as a zero-inflated Poisson distribution, as shown below: (2) (3) Example 2: Sample Data and Training The experiments in this example were performed on publicly available DEL data from Gerry, C. et al., “DNA barcoding a complete matrix of stereoisomeric small molecules”, Journal of the American Chemical Society, 2019, 141, 10225-10235, the entire contents of which are incorporated herein by reference. Gerry et al. describe panning data for two targets: carbonic anhydrase IX (CA-IX) and horseradish peroxidase (HRP). Their DEL is a trisynonym library consisting of 8 synthons at position A, 114 synthons at position B, and 118 synthons at position C (a total of 107,616 molecules), selected to promote molecular chemical diversity. Their data include target experiments with read counts and off-target read counts collected only with beads. For CAIX, the dataset includes 2 replicates of off-target control data and 4 replicates of target experiments; while for HRP, the dataset includes 4 replicates of off-target control data and 2 replicates of target data. In addition, there is data collected on pre-selected DELs, which is an indicator of the relative abundance of different DEL members.

[0219] Both CA-IX and HRP have known pharmacophores. The benzene-sulfonamide motif is known to promote binding to CA-IX. In this dataset, there are two synthons containing benzene-sulfonamide at the C position, one meta-substituted relative to the aryl group and the other para-substituted. The para-substituted benzene-sulfonamide generally exhibits higher activity towards CA-IX. Meanwhile, HRP, a protein historically well-studied in the DEL context, appears to have high affinity for compounds containing sulfonyl chloride-derived Michael receptors. In this dataset, three such synthons at the B position show high activity; these three synthons are considered the "gold standard" label for HRP. These structures are listed in descending order of binding affinity. Figure 9A and 9B Visualization within the text. Specifically, Figure 9A The known pharmacophores of carbonic anhydrase IX (CA-IX) are shown in order of binding affinity. Figure 9B The known pharmacophores of horseradish peroxidase (HRP) are shown in order of binding affinity. For CA-IX, benzene-sulfonamide is the known structure that induces activity. Substitution of sulfonamides affects the reactivity of chemical species, with para-substituted components found to be more active. For HRP, the electrophilic Michael receptor is the known pharmacophore. In this dataset, three Michael receptors are active.

[0220] Training settings Several training settings were included to validate the model's performance. At the most basic level, the model's performance was evaluated on a reserved test set of data. For this, the data was randomly split into five distinct partitions: 80% / 10% / 10% for training / validation / test. The model was trained on the training set, selected based on the validation set, and finally tested on the reserved test set. Where applicable, the results were averaged across the five partitions.

[0221] Random partitioning is not always an ideal method for testing molecular datasets. To test the generalization ability of molecular representations, many methods attempt to partition molecules by their molecular skeleton. For DEL, synthons provide a natural grouping and separation of chemical spaces, rather than using a general molecular skeleton strategy. By using synthons to partition data, the generalization ability of the model can be tested on unseen chemical structures.

[0222] In this dataset, known pharmacophores conveniently locate specific synthons, allowing for the development of intuitive segmentation strategies. Most signals are captured by these pharmacophores, so not all of these molecules are retained from training. Instead, segmentation is performed at synthon positions that do not include these individual pharmacophores. Specifically, for CA-IX, benzene-sulfonamide is at position C, so synthon segmentation is created by segmenting at position B. For HRP, the electrophilic Michael receptor is at position B, so the data is segmented at position C. To better understand the model, a third setting is introduced to test its adaptability under low-resource conditions. Since most signals reside in molecules with known pharmacophores for their respective targets, the model's performance is investigated as the amount of data provided to the model varies. These experiments provide a good way to compare different representation modalities, as factorization methods are expected to learn faster under resource-constrained conditions.

[0223] index Several well-motivated metrics were used to evaluate the model's performance without requiring additional data (i.e., on the DNA of the DEL molecule). (Data). Observed data is viewed by predicting the most likely count distribution, and performance is measured by the model loss, i.e., the negative log-likelihood of the test set. This is a typical metric for measuring the overall fitness of a probabilistic model. However, there are potential pitfalls in using likelihood metrics, as likelihood does not indicate the usefulness of the learned representation. Since there is interest in the quality of the latent variables being learned, metrics have been developed to capture their ability to capture useful signals in the learned data. The latent variables in the model are used as the mean of a zero-inflated Poisson distribution. The expected mean of the predicted distribution is used as the computational enrichment of the model, i.e. ,in It is a zero probability prediction. This is the predicted potential score of the molecule. The model predicts the count distribution for both control and target experiments; however, the former is primarily used to calibrate the molecule's affinity for the protein target.

[0224] The performance of the factorization model was evaluated at the synthon polymerization level, as both of our datasets, CA-IX and HRP, contain known pharmacophores. A new metric was further developed to assess the quality of the factorization model's predictions by its ability to separate molecules into different classes. CA-IX has three distinct groups. The groups, in order of protein activity, are para-substituted sulfonamides, meta-substituted sulfonamides, and other molecules. HRP has four distinct groups. The proteins, in order of activity, are three different Michael receptor electrophiles and other molecules.

[0225] To evaluate the model, a multi-class one-to-one area under the curve (OvO AUC) was constructed for the precision-recall (PR) curve to assess the model's ability to distinguish different molecular classes. Let... For use As a positive class and PR-AUC is calculated as a negative class. Since the expected ordering of these molecular classes is known (i.e., The AUC of each pair was calculated, and then an unweighted average of all such pairs was calculated. Because the data is heavily biased towards molecules that do not show significant activity toward the protein target, each molecular category was assigned equal weight.

[0226] These AUC calculations are noted in equations (4) and (5): (4) (5) Example 3: The published model achieves improved performance relative to the baseline and enriches important pharmacophores. Figure 10A-10D The predicted mean marginal enrichment of control and target counts for CA-IX and HRP is shown. Here, the graph shows the mean marginal enrichment of synthons predicted by the model on the test set. For both protein targets, the model correctly enriched the important synthons, namely the benzene-sulfonamide of CA and the Michael receptor electrophile of HRP. Furthermore, the model correctly predicted the ordering of these different groups. Of particular note is that the model enriched synthon #39 in the control experiment for CA (and to a lesser extent, synthon #97), but this synthon was not significantly enriched in the target, which was the expected result. This indicates that the model correctly distinguished synthons that might have high noise (i.e., off-target binding).

[0227] The depth probabilistic method was compared with several baselines that compute enrichment solely based on counts. Poisson enrichment computed the maximum target and control counts. Likelihood Poisson distribution, then calculate The target at the lower limit of the 95% confidence interval (CI) and the control at... The ratio of the 95% CI upper limit.

[0228] - Difference enrichment: Score =

[0229] -Ratio enrichment: Fraction = [(

[0230] - Poisson enrichment: score = CI lower95 [Poisson(Target)] / CI upper95 [Poisson(Control)] Since these baselines are not training models but rather explicit functions of the count data, it is impossible to compare these metrics with factorized models in terms of predictive likelihood. However, all methods provide an ordering of the test molecules from which the aforementioned multi-class PR-AUC can be computed. Therefore, as shown in Table 1, the performance of the models and baselines was compared on random and synthetic segmentation of the two targets. In terms of likelihood, the disclosed factorized models that simultaneously incorporate loading and repetition factors outperform all ablation experiments. Furthermore, the negative log-likelihood (NLL) scores for synthetic segmentation are generally higher, demonstrating that they are more difficult to model. Interestingly, loading factors are more useful for the target data of CA-IX, while repetition factors are more useful for HRP. This perhaps highlights the variance of the data even in experiments conducted under the same conditions.

[0231] Comparing the results of the enriched baseline with those of the factorized model, the variant of the factorized model outperformed the baseline in terms of multi-class PRC-AUC. Baseline metrics were not included in the loaded data, but even the basic factorized metrics... The model also outperforms these baselines in most cases. Since the baseline models have predictive access to real-world data, this indicates that the factorized model is capturing important aspects of the chemical data. Further noteworthy is that the model achieves the best multi-class PRC-AUC in the case of synthon splitting (a more challenging learning scenario). This suggests that incorporating the correct variation factors is crucial for generalization in challenging settings.

[0232] Table 1: Comparison of enrichment metrics between different ablation experiments and baseline using the factorized model. Metrics were averaged across five different splits on the test set, either random or based on synthetic sub-segments. Differences, ratios, and Poisson enrichments used actual counts, while the model made predictions on the test set.

[0233]

[0234] Example 4: The published model exhibits similar or better performance to the whole-molecule model even under low-resource conditions. One of the main advantages of using factorized models is that it avoids building a complex enumeration engine for DEL. However, while this is beneficial, factorized models were evaluated to demonstrate that their performance is competitive, even superior, to models using the full molecular representation. To this end, an in-depth investigation was conducted by training two versions of the model under different data constraints. (See references below.) Figure 11A-11D This paper compares the performance of models using molecular factorization representations with those using full molecular representations. Each model was trained using a different percentage of retained data. Here, model performance is compared as a function of the amount of data provided during training. For both CA-IX and HRP, the multi-class PR-AUC of the factorized model outperforms the whole-molecule model at every point. Furthermore, the test likelihoods of both models as a function of the amount of data provided are very comparable. These results support the use of factorized representations as a methodology because it incorporates the correct inductive biases into the model, serving as a more effective learning medium.

[0235] Example 5: The disclosed model provides insights into the data. Factorization models further provide interpretable insights into the data. Because they use a zero-inflation distribution as the output distribution, this zero probability can be intuitively used as a measure of data noise. Figure 12A and 12B It has been demonstrated that the zero probability of prediction is a good measure of the prediction noise of CA-IX and HRP.

[0236] exist Figure 12A and 12B In, the potential score of learning It is plotted as a function of the predicted zero probability. For HRP, all molecules with known active substructures have high prediction scores and low zero probabilities—strong signals and low noise. This can also be... Figure 11A-11D As seen in the diagram, the model trained on a small subset of the HRP data quickly reaches its optimal performance. However, the CA-IX plot reveals many molecules with high learning scores but also high zero probabilities—this region of distribution may contain more noise. Compared to the HRP data, the predicted scores for benzene-sulfonamide molecules in CA-IX exhibit some uncertainty, as evidenced by their predicted zero probabilities.

[0237] Furthermore, the use of attention mechanisms provides the model with good interpretability and insight. This can be valuable for the purpose of synthesizing compounds, for example, including or excluding certain synthons during synthetic activities. Figure 13A and 13B The attention distributions of the models on the CA-IX and HRP datasets are depicted respectively. For CA-IX, the attention probabilities are mainly concentrated in... On the bisynthetoid, and for HRP, the attention probabilities are mainly distributed in On single synthons. Since the sulfonamide is at the C position of CA-IX and the electrophile at the B position of HRP, these results suggest that both models choose to place the highest weights on synthon embeddings that contain the correct synthon positions. Interestingly, the CA-IX model chooses dual synthons. This indicates enrichment in certain bisynthetics, not just monosynthetics at the C position. This establishes the value of the published model, which considers higher-order synthon representations beyond monosynthetics.

[0238] Example 6: Generalization of the published model to the kinase inhibitor DNA-encoded library dataset (KinDEL) In a separate set of experiments, the disclosed method was applied to two protein kinase targets (disc domain receptor tyrosine kinase 1 (DDR1) and mitogen-activated protein kinase 14 (MAPK14)).

[0239] Training settings In this example, the kinase inhibitor DNA-encoded library (KinDEL), a library containing 100 million small molecules, was tested against two kinase targets, MAPK14 and DDR1. Various benchmark tasks were developed and implemented to demonstrate the efficacy of using DEL data to gain therapeutic insights. Furthermore, the validation of these computational methods was determined using biophysical assay data.

[0240] The dataset used in this example consists of roughly three main parts: 1. The synthesis of DEL 2. Selecting experiments, and 3. Biophysical validation data.

[0241] Typically, when using DEL to select experiments, at least one blank control is run, and a blank control is also run in this example.

[0242] For step 1 (DEL synthesis), DEL was designed as a trisynton library, comprising 382 synthons from the first step, 192 synthons from the second step, and 1152 synthons from the terminal or capping step (totaling approximately 85 million molecules). The first two steps were accomplished either by acylation with N-protected amino acids followed by deprotection, or by immobilizing DNA on a solid support followed by a series of chemical transformations for acylation. In the final step, the downstream amino group reacts with a monofunctional acid or aldehyde.

[0243] For step 2 (DEL selection), selection experiments were then performed using the synthesized library. Biotinylated proteins DDR1 and MAPK1 were immobilized on the Phynexus tip. The library was bound to the immobilized proteins, and the mixture was washed multiple times to progressively remove any weak bindings. Afterward, the bindings were eluted with hot water, followed by amplification and sequencing using the Novaseq S4 platform.

[0244] For step 3 (biophysical assay validation), to complement the DEL data which may trade volume for mass, a small amount of molecular biophysical data was also collected—both on and outside DNA. Fluorescence polarization (FP) was used on DNA to measure binding events in solution using polarized light. Surface plasmon resonance (SPR) was used on the outside DNA, also employing light to measure molecular interactions.

[0245] The KinDEL dataset contains approximately 85 million molecules with unique sequence counts of 3 distinct repeats for each experimental condition. Figure 14 The distribution of chemical properties in the KinDEL dataset is shown. These selected properties are commonly used to assess the drug-likeness of molecules. Furthermore, Figure 15 This is a 3D cube visualization of the dataset, where each axis corresponds to a different cycle in the DEL. Using Poisson enrichment, the points on the graph represent the most enriched molecules. Furthermore, Figure 15 The linear pattern in the data can represent enriched bisynthetics (e.g., a combination of two synthons that bind to protein targets such as MAPK14 and DDR1).

[0246] The KinDEL dataset was used to build predictive models of binding affinity. To this end, various benchmark models were investigated and their performance in modeling binding affinity was compared.

[0247] Benchmarking and Model Performance Evaluation Benchmarking involved two biological targets: MAPK14 and DDR1. For each target, a set of retention test compounds were selected from the DEL library and resynthesized both on and outside the DNA to create an in-library retention test set. Additional compounds were added from outside the library to create an extended retention test set. The binding affinity (Kd or K) of all molecules in the retention set was measured in a biophysical assay. D The performance of the retained set model (DEL-Compose) reported in the benchmark tests is the model prediction versus experimental K. D The Spearman correlation coefficient between them. Furthermore, model performance on the internal test set is reported as the value of the loss function: MSE for all models except DEL-Compose, and negative log-likelihood for DEL-Compose.

[0248] The KinDEL dataset uses two splitting strategies to ensure that all retained compounds are placed in the test set and not used for training. The first splitting type is random splitting, where 10% of randomly selected compounds are placed in the validation set and the other 10% in the test set. The second splitting type is bisynthesis splitting, where pairs of blocks B and C are randomly sampled, and all compounds containing this combination are placed in the same subset using the same 80-10-10 ratio, used for the training, validation, and test sets respectively. Each dataset is split five times for each splitting strategy, and the reported model performance is the sum of the five training runs.

[0249] To benchmark the model, commonly used models were compared with DEL data modeling. Two non-machine learning (ML) baselines were computed to evaluate the DEL screening results against the experimental K. DConsistency between data. The first baseline is the Spearman correlation of the sum of molecular sequence counts bound to the target across three replicate experiments. The second baseline is the Poisson enrichment of molecules (Gerry et al., 2019), which also considers the count of molecules bound to the matrix rather than the target.

[0250] In this example, six machine learning (ML) models are compared. Random Forest (RF), XGBoost, k-Nearest Neighbors (kNN), and Deep Neural Network (DNN) use Morgan fingerprints (radius=2, length=2048) as input features and are trained to predict Poisson enrichments. Graph Isomorphic Network (GIN) is a graph neural network that uses molecular graphs as input and predicts Poisson enrichments. DEL-Compose refers to the probabilistic model disclosed in this paper, which uses Morgan fingerprints as input and predicts the parameters of a zero-inflated Poisson distribution fitted to sequence count data. DEL-Compose is further distinguished as a model that runs using a fully enumerated molecular structure (DEL-compose). (M) ), and another model that runs using a synthetic substructure (DEL-compose) (S) ).

[0251] The architecture of the neural network model follows the implementation in the original publication. The DNN architecture consists of multiple linear layers, each followed by ReLU activation, batch normalization, and dropout (except for the last layer). All neural networks are trained using the Adam optimizer until convergence, with early stopping performed when the validation loss has not improved for more than 5 epochs.

[0252] result Tables 2 and 3 show the performance of the above models on MAPK14 and DDR1, respectively. The Poisson enrichment baseline, calculated directly from the sequence count data, is compared with the experimental K... DThe consistency between the estimates is interesting. Interestingly, in the case of selecting compounds for extra-DNA synthesis for MAPK14, the enrichment baseline is lower than the predictions of ML models trained using enrichment as the objective. This indicates that the ML model has denoising capabilities, making it suitable for compound selection in DEL screening experiments. The results show that DEL-Compose, which views the data probabilistically, performs well compared to other baseline models estimating output data points. Since DEL data is noisy, capturing uncertainties in the data, such as parameterizing DEL-Compose using a zero-inflated Poisson distribution, is valuable. Bisynthesis segmentation is a more challenging task because the structures are completely removed from the training data, and the model must infer based on the chemical structures. Data shows that for MAPK14, the model generally performs poorly in bisynthesis segmentation, while for DDR1, it performs comparably in this novel data segmentation. Overall, the results demonstrate that DEL-Compose outperforms benchmark algorithms on multiple benchmark metrics in predicting the binding affinity of MAPK14 and DDR1.

[0253] For MAPK14, on the Extended DNA dataset (Extended Preserved Test Set), the Spearman correlation coefficient between the DEL-compose model's predictions and experimental Kd was higher than the other five machine learning models for both random splitting and dual synthon splitting. This dataset contains additional compounds not present in the model training and may represent a more challenging and / or more diverse test dataset. Nevertheless, DEL-compose's generalization ability is highlighted by the higher Spearman correlation coefficient. Furthermore, DEL-compose exhibits superior denoising capabilities compared to the other five models, as evidenced by the higher Spearman correlation coefficient values ​​under the "in-library, out-of-DNA" conditions for both random splitting and synthon splitting.

[0254] For DDR1, random splitting and dual-synthesis splitting in both in-library DNA and out-of-DNA datasets, the DEL-compose model predictions versus experimental K... D The Spearman correlation coefficients between them were higher than those of the other five machine learning models, except that kNN performed similarly to the DEL-compose model in random segmentation.

[0255] In summary, the examples provided herein represent one of many exemplary use cases that offer technical advantages and improvements in predicting the binding affinity of DEL with a variety of biological targets, such as MAPK14, DDR1, CA-IX, and HRP.

[0256] Table 2: Model performance evaluation of MAPK14. The test loss column ("test MSE") contains the loss function values ​​calculated on the internal test set. The performance of on-DNA and off-DNA compounds is compared with experimental K. D The Spearman correlation coefficient between them.

[0257]

[0258] Table 3: Model Performance Evaluation for DDR1. The Test Loss column ("Test MSE") contains the loss function values ​​calculated on the internal test set. The performance of on-DNA and off-DNA compounds is compared with experimental K. D The Spearman correlation coefficient between them.

[0259]

Claims

1. A method for molecular screening of compounds bound to a target, characterized in that: The method includes: Obtain multiple synthons that form the compound; Convert the plurality of synthesizers into a plurality of synthesizer representations; The multiple synthons are represented as molecular embeddings; Using a machine learning model, the molecular embedding is analyzed to generate at least a target enrichment prediction representing a measure of binding between the compound and the target.

2. The method according to claim 1, characterized in that: The method further includes performing probability modeling using at least the target enrichment prediction by applying a probability density function that models the experimental target counts.

3. The method according to claim 2, characterized in that: The probability density function is represented by any one of the following: Poisson distribution, binomial distribution, gamma distribution, binomial-Poisson distribution, or negative binomial distribution.

4. The method according to claim 3, characterized in that: The Poisson distribution is a zero-inflated Poisson distribution.

5. The method according to any one of claims 1-4, characterized in that: The machine learning model is used to analyze the molecular embeddings and also generate covariate predictions.

6. The method according to any one of claims 1-5, characterized in that: The method does not include the step of enumerating the compounds from the plurality of synthons.

7. The method according to any one of claims 1-6, characterized in that: Converting the plurality of synthesizers into a plurality of synthesizer representations includes generating one or more single synthesizer representations from the plurality of synthesizers.

8. The method according to claim 7, characterized in that: Generating one or more single-synthetic representations from the plurality of synthesizers includes analyzing the plurality of synthesizers using a learned representation model, optionally wherein the learned representation model is a multilayer perceptron.

9. The method according to claim 7 or 8, characterized in that: Converting the plurality of synthesizers into a plurality of synthesizer representations also includes generating one or more bisynthetic representations from the one or more single synthesizer representations.

10. The method according to claim 9, characterized in that: Generating one or more dual-synthetic representations from the one or more single-synthetic representations includes analyzing the one or more single-synthetic representations using a learned representation model, optionally wherein the learned representation model is a multilayer perceptron.

11. The method according to claim 9 or 10, characterized in that: Converting the plurality of synthesizers into a plurality of synthesizer representations also includes generating one or more tri-synthesizer representations from the one or more bi-synthesizer representations.

12. The method according to claim 11, characterized in that: Generating one or more tri-synthetic representations from the one or more bisynthetic representations includes analyzing the one or more bisynthetic representations using a learned representation model, optionally wherein the learned representation model is a multilayer perceptron.

13. The method according to any one of claims 1-6, characterized in that: The plurality of synthesizer representations include one or more single synthesizer representations.

14. The method according to any one of claims 1-6, characterized in that: The plurality of synthon representations includes one or more bisynthetic representations.

15. The method according to any one of claims 1-6, characterized in that: The plurality of synthon representations includes one or more ternary synthon representations.

16. The method according to any one of claims 1-6, characterized in that: The plurality of synthesizer representations include one or more tetras synthesizer representations.

17. The method according to any one of claims 1-6, characterized in that: The plurality of synthesizer representations include one or more single synthesizer representations, one or more double synthesizer representations, and one or more triple synthesizer representations.

18. The method according to any one of claims 1-6, characterized in that: The plurality of synthesizer representations include three single synthesizer representations, three double synthesizer representations, and one triple synthesizer representation.

19. The method according to any one of claims 1-18, characterized in that: The machine learning model includes neural networks.

20. The method according to claim 19, characterized in that: The neural network includes a feedforward artificial neural network.

21. The method according to claim 19, characterized in that: The neural network includes a multilayer perceptron (MLP).

22. The method according to any one of claims 1-21, characterized in that: The machine learning model includes one or more parameters learned through supervised training techniques.

23. The method according to any one of claims 1-22, characterized in that: The method further includes using the target enrichment prediction to determine the binding affinity value between the compound and the target.

24. The method according to any one of claims 1-23, characterized in that: The method further includes sorting the compounds based on at least the target enrichment predictions.

25. The method according to any one of claims 1-24, characterized in that: Combining the plurality of synthesizer representations into a molecular embedding includes implementing a multi-head attention mechanism across the plurality of synthesizer representations.

26. The method according to claim 25, characterized in that: Implementing the multi-head attention mechanism involves using one or more learned attention weights represented by the plurality of synthesizers.

27. The method according to claim 26, characterized in that: The method further includes ranking the ability of the plurality of synthesizers to bind to the target using the one or more learned attention weights.

28. The method according to any one of claims 2-27, characterized in that: The covariate predictions are derived from one or more covariates, including either nonspecific binding or noise.

29. The method according to claim 28, characterized in that: Nonspecific binding includes one or more of the following: binding to beads, binding to the matrix, binding to streptavidin in beads, binding to biotin, binding to gel, binding to the surface of DNA-encoded library containers, or binding to tags.

30. The method according to claim 28, characterized in that: The noise includes one or more of the following: loading bias, duplication bias, enrichment in other negative control panning, enrichment in other target panning, confounding, compound synthesis yield, reaction type, initiation tag imbalance, initial loading population, experimental conditions, chemical reaction yield, byproducts and truncated products, library synthesis errors, DNA affinity for the target, sequencing depth, and sequencing noise.

31. The method according to any one of claims 2-27, characterized in that: The covariate predictions are derived from loading noise.

32. The method according to any one of claims 2-28, characterized in that: The covariate predictions are derived from repeating noise.

33. The method according to any one of claims 2-27, characterized in that: The machine learning model is used to analyze the molecular embeddings and also generate a second covariate prediction.

34. The method according to claim 33, characterized in that: The covariate prediction and the second covariate prediction are each independently selected from nonspecific binding or noise.

35. The method according to claim 34, characterized in that: The covariate prediction is derived from loading noise, and the second covariate prediction is derived from repeating noise.

36. The method according to any one of claims 1-30, characterized in that: Converting the plurality of synthesizers into the plurality of synthesizer representations includes applying one or more trained, learned representation models.

37. The method according to any one of claims 1-36, characterized in that: The machine learning model is trained using one or more training compounds with corresponding DNA-encoded library outputs.

38. The method according to claim 37, characterized in that: The corresponding DNA-coding library output for the training compounds includes: The experimental control counts determined by the first panning experiment; and The number of experimental targets determined by the second panning experiment.

39. The method according to claim 38, characterized in that: For one of the training compounds, the machine learning model is trained in the following manner: The machine learning model generates target enrichment predictions and covariate predictions from molecular embeddings, which are generated by combining multiple synthetic representations derived from multiple synthetics that form the training compound. Combine the target enrichment prediction and the covariate prediction to generate a predicted target count; and The loss value is determined based on at least the predicted target count and the experimental target count, according to the loss function.

40. The method according to claim 39, characterized in that: The method trains the machine learning model based on the determined loss value.

41. The method according to claim 40, characterized in that: The method further includes jointly training the machine learning model with one or more learned representation models based on the determined loss value.

42. The method according to claim 39 or 40, characterized in that: The loss value is also determined based on the covariate predictions and the experimental control counts.

43. The method according to any one of claims 39-42, characterized in that: The loss function is any one of the following: negative log-likelihood loss, binary cross-entropy loss, focal loss, arc loss, cosface loss, cosine-based loss, or loss function based on the BEDROC metric.

44. The method according to any one of claims 39-43, characterized in that: Combining the target enrichment prediction and the covariate prediction to generate a predicted target count includes applying a probability density function that models the experimental target count.

45. The method according to claim 44, characterized in that: The probability density function is represented by any one of the following: Poisson distribution, binomial distribution, gamma distribution, binomial-Poisson distribution, or negative binomial distribution.

46. ​​The method according to claim 45, characterized in that: The Poisson distribution is a zero-inflated Poisson distribution.

47. The method according to any one of claims 39-46, characterized in that: The machine learning model is also trained in the following ways: Predicted control counts are generated from the covariates by applying a probability density function that models the experimental control counts.

48. The method according to claim 47, characterized in that: The probability density function used to model the experimental control counts is represented by any one of the following distributions: Poisson, binomial, gamma, binomial-Poisson, or negative binomial.

49. The method according to claim 48, characterized in that: The Poisson distribution is a zero-inflated Poisson distribution.

50. The method according to any one of claims 1-49, characterized in that: The binding measure is any one of binding affinity, DEL count, DEL reading, or DEL index.

51. The method according to any one of claims 1-50, characterized in that: The molecular screening mentioned is virtual molecular screening.

52. The method according to any one of claims 1-51, characterized in that: The compounds were obtained from a virtual compound library.

53. The method according to any one of claims 1-52, characterized in that: The targets include protein targets.

54. The method according to claim 53, characterized in that: The protein targets are human carbonic anhydrase IX protein targets, mitogen-activated protein kinase 14 protein targets, discoid domain receptor tyrosine kinase 1 protein targets, or horseradish peroxidase protein targets.

55. The method according to any one of claims 1-54, characterized in that: The method further includes: Identify common binding motifs in a subset of one or more compounds, wherein the compounds in the subset have a predicted binding metric above a threshold binding value.

56. A method for generating molecular intercalation of a compound, characterized in that: The method includes: Obtain multiple synthons that form the compound; Converting the plurality of synthesizers into a plurality of synthesizer representations, wherein the conversion includes: One or more single-composite representations are generated by analyzing the plurality of composites using a first learned representation model; One or more bisynthetic representations are generated by analyzing the one or more single-synthetic representations using a second learned representation model; One or more tri-synthetic representations are generated by analyzing the one or more bisynthetic representations using a third learned representation model; and The plurality of synthons are represented as combined into a molecular embedding.

57. The method according to claim 56, characterized in that: Combining the plurality of synthesizer representations into a molecular embedding includes implementing a multi-head attention mechanism across the plurality of synthesizer representations.

58. The method according to claim 56 or 57, characterized in that: Converting the plurality of synthesizers into a plurality of synthesizer representations further includes generating one or more N-synthesizer representations, where N is 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19 or 20.

59. The method according to any one of claims 56-58, characterized in that: The first learned representation model includes a multilayer perceptron.

60. The method according to any one of claims 56-59, characterized in that: The second learned representation model includes a multilayer perceptron.

61. The method according to any one of claims 56-60, characterized in that: The third learned representation model includes a multilayer perceptron.

62. The method according to any one of claims 56-61, characterized in that: The plurality of synthesizer representations include one or more single synthesizer representations, one or more double synthesizer representations, one or more triple synthesizer representations, or one or more quad synthesizer representations.

63. The method according to any one of claims 56-61, characterized in that: The plurality of synthesizer representations include one or more single synthesizer representations, one or more double synthesizer representations, and one or more triple synthesizer representations.

64. The method according to any one of claims 56-61, characterized in that: The plurality of synthesizer representations include three single synthesizer representations, three double synthesizer representations, and one triple synthesizer representation.

65. A method for predicting experimental counts in a DNA-coding library, characterized in that: The method includes: To obtain molecular intercalations of compounds, wherein the molecular intercalations are generated by a plurality of synthons of the compounds; Using a machine learning model, the molecular embeddings are analyzed to generate (A) target enrichment predictions representing a measure of binding between the compound and the target, and (B) predictions for one or more covariates; By applying a probability density function that models the experimental target count, the target enrichment prediction and the one or more covariate predictions are combined to generate a predicted target count.

66. The method according to claim 65, characterized in that: The probability density function is represented by any one of the following: Poisson distribution, binomial distribution, gamma distribution, binomial-Poisson distribution, or negative binomial distribution.

67. The method according to claim 66, characterized in that: The Poisson distribution is a zero-inflated Poisson distribution.

68. The method according to any one of claims 65-67, characterized in that: The predictions of the one or more covariates are derived from one or more covariates including either nonspecific binding or noise.

69. The method according to claim 68, characterized in that: Nonspecific binding includes one or more of the following: binding to beads, binding to the matrix, binding to streptavidin in beads, binding to biotin, binding to gel, binding to the surface of the DEL container, or binding to the tag.

70. The method according to claim 68, characterized in that: The noise includes one or more of the following: loading bias, duplication bias, enrichment in other negative control panning, enrichment in other target panning, confounding, compound synthesis yield, reaction type, initiation tag imbalance, initial loading population, experimental conditions, chemical reaction yield, byproducts and truncated products, library synthesis errors, DNA affinity for the target, sequencing depth, and sequencing noise.

71. The method according to claim 65, characterized in that: At least one of the one or more covariate predictions originates from loading noise.

72. The method according to claim 65, characterized in that: At least one of the predictions of the one or more covariates originates from repeating noise.

73. The method according to any one of claims 65-70, characterized in that: The first covariate prediction originates from loading noise, and the second covariate prediction originates from repeating noise.

74. The method according to any one of claims 65-70, characterized in that: The one or more covariate predictions include two, three, four, five, six, seven, eight, nine, ten, eleven, twelve, thirteen, fourteen, fifteen, sixteen, seventeen, eighteen, nineteen, or twenty covariate predictions.

75. A method for predicting experimental counts in a DNA-coding library, characterized in that: The method includes: Obtain target enrichment predictions that represent the degree of binding between compounds and targets; Probabilistic modeling is performed using at least the target enrichment prediction by applying a probability density function to the experimental target counts of the DNA-coding library.

76. The method according to claim 75, characterized in that: The probability modeling includes implementing any one of the following distributions: Poisson, binomial, gamma, binomial-Poisson, or negative binomial.

77. The method according to claim 76, characterized in that: The Poisson distribution is a zero-inflated Poisson distribution.

78. The method according to any one of claims 75-77, characterized in that: The method further includes obtaining covariate predictions, and the covariate predictions are also used when performing the probabilistic modeling.

79. The method according to claim 78, characterized in that: The covariate predictions are derived from one or more covariates, including either nonspecific binding or noise.

80. The method according to claim 80, characterized in that: Nonspecific binding includes one or more of the following: binding to beads, binding to the matrix, binding to streptavidin in beads, binding to biotin, binding to gel, binding to the surface of DNA-encoded library containers, or binding to tags.

81. The method according to claim 79, characterized in that: The noise includes one or more of the following: loading bias, duplication bias, enrichment in other negative control panning, enrichment in other target panning, confounding, compound synthesis yield, reaction type, initiation tag imbalance, initial loading population, experimental conditions, chemical reaction yield, byproducts and truncated products, library synthesis errors, DNA affinity for the target, sequencing depth, and sequencing noise.

82. The method according to any one of claims 78-81, characterized in that: The covariate predictions are derived from loading noise.

83. The method according to any one of claims 78-81, characterized in that: The covariate predictions are derived from repeating noise.

84. The method according to any one of claims 75-83, characterized in that: The target enrichment prediction is generated by a machine learning model trained using one or more training compounds with corresponding DNA-encoded library outputs.

85. The method according to claim 84, characterized in that: The machine learning model includes neural networks.

86. The method according to claim 85, characterized in that: The neural network includes a feedforward artificial neural network.

87. The method according to claim 86, characterized in that: The neural network includes a multilayer perceptron (MLP).

88. The method according to any one of claims 84-87, characterized in that: The machine learning model includes one or more parameters learned through supervised training techniques.

89. The method according to claims 84-88, characterized in that: The machine learning model generates the target enrichment prediction in the following manner: Analyze molecular embeddings to generate at least a target enrichment prediction representing a measure of binding between the compound and the target.

90. The method according to any one of claims 84-89, characterized in that: The corresponding DNA-coding library output for the training compounds includes: The experimental control counts determined by the first panning experiment; and The number of experimental targets determined by the second panning experiment.

91. The method according to any one of claims 84-90, characterized in that: For one of the training compounds, the machine learning model is trained in the following manner: The machine learning model generates target enrichment predictions and covariate predictions from molecular embeddings, which are generated by combining multiple synthetic representations derived from multiple synthetics of the training compound. Combine the target enrichment prediction and the covariate prediction to generate a predicted target count; and The loss value is determined based on at least the predicted target count and the experimental target count, according to the loss function.

92. The method according to claim 91, characterized in that: The method includes training the machine learning model based on the determined loss value.

93. The method according to claim 92, characterized in that: The method further includes jointly training the machine learning model with one or more learned representation models based on the determined loss value.

94. The method according to claim 91 or 92, characterized in that: The loss value is also determined based on the covariate predictions and the experimental control counts.

95. The method according to any one of claims 91-94, characterized in that: The loss function is any one of the following: negative log-likelihood loss, binary cross-entropy loss, focal loss, arc loss, cosface loss, cosine-based loss, or loss function based on the BEDROC metric.

96. The method according to any one of claims 91-95, characterized in that: Combining the target enrichment prediction and the covariate prediction to generate a predicted target count includes applying a probability density function that models the experimental target count.

97. The method according to claim 96, characterized in that: The probability density function is represented by any one of the following: Poisson distribution, binomial distribution, gamma distribution, binomial-Poisson distribution, or negative binomial distribution.

98. The method according to claim 97, characterized in that: The Poisson distribution is a zero-inflated Poisson distribution.

99. The method according to any one of claims 91-98, characterized in that: The machine learning model is also trained in the following ways: Predicted control counts are generated from the covariates by applying a probability density function that models the experimental control counts.

100. The method according to claim 99, characterized in that: The probability density function used to model the experimental control counts is represented by any one of the following distributions: Poisson, binomial, gamma, binomial-Poisson, or negative binomial.

101. The method according to claim 100, characterized in that: The Poisson distribution is a zero-inflated Poisson distribution.

102. The method according to any one of claims 75-101, characterized in that: The binding metric is any one of binding affinity, DNA-coding library count, DNA-coding library reading, or DNA-coding library index.

103. The method according to any one of claims 75-102, characterized in that: The targets include protein targets.

104. The method according to claim 103, characterized in that: The protein targets are human carbonic anhydrase IX protein targets, mitogen-activated protein kinase 14 protein targets, discoid domain receptor tyrosine kinase 1 protein targets, or horseradish peroxidase protein targets.

105. A non-transitory computer-readable medium, characterized in that: The non-transitory computer-readable medium includes instructions that, when executed by a processor, cause the processor to perform the method of any one of claims 1-104.

106. A system, characterized in that: The system includes: processor; and A non-transitory computer-readable medium comprising instructions, when executed by the processor, causing the processor to perform the method of any one of claims 1-104.

Images