Computation-based methods for improved protein purification

Abstract

The present invention provides a method performed by one or more computer devices, the method comprising: accessing a matrix of molecular descriptors representing a set of amino acid sequences corresponding to a protein; and refining a set of hyper-parameters associated with a machine learning model trained to generate a prediction of molecular binding characteristics of the protein. Refining the hyper-parameter set includes iteratively performing a process until a desired accuracy is achieved, the process including: reducing the molecular descriptor matrix by selecting a representative feature vector for each of a plurality of feature vector clusters; determining the most predictive feature vector in the selected representative feature vectors based on the correlation; calculating a cross validation loss based on the most predictive feature vector and a predetermined batch of binding data; and updating the hyper-parameter set based on the cross validation loss. The method further includes generating the prediction of the molecular binding characteristic of the protein.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to U.S. Provisional Application No. 63 / 398,168, filed on August 15, 2022, entitled “COMPUTATIONAL-BASED METHODS FORIMPROVING PROTEIN PURIFICATION,” and the disclosure of that U.S. Provisional Application is hereby incorporated by reference in its entirety. Technical Field

[0003] The present application relates generally to protein purification and, more particularly, to computational-based methods for improving protein purification. Background Art

[0004] Cell cultures utilizing engineered mammalian or bacterial cell lines can be used to produce the target protein of interest by, for example, inserting a recombinant plasmid containing a gene for the target protein. Since the cell line itself is a living organism, the cell line produces other proteins in addition to the target protein, and may require a complex growth medium (including, for example, various sugars, amino acids, and growth factors). Even if not required, it is usually desirable to obtain a high-purity composition of the target protein, especially when the target protein is to be used as a therapeutic agent, such as when the target protein is a therapeutic antibody. Therefore, the target protein produced needs to be purified from these other components in the cell culture, which may involve a series of complex processes, each of which involves many variables, such as chromatographic stationary phases, mobile phases, salt concentrations, pH, and other operating conditions (such as temperature).

[0005] For example, a series of protein purification processes may include: (a) obtaining a cell culture sample containing the target protein; (b) one or more capture steps, such as an affinity capture step using, for example, protein A; (c) one or more conditioning steps; (d) one or more depth filtration steps; (e) one or more ion exchange chromatography steps, such as cation exchange or anion exchange chromatography, or mixed modes thereof, optionally including hydrophobic interaction chromatography; (f) one or more hydrophobic interaction chromatography steps, or mixed modes thereof; (g) a virus filtration step; and (h) one or more ultrafiltration steps. Each of these column chromatography techniques may include various conditions for purifying the target protein. Specifically, the purification technique includes many variables that are critical to the efficient production of a highly pure composition of the target protein, including, for example, chromatographic stationary phase, mobile phase, salt concentration, pH, and other operating conditions (such as temperature) in addition to considerations regarding the target protein itself.

[0006] At present, purification techniques and such operating conditions are determined experimentally in the laboratory. Therefore, it may be very cumbersome and time-consuming to determine how to purify the target protein. In addition, when the target protein is separated and isolated only by the aforementioned experimental purification techniques, useful feedback about which candidate proteins may be difficult to purify may not be determined without difficult experiments. Therefore, this may lead to high costs and low efficiency in the development and manufacturing process of therapeutic antibodies or other similar immunotherapies. Accordingly, it may be useful to provide a technology for optimizing and simplifying the protein purification process to identify target proteins (e.g., antibodies) and to speed up the selection process of therapeutic antibody candidates. Summary of the invention

[0007] Embodiments of the present disclosure relate to one or more computing devices, methods, and non-transitory computer-readable media that can utilize iteratively trained machine learning models to generate predictions of molecular binding properties for one or more target proteins as part of a simplified protein purification process for identifying target proteins (e.g., antibodies) and accelerating the selection process of therapeutic antibody candidates or other immunotherapy candidates by early identification of the most promising therapeutic antibody candidates. This simplified process of identifying target proteins (e.g., antibodies) on a computer can, for example, facilitate and accelerate the downstream development and manufacturing of one or more therapeutic monoclonal antibodies (mAbs), bispecific antibodies (bsAbs), trispecific antibodies (tsAbs), or other similar immunotherapies that can be used to treat various patient diseases. In some embodiments, the machine learning model includes an integrated machine learning model that includes multiple models.

[0008] For example, once trained, a machine learning model (e.g., a "boosting" ensemble learning model) can be used to generate predictions of molecular binding properties of one or more proteins (e.g., predictions of the percentage of protein bound at one or more specific pH values ​​and specific salt concentrations and / or specific salt types and chromatographic resins) by utilizing hyperparameters (e.g., regular parameters, booster parameters, learning task parameters) and learned learnable parameters (e.g., regression model weights, decision variables) optimized during machine learning model training and a selected k-best matrix of eigenvectors representing a molecular descriptor matrix corresponding to a set of amino acid sequences of one or more target proteins.

[0009] Specifically, according to the embodiments disclosed herein, once trained, the machine learning model can utilize the hyperparameters optimized during training (e.g., conventional parameters, booster parameters, learning task parameters) and the learned learnable parameters (e.g., regression model weights, decision variables) to predict the protein binding percentage of one or more target proteins (e.g., predicting the percentage of the protein set bound to the ligand in the solution for a given pH value and salt concentration) based only on the k-best matrix of selected feature vectors of the molecular descriptor matrix representing the amino acid sequence set corresponding to the one or more target proteins as input and one or more pH values ​​and salt concentration sets associated with the binding properties of the one or more target proteins.

[0010] In this way, molecular binding characteristics and elution characteristics of one or more target proteins can be determined by providing computational model-based predictions of percent binding to one or more target proteins without requiring extensive upstream experimentation. That is, desired proteins in one or more target proteins can be identified in silico and distinguished from undesirable proteins in the one or more target proteins, and those desired proteins identified in silico can be further utilized to facilitate and accelerate downstream development and manufacturing of one or more therapeutic mAbs, bsAbs, tsAbs, or other similar immunotherapies that can be used to treat various patient diseases (e.g., by reducing upstream experimental duration and experimental inefficiencies, and providing in silico feedback on which candidate proteins may be difficult to purify and, by extension, ultimately difficult to manufacture).

[0011] For example, during the training of a machine learning model (e.g., a "boosting" ensemble learning model), hyperparameters (e.g., conventional parameters, booster parameters, learning task parameters) and learnable parameters (e.g., regression model weights, decision variables) can be iteratively refined and optimized. Iterations can include 1) reducing a molecular descriptor matrix representing a set of amino acid sequences by clustering similar eigenvectors of the molecular descriptor matrix based on a distance metric. For example, the distance metric can be calculated based on Pearson correlation, mutual information, or maximum information coefficient (MIC) or other distance metrics. Other eigenvectors except the eigenvector closest to the cluster centroid can be discarded to generate a simplified molecular descriptor matrix. Iterations can then include determining the k-best most predictive eigenvectors of the reduced molecular descriptor matrix based on a k-best process and a maximum information coefficient (MIC) to determine the correlation between the eigenvectors of the reduced molecular descriptor matrix and the experimentally determined protein binding percentage and / or the first principal component (PC) value for one or more specific pH values ​​and salt concentrations. The iteration may then include calculating the n-number of cross validation losses based on the k-best most predictive feature vectors and the experimentally determined protein binding percentage and / or the first PC value. Finally, the iteration may include updating the hyperparameters (e.g., conventional parameters, booster parameters, learning task parameters) based on the n-number of cross validation losses.

[0012] In certain embodiments, reducing the molecular descriptor matrix (which may include a large set of amino acid sequence-based descriptors) through the aforementioned feature dimension reduction and feature selection techniques can ensure that the regression model successfully converges to an accurately trained regression model rather than overfitting due to redundant or noisy descriptors. Further, it should be understood that in some embodiments, in addition to using MIC as part of determining the correlation between the eigenvectors of the reduced molecular descriptor matrix and the experimentally determined protein binding percentage and / or the first PC value, distance correlation, mutual information, or other similar nonlinear correlation measures or linear correlation measures (e.g., Pearson correlation) can also be used.

[0013] In certain embodiments, one or more computing devices, methods, and non-transitory computer-readable media may access a molecular descriptor matrix representing a set of amino acid sequences corresponding to one or more proteins. In some embodiments, the molecular descriptor matrix may be generated by a first machine learning model (e.g., a matrix generating machine learning model) that is different from a machine learning model (e.g., an ensemble learning model). In some embodiments, the first machine learning model is trained to generate a molecular descriptor matrix based on the amino acid sequence set. For example, in some embodiments, the first machine learning model may include a neural network that is trained to generate an M x N descriptor matrix representing the amino acid sequence set, where N includes the number of amino acid sequence sets and M includes the number of nodes in the output layer of the neural network.

[0014] In certain embodiments, one or more computing devices may then refine a set of hyperparameters associated with a machine learning model that is trained to generate predictions of molecular binding properties of one or more proteins. In some embodiments, the machine learning model may include one or more of a gradient boosting model, an adaptive boosting (AdaBoost) model, an extreme gradient boosting (XGBoost) model, a lightweight gradient boosting machine (LightGBM) model, or a classification boosting (CatBoost) model.

[0015] In some embodiments, the prediction of the molecular binding properties of the one or more proteins can be generated by a computational model-based column process. In some embodiments, the computational model-based chromatography process can include one or more of a computational model-based affinity chromatography process, an ion exchange chromatography (IEC) process, a hydrophobic interaction chromatography (HIC) process, or a mixed mode chromatography (MMC) process.

[0016] The methods provided herein can be used to predict the molecular binding properties of one or more proteins in any chromatography setting (including any series of chromatographic techniques or combinations of specific operating conditions). In general, chromatographic techniques involve stationary phases and mobile phases. The stationary phase may include a portion designed to interact with the target protein (such as in a binding and elution mode chromatography) or not interact with the target protein (such as in a flow-through chromatography). The mobile phase used in the chromatographic technique (such as a loading, washing or elution mobile phase) may have many variables, including the concentration of one or more salts, pH and solvent gradients. In addition, the chromatographic technique may be carried out under various conditions (such as at an elevated temperature).

[0017] In some embodiments, the chromatographic process based on the computational model comprises an affinity chromatography process. In some embodiments, the affinity chromatography process may include an affinity ligand, such as according to any one of protein A chromatography, protein G chromatography, protein A / G chromatography, protein L chromatography, and κ chromatography. In some embodiments, the affinity chromatography process may include an elution mobile phase, such as a mobile phase with a set pH.

[0018] In some embodiments, the chromatographic process based on the computational model may include an ion exchange chromatography process. Ion exchange chromatography allows separation based on electrostatic interactions (anions and cations) between the ligands of the ion exchange stationary phase and the components of the sample (e.g., target proteins or non-target proteins). In some embodiments, the ion exchange chromatography process processes a cation exchange (CEX) stationary phase. In some embodiments, the ion exchange chromatography may include a strong CEX stationary phase. In some embodiments, the ion exchange chromatography may include a weak CEX stationary phase. In some embodiments, the ion exchange chromatography resin may be functionalized with a ligand containing anionic functional groups (such as carboxyl groups or sulfonate groups).

[0019] In some embodiments, the ion exchange chromatography stationary phase may include an anion exchange (AEX) stationary phase. In some embodiments, the ion exchange chromatography may include a strong AEX stationary phase. In some embodiments, the ion exchange chromatography may include a weak AEX stationary phase. In some embodiments, the ion exchange chromatography resin may be functionalized with a ligand containing a cationic functional group (such as a quaternary amine). In some embodiments, the ion exchange chromatography may include a multimodal ion exchange (MMIEX) stationary phase. The MMIEX chromatography stationary phase may include both cation exchange and anion exchange components and / or features. In some embodiments, the MMIEX stationary phase may include a multimodal anion / cation exchange (MM-AEX / CEX) stationary phase.

[0020] In some embodiments, the ion exchange chromatography may include a ceramic hydroxyapatite chromatography stationary phase. In some embodiments, the ion exchange chromatography stationary phase may be selected from the group consisting of: sulfopropyl (SP) Fast Flow (SPSFF), Quaternary Ammonium (Q) Fast Flow(QSFF), SP XL(SPXL), Streamline TM SPXL, ABx TM (MM-AEX / CEX medium), Poros TM XS, Poros TM 50HS, diethylaminoethyl (DEAE), dimethylaminoethyl (DMAE), trimethylaminoethyl (TMAE), quaternary aminoethyl (QAE), mercaptoethylpyridine (MEP) - Hypercel TM 、HiPrep TM Q XL, Q XL and HiPrep TM SP XL. In some embodiments, the ion exchange chromatography method may include an elution step mobile phase that includes an increased salt concentration, such as an increase relative to a binding or wash mobile phase.

[0021] In certain embodiments, the chromatographic process based on the computational model may include a mixed mode chromatographic process. The mixed mode chromatographic process may include combining the stationary phase based on charge (i.e., ion exchange chromatography characteristics) and based on hydrophobic elements. In certain embodiments, the mixed mode chromatographic process may include combining and eluting operation modes. In certain embodiments, the mixed mode chromatographic process may include flowing through the operation mode. In certain embodiments, the mixed mode chromatographic process may include selecting the stationary phase of the group consisting of CaptoMMC and Capto Adhere.

[0022] In some embodiments, the chromatographic process based on the computational model may include a hydrophobic interaction chromatography (HIC) process. The hydrophobic interaction chromatography process may include a hydrophobic stationary phase. In some embodiments, a mixed mode chromatography process may include a binding and elution mode of operation. In some embodiments, the hydrophobic interaction chromatography process may include a flow-through mode of operation. In some embodiments, the hydrophobic interaction chromatography process may include a stationary phase comprising a matrix, such as an inert matrix, for example, cross-linked agarose, Sepharose or a resin matrix. In some embodiments, at least a portion of the substrate of the hydrophobic interaction chromatography stationary phase may include a surface modification comprising a hydrophobic ligand.

[0023] In some embodiments, the hydrophobic interaction chromatography ligand is a ligand comprising between about 1 and 18 carbons. In some embodiments, the hydrophobic interaction chromatography ligand may comprise 1 or more carbons, such as any of 2 or more carbons, 3 or more carbons, 4 or more carbons, 5 or more carbons, 6 or more carbons, 7 or more carbons, 8 or more carbons, 9 or more carbons, 10 or more carbons, 11 or more carbons, 12 or more carbons, 13 or more carbons, 14 or more carbons, 15 or more carbons, 16 or more carbons, 17 or more carbons, or 18 or more carbons. In some embodiments, the hydrophobic interaction chromatography ligand may include any of 1 carbon, 2 carbons, 3 carbons, 4 carbons, 5 carbons, 6 carbons, 7 carbons, 8 carbons, 9 carbons, 10 carbons, 11 carbons, 12 carbons, 13 carbons, 14 carbons, 15 carbons, 16 carbons, 17 carbons, or 18 carbons. In some embodiments, the hydrophobic ligand is selected from the group consisting of an ether group, a methyl group, an ethyl group, a propyl group, an isopropyl group, a butyl group, a tert-butyl group, a hexyl group, an octyl group, a phenyl group, and a polypropylene glycol group.

[0024] In some embodiments, the HIC medium is a hydrophobic charge induction chromatography medium. In some embodiments, the hydrophobic interaction chromatography process can include a mobile phase that includes high salt conditions. For example, high salt conditions can be used to reduce the solvation of the target, thereby exposing hydrophobic regions, which can then interact with the hydrophobic interaction chromatography stationary phase. In some embodiments, the hydrophobic interaction chromatography process can include a mobile phase that includes low salt conditions, for example, without salt or without added salt. In some embodiments, the hydrophobic interaction chromatography stationary phase is selected from the group consisting of: Bakerbond WP HI-Propyl TM , Phenyl Fast Flow (Phenyl-SFF), Phenyl Fast Flow Hi-sub(Phenyl-SFF HS)、 Hexyl-650、Poros TM Benzyl Ultra and In some embodiments, Hexyl-650 is Hexyl-650M. In some embodiments, Hexyl-650 is Hexyl-650C. In some embodiments, Hexyl-650 is Hexyl-650S.

[0025] In certain embodiments, the prediction of the molecular binding properties of the one or more proteins may include the identification of the target protein in the one or more proteins. In another embodiment, the prediction of the molecular binding properties of the one or more proteins may be modeled using a quantitative structure-property relationship (QSPR) or a quantitative structure-activity relationship (QSAR) to the one or more proteins. In another embodiment, the prediction of the molecular binding properties of the one or more proteins may include the prediction of the molecular binding properties of each amino acid sequence in the amino acid sequence set corresponding to the one or more proteins. In another embodiment, the prediction of the molecular binding properties of each amino acid sequence may include separating the desired amino acid molecules from the undesired amino acid molecules based on the computational model. In one embodiment, a machine learning model (e.g., an integrated learning model) may be further trained to generate a prediction of the molecular elution properties of one or more proteins. In another embodiment, the machine learning model may be further trained to generate a prediction of the flow-through properties of one or more proteins.

[0026] In certain embodiments, one or more computing devices may iteratively refine the hyperparameter set by executing a process until the desired accuracy is reached. For example, in certain embodiments, one or more computing devices may reduce the molecular descriptor matrix by first selecting a representative feature vector for each of a plurality of feature vector clusters to perform the process. In one embodiment, each of the feature vector clusters includes similar feature vectors. For example, in some embodiments, reducing the molecular descriptor matrix may include clustering using a correlation distance metric (e.g., calculated based on the Pearson correlation of the feature vectors of the molecular descriptor matrix) to generate a plurality of feature vector clusters. In some cases, the clustering of the descriptor set may be based on the correlation distance between the descriptors, and the correlation distance may be calculated according to the Pearson correlation (e.g., 1–abs (Pearson correlation)). In one embodiment, a representative feature vector selected for each of the plurality of feature vector clusters may include a centroid feature vector for each of the plurality of feature vector clusters, for representing two or more of the similar feature vectors.

[0027] In some embodiments, one or more computing devices may then perform the process in the following manner: determine one or more most predictive feature vectors in the selected representative feature vectors for each of the plurality of feature vector clusters based on the correlation between the selected representative feature vectors and the predetermined batch binding data associated with the one or more proteins. For example, in some embodiments, determining the one or more representative feature vectors in the selected representative feature vectors for each of the plurality of feature vector clusters may include: selecting a k-best matrix of feature vectors in the selected representative feature vectors for each of the plurality of feature vector clusters. In some embodiments, the k-best matrix of feature vectors in the selected representative feature vectors is determined based on a predetermined k-best process. In some embodiments, the correlation between the selected representative feature vector and the predetermined batch binding data is determined based on a Pearson correlation, mutual information, maximum information coefficient (MIC), or other metric between the selected representative feature vector and the predetermined batch binding data. In other embodiments, in addition to using MIC as part of determining the correlation, distance correlation, mutual information, or other similar nonlinear correlation metrics and / or linear correlation metrics may also be used.

[0028] In some embodiments, one or more computing devices may then perform the process by calculating one or more cross-validation losses based at least in part on the one or more most predictive feature vectors and the predetermined batch of combined data. For example, in some embodiments, calculating the one or more cross-validation losses may further include evaluating a cross-validation loss function based on the one or more most predictive feature vectors, the predetermined batch of combined data, a hyperparameter set, and a learnable parameter set associated with the machine learning model; and further minimizing the cross-validation loss function by varying the learnable parameter set while keeping the one or more most predictive feature vectors, the predetermined batch of combined data, and the hyperparameter set constant.

[0029] For example, in some embodiments, minimizing the cross validation loss function may include optimizing a hyperparameter set. For example, in one embodiment, the hyperparameter set may include one or more of a conventional parameter set, a booster parameter set, or a learning task parameter set. In some embodiments, minimizing the cross validation loss function includes minimizing the loss between the prediction of the protein binding percentage of the one or more proteins and the experimentally determined protein binding percentage for the one or more proteins. In one embodiment, the predetermined batch binding data may include experimentally determined protein binding percentages for one or more pH values ​​and salt concentrations associated with the molecular binding properties of the one or more proteins. In one embodiment, the learnable parameter set includes one or more weights or decision variables determined by the machine learning model at least in part based on the one or more most predictive feature vectors and the predetermined batch binding data.

[0030] In some embodiments, calculating one or more cross validation losses may include calculating n number of cross validation losses, where n includes an integer from 1 to n. In some embodiments, calculating one or more cross validation losses may include determining n number of separate training test splits based on the one or more most predictive feature vectors and a predetermined batch binding data, and where n includes an integer from 1 to n. In some embodiments, calculating one or more cross validation losses may include calculating n number of cross validation losses and generating a prediction of a molecular binding property of one or more proteins based on an average of the n number of cross validation losses.

[0031] In some embodiments, the one or more computing devices may then perform the process by updating the hyperparameter set based on one or more cross-validation losses. For example, the updated hyperparameter set may include one or more of an updated conventional parameter set, an updated booster parameter set, or an updated learning task parameter set. In some embodiments, after refining the hyperparameter set, the one or more computing devices may output a prediction of the molecular binding properties of the one or more proteins by the machine learning model based at least in part on the updated hyperparameter set.

[0032] In certain embodiments, after refining the hyperparameter set, the one or more computing devices may further access a second molecular descriptor matrix representing a second set of amino acid sequences corresponding to one or more second proteins; reduce the second molecular descriptor matrix by selecting a representative feature vector for each of a second plurality of feature vector clusters of the second molecular descriptor matrix; determine one or more second most predictive feature vectors of the selected representative feature vectors for each feature vector cluster based on a second correlation between the selected representative feature vectors and a predetermined batch of binding data associated with the one or more second proteins. Input the one or more second most predictive feature vectors into a machine learning model that is trained to generate a prediction of the molecular binding properties of the one or more second proteins; and output, by the machine learning model, a prediction of the molecular binding properties of the one or more second proteins based at least in part on the updated hyperparameter set. For example, the second prediction of the molecular binding properties of the one or more second proteins may include a prediction of a second protein binding percentage for the one or more second proteins.

[0033] In certain embodiments, one or more computing devices may further optimize the machine learning model based on a Bayesian model optimization process. In some embodiments, one or more computing devices may then utilize grouped K-fold cross validation to train and evaluate the optimized machine learning model based on the one or more most predictive feature vectors, predetermined batch combined data, hyperparameter sets, and learnable parameter sets. In some embodiments, the grouped K-fold cross validation may be stratified to ensure that the cross-validation training and evaluation segments include different ranges of regression target values. In some embodiments, stratification may be accomplished using labels generated by binning the regression target values ​​into multiple quantiles.

[0034] In certain embodiments, one or more computing devices, methods, and non-transitory computer-readable media may access a molecular descriptor matrix representing a set of amino acid sequences corresponding to one or more proteins; and obtain predictions of molecular binding properties of the one or more proteins by a machine learning model based at least in part on the molecular descriptor matrix, wherein the machine learning model is trained by: accessing a training molecular descriptor matrix representing a training set of amino acid sequences corresponding to one or more empirically evaluated proteins; and iteratively performing a process to refine a set of hyperparameters associated with the machine learning model until a desired accuracy is achieved, the process comprising: reducing the training molecular descriptor matrix by selecting a representative feature vector for each of a plurality of feature vector clusters, wherein each feature vector cluster includes similar feature vectors; determining one or more most predictive feature vectors among the selected representative feature vectors for each feature vector cluster based on correlations between the selected representative feature vectors and predetermined batch binding data associated with the one or more empirically evaluated proteins; and calculating one or more cross-validation losses based at least in part on the one or more most predictive feature vectors and the predetermined batch binding data; and updating the set of hyperparameters based on the one or more cross-validation losses. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] This patent or application file contains at least one drawing executed in color. Copies of this patent or patent application publication with one or more color drawing(s) will be provided by the Office upon request and payment of the necessary fee.

[0036] Figure 1 A diagram showing an experimental example for performing one or more protein purification processes compared to a computational model-based example for performing one or more protein purification processes according to various embodiments.

[0037] Figure 2 A high-level workflow diagram is shown for performing feature generation, feature dimensionality reduction, regression model optimization, and feature selection based on model output according to various embodiments.

[0038] Figure 3A A workflow diagram is shown for optimizing hyperparameters and learnable parameters of a machine learning model for performing one or more computational model-based protein purification processes, according to various embodiments.

[0039] Figure 3B A workflow diagram is shown for optimizing a machine learning model for performing one or more computational model-based protein purification processes according to various embodiments.

[0040] Figure 4A flow chart showing a method for generating predictions of molecular binding properties for one or more target proteins as part of a simplified process for protein purification to identify target proteins, according to various embodiments.

[0041] Figure 5 An exemplary computing system is shown in accordance with various embodiments.

[0042] Figure 6 Shown according to various embodiments is included as Figure 5 Diagram of an exemplary artificial intelligence (AI) architecture that is a portion of an exemplary computing system.

[0043] Figure 7 Another high-level workflow diagram for performing feature generation, feature dimensionality reduction, regression model optimization, and feature selection based on model output according to various embodiments is shown.

[0044] Figure 8 Another workflow diagram for optimizing hyperparameters and learnable parameters of a machine learning model for performing one or more computational model-based protein purification processes is shown in accordance with various embodiments.

[0045] Fig. 9 A process for training a machine learning model to detect molecular binding properties is shown according to various embodiments.

[0046] FIG. 10A to FIG. 10D An exemplary graph illustrating how principal component analysis can be used to predict molecular binding properties is shown in accordance with various embodiments.

[0047] FIG. 11A to FIG. 11F 1 and 1 are exemplary heat maps showing experimental conditions and experimental K, respectively, according to various embodiments. p The values ​​between experimental conditions and modeling K p The relationship between values.

[0048] Fig.12 A flow chart showing a method for generating predictions of molecular binding properties for one or more target proteins as part of another simplified process for protein purification for identifying target proteins, according to various embodiments. DETAILED DESCRIPTION

[0049] Embodiments of the invention relate to one or more computing devices, methods, and non-transitory computer-readable media that can utilize iteratively trained machine learning models to generate predictions of molecular binding properties of one or more target proteins as part of a simplified protein purification process for identifying target proteins (e.g., antibodies) and accelerating the selection process of therapeutic antibody candidates or other immunotherapy candidates by early identification of the most promising therapeutic antibody candidates. This simplified process by identifying target proteins (e.g., antibodies) on a computer can, for example, facilitate and accelerate the downstream development and manufacturing of one or more therapeutic monoclonal antibodies (mAbs), bispecific antibodies (bsAbs), trispecific antibodies (tsAbs), or other similar immunotherapies that can be used to treat various diseases.

[0050] For example, once trained, a machine learning model (e.g., an ensemble learning model or a "boosted" ensemble learning model) can be used to generate predictions of molecular binding properties of one or more proteins (e.g., predictions of the percentage of protein bound at one or more specific pH values ​​and specific salt concentrations and / or specific salt types and chromatography resins) by utilizing hyperparameters (e.g., regular parameters, booster parameters, learning task parameters) and learned learnable parameters (e.g., regression model weights, decision variables) optimized during machine learning model training and a selected k-best matrix of eigenvectors representing a molecular descriptor matrix corresponding to a set of amino acid sequences of one or more target proteins.

[0051] Specifically, according to the disclosed embodiments, once trained, the machine learning model can utilize the hyperparameters optimized during training (e.g., regular parameters, booster parameters, learning task parameters) and the learned learnable parameters (e.g., regression model weights, decision variables) to predict, for one or more target proteins, (i) the percentage of protein binding for a given pH value and salt concentration or a combination of multiple different pH values ​​and salt concentrations (e.g., predicting the percentage in the protein group that binds to a ligand in a solution); (ii) predicting the percentage of protein binding for a set of pH values ​​and salt concentrations (e.g., predicting the percentage in the protein set that binds to a ligand in a solution); and / or (iii) predicting the principal component (PC) representing the set of pH values ​​and salt concentrations.

[0052] In this way, by providing a prediction based on a computational model for binding percentage and / or PC value for one or more target proteins, the molecular binding characteristics and elution characteristics of the one or more target proteins can be determined without a large number of upstream experiments. That is, the desired protein in one or more target proteins can be identified in silico and distinguished from the undesirable protein in the one or more target proteins, and those desired proteins identified in silico can be further utilized to accelerate and promote the downstream development of one or more therapeutic mAbs, bsAbs, tsAbs or other similar immunotherapies that can be used to treat various diseases (e.g., by reducing upstream experimental duration and experimental inefficiency, and providing feedback on which candidate proteins may be difficult to purify and, by extension, difficult to manufacture in the end).

[0053] For example, during training of a machine learning model (e.g., a "boosting" ensemble learning model), hyperparameters (e.g., regular parameters, booster parameters, learning task parameters) and learnable parameters (e.g., regression model weights, decision variables) can be iteratively refined and optimized by: 1) reducing a molecular descriptor matrix representing a set of amino acid sequences by clustering similar feature vectors of the molecular descriptor matrix based on a correlation distance metric (which can be calculated based on Pearson correlation) and discarding feature vectors other than those closest to the cluster centroid; 2) determining the k-best most predictive feature vectors of the reduced molecular descriptor matrix based on a k-best procedure and a correlation coefficient (e.g., maximum information coefficient (MIC)) for determining a correlation between the feature vectors of the reduced molecular descriptor matrix and experimentally determined protein binding percentages for one or more specific pH values ​​and salt concentrations; 3) calculating a cross-validation loss of n number based on the k-best most predictive feature vectors and the experimentally determined protein binding percentages; and 4) updating the hyperparameters (e.g., regular parameters, booster parameters, learning task parameters) based on the n number of cross-validation losses. In one or more examples, clustering of the one or more descriptor sets can be based on correlation distances between descriptors calculated according to a Pearson correlation (eg, 1-abs(Pearson correlation)).

[0054] In certain embodiments, reducing the molecular descriptor matrix (which may include a large set of amino acid sequence-based descriptors) through the aforementioned feature dimensionality reduction and feature selection techniques can ensure that the regression model successfully converges to an accurately trained regression model rather than overfitting due to redundant or noisy descriptors. Further, it should be understood that in some embodiments, in addition to using MIC as part of determining the correlation between the eigenvectors of the reduced molecular descriptor matrix and the experimentally determined protein binding percentage, distance correlation, mutual information or other similar non-linear correlation measures and / or linear correlation measures (e.g., Pearson correlation, f-statistic-based measures) can also be used.

[0055] As used herein, the terms "polypeptide" and "protein" refer interchangeably to polymers of amino acid residues and are not limited to a minimum length. For example, such polymers of amino acid residues may contain natural or non-natural amino acid residues, and include, but are not limited to, peptides, oligopeptides, dimers, trimers, and multimers of amino acid residues. For example, the definition encompasses both full-length proteins and fragments thereof. The terms "polypeptide" and "protein" may also include post-translational modifications of the polypeptide, such as glycosylation, sialylation, acetylation, phosphorylation, and the like.

[0056] Figure 1 A diagram 100 is shown of an experimental instance 102 for performing one or more protein purification processes in accordance with the disclosed embodiments compared to a computational model-based instance 104 for performing one or more protein purification processes. As shown, in one aspect, the experimental duration of the experimental instance 102 for performing one or more protein purification processes may span several weeks. On the other hand, the execution time of the computational model-based instance 104 for performing one or more protein purification processes may be only a few minutes.

[0057] For example, an experimental example 102 for performing one or more protein purification processes may include: receiving an amino acid sequence at box 106; selecting a plasmid at box 108; engineering the protein through a cell line and cell culture at boxes 110 and 112, respectively; performing one or more chromatography processes (e.g., an affinity chromatography process, an ion exchange chromatography (IEX) process, a hydrophobic interaction chromatography (HIC) process, or a mixed mode chromatography (MMC) process) at box 114; and performing high throughput screening (HTS) and calculating a partition coefficient (K) at box 116. p ) to quantify protein binding, all of which are integral to the tedious and time-consuming protein purification process. Molecular evaluation of one or more target proteins can then be performed at box 118.

[0058] In certain embodiments, according to the technology disclosed in the present invention, a computational model-based instance 104 for performing one or more protein purification processes may include: accessing amino acid sequences corresponding to one or more target proteins at box 106; generating molecular descriptors based on the amino acid sequences and reducing the molecular descriptor matrix at box 120; and generating predictions of molecular binding properties of one or more target proteins using a machine learning model (e.g., an ensemble learning model) at box 122, which is used as a component of an optimized and simplified protein purification process for identifying target proteins (e.g., antibodies) and accelerating the selection process of therapeutic antibody candidates or other immunotherapy candidates according to embodiments disclosed in the present invention.

[0059] In fact, as mentioned below Figures 2 to 4 As discussed in further detail, the machine learning model can utilize the hyperparameters optimized during training (e.g., conventional parameters, booster parameters, learning task parameters) and the learned learnable parameters (e.g., regression model weights, decision variables) to predict the protein binding percentage of one or more target proteins (e.g., predicting the percentage of a set of proteins bound to a ligand in solution for a given pH value and salt concentration) based only on the k-best matrix of the selected molecular descriptor matrix generated at block 120 as input and one or more pH values ​​and salt concentration sets associated with the binding properties of the one or more proteins of interest. Molecular assessment of the one or more target proteins can then be performed at block 118 without the need to perform extensive upstream experiments (e.g., compared to experimental example 102 for performing one or more protein purification processes). That is, desired proteins in one or more target proteins can be identified in silico and distinguished from undesirable proteins in the one or more target proteins, and those desired proteins identified in silico can be further utilized to accelerate and promote downstream development of one or more therapeutic mAbs, bsAbs, tsAbs, or other similar immunotherapies that can be used to treat various diseases (e.g., by reducing upstream experimental duration and experimental inefficiency, and providing feedback in silico about which candidate proteins may be difficult to purify and, by extension, ultimately difficult to manufacture). As an example, based at least in part on a molecular descriptor matrix, a machine learning model can be configured to obtain a prediction of molecular binding properties for one or more proteins. Desired proteins can be identified based on molecular binding properties.

[0060] Figure 2A high-level workflow diagram 200 is shown for performing feature generation 202, feature dimensionality reduction 204, feature selection 206 based on model output, and regression model optimization 208 according to the disclosed embodiments. In particular, it should be understood that a high-level example (such as referring to FIG. 200 ) for performing feature generation 202, feature dimensionality reduction 204, feature selection 206 based on model output, and regression model optimization 208 is provided. Figure 2 The discussion) can be found below Figure 3A and Figure 3B This is discussed in more detail and may be performed by combining machine learning (e.g., a matrix generation machine learning model) according to the embodiments disclosed herein with another machine learning model (e.g., an integrated learning model).

[0061] That is to say, as follows Figure 3A and Figure 3B As discussed, feature generation 202 can be performed by the machine learning model 301, feature dimensionality reduction 204 can be performed by feature dimensionality reduction models 307A, 307B of the machine learning models 302A, 302B, feature selection 206 based on model output can be performed by feature selection models 309A, 309B of the machine learning models 302A, 302B, and regression model optimization 208 can be performed by regression models 311A, 311B of the machine learning models 302A, 302B.

[0062] For example, as will be described in more detail below, performing feature generation 202 may include generating, for example, 1024 molecular descriptors (e.g., descriptors based on amino acid sequences). In certain embodiments, performing feature dimensionality reduction 204 may include, for example, clustering and reducing 1024 molecular descriptors (e.g., descriptors based on amino acid sequences) to remove redundant features or other features determined to be extremely similar. In certain embodiments, performing feature selection 206 based on model output may include generating a k-best feature matrix to reduce molecular descriptors to the k-best most predictive features of only those molecular descriptors. As an example, based on a specific model for generating descriptors, the number of molecular descriptors may be 1024. As another example, the number of molecular descriptors may be larger or smaller, such as 2048 descriptors, 320 descriptors, etc. In certain embodiments, performing regression model optimization 208 may include, for example, optimizing hyperparameters and learnable parameters associated with regression models 311A, 311B of machine learning models 302A, 302B. In certain embodiments, as will be discussed in more detail below, in some embodiments, feature dimensionality reduction 204 and model output-based feature selection 206 may be provided to filter a large set of amino acid sequence-based descriptors that may be generated as part of feature generation 202. In this way, reducing the large set of amino acid sequence-based descriptors through feature dimensionality reduction 204 and model output-based feature selection 206 may ensure that the regression model successfully converges to an accurately trained regression model rather than being overfitted due to redundant or noisy descriptors.

[0063] Figure 3A A detailed workflow diagram 300A is shown for optimizing hyperparameters and learnable parameters of a machine learning model 302A (e.g., an ensemble learning model) and utilizing the machine learning model 302A to generate predictions of molecular binding properties of one or more target proteins as part of a simplified protein purification process for identifying target proteins (e.g., antibodies) and accelerating the selection process of therapeutic antibodies or other immunotherapy candidates by early identification of the most promising therapeutic antibody candidates, according to the disclosed embodiments. In certain embodiments, as Figure 3A As shown, the workflow diagram 300A can be jointly performed by a machine learning model 301 (e.g., a matrix generation machine learning model) and a machine learning model 302A (e.g., as shown by the dashed line), wherein the machine learning model utilizes one or more processing devices (e.g., as will be referred to below) Figure 5 and Figure 6The computing device 500 and artificial intelligence architecture 600 discussed herein may be performed by hardware (e.g., a general purpose processor, a graphics processing unit (GPU), an application specific integrated circuit (ASIC), a system on a chip (SoC), a microcontroller, a field programmable gate array (FPGA), a central processing unit (CPU), an application processor (AP), a visual processing unit (VPU), a neural processing unit (NPU), a neural decision processor (NDP), a deep learning processor (DLP), a tensor processing unit (TPU), a neuromorphic processing unit (NPU), and / or any other processing device that may be suitable for processing genomics data, proteomics data, metabolomics data, metagenomics data, transcriptomics data, or other omics data and making one or more decisions based thereon), software (e.g., instructions running / executed on one or more processors), firmware (e.g., microcode), or some combination thereof.

[0064] In some embodiments, the machine learning model 302A may include, for example, any number of separate machine learning models or other prediction models (e.g., feature reduction model 307A, feature selection model 309A, and regression model 311A), which may be jointly trained and executed (e.g., trained and / or executed serially, in parallel, or end-to-end) to make one or more predictions in sequence, such that the output of one or more initial models in the pipeline is used as input to one or more subsequent models in the collection until a final overall prediction (e.g., "boosting") is output. For example, in some embodiments, the machine learning model 302A may include a gradient boosting model, an adaptive boosting (AdaBoost) model, an extreme gradient boosting (XGBoost) model, a lightweight gradient boosting machine (LightGBM) model, or a classification boosting (CatBoost) model.

[0065] In certain embodiments, such as Figure 3A As shown and as will be discussed in more detail below, the machine learning model 301 may perform one or more feature generation and data import tasks 303, and the machine learning model 302A may include a feature dimensionality reduction model 307A, a feature selection model 309A, and a regression model 311A. One or more hyperparameter optimization tasks 314 may be further performed to refine the hyperparameter set associated with the machine learning model 302A.

[0066] Workflow diagram 300A may begin at functional block 304, where machine learning model 301 imports amino acid sequences for a set of one or more P proteins. For example, in certain embodiments, machine learning model 301 may include one or more pre-trained artificial neural networks (ANNs), convolutional neural networks (CNNs), or other neural networks that may be suitable for generating a large set of amino acid sequence-based descriptors, for example, in a supervised, weakly supervised, semi-supervised, or unsupervised manner. According to embodiments disclosed herein, amino acid sequence-based descriptors may be utilized (e.g., rather than structure-based descriptors) because amino acid sequence-based descriptors may be more effective for training machine learning model 302A to generate predictions of molecular binding properties for one or more target proteins (e.g., compared to utilizing structure-based descriptors). In fact, as will be discussed in more detail below, in some embodiments, feature dimensionality reduction models 307A, 307B and feature selection models 309A, 309B may be provided to filter the large set of amino acid sequence-based descriptors that may be output by machine learning model 301. In this way, reducing a large set of amino acid sequence-based descriptors through feature dimensionality reduction models 307A, 307B and feature selection models 309A, 309B can ensure that regression models 311A, 311B successfully converge to accurately trained regression models rather than overfitting due to redundant or noisy descriptors.

[0067] At function block 305, predetermined batch binding data for a set of one or more P proteins may also be imported for use by the machine learning model 302A. For example, in certain embodiments, the predetermined batch binding data may include data for one or more specific pH values ​​and salt concentrations (e.g., sodium chloride (NaCl) concentration, phosphate (PO4 3- ) concentration) and / or salt type (e.g., sodium acetate (CH3COONa) type, sodium phosphate (Na3PO4) type) and the experimentally determined protein binding percentage of the chromatography resin. The workflow diagram 300A can then continue at functional block 306, where the machine learning model 301 generates a molecular descriptor matrix of size M×N. For example, in certain embodiments, based on the amino acid sequence of each protein in the set of one or more P proteins, the machine learning model 301 (e.g., a neural network, a convolutional neural network (CNN), a deep neural network (DNN)) can generate a molecular descriptor matrix of size M×N, where M is the number of descriptors (M=1024), and N is the number of amino acids in a given protein in the set of one or more P proteins.

[0068] In certain embodiments, the workflow diagram 300A may then continue at functional block 308 to generate a weighted average (M) of the descriptors across all amino acids (N) in the molecular descriptor matrix. For example, in certain embodiments, the weighted average (M) of the descriptors across all amino acids (N) in the molecular descriptor matrix may be calculated to obtain a descriptor vector of size M×1 for each protein in the set of one or more P proteins. For example, in some embodiments, the machine learning model 301 may generate one or more M×1 vectors for each protein in the set of one or more P proteins. In certain embodiments, the workflow diagram 300A may then continue at functional block 310 to represent the descriptor vectors for all proteins (P) as a protein descriptor matrix of size M×P.

[0069] In certain embodiments, functional block 312 of workflow diagram 300A may illustrate iterations of a trained machine learning model 302A, wherein a set of hyperparameters (e.g., conventional parameters, booster parameters, learning task parameters) and a set of learnable parameters (e.g., regression model weights, decision variables) are identified during the training of the machine learning model 302A. Specifically, as part of one or more hyperparameter optimization tasks 314 performed with respect to the machine learning model 302A, at functional block 316, hyperparameters are determined that minimize the average score of the n-cycle regression-based model of the machine learning model 302A.

[0070] For example, in one embodiment, a baseline set of hyperparameters (e.g., conventional parameters, booster parameters, learning task parameters) may be selected and then iteratively updated to minimize the average score of the machine learning model 302A based on 10 cycles of regression. In certain embodiments, at function block 318, the machine learning model 302A may be iteratively trained until the desired accuracy is achieved, refining the hyperparameter set by updating the selected hyperparameters (e.g., conventional parameters, booster parameters, learning task parameters) with each successive iteration. The selected hyperparameters may be updated based on one or more cross-validation losses. For example, in one embodiment, when the selected given hyperparameter set minimizes one or more cross-validation losses (e.g., reaches the lowest possible value or error in the range of 0.0 to 1.0), the desired accuracy is achieved. For example, as will be further described below, in some embodiments, minimizing one or more cross-validation losses may include minimizing the loss between the predicted protein binding percentage and the experimentally determined protein binding percentage. Therefore, in some embodiments, when the selected given hyperparameter set minimizes the loss between the predicted protein binding percentage and the experimentally determined protein binding percentage, the desired accuracy of the machine learning model 302A is achieved.

[0071] According to the technology disclosed in the present invention, in certain embodiments, at function block 320, hyperparameters can be optimized by evaluating a cross-validation loss function based on the k-best best feature vectors that are most predictive for predetermined batch binding data, predetermined batch binding data (e.g., experimentally determined protein binding percentages for one or more specific pH values ​​and salt concentrations and / or salt types and chromatography resins), a baseline hyperparameter set (e.g., conventional parameters, booster parameters, learning task parameters), and a set of learnable parameters (e.g., regression model weights, decision variables) associated with and determined by the machine learning model 302A. The machine learning model 302A can then minimize the cross-validation loss function by varying the set of learnable parameters while keeping the k-best most predictive feature vectors, predetermined batch binding data, and the set of hyperparameters constant.

[0072] In some embodiments, as described above, the machine learning model 302A may include a feature dimension reduction model 307A, a feature selection model 309A, and a regression model 311A. The feature dimension reduction task may reduce the molecular descriptor matrix by selecting a representative feature vector for each of a plurality of feature vector clusters. Starting from the feature dimension reduction model 307A, in some embodiments, the workflow diagram 300A may continue at function block 322, where the machine learning model 302A evaluates the similarity of different descriptors by comparing a set of M feature vectors of size 1×P.

[0073] For example, in some embodiments, for a protein descriptor matrix (M×P), the similarity of different descriptors may be evaluated by comparing a set of M feature vectors of size 1×P. In some embodiments, the workflow diagram 300A may then continue at function block 324, where the machine learning model 302A calculates the correlation between the feature vectors (of size 1×P). For example, in some embodiments, the machine learning model 302A may calculate a correlation distance metric between each of the feature vectors (of size 1×P), which may be calculated, for example, using a Pearson correlation. In one or more instances, clustering of the descriptors may be based on correlation distances between the descriptors calculated according to a Pearson correlation (e.g., 1-abs(Pearson correlation)).

[0074] In some embodiments, the workflow diagram 300A may then continue at function block 326, where the machine learning model 302A clusters the feature vectors so that redundant features that capture similar information are grouped together. For example, in some embodiments, using an aggregate clustering process and a calculated distance correlation metric (which may be calculated based on a Pearson correlation), the machine learning model 302A may cluster the feature vectors so that any and all redundant features (similar feature vectors) that include similar information are grouped together. In some embodiments, the workflow diagram 300A may then continue at function block 328, where the machine learning model 302A determines the centroid of each cluster as a representative of the cluster, which is valuable for feature selection. In addition, the selection of the centroid of each cluster may select a set of orthogonal features that can reduce multicollinearity. In some embodiments, the workflow diagram 300A may then continue at function block 330, where the machine learning model 302A iteratively evaluates the number of clusters (C) to determine how many clusters give the machine learning model 302A the best performance.

[0075] In some embodiments, the machine learning model 302A may also include a feature selection model 309A that may determine one or more most predictive feature vectors from the selected representative feature vectors for each feature vector cluster. The workflow diagram 300A may continue at function block 332, where the machine learning model 302A starts with a reduced descriptor matrix (of size C×P) and calculates the correlation between the feature vectors (1×P) in the reduced descriptor matrix (C×P) and the predetermined batch of binding data at function block 334. For example, in some embodiments, the machine learning model 302A may calculate the correlation between the selected representative feature vectors (1×P) in the reduced descriptor matrix (C×P) and the predetermined batch of binding data (associated with one or more proteins) using a non-linear correlation metric (e.g., maximum information coefficient (MIC), distance correlation, mutual information, or other similar non-linear correlation metrics), and / or using a linear correlation metric (e.g., Pearson correlation) to rank which features and / or descriptors capture information suitable for predicting an output.

[0076] In certain embodiments, the workflow diagram 300A may then continue at functional block 336, where the machine learning model 1 302A determines the top K feature vectors (1×P) that are most predictive of the predetermined batch of combined data to generate a k-best feature matrix (K×P). For example, in certain embodiments, using a k-best process, the machine learning model 302A may select the top K feature vectors (1×P) that are most predictive of the predetermined batch of combined data (e.g., as scored according to MIC, distance correlation, mutual information, or other similar nonlinear correlation metrics) to generate a k-best feature matrix (K×P). Specifically, in one embodiment, the k-best feature matrix (K×P) may maintain the top K feature vectors (1×P), where K is an integer value indicating the number of feature vectors maintained. In another embodiment, the k-best feature matrix (K×P) may maintain the top K feature vectors, where K is a percentage value indicating the percentage of feature vectors maintained. In certain embodiments, the workflow diagram 300A may then continue at functional block 338 where the machine learning model 302A iteratively evaluates the K feature vectors to determine how many feature vectors result in optimal performance of the machine learning model 302A.

[0077] In certain embodiments, the machine learning model 302A may also include a regression model 311. For example, the workflow diagram 300A may continue at functional block 340, where the machine learning model 302A starts with the baseline hyperparameters selected and updated as part of the hyperparameter optimization task 314, and the machine learning model 302A may then be cross-validated (e.g., grouped K-fold cross-validation, stratified K-fold cross-validation) using n unique training test splits. The cross-validation may include calculating one or more cross-validation losses based at least in part on one or more most predictive feature vectors and a predetermined batch of combined data.

[0078] For example, starting with the baseline hyperparameters, k-best feature matrix, and predetermined batch of combined data selected and updated as part of the hyperparameter optimization task 314, the machine learning model 302A may be cross-validated using 10 unique training-test splits of the k-best feature matrix and the predetermined batch of combined data (e.g., training data set). In one embodiment, the machine learning model 302A may be cross-validated using 2 or more, 5 or more, 10 or more, or other number of unique training-test splits of the k-best feature matrix and the predetermined batch of combined data (e.g., training data set) to, for example, reduce the likelihood of overfitting or miscalculating the accuracy of the machine learning model 302A due to the training-test splits. In other embodiments, the machine learning model 302A may be cross-validated using any integer number n of unique training-test splits, as long as the integer n is less than or equal to the number of data points corresponding to, for example, the training data set.

[0079] In certain embodiments, the workflow diagram 300A may then continue at functional block 342, where the machine learning model 302A adjusts the weight given to the data of the predetermined batch binding data (e.g., protein binding percentage at various pH values ​​and salt concentrations and / or salt types and chromatography resins) to weight data with higher importance in the transition region. For example, the machine learning model 302A may adjust the weight given to each point in the predetermined batch binding data to the weight of data in the transition region (e.g., partially bound proteins) that are more important than those fully bound proteins or fully unbound proteins. In certain embodiments, in the workflow diagram 344, the machine learning model 302A predicts the protein binding percentage for the protein set P, and optimizes the machine learning model 302A by minimizing the loss between the predicted protein binding percentage and the experimentally determined protein binding percentage. In certain embodiments, the workflow diagram 300A may then continue at functional block 346, where the machine learning model 302 repeats the model optimization n times using a unique training test split and reports the average score.

[0080] Specifically, the regression task of the machine learning model 302A may include receiving a predetermined batch binding data and a k-best feature matrix, and predicting (at function block 346) a protein binding percentage for a set of proteins P based on the predetermined batch binding data and the k-best feature matrix. The machine learning model 302A may then predict the protein binding percentage for a set of proteins P by using the protein binding data for one or more specific pH values ​​and salt concentrations (e.g., sodium chloride (NaCl) concentration, phosphate (PO4 3-) concentration) and / or salt type (e.g., sodium acetate (CH3COONa) type, sodium phosphate (Na3PO4) type) and chromatography resin) and minimize the loss (e.g., the sum of squared errors (SSE)) between the predicted protein binding percentage and the experimentally determined protein binding percentage (at functional block 346). In some embodiments, pH and salt concentration and / or salt type and chromatography resin can be correlated with molecular binding properties of one or more proteins.

[0081] Therefore, if Figure 3A As shown in the workflow diagram 300A of , the machine learning model 302A can be iteratively trained to generate predictions of molecular binding properties of one or more target proteins as part of a simplified protein purification process for identifying target proteins (e.g., antibodies) and accelerating the selection process of therapeutic antibody candidates or other immunotherapy candidates by early identification of the most promising therapeutic antibody candidates. This simplified process of identifying target proteins (e.g., antibodies) on a computer can, for example, facilitate and accelerate the downstream development and manufacturing of one or more therapeutic monoclonal antibodies mAb, bsAb, tsAb, 2+1Ab or other similar immunotherapies that can be used to treat various diseases.

[0082] For example, once trained, the machine learning model 302A (e.g., a "boosted" machine learning model) can be used to generate predictions of molecular binding properties of one or more proteins (e.g., predictions of the percentage of protein bound at one or more specific pH values ​​and specific salt concentrations and / or specific salt types and chromatographic resins) by utilizing the hyperparameters (e.g., regular parameters, booster parameters, learning task parameters) and the learnable parameters (e.g., regression model weights, decision variables) learned during the training of the machine learning model 302A and a selected k-best matrix of eigenvectors representing a molecular descriptor matrix corresponding to a set of amino acid sequences of one or more target proteins.

[0083] Specifically, according to the disclosed embodiments, once trained, the machine learning model 302A can utilize the hyperparameters optimized during training (e.g., conventional parameters, booster parameters, learning task parameters) and the learned learnable parameters (e.g., regression model weights, decision variables) to predict the protein binding percentage of one or more target proteins (e.g., for a given pH value and salt concentration and / or salt type and chromatography resin, predict the percentage of the protein set bound to the ligand in the solution) and / or Log(K) based only on the k-best matrix of selected feature vectors of the molecular descriptor matrix representing the amino acid sequence set corresponding to the one or more target proteins as input and one or more pH values ​​and salt concentration sets associated with the binding properties of the one or more target proteins.p ) values ​​(combined with the logit transformation of the percentages). In some embodiments, as described below with reference to Figure 7 to Figure 1 As described above, instead of predicting the percent binding for a given pH and salt concentration, one can predict Log(K) based on data within the design space for a given resin (a set of data points covering a range of pH / salt concentrations). p ) values ​​(combined with the logit transformation of the percentages).

[0084] In this way, the molecular binding properties and elution characteristics of the one or more target proteins can be determined by providing a prediction based on a computational model for binding percentages for one or more target proteins, without the need for a large amount of upstream experiments. That is, the desired protein in one or more target proteins can be identified on a computer and distinguished from the undesirable protein in the one or more target proteins, and those desired proteins identified on a computer can be further utilized to accelerate and promote the downstream development of one or more therapeutic mAbs, bsAbs, tsAbs or other similar immunotherapies that can be used to treat various diseases (for example, by reducing the upstream experimental duration and experimental inefficiency, and providing feedback on which candidate proteins may be difficult to purify and extend to the final difficulty in manufacturing on the computer). As an example, based at least in part on a molecular descriptor matrix, a machine learning model can be configured to obtain a prediction of the molecular binding properties of one or more proteins. The desired protein can be identified according to the molecular binding properties. Although embodiments of the present invention are discussed herein primarily with reference to a machine learning model 302A that generates predictions of molecular binding properties of one or more target proteins, it should be understood that according to embodiments disclosed herein, the trained machine learning model 302A may also generate predictions of elution properties of one or more proteins or generate predictions of flow-through properties of one or more proteins.

[0085] Figure 3B A detailed workflow diagram 300B is shown according to the disclosed embodiment, which is used to optimize the above-referenced Figure 3A The machine learning model 302A discussed, and using the optimized machine learning model 302B to generate predictions of molecular binding properties for one or more target proteins, as part of a simplified protein purification process, the protein purification process is used to identify target proteins (e.g., antibodies) and accelerate the selection process of therapeutic antibody candidates or other immunotherapy candidates by early identification of the most promising therapeutic antibody candidates. Specifically, as will be further understood below, the workflow diagram 300B can represent the above referenced Figure 3AImprovements of the discussed workflow diagram 300A. For example, as discussed below, the workflow diagram 300B may include performing one or more Bayesian optimization processes (e.g., sequential model optimization (SMBO), expected improvement (EI)) to iteratively optimize and evaluate the machine learning model 302B by, for example, selectively determining the functional blocks of the feature dimensionality reduction model 307B, the feature selection model 309B, and the regression model 311B to be executed; and the order in which the determined functional blocks of the feature dimensionality reduction model 307B, the feature selection model 309B, and the regression model 311B are to be executed.

[0086] In certain embodiments, such as Figure 3B As shown, the workflow diagram 300B may utilize one or more processing devices (for example, as will be referred to below) Figure 5 and Figure 6 The computing device 500 and artificial intelligence architecture 600 discussed herein may be performed by hardware (e.g., a general purpose processor, a graphics processing unit (GPU), an application specific integrated circuit (ASIC), a system on a chip (SoC), a microcontroller, a field programmable gate array (FPGA), a central processing unit (CPU), an application processor (AP), a visual processing unit (VPU), a neural processing unit (NPU), a neural decision processor (NDP), a deep learning processor (DLP), a tensor processing unit (TPU), a neuromorphic processing unit (NPU), and / or any other processing device that may be suitable for processing genomics data, proteomics data, metabolomics data, metagenomics data, transcriptomics data, or other omics data and making one or more decisions based thereon), software (e.g., instructions running / executed on one or more processors), firmware (e.g., microcode), or some combination thereof.

[0087] In certain embodiments, the workflow diagram 300B may begin at function block 348, where amino acid sequences for a set of one or more P proteins are imported. For example, in certain embodiments, at function block 348, one or more distribution systems (K) for experimental amino acid sequences for a set of one or more P proteins and / or molecular amino acid sequences for a set of one or more P proteins may be imported. p ) screening. The workflow diagram 300B may then continue at functional block 350, where the amino acid sequences for the set of one or more P proteins are formatted and a molecular descriptor matrix of size M x N is generated.

[0088] In certain embodiments, M may be the number of descriptors (M=1024) and N may be the number of amino acids in the light chain (LC) and heavy chain (HC) amino acid sequences of a given protein in the set of one or more P proteins. At functional block 350, the workflow diagram 300B may also include generating a weighted average of the descriptors (M) across all amino acids (N) in the molecular descriptor matrix. For example, in certain embodiments, the weighted average (M) of the descriptors across all amino acids (N) in the molecular descriptor matrix may be calculated to obtain a descriptor vector of size M×1 for each protein in the set of one or more P proteins. For example, in some embodiments, the machine learning model 301 (as described above with reference to Figure 3A The method) may generate one or more M×1 vectors for each protein in a set of one or more P proteins.

[0089] In certain embodiments, the workflow diagram 300B may then continue at functional block 352, where the descriptor vector is pre-processed by removing amino acid sequence data containing precipitation at high salt concentrations and weighting the experimental data to prioritize binding transition regions (e.g., -2 <Log[K p ]<+2, or -0.5 <Log[K p ]<+2). In certain embodiments, as previously discussed, a workflow diagram 300B may be provided for optimizing the above referenced Figure 3A The machine learning model 302A discussed, can then utilize the optimized machine learning model 302B to generate predictions of molecular binding properties of one or more target proteins according to embodiments disclosed herein. For example, the workflow diagram 300B can continue at function block 354, selectively determining which function blocks of the feature dimensionality reduction model 307B, the feature selection model 309B, and the regression model 311B are to be executed, and the order in which the function blocks of the feature dimensionality reduction model 307B, the feature selection model 309B, and the regression model 311B are to be executed.

[0090] For example, in certain embodiments, as part of a process of optimizing a machine learning model 302B (e.g., an ensemble learning model), the workflow diagram 300B may perform one or more Bayesian optimization processes (e.g., sequential model optimization (SMBO), expected improvement (EI)) at functional block 354 to optimize and evaluate the machine learning model 302B. For example, in one embodiment, the Bayesian optimization process (e.g., SMBO, EI) may include, for example, one or more probability-based objective functions that may be constructed and utilized to select the most predictive or promising functional blocks of the feature dimensionality reduction model 307B, feature selection model 309B, and regression model 311B to be executed and / or the order in which to execute these functional blocks. These functional blocks of the feature dimensionality reduction model 307B, feature selection model 309B, and regression model 311B are discussed below.

[0091] In some embodiments, based on the function blocks of selecting the feature dimensionality reduction model 307B, the feature selection model 309B, and the regression model 311B to be performed, the workflow diagram 300B at function block 354 may further utilize, for example, nested cross validation including grouped K-fold cross validation to estimate the accuracy of the machine learning model 302B. In this way, the workflow diagram 300B may optimize the machine learning model 302B to compare, for example, to the above referenced Figure 3A The machine learning model 302A discussed generates predictions of molecular binding properties of one or more target proteins more efficiently (e.g., shortening the execution time of the machine learning model 302B and reducing the size of a database suitable for storing the machine learning model 302B).

[0092] In certain embodiments, the workflow diagram 300B may then continue at functional block 356, where the optimized machine learning model 302B is trained and evaluated. For example, in some embodiments, the optimized machine learning model 302B (e.g., as optimized at functional block 354) may be trained and evaluated based on the descriptor vector representing the amino acid sequence for the set of one or more P proteins (e.g., as calculated at functional block 352) and the functional blocks for selecting the feature dimensionality reduction model 307B, the feature selection model 309B, and the regression model 311B to be performed.

[0093] In some embodiments, the workflow diagram 300B at function block 356 may further include combining the optimized set of hyperparameters (e.g., conventional parameters, booster parameters, learning task parameters) and the optimized set of learnable parameters (e.g., regression model weights, decision variables) (e.g., as described above with reference to Figure 3AThe optimized machine learning model 302B is iteratively optimized and discussed in the workflow diagram 300A of the invention, and the optimized machine learning model 302B is used to generate predictions of molecular binding properties of one or more target proteins according to the embodiments disclosed herein. In certain embodiments, the workflow diagram 300B can then end at functional block 358, where the optimized machine learning model 302B, the optimized set of hyperparameters (e.g., conventional parameters, booster parameters, learning task parameters), and the optimized set of learnable parameters (e.g., regression model weights, decision variables) are stored to be used for subsequent predictions of molecular binding properties of one or more target proteins.

[0094] In certain embodiments, for example, during the inference phase (e.g., after the optimized machine learning model 302B is trained and stored with the optimized set of hyperparameters and learnable parameter sets, as described above), Figure 3B As further shown, the feature dimension reduction model 307B of the machine learning model 302B can receive or import a molecular descriptor matrix and scale and normalize one or more descriptor sets of the descriptor matrix. For example, the molecular descriptor matrix can represent a set of amino acid sequences corresponding to a set of P proteins. In certain embodiments, the feature dimension reduction model 307B can then cluster one or more descriptor sets by determining a correlation distance (e.g., 1-abs (Pearson correlation)) between descriptors, and then can store only the descriptors closest to the centroid. For example, in some embodiments, using a calculated correlation distance metric (which can be calculated based on Pearson correlation), the feature dimension reduction model 307B can cluster feature vectors so that any and all redundant features that include similar information (similar feature vectors) are grouped together, and the centroid of each cluster is determined as a representative of the cluster. In certain embodiments, the feature dimension reduction model 307B can then optimize the number of selected descriptors.

[0095] In certain embodiments, the feature selection model 309B may then calculate the nonlinear correlation between the descriptors and output the protein binding percentage using the MIC metric, distance correlation, mutual information, or other similar nonlinear correlation metrics and / or other linear correlation metrics (e.g., Pearson correlation). In one or more other embodiments, the feature selection model 309B may calculate the nonlinear correlation between the descriptors and output the protein binding percentage using the distance correlation, mutual information, or other similar nonlinear correlation metrics. For example, the feature selection model 309B may determine the k-best most predictive feature vectors of the reduced molecular descriptor matrix based on the k-best process and the MIC, for determining the correlation between the feature vectors of the reduced molecular descriptor matrix and the experimentally determined protein percentage bound to one or more specific pH values ​​and salt concentrations and / or salt types and chromatography resins. In other embodiments, in addition to using the MIC as part of determining the correlation, distance correlation, mutual information, or other similar nonlinear correlation metrics may also be used. The feature selection model 309B may then select highly correlated descriptors and optimize the selected descriptors. In certain embodiments, the feature selection model 309B may then select a set of descriptors based on their impact on the overall performance (e.g., processing speed, storage capacity) of the machine learning model 302B. For example, in some embodiments, the feature selection model 309B may iteratively evaluate the K descriptors to determine how many descriptors result in the best performance of the machine learning model 302B. In some embodiments, the feature selection model 309B may select a set of descriptors entirely selectively based on the impact on overall performance.

[0096] For example, in other embodiments, the feature selection model 309B may perform, for example, one or more Boruta feature selection algorithms, one or more SHapley Additive exPlanations (SHAP) feature selection algorithms, or other similar recursive feature elimination algorithms to select K descriptors and optimize the percentage of the number of selected K descriptors. In certain embodiments, the regression model 311B of the machine learning model 302B may then receive pH, salt concentration, and sequence-based descriptors as inputs, and may then output a prediction of the protein binding percentage for the set of proteins P, and optimize the machine learning model 302B by minimizing the loss between the predicted protein binding percentage and the experimentally determined protein binding percentage.

[0097] Therefore, if Figure 3BAs further illustrated in the workflow diagram 300B of , the machine learning model 302B is iteratively trained to generate predictions of molecular binding properties of one or more target proteins as part of a simplified protein purification process for identifying target proteins and accelerating the selection process of therapeutic antibody candidates or other immunotherapy candidates by early identification of the most promising therapeutic antibody candidates. This simplified process of identifying target proteins (e.g., antibodies) on a computer can, for example, facilitate and accelerate the downstream development and manufacturing of one or more therapeutic monoclonal antibodies mAb, bsAb, tsAb or other similar immunotherapies that can be used to treat various diseases.

[0098] For example, once trained, the machine learning model 302B (e.g., a "boosted" machine learning model) can be used to generate predictions of molecular binding properties of one or more proteins (e.g., predictions of the percentage of protein bound at one or more specific pH values ​​and specific salt concentrations and / or specific salt types and chromatographic resins) by utilizing the hyperparameters (e.g., regular parameters, booster parameters, learning task parameters) and the learned learnable parameters (e.g., regression model weights, decision variables) optimized during the training of the machine learning model 302B and a selected k-best matrix of eigenvectors representing a molecular descriptor matrix corresponding to a set of amino acid sequences of one or more target proteins.

[0099] Specifically, according to the disclosed embodiments, once trained, the machine learning model 302B can utilize the hyperparameters optimized during training (e.g., conventional parameters, booster parameters, learning task parameters) and the learned learnable parameters (e.g., regression model weights, decision variables) to predict the protein binding percentage of one or more target proteins (e.g., predicting the percentage of the protein set bound to the ligand in the solution for a given pH value and salt concentration and / or salt type and chromatography resin) based only on the k-best matrix of the selected feature vectors of the molecular descriptor matrix representing the amino acid sequence set corresponding to the one or more target proteins as input and one or more pH values ​​and salt concentration sets associated with the binding properties of the one or more target proteins. In addition, according to the disclosed embodiments, the ensemble learning 302B can be further optimized using one or more Bayesian optimization processes to more efficiently generate predictions of molecular binding properties (e.g., predictions of protein binding percentages at one or more more specific pH values ​​and specific salt concentrations and / or specific salt types and chromatography resins).

[0100] In this way, by providing an optimization prediction based on a computational model for the protein binding percentage of one or more target proteins, the molecular binding properties and elution characteristics of the one or more target proteins can be determined without a large amount of upstream experiments. That is, the desired protein in one or more target proteins can be identified on a computer and distinguished from the undesirable protein in the one or more target proteins, and those desired proteins identified on a computer can be further utilized to accelerate and promote the downstream development of one or more therapeutic mAbs, bsAbs, tsAbs or other similar immunotherapies that can be used to treat various diseases (for example, by reducing the upstream experimental duration and experimental inefficiency, and providing feedback on which candidate proteins may be difficult to purify and extend to the final difficulty in manufacturing on the computer). As an example, based at least in part on a molecular descriptor matrix, a machine learning model can be configured to obtain a prediction of the molecular binding properties of one or more proteins. Desired proteins can be identified according to molecular binding properties. Although embodiments of the present invention are discussed herein primarily with reference to a machine learning model 302B that generates predictions of molecular binding properties of one or more target proteins, it should be understood that according to embodiments disclosed herein, the trained machine learning model 302B may also generate predictions of elution properties of one or more proteins or generate predictions of flow-through properties of one or more proteins.

[0101] Figure 4 A flow chart of a method 400 for generating predictions of molecular binding properties for one or more target proteins as part of a simplified protein purification process for identifying target proteins (e.g., antibodies) and accelerating the selection process of therapeutic antibody candidates or other immunotherapy candidates by early identification of the most promising therapeutic antibody candidates is shown in accordance with the disclosed embodiments. The method 400 may utilize one or more processing devices (e.g., described below with reference to Figure 5 and Figure 6The present invention relates to a computing device and artificial intelligence architecture discussed herein, which may include hardware (e.g., a general-purpose processor, a graphics processing unit (GPU), an application-specific integrated circuit (ASIC), a system on a chip (SoC), a microcontroller, a field-programmable gate array (FPGA), a central processing unit (CPU), an application processor (AP), a visual processing unit (VPU), a neural processing unit (NPU), a neural decision processor (NDP), a deep learning processor (DLP), a tensor processing unit (TPU), a neuromorphic processing unit (NPU), or any other processing device that may be suitable for processing genomics data, proteomics data, metabolomics data, metagenomics data, transcriptomics data, or other omics data and preparing one or more processors), firmware (e.g., microcode), or some combination thereof.

[0102] Method 400 may begin at block 402, where one or more processing devices access a molecular descriptor matrix representing a set of amino acid sequences corresponding to one or more proteins. Method 400 may then continue at block 404, where one or more processing devices refine a set of hyperparameters associated with a machine learning model that is trained to generate predictions of molecular binding properties for one or more proteins. As shown, method 400 may then proceed to an iterative subprocess that optimizes a set of hyperparameters (e.g., by performing the subprocess iteratively until the machine learning model reaches a desired accuracy). Figure 4 4. As shown in dashed lines around a portion of method 400.

[0103] For example, method 400 may continue at frame 406, where one or more processing devices reduce the molecular descriptor matrix by selecting a representative feature vector for each of a plurality of feature vector clusters, wherein each of the feature vector clusters includes similar feature vectors. Method 400 may then continue at frame 408, where one or more processing devices determine one or more most predictive feature vectors in the selected representative feature vector for each of the plurality of feature vector clusters based on the correlation between the selected representative feature vector and the predetermined batch binding data associated with one or more proteins. Method 400 may then continue at frame 410, where one or more processing devices calculate one or more cross-validation losses based at least in part on the one or more most predictive feature vectors and the predetermined batch binding data. Method 400 may then end at frame 412, where one or more processing devices update the hyperparameter set based on one or more cross-validation losses.

[0104] Figure 5One or more computing devices 500 are shown as examples, which can be used to generate predictions of molecular binding properties of one or more target proteins, as part of a simplified protein purification process, which is used to identify target proteins (e.g., antibodies) and to speed up the selection process of therapeutic antibody candidates or other immunotherapy candidates by early identification of promising therapeutic antibody candidates. In certain embodiments, one or more computing devices 500 can perform one or more steps of one or more methods described or shown herein. In certain embodiments, one or more computing devices 500 provide functionality described or shown herein. In certain embodiments, software running on one or more computing devices 500 performs one or more steps of one or more methods described or shown herein, or provides functionality described or shown herein. Certain embodiments include one or more parts of one or more computing devices 500.

[0105] The present disclosure contemplates any suitable number of computing systems 500. The present disclosure contemplates one or more computing devices 500 in any suitable physical form. By way of example and not limitation, one or more computing devices 500 may be an embedded computing system, a system on a chip (SOC), a single board computer system (SBC) (e.g., a computer on a module (COM) or a system on a module (SOM)), a desktop computer system, a laptop or notebook computer system, an interactive kiosk, a mainframe, a computer system network, a mobile phone, a personal digital assistant (PDA), a server, a tablet computer system, an augmented / virtual reality device, or a combination of two or more thereof. Where appropriate, one or more computing devices 500 may be: integrated or distributed; span multiple locations; span multiple machines; span multiple data centers; or reside in a cloud, which may include one or more cloud components in one or more networks.

[0106] Where appropriate, one or more computing devices 500 may perform one or more steps of one or more methods described or illustrated herein without substantial spatial or temporal limitations. By way of example and not limitation, one or more computing devices 500 may perform one or more steps of one or more methods described or illustrated herein in real time or in batch mode. Where appropriate, one or more computing devices 500 may perform one or more steps of one or more methods described or illustrated herein at different times or in different locations.

[0107] In some embodiments, one or more computing devices 500 include a processor 502, a memory 504, a database 506, an input / output (I / O) interface 508, a communication interface 510, and a bus 512. Although the present disclosure describes and illustrates a particular computer system having a particular number of particular components in a particular arrangement, the present disclosure contemplates any suitable computer system having any suitable number of any suitable components in any suitable arrangement. In some embodiments, the processor 502 includes hardware for executing instructions, such as those constituting a computer program. As an example and not limitation, to execute instructions, the processor 502 may: retrieve (or extract) instructions from an internal register, an internal cache, a memory 504, or a database 506; decode and execute these instructions; and then write one or more results to an internal register, an internal cache, a memory 504, or a database 506. In some embodiments, the processor 502 may include one or more internal caches for data, instructions, or addresses. Where appropriate, the present disclosure contemplates a processor 502 including any suitable number of any suitable internal caches. As an example and not limitation, processor 502 may include one or more instruction caches, one or more data caches, and one or more translation lookaside buffers (TLBs). The instructions in the instruction cache may be a copy of the instructions in memory 504 or database 506, and the instruction cache may speed up the retrieval of these instructions by processor 502.

[0108] The data in the data cache may be: a copy of the data in the memory 504 or the database 506 for operation by the instructions executed at the processor 502; the result of the previous instruction executed at the processor 502, for access or writing to the memory 504 or the database 506 by the subsequent instructions executed at the processor 502; or other suitable data. The data cache may accelerate the read or write operation of the processor 502. The TLB may accelerate the virtual address translation of the processor 502. In some embodiments, the processor 502 may include one or more internal registers for data, instructions, or addresses. Where appropriate, the present disclosure contemplates a processor 502 including any suitable number of any suitable internal registers. Where appropriate, the processor 502 may include one or more arithmetic logic units (ALUs); may be a multi-core processor; or may include one or more processors 502. Although the present disclosure describes and illustrates a specific processor, the present disclosure contemplates any suitable processor.

[0109] In some embodiments, the memory 504 includes a main memory for storing instructions for the processor 502 to execute or data for the processor 502 to operate on. As an example and not limitation, one or more computing devices 500 can load instructions from a database 506 or another source (such as, for example, another one or more computing devices 500) into the memory 504. The processor 502 can then load the instructions from the memory 504 into an internal register or an internal cache. In order to execute the instructions, the processor 502 can retrieve the instructions from the internal register or the internal cache and decode them. During or after the execution of the instructions, the processor 502 can write one or more results (which can be intermediate results or final results) to the internal register or the internal cache. The processor 502 can then write one or more of those results to the memory 504.

[0110] In some embodiments, the processor 502 executes only instructions in one or more internal registers, internal caches, or memory 504 (rather than database 506 or elsewhere), and operates only on data in one or more internal registers, internal caches, or memory 504 (rather than database 506 or elsewhere). One or more memory buses (which may each include an address bus and a data bus) may couple the processor 502 to the memory 504. The bus 512 may include one or more memory buses, as described below. In some embodiments, one or more memory management units (MMUs) reside between the processor 502 and the memory 504 and facilitate access to the memory 504 requested by the processor 502. In some embodiments, the memory 504 includes a random access memory (RAM). Where appropriate, the RAM may be a volatile memory. Where appropriate, the RAM may be a dynamic RAM (DRAM) or a static RAM (SRAM). In addition, where appropriate, the RAM may be a single-port or multi-port RAM. The present disclosure contemplates any suitable RAM. Where appropriate, the memory 504 may include one or more memory devices 504. Although this disclosure describes and illustrates particular memory, this disclosure contemplates any suitable memory.

[0111] In some embodiments, database 506 includes a mass storage device for data or instructions. As an example and not limitation, database 506 may include a hard disk drive (HDD), a floppy disk drive, a flash memory, an optical disk, a magneto-optical disk, a tape, or a universal serial bus (USB) drive, or a combination of two or more thereof. Where appropriate, database 506 may include removable or non-removable (or fixed) media. Where appropriate, database 506 may be inside or outside one or more computing devices 500. In some embodiments, database 506 is a non-volatile solid-state memory. In some embodiments, database 506 includes a read-only memory (ROM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically rewritable ROM (EAROM), a flash memory, or a combination of two or more thereof. The present disclosure contemplates a mass database 506 in any suitable physical form. Where appropriate, database 506 may include one or more storage control units that facilitate communication between processor 502 and database 506. Where appropriate, database 506 may include one or more databases 506. Although this disclosure describes and illustrates particular storage devices, this disclosure contemplates any suitable storage devices.

[0112] In some embodiments, the I / O interface 508 includes: hardware, software, or both, which provides one or more interfaces for communicating between one or more computing devices 500 and one or more I / O devices. Where appropriate, one or more computing devices 500 may include one or more of these I / O devices. One or more of these I / O devices can enable communication between a person and one or more computing devices 500. As an example and not limitation, the I / O device may include a keyboard, a keypad, a microphone, a monitor, a mouse, a printer, a scanner, a speaker, a still camera, a stylus, a tablet computer, a touch screen, a trackball, a camera, another suitable I / O device, or a combination of two or more thereof. The I / O device may include one or more sensors. The present disclosure contemplates any suitable I / O device and any suitable I / O interface 508 for them. Where appropriate, the I / O interface 508 may include one or more devices or software drivers so that the processor 502 can drive one or more of these I / O devices. Where appropriate, the I / O interface 508 may include one or more I / O interfaces 508. Although this disclosure describes and illustrates a particular I / O interface, this disclosure contemplates any suitable I / O interface.

[0113] In some embodiments, the communication interface 510 includes: hardware, software, or both, which provides one or more interfaces for communication (such as, for example, packet-based communication) between one or more computing devices 500 and one or more other computing devices 500 or one or more networks. As an example and not limitation, the communication interface 510 may include a network interface controller (NIC) or network adapter for communicating with an Ethernet or other wire-based network, or a wireless NIC (WNIC) or wireless adapter for communicating with a wireless network (such as a WI-FI network). The present disclosure contemplates any suitable network and any suitable communication interface 510 for it.

[0114] By way of example and not limitation, one or more computing devices 500 may communicate with a self-organizing network, a personal area network (PAN), a local area network (LAN), a wide area network (WAN), a metropolitan area network (MAN), one or more parts of the Internet, or a combination of two or more thereof. One or more parts of one or more of these networks may be wired or wireless. As an example, one or more computing devices 500 may communicate with a wireless PAN (WPAN) (such as, for example, a BLUETOOTH WPAN), a WI-FI network, a WI-MAX network, a cellular phone network (such as, for example, a global system for mobile communications (GSM) network), other suitable wireless networks, or a combination of two or more thereof. Where appropriate, one or more computing devices 500 may include any suitable communication interface 510 for any of these networks. Where appropriate, the communication interface 510 may include one or more communication interfaces 510. Although the present disclosure describes and illustrates a specific communication interface, the present disclosure encompasses any suitable communication interface.

[0115] In some embodiments, bus 512 includes hardware, software, or both that couples one or more components of computing device 500 to each other. By way of example and not limitation, bus 512 may include an accelerated graphics port (AGP) or other graphics bus, an enhanced industry standard architecture (EISA) bus, a front side bus (FSB), a hypertransport (HT) interconnect, an industry standard architecture (ISA) bus, an INFINIBAND interconnect, a low pin count (LPC) bus, a memory bus, a micro channel architecture (MCA) bus, a peripheral component interconnect (PCI) bus, a PCI-Express (PCIe) bus, a serial advanced technology attachment (SATA) bus, a video electronics standard association local (VLB) bus, another suitable bus, or a combination of two or more thereof. Where appropriate, bus 512 may include one or more buses 512. Although the present disclosure describes and illustrates a particular bus, the present disclosure contemplates any suitable bus.

[0116] Herein, one or more computer-readable non-transitory storage media may include one or more semiconductor-based or other integrated circuits (ICs) (such as, for example, field programmable gate arrays (FPGAs) or application-specific ICs (ASICs)), hard disk drives (HDDs), hybrid hard disk drives (HHDs), optical disks, optical disk drives (ODDs), magneto-optical disks, magneto-optical disk drives, floppy disks, floppy disk drives (FDDs), magnetic tapes, solid-state drives (SSDs), RAM drives, SECURE DIGITAL cards or drives, any other suitable computer-readable non-transitory storage media, or any suitable combination of two or more thereof. Where appropriate, the computer-readable non-transitory storage medium may be a volatile storage medium, a non-volatile storage medium, or a combination of a volatile storage medium and a non-volatile storage medium.

[0117] Figure 6 An exemplary artificial intelligence (AI) architecture 602 is shown (which may be used as referenced above). Figure 5 600 of the present invention (a portion of one or more computing devices 500 discussed is included), the artificial intelligence architecture can be used to generate predictions of molecular binding properties of one or more target proteins as part of a simplified protein purification process for identifying target proteins (e.g., antibodies) and accelerating the selection process of therapeutic antibody candidates or other immunotherapy candidates by early identification of the most promising therapeutic antibody candidates. In certain embodiments, the AI ​​architecture 602 can be implemented using, for example, one or more processing devices, which can include hardware (e.g., a general purpose processor, a graphics processing unit (GPU), an application specific integrated circuit (ASIC), a system on a chip (SoC), a microcontroller, a field programmable gate array (FPGA), a central processing unit (CPU), an application processor (AP), a visual processing unit (VPU), a neural processing unit (NPU), a neural decision processor (NDP), a deep learning processor (DLP), a tensor processing unit (TPU), a neuromorphic processing unit (NPU), and / or other processing devices that can be suitable for processing various molecular data and making one or more decisions based thereon), software (e.g., instructions running / executing on one or more processing devices), firmware (e.g., microcode), or some combination thereof.

[0118] In certain embodiments, such as Figure 6As depicted, AI architecture 602 may include machine learning (ML) algorithms and functions 604, natural language processing (NLP) algorithms and functions 606, expert systems 608, computer-based vision algorithms and functions 610, speech recognition algorithms and functions 612, planning algorithms and functions 614, and robotics algorithms and functions 616. In certain embodiments, ML algorithms and functions 604 may include any statistically based algorithms that may be suitable for finding patterns in large amounts of data (e.g., “big data” such as genomic data, proteomic data, metabolomic data, metagenomic data, transcriptomic data, or other omics data). For example, in certain embodiments, ML algorithms and functions 604 may include deep learning algorithms 618, supervised learning algorithms 620, and unsupervised learning algorithms 622.

[0119] In some embodiments, the deep learning algorithm 618 may include any artificial neural network (ANN) that can be used to learn deep representations and abstract concepts from large amounts of data. For example, the deep learning algorithm 618 may include ANNs such as perceptrons, multi-layer perceptrons (MLPs), autoencoders (AEs), convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memories (LSTMs), gated recurrent units (GRUs), restricted Boltzmann machines (RBMs), deep belief networks (DBNs), bidirectional recurrent deep neural networks (BRDNNs), generative adversarial networks (GANs) and deep Q-networks, neural autoregressive distribution estimation (NADEs), adversarial networks (ANs), attention models (AMs), spiking neural networks (SNNs), deep reinforcement learning, and the like.

[0120] In some embodiments, the supervised learning algorithm 620 may include any algorithm that can be used to apply what has been learned in the past, for example, to new data to predict future events using labeled examples. For example, starting from an analysis of a known training data set, the supervised learning algorithm 620 can generate an inference function to make predictions about output values. The supervised learning algorithm 500 can also compare its output with the correct and expected output and find errors so that the supervised learning algorithm 620 can be modified accordingly. On the other hand, the unsupervised learning algorithm 622 may include any algorithm that can be applied, for example, when the data used to train the unsupervised learning algorithm 622 is neither classified nor labeled. For example, the unsupervised learning algorithm 622 can study and analyze how the system infers functions that describe hidden structures from unlabeled data.

[0121] In some embodiments, NLP algorithms and functions 606 may include any algorithm or function suitable for automatically manipulating natural language (such as speech and / or text). For example, in some embodiments, NLP algorithms and functions 606 may include content extraction algorithms or functions 624, classification algorithms or functions 626, machine translation algorithms or functions 628, question answering (QA) algorithms or functions 630, and text generation algorithms or functions 632. In some embodiments, content extraction algorithms or functions 624 may include means for extracting text or images from electronic documents (e.g., web pages, text editor documents, etc.) for use in, for example, other applications.

[0122] In some embodiments, the classification algorithm or function 626 may include any algorithm that can learn from the data input of the supervised learning model using a supervised learning model (e.g., logistic regression, naive Bayes, stochastic gradient descent (SGD), k-nearest neighbor, decision tree, random forest, support vector machine (SVM), etc.), and make new observations or classifications based on this. The machine translation algorithm or function 628 may include any algorithm or function that can be suitable for automatically converting a source text in one language into text in another language, for example. The QA algorithm or function 630 may include any algorithm or function that can be suitable for automatically answering questions asked by humans in natural language, such as algorithms or functions performed by voice-controlled personal assistant devices. The text generation algorithm or function 632 may include any algorithm or function that can be suitable for automatically generating natural language text.

[0123] In some embodiments, the expert system 608 may include any algorithm or function that may be suitable for simulating the judgment and behavior of a human or organization with specialized knowledge and experience in a particular field (e.g., stock trading, medicine, sports statistics, etc.). The computer-based vision algorithms and functions 610 may include any algorithm or function that may be suitable for automatically extracting information from an image (e.g., a photo image, a video image). For example, the computer-based vision algorithms and functions 610 may include an image recognition algorithm 634 and a machine vision algorithm 636. The image recognition algorithm 634 may include any algorithm that may be suitable for automatically identifying and / or classifying, for example, objects, places, people, etc. that may be included in one or more image frames or other display data. The machine vision algorithm 636 may include any algorithm that may be suitable for allowing a computer to "see" or, for example, rely on an image sensor camera with specialized optical elements to acquire images for processing, analysis, and / or measurement of various data characteristics for decision-making purposes.

[0124] In some embodiments, speech recognition algorithms and functions 612 may include any algorithm or function that may be suitable for recognizing spoken language and translating it into text, such as through automatic speech recognition (ASR), computer speech recognition, speech to text (STT) 638, or text to speech (TTS) 640, for example, in order to compute communication with one or more users via speech. In some embodiments, planning algorithms and functions 614 may include any algorithm or function that may be suitable for generating a sequence of actions, where each action may include its own set of prerequisites to be satisfied before performing the action. Examples of AI planning may include classical planning, reduction to other problems, temporal planning, probabilistic planning, preference-based planning, conditional planning, and the like. Finally, robotics algorithms and functions 616 may include any algorithm, function, or system that may enable one or more devices to replicate human behavior through, for example, motion, posture, performance tasks, decision making, emotions, and the like.

[0125] Described herein include processes associated with predicting molecular binding properties of one or more proteins, as described above. This may include importing the amino acid sequence of a protein and generating a molecular descriptor matrix based on the amino acid sequence. Protein molecules are formed by amino acid sequences. The amino acid sequence can be represented by a string (e.g., a string of letters). In one or more instances, the amino acid sequence can be input into a machine learning model (e.g., a neural network) to generate a molecular descriptor matrix. In one or more instances, the machine learning model can be pre-trained using the amino acid sequence. For example, the machine learning model may include a protein language model. In another example, the machine learning model can be pre-trained in an unsupervised manner. In some embodiments, the machine learning model can be configured to generate structure-based descriptors representing sequences used to generate protein structures.

[0126] The generated molecular feature matrix can be used to predict the molecular binding properties of the corresponding protein. In some embodiments, the molecular descriptor matrix can be a multidimensional matrix (i.e., a tensor) composed of multiple feature vectors representing the descriptors of each amino acid in the sequence of each protein. In order to determine the predicted molecular binding properties, in one or more instances, the dimension of the multidimensional molecular descriptor matrix (e.g., a descriptor tensor) can be reduced. In some embodiments, the multidimensional molecular descriptor matrix (having each amino acid feature vector for each molecule) can be reduced to a two-dimensional molecular feature matrix (having a molecular feature vector for each molecule) by averaging the feature vectors of all amino acids in each molecule. In some embodiments, the feature dimension reduction technique for reducing the number of feature vectors of the molecular descriptor matrix can particularly include removing redundant feature vectors after averaging. For example, since some feature vectors (and / or the features included therein) may be highly correlated, a single representative feature vector can be identified to represent a set of highly correlated feature vectors. In some embodiments, clustering techniques (e.g., hierarchical / aggregate clustering techniques) can be used to identify similar feature vectors (e.g., their corresponding embeddings are less than a threshold distance from each other in the embedding space). Based on the identified feature vectors, one or more representative feature vectors may be selected from each cluster of similar feature vectors as "representatives" of the cluster.

[0127] As described above, the representative feature vectors can be input into a machine learning model to obtain predictions of the molecular binding properties of proteins. These proteins can be proteins of interest for potential drug discovery assays. The machine learning model can be trained to receive as input one or more representative feature vectors describing one or more proteins, and output predictions of the molecular binding properties of proteins based on the representative feature vectors.

[0128] In some embodiments, the molecular descriptors (from Figure 3A The machine learning model 301 is based on the training molecular descriptor matrix generated based on the amino acid sequence of the imported one or more empirically evaluated proteins and the predetermined batch binding data associated with the empirically evaluated proteins to train the machine learning model. After alignment, supervised regression can be performed to train the machine learning model. In one or more instances, the regressor used may include a bagged decision tree, a bagged linear model, a non-bagged linear model, a random forest, a linear forest, or another type of regressor or a combination thereof.

[0129] In some embodiments, a part of the training step includes optimizing the hyperparameter set of the machine learning model. In one or more instances, the hyperparameter may include a regularization parameter, the number of estimators, the maximum tree depth, etc. In one or more instances, the pipeline (e.g., feature dimensionality reduction model 307A, feature selection model 309A, 309B and regression model 311A, 311B) may be jointly optimized. The feature dimensionality reduction model may be configured to use correlation clustering, recursive feature elimination and / or other techniques to reduce the number of eigenvectors of the molecular descriptor matrix.

[0130] In some embodiments, the training step may also include a cross-validation step, in which an optimized set of learnable parameters for the machine learning model is identified, and then cross-validation testing is iteratively performed until an optimized set of learnable parameters is determined. For example, if the machine learning model includes a decision tree structure (e.g., a random forest), the number of learnable parameters may include the number of trees and / or the depth of the trees. The optimized set of learnable parameters is selected so that they optimize the performance of the machine learning model. Using the optimized set of learnable parameters, the machine learning model can be trained to generate predictions of molecular binding properties for new amino acid sequences that are not part of the training set.

[0131] In some embodiments, one or more additional steps may be performed to predict molecular binding properties of one or more protein molecules based on the amino acid sequence. One of the goals of the disclosed technology includes predicting the properties of a molecule to be assessed. In particular, the degree of binding of a protein molecule to a resin provides valuable clinical information and / or valuable manufacturing process developability information that can be used in the development of new therapeutic drugs. Additional / alternative step sets to the aforementioned steps are described above that can be performed to predict molecular binding properties based on amino acid sequence.

[0132] The technology described herein achieves many technical advantages. For example, Figure 1 As shown, testing the binding properties of molecules (such as for drug discovery purposes) is a complex and time-consuming process. Figure 1 In the experimental example 102 of , the duration of the experiment for performing one or more protein purification processes may span several weeks. On the other hand, the execution time of the computational model-based instance 104 (e.g., the machine learning model described herein) for performing one or more protein purification processes may be only a few minutes. Therefore, the experimental example 102 describes a non-ideal process for testing each potential molecule. The machine learning model described herein can reduce the amount of time spent on testing by increasing the number of molecules that can be screened in a given time, or the number of molecules screened by a given researcher as the goal of the model. As another example, various existing protein language models (e.g., Figure 1120) to generate a molecular descriptor matrix. Therefore, existing techniques can be used to generate inputs to the machine learning model, thereby reducing the amount of additional data that needs to be collected and reducing the amount of additional model training required. As another example, less data can be used to train the machine learning model described herein while maintaining or improving the accuracy of the model. For example, instead of inputting a molecular descriptor matrix into a machine learning model (which may include a very large number of descriptors), the molecular descriptor matrix can be reduced to determine (and used as input to the machine learning model) the most predictive feature vectors. This descriptor reduction process can further optimize the training process of the machine learning model. For example, each training molecular descriptor matrix can be reduced by determining the most predictive feature vector, and the model can be trained based on the most predictive feature vector.

[0133] Figure 7 Another high-level workflow diagram 700 is shown for performing feature generation 202, feature dimensionality reduction 204, feature filtering 206, recursive model-based feature elimination 207, and regression model optimization 208 according to various embodiments. The description of feature generation 202, feature dimensionality reduction 204, feature filtering 206, and regression model optimization 208 is equally applicable here.

[0134] However, unlike graph 200, graph 700 may further include recursive model-based feature elimination 207. Recursive model-based feature elimination 207 may include additional models for further reducing the number of features in the feature set. In particular, recursive model-based feature elimination 207 may help prevent or reduce the possibility of overfitting. As an example, refer to Figure 8 , feature elimination based on recursive model 207 can be implemented Figure 8 Machine learning model 820. Figure 8 Included with Figure 3A Similar components are described herein, and similar labels are used to refer to those components. For example, workflow 800 may include model 301 and machine learning model 820. Machine learning model 820 may include feature dimensionality reduction model 307A, feature filtering model 309A, recursive feature elimination model 801, and regression model 311A. Workflow 800 may follow a similar path as workflow 300A, except that the most predictive feature vectors may include feature vectors that have been reduced by recursive feature elimination model 801. As an example, determining one or more most predictive feature vectors may further include implementing recursive feature elimination model 801 to further reduce the number of feature vectors. In particular, some embodiments include a number of feature vectors included in the further reduced number of feature vectors that is equal to or less than the number of training items.

[0135] In some embodiments, at function block 802, the recursive feature elimination model 801 may be configured to fit a model to a representative feature vector. For example, the model may be a regression model. At function block 804, feature importance scores may be calculated based on the fitted model. The feature importance scores may indicate the importance of each representative feature vector. At function block 806, one or more feature vectors in the representative feature vector may be removed based on the feature importance scores of each of the representative feature vectors to obtain a representative feature vector subset. For example, the least important features or feature vectors may be removed from the representative feature vector. In one or more instances, the most predictive feature vector may include one or more feature vectors from the representative feature vector subset. At function block 808, the recursive feature elimination model 801 may iteratively perform blocks 802 to 806 until the number of feature vectors included in the subset meets the feature quantity criterion. For example, the feature quantity criterion that is met includes the number of feature vectors included in the representative feature vector subset being less than or equal to a threshold number of feature vectors. In some instances, the threshold number of feature vectors may include features of the same or similar number from the training data used to train the machine learning model 820. In some instances, the number of feature vectors included in the representative feature vector subset may include a hyperparameter in the hyperparameter set. In some examples, the number of feature vector clusters included in the plurality of feature vector clusters may include a hyperparameter in the hyperparameter set.

[0136] Fig. 9 The process for training a machine learning model to detect molecular binding properties according to various embodiments is shown. As an example, the above reference FIG. 3A to FIG. 4 The process described is compared to that described in this paper Fig. 9The process 900 can organize the data used to train the regression model (e.g., at step 930) in different ways. As an example, in the process described above, for each empirically evaluated protein, the data used to train the machine learning model (e.g., machine learning pipeline 908) includes a predetermined number of experimental conditions. The experimental conditions may specify the molecular binding properties of the protein for a given set of experimental conditions. For example, the data may include the molecular binding level of the protein measured at a first salt concentration and a first pH level, the molecular binding level of the protein measured at a second salt concentration and a first pH level, the molecular binding level of the protein measured at a first salt concentration and a second pH level, and the change in the molecular binding level of the protein measured at a first salt concentration and a second pH level, etc. In one or more instances, the number of predetermined experimental conditions for predetermined batch binding may include 12 or more experimental conditions (e.g., 4 salt concentrations, 3 pH levels), 24 or more experimental conditions (e.g., 6 salt concentrations, 4 pH levels), etc. As described above, the trained machine learning model can use experimental conditions (e.g., pH level and salt concentration) as inputs in addition to the molecular descriptor matrix to predict the molecular binding properties of one or more proteins.

[0137] In some embodiments, the experimental conditions may not need to be input into the machine learning model, but the predicted molecular binding properties can be determined for a series of experimental conditions. However, to do this, the training data and training process can be adjusted, such as Fig. 9 shown.

[0138] Fig. 9 A workflow diagram of a process 900 for optimizing hyperparameters and learnable parameters of a machine learning model for performing one or more computational model-based protein purification processes according to various embodiments is shown. Process 900 is similar to the process described above with reference to FIG. 3A to FIG. 4 The difference in the described process is that a transformed representation of the molecular binding properties of a protein evaluated by training experience can be used to train a machine learning model. The trained machine learning model can output a value corresponding to the transformed representation of the molecular binding properties, which in turn can be used to predict all binding conditions for all experimental conditions for a given protein molecule. Therefore, the amount of training data required to train a machine learning model can be reduced from N empirically derived binding measures for N different experimental conditions (e.g., salt concentration levels and pH levels) to a single transformed binding measure that can be used to solve the N empirically derived binding measures.

[0139] In process 900, sequence data 902 corresponding to one or more amino acid sequences of protein P may be provided to a matrix generation machine learning (ML) model 904. In some embodiments, machine learning model 904 may be used in conjunction with Figure 3AThe machine learning model 301 of FIG. 1 is the same or similar to that of FIG. 1 , and the previous description may apply. In some embodiments, the matrix generation ML model 904 may be trained to generate a molecular descriptor matrix 906 based on the sequence data 902 representing the amino acid sequence of the P protein. The matrix generation ML model 904 may include a neural network that may generate features X structured as the molecular descriptor matrix 906. The molecular descriptor matrix 906 may be the same as that in FIG. Figure 3A The molecular descriptor matrix 906 may be the same or similar to the molecular descriptor matrix generated at functional block 306 of . In one or more examples, the molecular descriptor matrix 906 may include 100 or more features, 500 or more features, 1,000 or more features, 2,000 or more features, 10,000 or more features, or another number of features. The features of the molecular descriptor matrix 906 may then be analyzed to determine which features, if any, are associated with the molecular binding properties of the corresponding protein molecules.

[0140] The molecular descriptor matrix 906 may have dimensions of the number of molecules M times the number of descriptors (eg, features) N. For a given molecule, the amino acid sequence may be represented using a string of characters (eg, letters) that form the protein being tested.

[0141] In some embodiments, sequence 902 may also be analyzed by experiment. The experiment may generate empirically derived protein binding data 912. The empirically derived protein binding data 912 may include molecular binding property values ​​for a set of experimental conditions 914. For example, the empirically derived protein binding data 912 may indicate that for a given sequence (e.g., sequence A) and a first experimental condition (e.g., a first salt concentration level and a first pH level), the molecular binding property is Y1. Similarly, the empirically derived protein binding data 912 may indicate that for a sequence (e.g., sequence A) and a second experimental condition (e.g., a second salt concentration level and a first pH level), the molecular binding property is Y2. Further, the empirically derived protein binding data 912 may indicate that for a sequence (e.g., sequence A) and a third experimental condition (e.g., a first salt concentration level and a second pH level), the molecular binding property is Y3. In some embodiments, the predetermined batch binding data may be expressed in the form of molecules as rows and experimental conditions 914 as columns.

[0142] Process 900 can be configured to train a machine learning model (e.g., machine learning model 820) to predict the molecular binding properties of proteins for a set of experimental conditions. Testing the binding properties of molecules (such as for drug discovery purposes) is a complex and time-consuming process (e.g., it takes 2 to 6 weeks to culture the molecules, purify, and then test, so it may take several weeks to fully evaluate each molecule). Testing each potential molecule is not ideal. Therefore, increasing the number of molecules that can be screened in a given amount of time, or the number of molecules screened by a given researcher, is a goal of the model. Another goal is to increase the number of molecules that can be screened without causing time delays or additional experimental burdens.

[0143] The process 900 can be trained using a small number of training examples (e.g., a few molecules) and a large number of descriptors (e.g., 100 or more features, 500 or more features, 1,000 or more features, 2,000 or more features, 10,000 or more features, etc.). The process 900 can rank the descriptors in a systematic manner to train a machine learning model to predict molecular binding properties 910. In addition, the process 900 can utilize descriptors related to one or more physical properties of the protein. The machine learning pipeline 908 can thus be configured to find descriptors (e.g., features) that best predict the molecular binding properties of the protein based on the molecular descriptor matrix. The ML model can then try and determine which descriptors are most predictive.

[0144] In one illustrative example, predetermined batch binding data 912 includes empirically measured binding properties for each analyzed protein for the set of experimental conditions. In some embodiments, process 900 may include performing a set of (e.g., using) Figure 5 The process 900 may transform the empirically measured binding characteristic of percent binding stored in the predetermined batch binding data 912 into a linearized or pseudo-linear representation of the empirically measured binding characteristic. For example, a logit transformation operation may be performed.

[0145] The logit transformation involves calculating the logarithm of the ratio of bound / unbound protein concentrations. The result of this transformation is that binding is transformed from 0.0 to 1.0 (i.e., 0% binding to 100% binding) to negative infinity to positive infinity (in log(K p ) space). After the data is linearized, a linear model with better convergence can be used, such as the PCA model.

[0146] In some embodiments, process 900 may include applying one or more dimensionality reduction techniques (e.g., principal component analysis (PCA) 918) to a linear representation of the empirically measured binding properties of each analyzed protein. PCA 918 may be configured to derive a first principal component (PC), a second principal component, etc., of a linearized transformation (e.g., logit transformation) of their empirically measured binding properties. The PCA 918 performed may represent the linear representation of the empirically measured protein binding properties as a more concise representation. Specifically, the number C of experimental conditions defines the number of data points in the predetermined batch binding data 912. PCA 918 may reduce the number of data points from C to less than or equal to C. For example, if 24 experimental conditions are used to obtain the predetermined batch binding data 912, PCA may allow the predetermined batch binding data 912 to be less than (or equal to) 24 data points. PCA 918 may be configured to output a transformed representation 920, which represents a transformed version of the empirically measured molecular binding properties. The number of molecules tested can be 1 or more, 5 or more, 10 or more, 20 or more, 50 or more, or other values.

[0147] The PCA model can decompose data (e.g., predetermined batch binding data 912) into a collection of lower dimensional vectors. For example, for 24 experimental conditions (e.g., 24 experimental data points), the PCA model can identify the first eigenvector of the data, which can capture multiple variances of the data set. Therefore, PCA enables the use of a lower dimensional projection to describe the behavior of the binding data. In one or more instances, if the average binding efficiency of a molecule is to be predicted, PCA provides a more representative and more valuable result than any individual experimental condition. In addition, PCA can succinctly (in a low-dimensional representation) summarize the trend in noisy multidimensional data, which may be useful for scientists. Those of ordinary skill in the art will recognize that PCA 918 can identify any number of principal components, including but not limited to the first principal component and / or the second principal component.

[0148] In some embodiments, the predicted molecular binding properties 910 can be compared to the transformed representations 920 of the empirically measured molecular binding properties. In one or more instances, a cross-validation loss can be calculated to determine how well the machine learning model 908 predicts the empirically measured molecular binding properties for a given protein. In particular, the prediction indicates how well the machine learning pipeline 908 predicts the transformed representations of the empirically measured molecular binding properties.

[0149] In some embodiments, at 930, a cross validation loss may be calculated. As previously described, one or more instances may use a k-fold cross validation technique. Additionally or alternatively, at 930, a stratified k-fold cross validation may be calculated. Stratified k-fold cross validation includes taking the molecules of the training set and ranking them in bins according to their molecular binding properties. For example, the bins may include a first bin corresponding to weakly binding proteins, a second bin corresponding to moderately binding proteins, a third bin corresponding to tightly binding proteins, and the like. The stratified k-fold cross validation may then evaluate the performance of the machine learning pipeline 908 by selecting a representative subset, such as an even number of representatives of each of the binned proteins.

[0150] FIG. 10A to FIG. 10D An exemplary graph is shown according to various embodiments, which shows how principal component analysis can be used to predict molecular binding properties. For example, Fig. 10A A graph 1000 showing the results of a principal component analysis of a set of molecules. In graph 1000, the X-axis corresponds to the first principal component value for each molecule in the set and the Y-axis corresponds to the second principal component for each molecule. The red and green ellipses represent the first and second standard deviations from the centroid of the data point cluster. As can be seen from graph 1000, the molecules (e.g., data points) are fairly evenly distributed about the x-axis.

[0151] Fig. 10B Isotherm graphs 1020 of a given molecule for various values ​​of the first principal component according to various embodiments are shown. In graph 1020, the x-axis represents the salt concentration level used to determine the protein binding properties of the molecule during the corresponding experiment, and y represents the protein binding level. Isotherms 1022 to 1030 correspond to different principal component (PC) values. For example, isotherm 1022 represents how the binding of the protein changes for the first PC (e.g., PC=-6) as the salt concentration level changes. Isotherm 1024 represents how the binding of the protein changes for the second PC (e.g., PC=-3) as the salt concentration level changes. Isotherm 1026 represents how the binding of the protein changes for the third PC (e.g., PC=0) as the salt concentration level changes. Isotherm 1028 represents how the binding of the protein changes for the fourth PC (e.g., PC=+3) as the salt concentration level changes. Isotherm 1030 represents how the binding of the protein changes for the fifth PC (e.g., PC=+6) as the salt concentration level changes.

[0152] The isotherms 1022 to 1030 of the graph 1020 can be calculated using a fixed pH level. As can be seen from the graph 1020, as the value of the first PC increases (e.g., -6 in the graph 1022) to a very large value (e.g., +6 in the graph 1030), the binding behavior changes. In the example of the graph 1020, for low salt concentrations, the percent binding is about 100%; for high salt concentration values, the percent binding is about 0%.

[0153] When making a protein, one or more protein purification steps may be performed to filter out molecules that are not the target protein. The protein purification step includes causing or otherwise promoting the target protein to bind to a resin (e.g., a chromatography column). Ideally, the resin will bind all the target protein. Then, in order to remove the protein from the resin, washing may be applied to deposit the target protein into a solution. Washing may include salts at a specific salt concentration level (and / or pH level). The salt concentration level may affect whether the protein is detached from the resin. For example, at a lower salt concentration level, the protein may remain bound to the resin, while a higher salt concentration level may cause the protein to detach from the resin. Once removed, the solution / protein may be assayed or otherwise studied.

[0154] In general, for all resins, molecules that are completely or mostly bound throughout the design space (the typical pH and salt range over which proteins remain stable) are likely incompatible due to an inability to elute the protein (resulting in low yield). For resins that are typically operated in bind and elute mode, molecules that are completely or mostly unbound throughout the design space are likely incompatible due to an inability to bind the protein. Therefore, it is desirable to have a protein that can express its binding variability, with binding transitions from a bound state (e.g., greater than 90% bound) to an unbound state (e.g., less than 10% bound). The first principal component (e.g., FIG. 10A to FIG. 10B ) uses a single value (e.g., a principal component) rather than a set of experimental conditions (e.g., 24 salt / pH combinations) to describe the transition of the protein from the bound state to the unbound state (as can be seen in isotherms 822 to 830).

[0155] In some embodiments, the first principal component (e.g. Fig. 10B The average binding can be intuitively described as a percentage of binding (as shown in the graph 1020 of FIG. 1024 ). For example, it can be seen from the isotherm 1022 that for the first principal component of -6, the protein can be tightly bound to the resin. The isotherm 1022 can be marked as problematic because, regardless of the salt concentration level, for a particular pH level and first principal component value, the protein under analysis is unlikely to be unbound from the resin.

[0156] observe Fig. 10CThe graph 1040 may be used to analyze different principal component value sets. The protein and pH levels used in the analysis in the example of the graph 1040 may be compared to Fig. 10B 1020 is similar, however not necessary. As can be seen from the graph 1040, isotherms 1042 to 1050 show how the binding percentage changes with the change of the salt concentration level of the washing solution for the value of the first principal component at different points. For example, isotherm 1042 shows how the binding of the protein changes with the change of the salt concentration level for the first PC (e.g., PC=-10). Isotherm 1044 shows how the binding of the protein changes with the change of the salt concentration level for the second PC (e.g., PC=-5). Isotherm 1046 shows how the binding of the protein changes with the change of the salt concentration level for the third PC (e.g., PC=0). Isotherm 1048 shows how the binding of the protein changes with the change of the salt concentration level for the fourth PC (e.g., PC=+5). Isotherm 1050 shows how the binding of the protein changes with the change of the salt concentration level for the fifth PC (e.g., PC=+10).

[0157] Observing isotherm 1042, the binding percentage of the protein does not change much as the salt concentration level changes. Isotherm 1042 can then also be marked as problematic because the protein binds to the resin and cannot be removed. As another example, isotherm 1050 can have a substantially static binding percentage, regardless of the salt concentration level. However, unlike isotherm 1042, the protein in this example may not be able to bind to the resin. Therefore, isotherm 1050 can also be marked as problematic because purification cannot be performed and all proteins are washed away. Isotherms 1044 to 1048 represent a more ideal state, in which the binding percentage changes from binding to unbinding when the salt concentration level changes.

[0158] In some embodiments, predicting the first principal component can enable determination of the percent binding for an unlimited number of salt concentrations (and / or pH). In contrast, without using PCA, percent binding predictions would need to be made for all experimental conditions (e.g., points along an isotherm). Thus, using PCA to predict the first principal component greatly simplifies the process of predicting molecular binding properties of a protein without sacrificing accuracy.

[0159] In some embodiments, PCA may output more than the first principal component. For example, a second principal component may also be determined and used to guide the decision making process. Fig. 10D, graph 1060 depicts isotherms 1062 to 1070 for the second principal component of the protein. Isotherms 1062 to 1070 show how the percentage of protein binding changes with changes in salt concentration levels for a set of second principal component values. For example, isotherm 1062 represents how the binding of the protein changes for a first PC value (e.g., 2PC value = -6) as the salt concentration level changes. Isotherm 1064 represents how the binding of the protein changes for a second PC value (e.g., 2PC value = -3) as the salt concentration level changes. Isotherm 1066 represents how the binding of the protein changes for a third PC value (e.g., 2PC value = 0) as the salt concentration level changes. Isotherm 1068 represents how the binding of the protein changes for a fourth PC value (e.g., 2PC value = +1) as the salt concentration level changes. Isotherm 1070 represents how the binding of the protein changes as the salt concentration level changes for a first PC value (e.g., 2nd PC value = +2). The first principal component may shift, where there is a transition from bound to unbound. In some examples, this transition may not even occur. The second principal component does not shift the transition as much, but changes the steepness of isotherms 1062 to 1070. For example, isotherm 1062 may include a 2nd PC value of -6, which is very steep as shown; in contrast, isotherm 1070 with a 2nd PC value of +2 is less steep (and does not reach a percent binding of about 0%). In some embodiments, other principal components may be used. Fig. 10D In the graph 1060 of , the isotherm 1066 may represent an "ideal" curve. In this example, the first principal component may be set to zero, while the second principal component is varied.

[0160] In some embodiments, the machine learning pipeline 908 may be trained to output a first principal component, a second principal component, other principal components, or a combination thereof. The machine learning pipeline 908 may output the principal components together or serially.

[0161] In some embodiments, process 900 can reduce the number of data points required to train a machine learning model. For example, the number of principal components can be limited by the number of data points for a protein that has been empirically measured. In one or more instances, the number of principal components can be less than or equal to the number of experimental conditions. For example, while FIG. 3A to FIG. 4 The described process may require N data points for N experimental conditions, but Fig. 9 Process 900 can reduce this number to 1 data point.

[0162] FIG. 11A to FIG. 11F Exemplary heatmaps 1100 to 1150 are shown, respectively, showing experimental conditions and experimental K according to various embodiments. p The values ​​between experimental conditions and modeling K pThe heat maps 1100 to 1150 include a color gradient that represents how tightly the proteins bind (in units of percent binding). The x-axis of the heat maps 1100 to 1150 depicts salt concentration levels and the y-axis represents pH levels. The "red" portions of the heat maps 1100 to 1150 represent higher log(K p ) value (e.g., molecular binding properties), and “green” indicates a lower log(K p ) values. Heatmaps 1100 to 1150 may be generated based on one or more empirically evaluated proteins. For example, FIG. 11A to FIG. 11B Heat maps 1100 to 1110 may be depicted that depict experimental K for ion exchange resins. p K for screening and model prediction p filter. FIG. 11C to FIG. 11D Heat maps 1120 to 1130 may be depicted, which depict experimental K for hydrophobic resins. p K for screening and model prediction p filter. FIG. 11E to FIG. 11F Heat maps 1140 to 1150 may be depicted that depict experimental K for mixed mode resins. p K for screening and model prediction p Screening. For example, at a pH level of pH = 5.5, the protein of interest may bind until the salt concentration level used in the experimental data reaches about 250 mM. On the other hand, the modeling data may indicate that for a pH level of pH = 5.5, the protein of interest may be bound until the salt concentration level reaches about 175 mM.

[0163] Fig.12 A flow chart showing a method 1200 for generating predictions of molecular binding properties for one or more target proteins according to various embodiments is shown, which is part of another simplified process for protein purification for identifying target proteins. According to the disclosed embodiments, the method 1200 can accelerate the selection process of therapeutic antibody candidates or other immunotherapy candidates by early identification of the most promising therapeutic antibody candidates. The method 1200 can utilize one or more processing devices (e.g., described below with reference to Figure 5 and Figure 6The present invention relates to a computing device and artificial intelligence architecture discussed herein, which may include hardware (e.g., a general-purpose processor, a graphics processing unit (GPU), an application-specific integrated circuit (ASIC), a system on a chip (SoC), a microcontroller, a field-programmable gate array (FPGA), a central processing unit (CPU), an application processor (AP), a visual processing unit (VPU), a neural processing unit (NPU), a neural decision processor (NDP), a deep learning processor (DLP), a tensor processing unit (TPU), a neuromorphic processing unit (NPU), or any other processing device that may be suitable for processing genomics data, proteomics data, metabolomics data, metagenomics data, transcriptomics data, or other omics data and preparing one or more processors), firmware (e.g., microcode), or some combination thereof.

[0164] In some embodiments, process 1200 may begin at block 1210. Block 1210 may form part of the steps performed by training machine learning pipeline 908. At block 1210, a training molecule descriptor matrix representing a training set of amino acid sequences corresponding to one or more empirically evaluated proteins may be accessed. For example, a training molecule matrix may be generated for proteins that have been experimentally evaluated under one or more experimental conditions (e.g., salt concentration levels, pH levels, etc.).

[0165] At frame 1220, an iterative process may be performed to refine the hyperparameter set associated with the ensemble learning model until the desired accuracy is reached. For example, the process may be repeated until the machine learning pipeline 908 predicts the molecular binding properties with a threshold level of accuracy. Frame 1220 may include the step performed during each iteration of frame 1220. For example, at step 1222, a representative feature vector may be selected for each of a plurality of feature vector clusters to reduce the training molecule descriptor matrix. Each feature vector cluster may include similar feature vectors. For example, two feature vectors (for example, in the embedding space) with a distance less than a threshold distance may be classified as "similar". The selected representative feature vector may represent all feature vectors included in a given feature vector cluster.

[0166] At step 1224, one or more most predictive feature vectors from the selected representative feature vectors for each feature vector cluster may be determined based on correlations between the selected representative feature vectors and predetermined bulk binding data associated with the empirically evaluated proteins. The most predictive feature vector may be determined based on a principal component analysis that identifies a first principal component.

[0167] At step 1226, one or more cross-validation losses may be calculated based at least in part on the most predictive feature vectors and the predetermined batch of combined data. A set of hyperparameters for the machine learning pipeline 908 may be updated based on the cross-validation losses. At step 1228, the set of hyperparameters may be updated based on the one or more cross-validation losses.

[0168] Blocks 1210 to 1220 (including steps 1222 to 1228) may include a "training" portion. The results of blocks 1210 to 1220 may include a trained machine learning model (e.g., machine learning model 908), which may be used during inference. For example, at block 1230, a training molecular descriptor matrix representing a set of amino acid sequences corresponding to one or more proteins may be accessed. At block 1240, a prediction of molecular binding properties of one or more proteins may be obtained by the trained ML model based at least in part on the molecular descriptor matrix.

[0169] In some embodiments, the protein may be a target protein. In one or more instances, a machine learning model (e.g., a protein language model implemented using a neural network) may be trained to receive data representing a set of amino acid sequences corresponding to the protein and generate a molecular descriptor matrix describing the amino acid sequence. In some instances, the molecular descriptor matrix may include a plurality of descriptors (e.g., features). The descriptors may be constructed as feature vectors.

[0170] In some embodiments, machine learning pipeline 908 can be trained to analyze the molecular descriptor matrix and perform dimensionality reduction. Dimensionality reduction can reduce the molecular descriptor matrix by selecting a representative eigenvector. The selected representative eigenvector can be selected from a cluster of similar eigenvectors of the molecular descriptor matrix. In one or more instances, each cluster can have a representative eigenvector. The most predictive eigenvector in the representative eigenvector can be determined. The most predictive eigenvector can then be used to generate a predicted molecular binding property. In one or more instances, the predicted molecular binding property can represent the first principal component.

[0171] As used herein, "or" is inclusive and not exclusive, unless expressly stated otherwise or the context indicates otherwise. Thus, as used herein, "A or B" means "A, B, or both," unless expressly stated otherwise or the context indicates otherwise. Furthermore, as used herein, "and" means both jointly and severally, unless expressly stated otherwise or the context indicates otherwise. Thus, as used herein, "A and B" means "A and B, jointly or severally," unless expressly stated otherwise or the context indicates otherwise.

[0172] As used herein, “automatically” and its derivatives mean “without human intervention” unless expressly indicated otherwise or the context indicates otherwise.

[0173] The embodiments disclosed herein are merely examples, and the scope of the present disclosure is not limited thereto. Embodiments according to the present disclosure are particularly disclosed in the attached claims for methods, storage media, systems, and computer program products, wherein any feature mentioned in one claim category (e.g., method) can be claimed for protection in another claim category (e.g., system). Dependents or references in the attached claims are selected only for formal reasons. However, any subject matter resulting from intentional reference to any preceding claim (particularly multiple dependents) may also be claimed, so any combination of claims and their features may be disclosed and claimed regardless of the dependents selected in the attached claims. The subject matter that may be claimed includes not only a combination of features as listed in the attached claims, but also any other combination of features in the claims, wherein each feature mentioned in the claims may be combined with any other feature or combination of other features in the claims. In addition, any embodiment and feature described or depicted herein may be claimed in a separate claim and / or in any combination with any embodiment or feature described or depicted herein or with any feature of the attached claims.

[0174] The scope of the present disclosure covers all changes, substitutions, variations, alterations and modifications to the exemplary embodiments described or shown herein that will be understood by those of ordinary skill in the art. The scope of the present disclosure is not limited to the exemplary embodiments described or shown herein. In addition, although the present disclosure describes and illustrates the corresponding embodiments herein as including specific parts, elements, features, functions, operations or steps, any of these embodiments may include any combination or arrangement of any parts, elements, features, functions, operations or steps described or shown anywhere herein that will be understood by those of ordinary skill in the art. In addition, in the appended claims, references to devices or systems or parts of devices or systems adapted to, arranged to, capable of, configured to, enabled to, operable to or operated to perform specific functions, cover the device, system, component, whether or not it or the specific function is activated, turned on or unlocked, as long as the device, system or component is so adapted, arranged, capable of, configured to, enabled to, operable or operated. In addition, although the present disclosure describes or illustrates the specific advantages provided by certain embodiments, certain embodiments may not provide, partially provide or provide all of these advantages.

[0175] Exemplary Embodiments

[0176] Embodiments disclosed herein may include:

[0177] 1. A method for predicting molecular binding properties of one or more proteins, the method comprising, by one or more computing devices: accessing a molecular descriptor matrix representing a set of amino acid sequences corresponding to one or more proteins; and refining a set of hyperparameters associated with a machine learning model, the machine learning model being trained to generate predictions of molecular binding properties of the one or more proteins, wherein refining the set of hyperparameters comprises iteratively performing a process until a desired accuracy is achieved, the process comprising: reducing the molecular descriptor matrix by selecting a representative feature vector for each of a plurality of feature vector clusters, wherein each feature vector cluster comprises similar feature vectors; determining one or more most predictive feature vectors of the selected representative feature vectors for each feature vector cluster based on correlations between the selected representative feature vectors and predetermined batch binding data associated with the one or more proteins; calculating one or more cross-validation losses based at least in part on the one or more most predictive feature vectors and the predetermined batch binding data; and updating the set of hyperparameters based on the one or more cross-validation losses; and outputting, by the machine learning model, a prediction of the molecular binding properties of the one or more proteins based at least in part on the updated set of hyperparameters.

[0178] 2. A method according to embodiment 1, wherein calculating the one or more cross-validation losses further includes: evaluating a cross-validation loss function based on the one or more most predictive feature vectors, a predetermined batch of combined data, a hyperparameter set, and a learnable parameter set associated with the machine learning model; and minimizing the cross-validation loss function by changing the learnable parameter set while keeping the one or more most predictive feature vectors, the predetermined batch of combined data, and the hyperparameter set constant.

[0179] 3. A method according to embodiment 2, wherein minimizing the cross-validation loss function includes optimizing a hyperparameter set, and wherein the hyperparameter set includes one or more of a regular parameter set, a booster parameter set, or a learning task parameter set.

[0180] 4. The method of embodiment 2 or 3, wherein minimizing the cross-validation loss function comprises minimizing the loss between the predicted protein binding percentage for the one or more proteins and the experimentally determined protein binding percentage for the one or more proteins.

[0181] 5. The method of any one of embodiments 2 to 4, wherein the predetermined batch binding data comprises experimentally determined protein binding percentages for one or more pH values ​​and salt concentrations associated with molecular binding properties of the one or more proteins.

[0182] 6. A method according to any one of embodiments 2 to 5, wherein the set of learnable parameters includes one or more weights or decision variables determined by a machine learning model at least in part based on the one or more most predictive feature vectors and a predetermined batch of combined data.

[0183] 7. A method according to any one of embodiments 1 to 6, further comprising: after refining the hyperparameter set: accessing a second molecular descriptor matrix representing a second set of amino acid sequences corresponding to one or more second proteins; reducing the second molecular descriptor matrix by selecting a representative feature vector for each of a second plurality of feature vector clusters of the second molecular descriptor matrix; determining one or more second most predictive feature vectors among the selected representative feature vectors for each feature vector cluster based on a second correlation between the selected representative feature vectors and predetermined batch binding data associated with the one or more second proteins; inputting the one or more second most predictive feature vectors into a machine learning model that is trained to generate predictions of molecular binding properties of the one or more second proteins; and outputting, by the machine learning model, predictions of molecular binding properties of the one or more second proteins based at least in part on the updated hyperparameter set.

[0184] 8. The method of embodiment 7, wherein the prediction of the molecular binding properties of the one or more second proteins comprises a prediction of the protein binding percentage for the one or more second proteins.

[0185] 9. A method according to any one of embodiments 1 to 8, wherein the updated hyperparameter set includes one or more of an updated regular parameter set, an updated booster parameter set, or an updated learning task parameter set.

[0186] 10. The method of any one of embodiments 1 to 9, wherein calculating the one or more cross-validation losses comprises calculating n number of cross-validation losses, and wherein n comprises an integer from 1 to n.

[0187] 11. A method according to any one of embodiments 1 to 10, wherein calculating the one or more cross-validation losses includes determining n number of separate training test splits based on the one or more most predictive feature vectors and a predetermined batch combination data, and wherein n includes an integer from 1 to n.

[0188] 12. A method according to any one of embodiments 1 to 11, wherein calculating the one or more cross-validation losses includes calculating n cross-validation losses, and the method further includes: generating a prediction of the molecular binding properties of the one or more proteins based on the average of the n cross-validation losses.

[0189] 13. The method of any one of embodiments 1 to 12, wherein the molecular descriptor matrix comprises 2 n feature vectors, and where n comprises the dimensions of the molecular descriptor matrix.

[0190] 14. A method according to any one of embodiments 1 to 13, wherein the molecular descriptor matrix is ​​generated by a first machine learning model that is different from the machine learning model.

[0191] 15. A method according to embodiment 14, wherein the first machine learning model is trained to generate a molecular descriptor matrix based on a set of amino acid sequences.

[0192] 16. A method according to embodiment 14 or 15, wherein the first machine learning model comprises a neural network trained to generate an M x N descriptor matrix representing a set of amino acid sequences, and wherein N comprises the number of the amino acid sequence sets and M comprises the number of nodes in the output layer of the neural network.

[0193] 17. A method according to any one of embodiments 1 to 16, wherein the machine learning model includes one or more of a gradient boosting model, an adaptive boosting (AdaBoost) model, an extreme gradient boosting (XGBoost) model, a lightweight gradient boosting machine (LightGBM) model, or a classification boosting (CatBoost) model.

[0194] 18. The method of any one of embodiments 1 to 17, wherein the machine learning model is further trained to generate predictions of the molecular elution characteristics of the one or more proteins.

[0195] 19. The method of any one of embodiments 1 to 18, wherein the machine learning model is further trained to generate predictions of the flow-through properties of the one or more proteins.

[0196] 20. The method of any one of embodiments 1 to 19, wherein reducing the molecular descriptor matrix comprises performing a Pearson correlation of eigenvectors of the molecular descriptor matrix to generate the plurality of eigenvector clusters.

[0197] 21. The method of embodiment 20, wherein the selected one representative feature vector for each of the plurality of feature vector clusters comprises a centroid feature vector for each of the plurality of feature vector clusters for representing two or more of the similar feature vectors.

[0198] 22. A method according to any one of embodiments 1 to 21, wherein determining the one or more representative feature vectors among the selected representative feature vectors for each of the multiple feature vector clusters includes: selecting a k-best matrix of feature vectors among the selected representative feature vectors for each of the multiple feature vector clusters.

[0199] 23. The method of embodiment 22, wherein a k-best matrix of feature vectors in the selected representative feature vectors is determined based on a predetermined k-best process.

[0200] 24. A method according to any one of embodiments 1 to 23, wherein the correlation between the selected representative feature vector and the predetermined batch combined data is determined based on a maximum information coefficient (MIC) between the selected representative feature vector and the predetermined batch combined data.

[0201] 25. The method of any one of embodiments 1 to 24, wherein said prediction of said molecular binding properties of said one or more proteins comprises a chromatographic process based on a computational model.

[0202] 26. The method of embodiment 25, wherein the computational model-based chromatography process comprises one or more of a computational model-based affinity chromatography process, an ion exchange chromatography (IEX) process, a hydrophobic interaction chromatography (HIC) process, or a mixed mode chromatography (MMC) process.

[0203] 27. The method of any one of embodiments 1 to 26, further comprising optimizing the machine learning model based on a Bayesian model optimization process.

[0204] 28. The method according to embodiment 27 further includes utilizing grouped K-fold cross validation to train and evaluate the optimized machine learning model based on the one or more most predictive feature vectors, the predetermined batch combined data, the hyperparameter set, and the learnable parameter set.

[0205] 29. The method of any one of embodiments 1 to 28, wherein the prediction of the molecular binding properties of the one or more proteins comprises identification of a target protein among the one or more proteins.

[0206] 30. The method of any one of embodiments 1 to 29, wherein the prediction of molecular binding properties of the one or more proteins comprises quantitative structure-property relationship (QSPR) or quantitative structure-activity relationship (QSAR) modeling of the one or more proteins.

[0207] 31. The method of any one of embodiments 1 to 30, wherein the prediction of the molecular binding properties of the one or more proteins comprises the prediction of the molecular binding properties of each amino acid sequence in the set of amino acid sequences corresponding to the one or more proteins.

[0208] 32. A method for predicting molecular binding properties of one or more proteins, the method comprising: accessing, by one or more computing devices, a molecular descriptor matrix representing a set of amino acid sequences corresponding to the one or more proteins; and obtaining, by a machine learning model, a prediction of the molecular binding properties of the one or more proteins based at least in part on the molecular descriptor matrix, wherein the machine learning model is trained by: accessing a training molecular descriptor matrix representing a training set of amino acid sequences corresponding to one or more empirically evaluated proteins; and iteratively performing a process to refine a set of hyperparameters associated with the machine learning model until a desired accuracy is achieved, the process comprising: reducing the training molecular descriptor matrix by selecting a representative feature vector for each of a plurality of feature vector clusters, wherein each feature vector cluster comprises similar feature vectors; determining one or more most predictive feature vectors from the selected representative feature vectors for each feature vector cluster based on correlations between the selected representative feature vectors and predetermined batch binding data associated with the one or more empirically evaluated proteins; and calculating one or more cross-validation losses based at least in part on the one or more most predictive feature vectors and the predetermined batch binding data; and updating the set of hyperparameters based on the one or more cross-validation losses.

[0209] 33. A method according to embodiment 32, wherein obtaining a prediction includes: reducing the molecular descriptor matrix by selecting a representative feature vector for each of multiple feature vector clusters of the molecular descriptor matrix; determining one or more most predictive feature vectors among the selected representative feature vectors for each feature vector cluster based on the correlation between the selected representative feature vectors and predetermined batch binding data associated with the one or more proteins; and inputting the one or more most predictive feature vectors into a machine learning model to obtain a prediction of the molecular binding properties of the one or more proteins.

[0210] 34. The method of any one of embodiments 32 to 33, wherein the prediction of the molecular binding properties of the one or more proteins comprises a prediction of the protein binding percentage for the one or more proteins.

[0211] 35. A method according to any one of embodiments 32 to 34, wherein calculating the one or more cross-validation losses further includes: evaluating a cross-validation loss function based on the one or more most predictive feature vectors, the predetermined batch combined data, the hyperparameter set, and the learnable parameter set associated with the machine learning model; and minimizing the cross-validation loss function by changing the learnable parameter set while keeping the one or more most predictive feature vectors, the predetermined batch combined data, and the hyperparameter set constant.

[0212] 36. A method according to embodiment 35, wherein minimizing the cross-validation loss function includes optimizing a hyperparameter set, and wherein the hyperparameter set includes one or more of a regular parameter set, a booster parameter set, or a learning task parameter set.

[0213] 37. A method according to embodiment 35 or 36, wherein minimizing the cross-validation loss function includes minimizing the loss between the predicted protein binding percentage for the one or more proteins and the experimentally determined protein binding percentage for the one or more proteins.

[0214] 38. The method of any one of embodiments 35 to 37, wherein the predetermined batch binding data comprises experimentally determined protein binding percentages for one or more pH values ​​and salt concentrations associated with molecular binding properties of the one or more proteins.

[0215] 39. A method according to any one of embodiments 35 to 38, wherein the set of learnable parameters includes one or more weights or decision variables determined by a machine learning model at least in part based on the one or more most predictive feature vectors and a predetermined batch of combined data.

[0216] 40. A method according to any one of embodiments 32 to 39, wherein the molecular descriptor matrix includes a first molecular descriptor matrix representing a first set of amino acid sequences corresponding to one or more first proteins, and the prediction of molecular binding properties includes a first prediction of molecular binding properties of the one or more first proteins, the method further comprising: accessing a second molecular descriptor matrix representing a second set of amino acid sequences corresponding to one or more second proteins; and obtaining a second prediction of molecular binding properties of the one or more second proteins by a machine learning model based at least in part on the second molecular descriptor matrix.

[0217] 41. A method according to embodiment 40, wherein the machine learning model is trained to: reduce the second molecular descriptor matrix by selecting a representative feature vector for each of a second plurality of feature vector clusters of the second molecular descriptor matrix; determine one or more second most predictive feature vectors among the selected representative feature vectors for each feature vector cluster based on a second correlation between the selected representative feature vectors and predetermined batch binding data associated with the one or more second proteins; and input the one or more second most predictive feature vectors into the machine learning model, which is trained to generate a second prediction.

[0218] 42. The method of embodiment 40 or 41, wherein the second prediction of the molecular binding properties of the one or more second proteins comprises a prediction of the protein binding percentage for the one or more second proteins.

[0219] 43. A method according to any one of embodiments 32 to 42, wherein the updated hyperparameter set includes one or more of an updated regular parameter set, an updated booster parameter set, or an updated learning task parameter set.

[0220] 44. The method of any one of embodiments 32 to 43, wherein the machine learning model used to generate predictions of molecular binding properties of the one or more proteins comprises an updated set of hyperparameters.

[0221] 45. A method according to any one of embodiments 32 to 44, wherein calculating the one or more cross-validation losses includes calculating n number of cross-validation losses, and wherein n includes an integer from 1 to n.

[0222] 46. ​​A method according to any one of embodiments 32 to 45, wherein calculating the one or more cross-validation losses includes determining n number of separate training test splits based on the one or more most predictive feature vectors and a predetermined batch of combined data, and wherein n includes an integer from 1 to n.

[0223] 47. A method according to any one of embodiments 32 to 46, wherein calculating the one or more cross-validation losses includes calculating n cross-validation losses, and the method further includes: generating a prediction of the molecular binding properties of the one or more proteins based on the average of the n cross-validation losses.

[0224] 48. A method according to any one of embodiments 32 to 48, wherein the molecular descriptor matrix is ​​generated by a first machine learning model that is different from the machine learning model.

[0225] 49. A method according to embodiment 48, wherein the first machine learning model is trained to generate a molecular descriptor matrix based on a set of amino acid sequences.

[0226] 50. The method of embodiment 49, wherein the first machine learning model comprises a neural network trained to generate an M x N descriptor matrix representing a set of amino acid sequences.

[0227] 51. A method according to embodiment 49 or 50, wherein N comprises the number of amino acid sequence sets and M comprises the number of nodes in the output layer of the neural network.

[0228] 52. A method according to any one of embodiments 32 to 51, wherein the machine learning model includes one or more of a gradient boosting model, an adaptive boosting (AdaBoost) model, an extreme gradient boosting (XGBoost) model, a lightweight gradient boosting machine (LightGBM) model, or a classification boosting (CatBoost) model.

[0229] 53. The method of any one of embodiments 32 to 52, wherein the machine learning model is further trained to generate predictions of the molecular elution characteristics of the one or more proteins.

[0230] 54. The method of any one of embodiments 32 to 53, wherein the machine learning model is further trained to generate predictions of the flow-through properties of the one or more proteins.

[0231] 55. The method of any one of embodiments 32 to 54, wherein reducing the molecular descriptor matrix comprises clustering similar feature vectors into a plurality of feature vector clusters based on correlation distance.

[0232] 56. A method according to embodiment 55, wherein Pearson correlation is used to calculate the correlation distance.

[0233] 57. A method according to embodiment 55 or 56, wherein the selected representative feature vector for each of the multiple feature vector clusters includes a centroid feature vector for each of the multiple feature vector clusters, used to represent two or more of the similar feature vectors.

[0234] 58. A method according to any one of embodiments 32 to 57, wherein determining the one or more predictive feature vectors in the selected representative feature vectors for each of the multiple feature vector clusters includes: selecting a k-best matrix of feature vectors in the selected representative feature vectors for each of the multiple feature vector clusters.

[0235] 59. The method of embodiment 58, wherein a k-best matrix of feature vectors in the selected representative feature vectors is determined based on a predetermined k-best process.

[0236] 60. A method according to any one of embodiments 32 to 59, wherein the correlation between the selected representative feature vector and the predetermined batch of combined data is determined based on a maximum information coefficient (MIC) between the selected representative feature vector and the predetermined batch of combined data.

[0237] 61. The method of any one of embodiments 32 to 60, wherein said prediction of said molecular binding properties of said one or more proteins comprises a chromatographic process based on a computational model.

[0238] 62. The method of embodiment 61, wherein the computational model-based chromatography process comprises one or more of a computational model-based affinity chromatography process, an ion exchange chromatography (IEX) process, a hydrophobic interaction chromatography (HIC) process, or a mixed mode chromatography (MMC) process.

[0239] 63. A method according to any one of embodiments 32 to 62, further comprising optimizing the machine learning model based on a Bayesian model optimization process.

[0240] 64. The method according to embodiment 63 further includes utilizing grouped K-fold cross validation to train and evaluate the optimized machine learning model based on the one or more most predictive feature vectors, the predetermined batch combined data, the hyperparameter set, and the learnable parameter set.

[0241] 65. The method according to embodiment 63 or 64 further includes utilizing stratified K-fold cross validation to train and evaluate the optimized machine learning model based on the one or more most predictive feature vectors, the predetermined batch combined data, the hyperparameter set, and the learnable parameter set.

[0242] 66. The method of any one of embodiments 32 to 65, wherein the prediction of the molecular binding properties of the one or more proteins comprises identification of a target protein among the one or more proteins.

[0243] 67. The method of any one of embodiments 32 to 66, wherein the prediction of the molecular binding properties of the one or more proteins comprises quantitative structure-property relationship (QSPR) or quantitative structure-activity relationship (QSAR) modeling of the one or more proteins.

[0244] 68. A method according to any one of embodiments 32 to 67, wherein the prediction of molecular binding properties of the one or more proteins includes predicting the molecular binding properties of each amino acid sequence in the set of amino acid sequences corresponding to the one or more proteins.

[0245] 69. The method of any one of embodiments 32 to 68, wherein for each of the one or more empirically evaluated proteins, a corresponding predetermined bulk binding is measured for each experimental condition in a set of experimental conditions.

[0246] 70. The method of embodiment 69, wherein the set of experimental conditions comprises 24 experimental conditions.

[0247] 71. The method of embodiment 70, wherein the set of experimental conditions comprises a first subset of salt concentrations and a second subset of pH values.

[0248] 72. A method according to embodiment 70 or 71, wherein a set of experimental conditions is input into a machine learning model together with a molecular descriptor matrix, and the prediction of the molecular binding properties of the one or more proteins includes the prediction of the molecular binding properties of the one or more proteins for each experimental condition in the set of experimental conditions.

[0249] 73. The method of any one of embodiments 32 to 72, further comprising: transforming the prediction of the molecular binding properties of the one or more proteins into a linear representation.

[0250] 74. A method according to embodiment 73, wherein a logit transformation is used to generate a linear representation.

[0251] 75. The method of embodiment 73 or 74, further comprising: performing a principal component analysis (PCA) on the linear representation to obtain at least a first principal component.

[0252] 76. A method according to any one of embodiments 32 to 75, wherein the predetermined batch binding data associated with the one or more empirically evaluated proteins includes experimentally determined binding values ​​measured for each of the one or more empirically evaluated proteins for each experimental condition in a set of experimental conditions.

[0253] 77. A method according to embodiment 76, wherein the correlation between the selected representative feature vector and the predetermined batch binding data includes: for each of the one or more empirically evaluated proteins and for each experimental condition in the set of experimental conditions: generating a linear representation of the experimentally determined binding value of the empirically evaluated protein based on a logit transformation of the experimentally determined binding value applied to the empirically evaluated protein; and performing a principal component analysis (PCA) on the linear representation of the experimentally determined binding values ​​of the one or more empirically evaluated proteins to obtain at least a first principal component.

[0254] 78. The method of embodiment 77, further comprising: generating training predictions for the molecular binding properties of the one or more empirically evaluated proteins using a machine learning model; and comparing the training predictions with the first principal component to calculate one or more cross-validation losses.

[0255] 79. The method of embodiment 77 or 78, wherein the first principal component describes the average batch binding value.

[0256] 80. The method of any one of embodiments 32 to 79, further comprising: generating a set of functions representing the behavior of the one or more proteins for a set of experimental conditions based on predictions; and selecting at least one of the one or more proteins for use in one or more drug discovery assays based on the behavior of the one or more proteins for the set of experimental conditions.

[0257] 81. A method according to any one of embodiments 32 to 80, wherein the correlation between the selected representative feature vector and predetermined batch binding data associated with the one or more empirically evaluated proteins includes: the correlation between the representative feature vector and a principal component calculated based on the predetermined batch binding data.

[0258] 82. The method of embodiment 81, wherein the one or more cross-validation losses are calculated based on predicted molecular binding properties and empirical molecular binding properties.

[0259] 83. The method of embodiment 82, wherein the predicted molecular binding properties comprise principal components calculated based on representative eigenvectors, and the empirical molecular binding properties comprise principal components calculated based on predetermined batch binding data.

[0260] 84. A method according to any one of embodiments 32 to 33, wherein determining the one or more most predictive feature vectors further comprises: (i) fitting a model to the representative feature vectors; (ii) calculating a feature importance score for each of the representative feature vectors based on the model; and (iii) removing one or more feature vectors from the representative feature vectors based on the feature importance score of each of the representative feature vectors to obtain a subset of representative feature vectors, wherein the one or more most predictive feature vectors include one or more feature vectors from the subset of representative feature vectors.

[0261] 85. The method according to embodiment 84 further comprises: iteratively performing steps (i) to (iii) until the number of feature vectors included in the subset meets the feature number criterion.

[0262] 86. A method according to embodiment 85, wherein the feature quantity criterion is satisfied including the number of feature vectors included in the representative feature vector subset being less than or equal to a threshold number of feature vectors.

[0263] 87. A method according to embodiment 86, wherein the threshold number of feature vectors includes the same or similar number of features from training data used to train the machine learning model.

[0264] 88. A method according to any one of embodiments 84 to 87, wherein the number of feature vectors included in the representative feature vector subset includes a hyperparameter in the hyperparameter set.

[0265] 89. A method according to any one of embodiments 32 to 88, wherein a hyperparameter in the hyperparameter set represents the number of feature vector clusters included in the multiple feature vector clusters.

[0266] 90. A system comprising one or more computing devices, the system further comprising: one or more non-temporary computer-readable storage media comprising instructions; and one or more processors coupled to the one or more storage media, the one or more processors being configured to execute instructions to implement a method according to any one of embodiments 1 to 89.

[0267] 91. A non-transitory computer-readable medium comprising instructions that, when executed by one or more processors of one or more computing devices, cause the one or more processors to perform operations including the method of any one of embodiments 1 to 89.

Claims

1. A method for predicting molecular binding properties of one or more proteins, the method comprising: accessing a molecular descriptor matrix representing a set of amino acid sequences corresponding to one or more proteins; as well as refining a set of hyperparameters associated with a machine learning model trained to generate predictions of molecular binding properties of the one or more proteins, wherein refining the set of hyperparameters comprises iteratively performing a process until a desired accuracy is achieved, the process comprising: reducing the molecular descriptor matrix by selecting a representative feature vector for each of a plurality of feature vector clusters, wherein each feature vector cluster includes similar feature vectors; determining one or more most predictive feature vectors from the selected representative feature vectors for each feature vector cluster based on correlations between the selected representative feature vectors and predetermined bulk binding data associated with the one or more proteins; calculating one or more cross validation losses based at least in part on the one or more most predictive feature vectors and the predetermined batch of combined data; and updating the set of hyperparameters based on the one or more cross-validation losses; and outputting, by the machine learning model, the prediction of the molecular binding property of the one or more proteins based at least in part on the updated set of hyperparameters.

2. The method of claim 1, wherein calculating the one or more cross-validation losses further comprises: evaluating a cross-validation loss function based on the one or more most predictive feature vectors, the predetermined batch of combined data, the set of hyperparameters, and the set of learnable parameters associated with the machine learning model; and The cross validation loss function is minimized by varying the set of learnable parameters while keeping the one or more most predictive feature vectors, the predetermined batch bound data, and the set of hyperparameters constant.

3. The method of claim 2, wherein minimizing the cross-validation loss function comprises optimizing the hyperparameter set, and wherein the hyperparameter set comprises one or more of a regular parameter set, a booster parameter set, or a learning task parameter set.

4. The method of claim 2, wherein minimizing the cross-validation loss function comprises minimizing the loss between the predicted protein binding percentage for the one or more proteins and the experimentally determined protein binding percentage for the one or more proteins.

5. The method of claim 2, wherein the predetermined bulk binding data comprises experimentally determined protein binding percentages for one or more pH values ​​and salt concentrations associated with the molecular binding properties of the one or more proteins.

6. The method of claim 2, wherein the set of learnable parameters comprises one or more weights or decision variables determined by the machine learning model based at least in part on the one or more most predictive feature vectors and the predetermined batch of combined data.

7. The method according to claim 1, further comprising: After refining the hyperparameter set: accessing a second molecular descriptor matrix representing a second set of amino acid sequences corresponding to one or more second proteins; reducing the second molecular descriptor matrix by selecting a representative eigenvector for each of a second plurality of eigenvector clusters of the second molecular descriptor matrix; determining one or more second most predictive feature vectors from the selected representative feature vectors for each feature vector cluster based on a second correlation between the selected representative feature vectors and predetermined bulk binding data associated with the one or more second proteins; inputting the one or more second most predictive feature vectors into the machine learning model trained to generate predictions of molecular binding properties of the one or more second proteins; as well as The prediction of the molecular binding property of the one or more second proteins is output by the machine learning model based at least in part on the updated set of hyperparameters.

8. The method of claim 7, wherein the prediction of the molecular binding properties of the one or more second proteins comprises a prediction of a protein binding percentage for the one or more second proteins.

9. The method of claim 1, wherein the updated hyperparameter set comprises one or more of an updated regular parameter set, an updated booster parameter set, or an updated learning task parameter set. 10 . The method of claim 1 , wherein calculating the one or more cross-validation losses comprises calculating n number of cross-validation losses, and wherein n comprises an integer from 1 to n.

11. The method of claim 1 , wherein calculating the one or more cross-validation losses comprises determining n number of separate training-test splits based on the one or more most predictive feature vectors and the predetermined batch combined data, and wherein n comprises an integer from 1 to n.

12. The method of claim 1, wherein calculating the one or more cross-validation losses comprises calculating n cross-validation losses, the method further comprising: The prediction of the molecular binding property of the one or more proteins is generated based on averaging the n number of cross-validation losses.

13. The method of claim 1, wherein the molecular descriptor matrix comprises 2 n feature vectors, and where n comprises the dimension of the molecular descriptor matrix.

14. The method of claim 1, wherein the molecular descriptor matrix is ​​generated by a first machine learning model different from the machine learning model.

15. The method of claim 14, wherein the first machine learning model is trained to generate the molecular descriptor matrix based on the set of amino acid sequences.

16. The method of claim 14, wherein the first machine learning model comprises a neural network trained to generate an M x N descriptor matrix representing the set of amino acid sequences, and wherein N comprises the number of the sets of amino acid sequences and M comprises the number of nodes in an output layer of the neural network.

17. The method of claim 1, wherein the machine learning model comprises one or more of a gradient boosting model, an adaptive boosting (AdaBoost) model, an extreme gradient boosting (XGBoost) model, a lightweight gradient boosting machine (LightGBM) model, or a classification boosting (CatBoost) model.

18. The method of claim 1, wherein the machine learning model is further trained to generate predictions of molecular elution characteristics of the one or more proteins.

19. The method of claim 1, wherein the machine learning model is further trained to generate predictions of flow-through properties of the one or more proteins.

20. The method of claim 1, wherein reducing the molecular descriptor matrix comprises performing a Pearson correlation of eigenvectors of the molecular descriptor matrix to generate the plurality of eigenvector clusters.

21. The method of claim 20, wherein the selected one representative feature vector for each of the plurality of feature vector clusters comprises a centroid feature vector for each of the plurality of feature vector clusters for representing two or more of the similar feature vectors.

22. The method of claim 1, wherein determining the one or more representative ones of the selected representative feature vectors for each of the plurality of feature vector clusters comprises: A k-best matrix of eigenvectors in the selected representative eigenvectors for each of the plurality of eigenvector clusters is selected.

23. The method of claim 22, wherein the k-best matrix of feature vectors in the selected representative feature vectors is determined based on a predetermined k-best process.

24. The method of claim 1, wherein the correlation between the selected representative feature vector and the predetermined batch of combined data is determined based on a maximum information coefficient (MIC) between the selected representative feature vector and the predetermined batch of combined data.

25. The method of claim 1, wherein the prediction of the molecular binding properties of the one or more proteins comprises a chromatographic process based on a computational model.

26. The method of claim 25, wherein the computational model-based chromatography process comprises one or more of a computational model-based affinity chromatography process, an ion exchange chromatography (IEX) process, a hydrophobic interaction chromatography (HIC) process, or a mixed mode chromatography (MMC) process.

27. The method of claim 1, further comprising optimizing the machine learning model based on a Bayesian model optimization process.

28. The method of claim 27, further comprising training and evaluating the optimized machine learning model based on the one or more most predictive feature vectors, the predetermined batch combined data, the set of hyperparameters, and the set of learnable parameters using grouped K-fold cross validation.

29. The method of claim 1, wherein the prediction of the molecular binding properties of the one or more proteins comprises identification of a target protein among the one or more proteins.

30. The method of claim 1, wherein the prediction of the molecular binding properties of the one or more proteins comprises quantitative structure-property relationship (QSPR) or quantitative structure-activity relationship (QSAR) modeling of the one or more proteins.

31. The method of claim 1, wherein the prediction of the molecular binding properties of the one or more proteins comprises a prediction of the molecular binding properties of each amino acid sequence in the set of amino acid sequences corresponding to the one or more proteins.

32. A system comprising one or more computing devices, the system further comprising: one or more non-transitory computer-readable storage media including instructions; as well as One or more processors coupled to the one or more storage media, the one or more processors configured to execute the instructions to: accessing a molecular descriptor matrix representing a set of amino acid sequences corresponding to one or more proteins; as well as refining a set of hyperparameters associated with a machine learning model trained to generate predictions of molecular binding properties of the one or more proteins, wherein refining the set of hyperparameters comprises iteratively performing a process until a desired accuracy is achieved, the process comprising instructions to: reducing the molecular descriptor matrix by selecting a representative eigenvector for each of a plurality of eigenvector clusters, wherein each of the eigenvector clusters includes similar eigenvectors; determining one or more representative feature vectors of the selected representative feature vectors for each of the plurality of feature vector clusters based on correlations between the selected representative feature vectors and predetermined bulk binding data associated with the one or more proteins; calculating one or more cross validation losses based at least in part on the one or more most predictive feature vectors and the predetermined batch of combined data; as well as updating the set of hyperparameters based on the one or more cross-validation losses; as well as The prediction of the molecular binding property of the one or more proteins is output by the machine learning model based at least in part on the updated set of hyperparameters.

33. The system of claim 32, wherein the instructions to calculate the one or more cross validation losses further comprise instructions to: evaluating a cross-validation loss function based on the one or more most predictive feature vectors, the predetermined batch of combined data, the set of hyperparameters, and the set of learnable parameters associated with the machine learning model; and The cross validation loss function is minimized by varying the set of learnable parameters while keeping the one or more most predictive feature vectors, the predetermined batch bound data, and the set of hyperparameters constant.

34. A system according to claim 33, wherein the instructions for minimizing the cross-validation loss function include instructions for optimizing the hyperparameter set, and wherein the hyperparameter set includes one or more of a regular parameter set, a booster parameter set, or a learning task parameter set.

35. The system of claim 33, wherein the instructions to minimize the cross-validation loss function include instructions to minimize a loss between a prediction of a protein binding percentage for the one or more proteins and an experimentally determined protein binding percentage for the one or more proteins.

36. The system of claim 33, wherein the predetermined bulk binding data comprises experimentally determined protein binding percentages for one or more pH values ​​and salt concentrations associated with the molecular binding properties of the one or more proteins.

37. The system of claim 33, wherein the set of learnable parameters comprises one or more weights or decision variables determined by the machine learning model based at least in part on the one or more most predictive feature vectors and the predetermined batch of combined data.

38. The system of claim 32, wherein the instructions further comprise instructions to: After refining the hyperparameter set: accessing a second molecular descriptor matrix representing a second set of amino acid sequences corresponding to one or more second proteins; reducing the second molecular descriptor matrix by selecting a representative eigenvector for each of a second plurality of eigenvector clusters of the second molecular descriptor matrix; determining one or more second most predictive feature vectors from the selected representative feature vectors for each feature vector cluster based on a second correlation between the selected representative feature vectors and predetermined bulk binding data associated with the one or more second proteins; inputting the one or more second most predictive feature vectors into the machine learning model trained to generate predictions of molecular binding properties of the one or more second proteins; as well as The prediction of the molecular binding property of the one or more second proteins is output by the machine learning model based at least in part on the updated set of hyperparameters.

39. The system of claim 38, wherein the prediction of the molecular binding property of the one or more second proteins comprises a prediction of a protein binding percentage for the one or more second proteins.

40. The system of claim 32, wherein the updated hyperparameter set comprises one or more of an updated regular parameter set, an updated booster parameter set, or an updated learning task parameter set.

41. The system of claim 32, wherein the instructions to calculate the one or more cross-validation losses further comprise instructions to calculate n number of cross-validation losses, and wherein n comprises an integer from 1 to n.

42. The system of claim 32, wherein the instructions for calculating the one or more cross-validation losses further comprise instructions for determining n number of separate training-test splits based on the one or more most predictive feature vectors and the predetermined batch combined data, and wherein n comprises an integer from 1 to n.

43. The system of claim 32, wherein the instructions to calculate the one or more cross-validation losses further include instructions to calculate n number of cross-validation losses, the instructions further include instructions to: The prediction of the molecular binding property of the one or more proteins is generated based on averaging the n number of cross-validation losses.

44. The system of claim 32, wherein the molecular descriptor matrix comprises 2 n feature vectors, and where n comprises the dimension of the molecular descriptor matrix.

45. The system of claim 32, wherein the molecular descriptor matrix is ​​generated by a first machine learning model different from the machine learning model.

46. ​​The system of claim 45, wherein the first machine learning model is trained to generate the molecular descriptor matrix based on the set of amino acid sequences.

47. The system of claim 45, wherein the first machine learning model comprises a neural network trained to generate an M x N descriptor matrix representing the set of amino acid sequences, and wherein N comprises the number of the sets of amino acid sequences and M comprises the number of nodes in an output layer of the neural network.

48. The system of claim 32, wherein the machine learning model comprises one or more of a gradient boosting model, an adaptive boosting (AdaBoost) model, an extreme gradient boosting (XGBoost) model, a lightweight gradient boosting machine (LightGBM) model, or a classification boosting (CatBoost) model.

49. The system of claim 32, wherein the machine learning model is further trained to generate predictions of molecular elution characteristics of the one or more proteins.

50. The system of claim 32, wherein the machine learning model is further trained to generate predictions of flow-through properties of the one or more proteins.

51. The system of claim 32, wherein the instructions to reduce the molecular descriptor matrix further comprise instructions to perform a Pearson correlation of eigenvectors of the molecular descriptor matrix to generate the plurality of eigenvector clusters.

52. The system of claim 51, wherein the selected one representative feature vector for each of the plurality of feature vector clusters comprises a centroid feature vector for each of the plurality of feature vector clusters for representing two or more of the similar feature vectors.

53. The system of claim 32, wherein the instructions to determine the one or more representative ones of the selected representative feature vectors for each of the plurality of feature vector clusters further comprise: Instructions to select a k-best matrix of eigenvectors in the selected representative eigenvectors for each of the plurality of eigenvector clusters.

54. The system of claim 53, wherein the k-best matrix of feature vectors in the selected representative feature vectors is determined based on a predetermined k-best process.

55. The system of claim 32, wherein the correlation between the selected representative feature vector and the predetermined batch of combined data is determined based on a maximum information coefficient (MIC) between the selected representative feature vector and the predetermined batch of combined data.

56. The system of claim 32, wherein the prediction of the molecular binding properties of the one or more proteins comprises a chromatographic process based on a computational model.

57. The system of claim 56, wherein the computational model-based chromatography process comprises one or more of a computational model-based affinity chromatography process, an ion exchange chromatography (IEX) process, a hydrophobic interaction chromatography (HIC) process, or a mixed mode chromatography (MMC) process.

58. The system of claim 32, wherein the instructions further comprise instructions for optimizing the machine learning model based on a Bayesian model optimization process.

59. The system of claim 58, wherein the instructions further comprise instructions for training and evaluating the optimized machine learning model based on the one or more most predictive feature vectors, the predetermined batch combined data, the set of hyperparameters, and the set of learnable parameters using grouped K-fold cross validation.

60. The system of claim 32, wherein the prediction of the molecular binding properties of the one or more proteins comprises identification of a target protein among the one or more proteins.

61. The system of claim 32, wherein the prediction of the molecular binding properties of the one or more proteins comprises quantitative structure-property relationship (QSPR) or quantitative structure-activity relationship (QSAR) modeling of the one or more proteins.

62. The system of claim 32, wherein the prediction of the molecular binding properties of the one or more proteins comprises a prediction of the molecular binding properties of each amino acid sequence in the set of amino acid sequences corresponding to the one or more proteins.

63. A non-transitory computer readable medium comprising instructions that, when executed by one or more processors of one or more computing devices, cause the one or more processors to: accessing a molecular descriptor matrix representing a set of amino acid sequences corresponding to one or more proteins; as well as refining a set of hyperparameters associated with a machine learning model trained to generate predictions of molecular binding properties of the one or more proteins, wherein refining the set of hyperparameters comprises iteratively performing a process until a desired accuracy is achieved, the process comprising instructions to: reducing the molecular descriptor matrix by selecting a representative eigenvector for each of a plurality of eigenvector clusters, wherein each of the eigenvector clusters includes similar eigenvectors; determining one or more representative feature vectors of the selected representative feature vectors for each of the plurality of feature vector clusters based on correlations between the selected representative feature vectors and predetermined bulk binding data associated with the one or more proteins; calculating one or more cross validation losses based at least in part on the one or more most predictive feature vectors and the predetermined batch of combined data; as well as updating the set of hyperparameters based on the one or more cross-validation losses; and outputting, by the machine learning model, the prediction of the molecular binding property of the one or more proteins based at least in part on the updated set of hyperparameters.

64. The non-transitory computer readable medium of claim 63, wherein the instructions to calculate the one or more cross validation losses further comprise instructions to: evaluating a cross-validation loss function based on the one or more most predictive feature vectors, the predetermined batch of combined data, the set of hyperparameters, and the set of learnable parameters associated with the machine learning model; and The cross validation loss function is minimized by varying the set of learnable parameters while keeping the one or more most predictive feature vectors, the predetermined batch bound data, and the set of hyperparameters constant.

65. A non-transitory computer-readable medium according to claim 64, wherein the instructions for minimizing the cross-validation loss function include instructions for optimizing the hyperparameter set, and wherein the hyperparameter set includes one or more of a regular parameter set, a booster parameter set, or a learning task parameter set.

66. The non-transitory computer-readable medium of claim 64, wherein the instructions to minimize the cross-validation loss function comprise instructions to minimize the loss between a prediction of a protein binding percentage for the one or more proteins and an experimentally determined protein binding percentage for the one or more proteins.

67. The non-transitory computer readable medium of claim 64, wherein the predetermined batch binding data comprises experimentally determined protein binding percentages for one or more pH values ​​and salt concentrations associated with the molecular binding properties of the one or more proteins.

68. The non-transitory computer-readable medium of claim 64, wherein the set of learnable parameters comprises one or more weights or decision variables determined by the machine learning model based at least in part on the one or more most predictive feature vectors and the predetermined batch of combined data.

69. The non-transitory computer readable medium of claim 63, wherein the instructions further comprise instructions to: After refining the hyperparameter set: accessing a second molecular descriptor matrix representing a second set of amino acid sequences corresponding to one or more second proteins; reducing the second molecular descriptor matrix by selecting a representative eigenvector for each of a second plurality of eigenvector clusters of the second molecular descriptor matrix; determining one or more second most predictive feature vectors from the selected representative feature vectors for each feature vector cluster based on a second correlation between the selected representative feature vectors and predetermined bulk binding data associated with the one or more second proteins; inputting the one or more second most predictive feature vectors into the machine learning model trained to generate predictions of molecular binding properties of the one or more second proteins; as well as The prediction of the molecular binding property of the one or more second proteins is output by the machine learning model based at least in part on the updated set of hyperparameters.

70. The non-transitory computer readable medium of claim 68, wherein the prediction of the molecular binding property of the one or more second proteins comprises a prediction of a second protein binding percentage for the one or more second proteins.

71. The non-transitory computer-readable medium of claim 63, wherein the updated hyperparameter set comprises one or more of an updated regular parameter set, an updated booster parameter set, or an updated learning task parameter set.

72. The non-transitory computer-readable medium of claim 63, wherein the instructions to calculate the one or more cross-validation losses further comprise instructions to calculate n number of cross-validation losses, and wherein n comprises an integer from 1 to n.

73. The non-transitory computer-readable medium of claim 63, wherein the instructions for calculating the one or more cross-validation losses further comprise instructions for determining n number of separate training-test splits based on the one or more most predictive feature vectors and the predetermined batch combined data, and wherein n comprises an integer from 1 to n.

74. The non-transitory computer readable medium of claim 63, wherein the instructions to calculate the one or more cross validation losses further comprise instructions to calculate n number of cross validation losses, the instructions further comprising instructions to: The prediction of the molecular binding property of the one or more proteins is generated based on averaging the n number of cross-validation losses.

75. The non-transitory computer readable medium of claim 63, wherein the molecular descriptor matrix comprises 2 n feature vectors, and where n comprises the dimension of the molecular descriptor matrix.

76. The non-transitory computer-readable medium of claim 63, wherein the molecular descriptor matrix is ​​generated by a first machine learning model different from the machine learning model.

77. The non-transitory computer-readable medium of claim 76, wherein the first machine learning model is trained to generate the molecular descriptor matrix based on the set of amino acid sequences.

78. The non-transitory computer-readable medium of claim 76, wherein the first machine learning model comprises a neural network trained to generate an M x N descriptor matrix representing the set of amino acid sequences, and wherein N comprises the number of the sets of amino acid sequences and M comprises the number of nodes in an output layer of the neural network.

79. A non-transitory computer-readable medium according to claim 63, wherein the machine learning model comprises one or more of a gradient boosting model, an adaptive boosting (AdaBoost) model, an extreme gradient boosting (XGBoost) model, a lightweight gradient boosting machine (LightGBM) model, or a classification boosting (CatBoost) model.

80. The non-transitory computer readable medium of claim 63, wherein the machine learning model is further trained to generate predictions of molecular elution characteristics of the one or more proteins.

81. The non-transitory computer readable medium of claim 63, wherein the machine learning model is further trained to generate predictions of flow-through properties of the one or more proteins.

82. The non-transitory computer readable medium of claim 63, wherein the instructions to reduce the molecular descriptor matrix further comprise instructions to perform a Pearson correlation of eigenvectors of the molecular descriptor matrix to generate the plurality of eigenvector clusters.

83. The non-transitory computer-readable medium of claim 82, wherein the selected one representative feature vector for each of the plurality of feature vector clusters comprises a centroid feature vector for each of the plurality of feature vector clusters representing two or more of the similar feature vectors.

84. The non-transitory computer-readable medium of claim 63, wherein the instructions to determine the one or more representative ones of the selected representative feature vectors for each of the plurality of feature vector clusters further comprise: Instructions to select a k-best matrix of eigenvectors in the selected representative eigenvectors for each of the plurality of eigenvector clusters.

85. The non-transitory computer-readable medium of claim 84, wherein the k-best matrix of eigenvectors in the selected representative eigenvectors is determined based on a predetermined k-best procedure.

86. The non-transitory computer-readable medium of claim 63, wherein the correlation between the selected representative feature vector and the predetermined batch of combined data is determined based on a maximum information coefficient (MIC) between the selected representative feature vector and the predetermined batch of combined data.

87. The non-transitory computer readable medium of claim 63, wherein the prediction of the molecular binding properties of the one or more proteins comprises a computational model-based column chromatography process.

88. The non-transitory computer readable medium of claim 87, wherein the computational model-based chromatography process comprises one or more of a computational model-based affinity chromatography process, an ion exchange chromatography (IEX) process, a hydrophobic interaction chromatography (HIC) process, or a mixed mode chromatography (MMC) process.

89. The non-transitory computer-readable medium of claim 63, wherein the instructions further comprise instructions for optimizing the machine learning model based on a Bayesian model optimization process.

90. The non-transitory computer-readable medium of claim 89, wherein the instructions further comprise instructions for training and evaluating the optimized machine learning model based on the one or more most predictive feature vectors, the predetermined batch combined data, the set of hyperparameters, and the set of learnable parameters using grouped K-fold cross validation.

91. The non-transitory computer readable medium of claim 63, wherein the prediction of the molecular binding properties of the one or more proteins comprises identification of a target protein among the one or more proteins.

92. The non-transitory computer-readable medium of claim 63, wherein the prediction of the molecular binding properties of the one or more proteins comprises quantitative structure-property relationship (QSPR) or quantitative structure-activity relationship (QSAR) modeling of the one or more proteins.

93. The non-transitory computer-readable medium of claim 63, wherein the prediction of the molecular binding properties of the one or more proteins comprises a prediction of the molecular binding properties of each amino acid sequence in the set of amino acid sequences corresponding to the one or more proteins.

94. A method for predicting molecular binding properties of one or more proteins, the method comprising: accessing a molecular descriptor matrix representing a set of amino acid sequences corresponding to one or more proteins; as well as and obtaining, by a machine learning model, predictions of molecular binding properties of the one or more proteins based at least in part on the molecular descriptor matrix, wherein the machine learning model is trained by: accessing a training molecular descriptor matrix representing a training set of amino acid sequences corresponding to one or more empirically evaluated proteins; as well as Iteratively performing a process to refine a set of hyperparameters associated with the machine learning model until a desired accuracy is achieved, the process comprising: reducing the training molecular descriptor matrix by selecting a representative feature vector for each of a plurality of feature vector clusters, wherein each feature vector cluster includes similar feature vectors; determining one or more most predictive feature vectors from the selected representative feature vectors for each feature vector cluster based on correlations between the selected representative feature vectors and predetermined bulk binding data associated with the one or more empirically evaluated proteins; and calculating one or more cross validation losses based at least in part on the one or more most predictive feature vectors and the predetermined batch of combined data; and The set of hyperparameters is updated based on the one or more cross-validation losses.

95. The method of claim 94, wherein obtaining the prediction comprises: reducing the molecular descriptor matrix by selecting a representative eigenvector for each of a plurality of eigenvector clusters of the molecular descriptor matrix; determining one or more most predictive feature vectors from the selected representative feature vectors for each feature vector cluster based on correlations between the selected representative feature vectors and predetermined bulk binding data associated with the one or more proteins; The one or more most predictive feature vectors are input into the machine learning model to obtain the prediction of the molecular binding property of the one or more proteins.

96. The method of claim 94, wherein the prediction of the molecular binding properties of the one or more proteins comprises a prediction of a percent protein binding to the one or more proteins.

97. The method of claim 94, wherein calculating the one or more cross validation losses further comprises: evaluating a cross-validation loss function based on the one or more most predictive feature vectors, the predetermined batch of combined data, the set of hyperparameters, and the set of learnable parameters associated with the machine learning model; and The cross validation loss function is minimized by varying the set of learnable parameters while keeping the one or more most predictive feature vectors, the predetermined batch bound data, and the set of hyperparameters constant.

98. The method of claim 97, wherein minimizing the cross-validation loss function comprises optimizing the hyperparameter set, and wherein the hyperparameter set comprises one or more of a regular parameter set, a booster parameter set, or a learning task parameter set.

99. The method of claim 97, wherein minimizing the cross-validation loss function comprises minimizing the loss between the predicted protein binding percentage for the one or more proteins and the experimentally determined protein binding percentage for the one or more proteins.

100. The method of claim 97, wherein the predetermined batch binding data comprises experimentally determined protein binding percentages for one or more pH values ​​and salt concentrations associated with the molecular binding properties of the one or more proteins.

101. A method according to claim 97, wherein the set of learnable parameters includes one or more weights or decision variables determined by the machine learning model based at least in part on the one or more most predictive feature vectors and the predetermined batch of combined data.

102. The method of claim 94, wherein the molecular descriptor matrix comprises a first molecular descriptor matrix representing a first set of amino acid sequences corresponding to one or more first proteins, and the prediction of the molecular binding property comprises a first prediction of a molecular binding property of the one or more first proteins, the method further comprising: accessing a second molecular descriptor matrix representing a second set of amino acid sequences corresponding to one or more second proteins; as well as A second prediction of molecular binding properties of the one or more second proteins is obtained by the machine learning model based at least in part on the second molecular descriptor matrix.

103. The method of claim 102, wherein the machine learning model is trained to: reduce the second molecular descriptor matrix by selecting a representative eigenvector for each of a second plurality of eigenvector clusters of the second molecular descriptor matrix; determining one or more second most predictive feature vectors from the selected representative feature vectors for each feature vector cluster based on a second correlation between the selected representative feature vectors and predetermined bulk binding data associated with the one or more second proteins; The one or more second most predictive feature vectors are input into the machine learning model, which is trained to generate the second prediction.

104. The method of claim 102, wherein the second prediction of the molecular binding property of the one or more second proteins comprises a prediction of a protein binding percentage for the one or more second proteins.

105. The method of claim 94, wherein the updated hyperparameter set comprises one or more of an updated regular parameter set, an updated booster parameter set, or an updated learning task parameter set.

106. The method of claim 94, wherein the machine learning model used to generate the prediction of the molecular binding properties of the one or more proteins comprises an updated set of hyperparameters.

107. The method of claim 94, wherein calculating the one or more cross-validation losses comprises calculating n number of cross-validation losses, and wherein n comprises an integer from 1 to n.

108. The method of claim 94, wherein calculating the one or more cross-validation losses comprises determining n number of separate training-test splits based on the one or more most predictive feature vectors and the predetermined batch combined data, and wherein n comprises an integer from 1 to n.

109. The method of claim 94, wherein calculating the one or more cross-validation losses comprises calculating n cross-validation losses, the method further comprising: The prediction of the molecular binding property of the one or more proteins is generated based on averaging the n number of cross-validation losses.

110. The method of claim 94, wherein the molecular descriptor matrix is ​​generated by a first machine learning model that is different from the machine learning model.

111. The method of claim 110, wherein the first machine learning model is trained to generate the molecular descriptor matrix based on the set of amino acid sequences.

112. The method of claim 111, wherein the first machine learning model comprises a neural network trained to generate an M x N descriptor matrix representing the set of amino acid sequences.

113. The method of claim 112, wherein N comprises the number of said sets of amino acid sequences and M comprises the number of nodes in the output layer of said neural network.

114. The method of claim 94, wherein the machine learning model comprises one or more of a gradient boosting model, an adaptive boosting (AdaBoost) model, an extreme gradient boosting (XGBoost) model, a lightweight gradient boosting machine (LightGBM) model, or a classification boosting (CatBoost) model.

115. The method of claim 94, wherein the machine learning model is further trained to generate predictions of molecular elution characteristics of the one or more proteins.

116. The method of claim 94, wherein the machine learning model is further trained to generate predictions of flow-through properties of the one or more proteins.

117. The method of claim 94, wherein reducing the molecular descriptor matrix comprises clustering the similar feature vectors into the plurality of feature vector clusters based on correlation distance.

118. A method according to claim 117, wherein the correlation distance is calculated using Pearson correlation.

119. The method of claim 117, wherein the selected one representative feature vector for each of the plurality of feature vector clusters comprises a centroid feature vector for each of the plurality of feature vector clusters for representing two or more of the similar feature vectors.

120. The method of claim 94, wherein determining the one or more predictive feature vectors in the selected representative feature vectors for each of the plurality of feature vector clusters comprises: A k-best matrix of eigenvectors in the selected representative eigenvectors for each of the plurality of eigenvector clusters is selected.

121. The method of claim 120, wherein the k-best matrix of feature vectors in the selected representative feature vectors is determined based on a predetermined k-best process.

122. The method of claim 94, wherein the correlation between the selected representative feature vector and the predetermined batch combined data is determined based on a maximum information coefficient (MIC) between the selected representative feature vector and the predetermined batch combined data.

123. The method of claim 94, wherein the prediction of the molecular binding properties of the one or more proteins comprises a chromatographic process based on a computational model.

124. The method of claim 123, wherein the computational model-based chromatography process comprises one or more of a computational model-based affinity chromatography process, an ion exchange chromatography (IEX) process, a hydrophobic interaction chromatography (HIC) process, or a mixed mode chromatography (MMC) process.

125. The method of claim 94, further comprising optimizing the machine learning model based on a Bayesian model optimization process.

126. The method of claim 125, further comprising training and evaluating the optimized machine learning model based on the one or more most predictive feature vectors, the predetermined batch combined data, the set of hyperparameters, and the set of learnable parameters using grouped K-fold cross validation.

127. The method of claim 125, further comprising training and evaluating the optimized machine learning model based on the one or more most predictive feature vectors, the predetermined batch combined data, the set of hyperparameters, and the set of learnable parameters using stratified K-fold cross validation.

128. The method of claim 94, wherein the prediction of the molecular binding properties of the one or more proteins comprises identification of a target protein among the one or more proteins.

129. The method of claim 94, wherein the prediction of the molecular binding properties of the one or more proteins comprises quantitative structure-property relationship (QSPR) or quantitative structure-activity relationship (QSAR) modeling of the one or more proteins.

130. The method of claim 94, wherein the prediction of the molecular binding properties of the one or more proteins comprises a prediction of the molecular binding properties of each amino acid sequence in the set of amino acid sequences corresponding to the one or more proteins.

131. The method of claim 94, wherein for each of the one or more empirically evaluated proteins, a corresponding predetermined bulk binding is measured for each experimental condition in a set of experimental conditions.

132. A method according to claim 131, wherein the set of experimental conditions includes 24 experimental conditions.

133. The method of claim 131, wherein the set of experimental conditions comprises a first subset of salt concentrations and a second subset of pH values.

134. A method according to claim 131, wherein the set of experimental conditions is input into the machine learning model together with the molecular descriptor matrix, and the prediction of the molecular binding properties of the one or more proteins includes a prediction of the molecular binding properties of the one or more proteins for each experimental condition in the set of experimental conditions.

135. The method of claim 94, further comprising: The prediction of the molecular binding properties of the one or more proteins is transformed into a linear representation.

136. The method of claim 135, wherein a logit transformation is used to generate the linear representation.

137. The method of claim 135, further comprising: A principal component analysis (PCA) is performed on the linear representation to obtain at least a first principal component.

138. A method according to claim 94, wherein the predetermined batch binding data associated with the one or more empirically evaluated proteins includes, for each of the one or more empirically evaluated proteins, an experimentally determined binding value measured for each experimental condition in the set of experimental conditions.

139. The method of claim 138, wherein the correlation between the selected representative feature vector and the predetermined batch combined data comprises: For each of the one or more empirically evaluated proteins and for each experimental condition in the set of experimental conditions: generating a linear representation of the experimentally determined binding value for the empirically evaluated protein based on a logit transformation applied to the experimentally determined binding value for the empirically evaluated protein; as well as A principal component analysis (PCA) is performed on said linear representation of said experimentally determined binding values ​​of said one or more empirically assessed proteins to obtain at least a first principal component.

140. The method of claim 139, further comprising: generating, using the machine learning model, trained predictions for molecular binding properties of the one or more empirically evaluated proteins; as well as The training predictions are compared to the first principal components to calculate the one or more cross validation losses.

141. The method of claim 139, wherein the first principal component describes an average batch binding value.

142. The method of claim 94, further comprising: generating a set of functions representing the behavior of the one or more proteins for a set of experimental conditions based on the predictions; as well as At least one of the one or more proteins is selected for use in one or more drug discovery assays based on the behavior of the one or more proteins for the set of experimental conditions.

143. The method of claim 94, wherein the correlation between the selected representative feature vector and the predetermined bulk binding data associated with the one or more empirically evaluated proteins comprises: The representative feature vector is correlated with a principal component calculated based on the predetermined batch combined data.

144. The method of claim 143, wherein the one or more cross-validation losses are calculated based on predicted molecular binding properties and empirical molecular binding properties.

145. The method of claim 144, wherein the predicted molecular binding properties comprise principal components calculated based on the representative feature vectors, and the empirical molecular binding properties comprise principal components calculated based on the predetermined batch binding data.

146. The method of claim 94, wherein determining the one or more most predictive feature vectors further comprises: (i) fitting a model to the representative feature vector; (ii) calculating a feature importance score for each of the representative feature vectors based on the model; as well as (iii) removing one or more feature vectors from the representative feature vectors based on the feature importance score of each of the representative feature vectors to obtain a subset of representative feature vectors, wherein the one or more most predictive feature vectors include one or more feature vectors from the subset of representative feature vectors.

147. The method of claim 146, further comprising: Steps (i) to (iii) are iteratively performed until the number of feature vectors included in the subset meets the feature quantity criterion.

148. The method of claim 147, wherein the feature quantity criterion is satisfied comprising the number of feature vectors included in the representative feature vector subset being less than or equal to a threshold number of feature vectors.

149. A method according to claim 148, wherein the threshold number of feature vectors includes the same or similar number of features from the training data used to train the machine learning model.

150. The method of claim 147, wherein the number of feature vectors included in the representative feature vector subset comprises a hyperparameter in the hyperparameter set.

151. A method according to claim 94, wherein a hyperparameter in the hyperparameter set represents the number of feature vector clusters included in the multiple feature vector clusters.

152. A system comprising one or more computing devices, the system further comprising: one or more non-transitory computer-readable storage media including instructions; as well as One or more processors coupled to the one or more storage media, the one or more processors configured to execute the instructions to: accessing a molecular descriptor matrix representing a set of amino acid sequences corresponding to one or more proteins; and obtaining, by a machine learning model, predictions of molecular binding properties of the one or more proteins based at least in part on the molecular descriptor matrix, wherein the machine learning model is trained by: accessing a training molecular descriptor matrix representing a training set of amino acid sequences corresponding to one or more empirically evaluated proteins; reducing the training molecular descriptor matrix by selecting a representative feature vector for each of a plurality of feature vector clusters, wherein each feature vector cluster includes similar feature vectors; determining one or more most predictive feature vectors from the selected representative feature vectors for each feature vector cluster based on correlations between the selected representative feature vectors and predetermined bulk binding data associated with the one or more empirically evaluated proteins; as well as Iteratively performing a process to refine a set of hyperparameters associated with the machine learning model until a desired accuracy is achieved, the process comprising: calculating one or more cross validation losses based at least in part on the one or more most predictive feature vectors and the predetermined batch of combined data; as well as The set of hyperparameters is updated based on the one or more cross-validation losses.

153. A non-transitory computer readable medium comprising instructions that, when executed by one or more processors of one or more computing devices, cause the one or more processors to perform operations comprising: accessing a molecular descriptor matrix representing a set of amino acid sequences corresponding to one or more proteins; and obtaining, by a machine learning model, predictions of molecular binding properties of the one or more proteins based at least in part on the molecular descriptor matrix, wherein the machine learning model is trained by: accessing a training molecular descriptor matrix representing a training set of amino acid sequences corresponding to one or more empirically evaluated proteins; reducing the training molecular descriptor matrix by selecting a representative feature vector for each of a plurality of feature vector clusters, wherein each feature vector cluster includes similar feature vectors; determining one or more most predictive feature vectors from the selected representative feature vectors for each feature vector cluster based on correlations between the selected representative feature vectors and predetermined bulk binding data associated with the one or more empirically evaluated proteins; as well as Iteratively performing a process to refine a set of hyperparameters associated with the machine learning model until a desired accuracy is achieved, the process comprising: calculating one or more cross validation losses based at least in part on the one or more most predictive feature vectors and the predetermined batch of combined data; as well as The set of hyperparameters is updated based on the one or more cross-validation losses.

Images