Methods and systems for predicting drug binding using synthetic data

Abstract

A method for predicting drug-target binding using synthetically enhanced data involves generating multiple phantom ligands for multiple proteins in a protein structure database; generating multiple DTI features for proteins and ligands in a drug-target interaction (DTI) database using the multiple phantom ligands; generating a machine learning model using the multiple DTI features; and using the machine learning model to predict the likelihood of interaction for a combination of a query protein and a query ligand.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority under 35 U.S.C. §119(e) to U.S. Provisional Application No. 62 / 788,682, filed on January 4, 2019, which has at least one inventor in common with the present application and is entitled “METHODS AND SYSTEMS FOR PREDICTING DRUG BINDING USING SYNTHETIC DATA.” U.S. Provisional Application No. 62 / 788,682 is incorporated herein by reference. Background Art

[0003] Computational methods exist to predict interactions between ligands and proteins. They are often categorized as either 'ligand-based' or 'structure-based', depending on the type of information used to make the predictions.

[0004] Protein-based predictions have the potential to learn biophysical compatibility and are therefore potentially more generalizable, but are highly data-constrained. Specifically, protein-based predictions use the 3D molecular structure of a ligand co-crystallized with a protein to assess or predict interactions. These methods are computationally demanding, and they are often trained on only 100s to 1000s of different proteins. Neural networks trained on these tend to have very high feature space to data ratios. As a result, the methods can produce a large number of false negatives and / or false positives in docking when applied to previously unseen protein systems or drug scaffolds.

[0005] Ligand-based predictions can be performed using drug-target interaction (DTI) databases with millions of records. Examples of publicly available DTI databases include ChEMBL, NCBI's Bioassay, and STITCH. However, these records tend to represent only about 2,000 of the 20,000 human proteins. Due to the high ligand-to-protein data ratio, a standard approach is to derive many different models for each of the 2,000 proteins. These are often successful, even exceeding high-throughput experimental results in many cases, but (1) only represent about 10% of human proteins, (2) can be weaker when there is not much chemical diversity between examples of individual proteins, and (3) individual models do not learn the physical properties of drug-protein compatibility and lack the benefit of the data used to generate other models. Summary of the invention

[0006] In general, in one aspect, one or more embodiments relate to a method for predicting drug-target binding using synthetically enhanced data, the method comprising: generating multiple phantom ligands for multiple proteins in a protein structure database; using the multiple phantom ligands to generate multiple DTI features for proteins and ligands in a drug-target interaction (DTI) database; using the multiple DTI features to generate a machine learning model; and using the machine learning model to predict the likelihood of interaction for a combination of a query protein and a query ligand.

[0007] In general, in one aspect, one or more embodiments relate to a non-transitory computer-readable medium comprising computer-readable program code for predicting drug-target binding using synthetically enhanced data, the computer-readable program code causing a computer system to: generate a plurality of phantom ligands for a plurality of proteins in a protein structure database; generate a plurality of DTI features for proteins and ligands in a drug-target interaction (DTI) database using the plurality of phantom ligands; generate a machine learning model using the DTI features; and use the machine learning model to predict the likelihood of interaction for a combination of a query protein and a query ligand.

[0008] In general, in one aspect, one or more embodiments relate to a system for differential drug discovery, the system comprising: a protein structure database; a phantom ligand recognition engine configured to generate multiple phantom ligands for multiple proteins in the protein structure database; a phantom ligand database storing multiple phantom ligands; a drug-target interaction (DTI) database storing proteins and ligands; a feature generation engine configured to generate multiple DTI features for proteins and ligands in a DTI database using multiple phantom ligands in the phantom ligand database; a machine learning model training engine configured to generate a machine learning model using the DTI features; and a DTI prediction engine configured to use the machine learning model to predict the likelihood of interaction of a combination of a query protein and a query ligand. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The present embodiments are depicted by way of example and are not intended to be limited by the figures of the accompanying drawings.

[0010] Figure 1A A block diagram of a system for predicting drug binding is shown in accordance with one or more embodiments.

[0011] Figure 1B A block diagram of a protein structure database is shown in accordance with one or more embodiments.

[0012] Figure 1C A block diagram of a phantom ligand database is shown in accordance with one or more embodiments.

[0013] Figure 1D A block diagram of a drug-target interaction database is shown in accordance with one or more embodiments.

[0014] Figure 1E A block diagram of a protein annotation database is shown in accordance with one or more embodiments.

[0015] Figure 2 A flow chart describing a method for training a machine learning model for predicting drug-target interactions according to one or more embodiments is shown.

[0016] Figure 3 A flow chart describing a method for generating a phantom ligand database in accordance with one or more embodiments is shown.

[0017] Figure 4 A flow chart describing a method for generating a drug-target interaction (DTI) signature according to one or more embodiments is shown.

[0018] Figure 5 A flow chart describing a method for generating a machine learning model for DTI prediction according to one or more embodiments is shown.

[0019] Figure 6 A flow chart describing a method for predicting an interaction between a query protein and a query ligand according to one or more embodiments is shown.

[0020] Fig. 7A Examples for generating phantom ligands according to one or more embodiments are shown.

[0021] Figure 7B A diagram of a concentric shell model for obtaining binding site characteristics is shown in accordance with one or more embodiments.

[0022] Figure 8 Generation of training data for a machine learning model according to one or more embodiments is shown.

[0023] Fig. 9 The performance comparison between the embodiments of the present disclosure and the conventional method is shown.

[0024] Fig. 10A and Fig. 10B A computing system is shown in accordance with one or more embodiments. DETAILED DESCRIPTION

[0025] Now will be described in detail with reference to the accompanying drawings the specific embodiments disclosed herein. For consistency, the same elements in the various figures may be represented by the same reference numerals and / or the same names.

[0026] The following detailed description is merely exemplary in nature and is not intended to limit the embodiments disclosed herein or the application and use of the embodiments disclosed herein. In addition, it is not intended to be bound by any express or implied theory presented in the aforementioned technical field, background technology, summary of the invention or the following detailed description.

[0027] In the following detailed description of some embodiments disclosed herein, many specific details are set forth in order to provide a more comprehensive understanding of the various embodiments disclosed herein. However, it will be apparent to one of ordinary skill in the art that the embodiments may be practiced without these specific details. In other cases, well-known features are not described in detail to avoid unnecessarily complicating the description.

[0028] Throughout the application, ordinal numbers (e.g., first, second, third, etc.) may be used as adjectives for elements (i.e., any noun in the application). The use of ordinal numbers does not imply or create any particular order of elements, nor does it limit any element to only a single element, unless explicitly disclosed, such as by using the terms "before," "after," "single," and other such terms. Instead, the use of ordinal numbers is to distinguish between elements. By way of example, a first element is different from a second element, and a first element may include more than one element and is after (or before) a second element in the order of the elements.

[0029] In one or more embodiments of the invention, elements of protein-based prediction methods and ligand-based prediction methods can be combined to obtain superior predictions of drug-target interactions. In one or more embodiments of the invention, machine learning models are used to predict drug-target interactions (DTIs).

[0030] There are a large number of data points about protein-ligand interactions in DTI databases. For example, the ChEMBL database contains about 15,000,000 records describing protein and ligand pairs as binding or non-binding (and often provides measures of affinity or confidence). However, there are certain limitations to using the content of these DTI databases to predict the interactions of new protein and ligand pairs, such as limited coverage of the human proteome and weak predictive power when the diversity between individual protein instances is limited. Therefore, the predictive quality of machine learning models that run only on DTI database records may be limited. Alternatively, machine learning models can operate on the 3D molecular structure of a ligand co-crystallized with a protein to capture the biophysics of the interaction between the protein and the ligand. There are some databases that capture protein-ligand interactions (e.g., sc-PDB), but these databases contain relatively few data points and are often highly redundant. In addition, these databases often lack protein structural diversity and chemical type diversity. Therefore, training machine learning models on these data points can be challenging due to insufficient data volume. In particular, machine learning models operating on the 3D molecular structures of ligands co-crystallized with proteins have a high-dimensional feature space, which, coupled with the limited availability of suitable training samples, makes machine learning models prone to overfitting.

[0031] In one or more embodiments of the present disclosure, a machine learning model uses local 3D features in conjunction with DTI recordings to achieve predictions superior to conventional protein-based and ligand-based predictions described above. More specifically, synthetic data is generated by projecting DTI recordings onto a 3D structural model of a known protein-ligand complex to generate local protein features that would otherwise be unavailable. Synthetic data generated in this manner can be used to train a machine learning model. The machine learning model can then be used to make predictions for a query protein and query ligand pair.

[0032] Go to Figure 1A , a system for predicting drug binding using synthetic data according to one or more embodiments is shown. The system (100) may include a phantom ligand identification engine (110), a feature generation engine (120), a machine learning model training engine (130), a drug-target interaction prediction engine (150), a protein structure database (160), a phantom ligand database (170), a drug-target interaction database (180), and a protein annotation database (190). Each of these components will be described later.

[0033] According to one or more embodiments, the phantom ligand recognition engine (110) includes instructions in the form of computer readable program code to perform Figure 2 and 3At least one of the steps described in to generate a phantom ligand database (170) of phantom ligands and associated confidence scores from proteins in a protein structure database (160). The phantom ligand identification engine (110) can obtain phantom ligands for a protein by structurally aligning known homologs and projecting the ligands of these known homologs onto the protein at the aligned sites. Although these phantom ligands may not interact with the protein, they can serve as placeholders indicating structural compatibility between the phantom ligand and the aligned sites of the protein. The phantom ligand identification engine (110) is operably connected to the protein structure database (160) and the phantom ligand database (170).

[0034] According to one or more embodiments, the feature generation engine (120) includes instructions in the form of computer-readable program code to perform Figure 2 and 4 At least one of the steps described in to generate drug-target interaction (DTI) features for training a machine learning model for DTI prediction. The feature generation engine (110) can generate features of proteins and ligands using data from a phantom ligand database (170), a drug-target interaction database (180), and a protein annotation database (190). Therefore, the feature generation engine (120) is operably connected to the phantom ligand database (170), the drug-target interaction database (180), and the protein annotation database (190).

[0035] continue Figure 1A According to one or more embodiments, the machine learning model training engine (130) includes instructions in the form of computer-readable program code to perform Figure 2 and 5 At least one of the steps described in to train a machine learning model (140) for DTI prediction. The machine learning model training engine (130) can be trained using the DTI features generated by the feature generation engine (120). Therefore, the machine learning model training engine (130) is operably connected to the feature generation engine (120). The final machine learning model (140) for DTI prediction can be any type of classifier that can predict the interaction between the query drug and the query protein. In one or more embodiments, the machine learning model (140) for DTI prediction is a deep neural network.

[0036] According to one or more embodiments, the drug-target interaction (DTI) prediction engine (150) includes instructions in the form of computer-readable program code to perform Figure 6At least one of the steps described in to predict drug-target interactions of the query drug and the query protein using the machine learning model (140). The DTI prediction engine (150) generates features of the query protein and the query ligand associated with the query drug that are compatible with the machine learning model (140), and then calculates the likelihood of interaction based on the features and using the same machine learning model (140) trained by the machine learning model training engine (130). In various embodiments, one or more of the same and / or different machine learning models may be used.

[0037] The protein structure database (160) according to one or more embodiments may be any type of storage unit and / or device for storing data (e.g., a file system, a database, a collection of tables, or any other storage mechanism). Figure 1B Describes the protein structure database (160).

[0038] According to one or more embodiments, the phantom ligand database (170) can be any type of storage unit and / or device for storing data (e.g., a file system, a database, a collection of tables, or any other storage mechanism). Figure 1C Describes the phantom ligand database (170).

[0039] The drug-target interaction database (180) according to one or more embodiments may be any type of storage unit and / or device for storing data (e.g., a file system, a database, a collection of tables, or any other storage mechanism). Figure 1D Describes a drug-target interaction database (180).

[0040] The protein annotation database (190) according to one or more embodiments may be any type of storage unit and / or device for storing data (e.g., a file system, a database, a collection of tables, or any other storage mechanism). Figure 1D Describes the protein annotation database (190).

[0041] Go to Figure 1B , shows a protein structure database (160) according to one or more embodiments. The protein structure database (160) can store 3D models (162A, 162B, 162N) of proteins. Each 3D model can be associated with a homolog model (164A, 164B, 164N) and / or an experimental model (166A, 166B, 166N). Examples of publicly available protein structure databases (160) include, but are not limited to, the Protein Data Bank (PDB) and SWISS-MODEL.

[0042] Go to Figure 1C, shows a phantom ligand database (170) according to one or more embodiments. For multiple 3D models of proteins (e.g., 172A, 172B, 172N), the phantom ligand database (170) can store the identified phantom ligands (e.g., 174A, 174B, 174N). In addition, for each identified phantom ligand, a confidence score (e.g., 176A, 176B, 176N) is included that can be based on similarity. The phantom ligand database can be as follows Figure 3 Established as described in.

[0043] Go to Figure 1D , a drug-target interaction database (180) according to one or more embodiments is shown. The drug-target interaction database (180) can store interaction confidences (e.g., 186A, 186B, 186N) for multiple pairs of drugs (e.g., 182A, 182B, 182N) and targets (e.g., 184A, 184B, 184N). Examples of publicly available drug-target interaction databases (180) include, but are not limited to, STITCH and ChEMBL. These databases may contain many data points (~15,000,000 ChEMBL).

[0044] Go to Figure 1E , shows a protein annotation database (190) according to one or more embodiments. The protein annotation database (190) can store related annotations (e.g., 194A, 192B, 194N) for multiple proteins (e.g., 192A, 192B, 192N). The annotations associated with the protein can include any available information about the protein and can be added to the protein annotation database manually or computationally. For example, the UniProt database can be used.

[0045] Figure 2 , Figure 3 , Figure 4 , Figure 5 and Figure 6 A flow chart is shown according to one or more embodiments. Figure 2 , Figure 3 , Figure 4 and Figure 5 A flowchart depicting the method used to train a machine learning model to predict drug-target interactions, and Figure 6 The flowchart of the method for using machine learning models to predict drug-target interactions can be found in the above reference Figure 1A The components of the system (100) discussed herein perform Figure 2 , Figure 3 , Figure 4 , Figure 5 and Figure 6In one or more embodiments, the steps may be omitted, repeated, Figure 2 , Figure 3 , Figure 4 , Figure 5 and Figure 6 One or more steps shown in and / or with Figure 2 , Figure 3 , Figure 4 , Figure 5 and Figure 6 The steps may be performed in a different order than shown. Additional steps may be further performed. Therefore, the scope of the present invention should not be considered to be limited to Figure 2 , Figure 3 , Figure 4 , Figure 5 and Figure 6 The specific arrangement of the steps shown.

[0046] Go to Figure 2 Flowchart showing a method for generating a machine learning model to predict drug-target interactions (DTI). Figure 2 The main steps in building a machine learning model are introduced here, but the flowchart discussed later provides a more detailed description. Figure 2 After the method, the obtained machine learning model can be used to make predictions, such as Figure 6 As described in.

[0047] In step 200, a phantom ligand database is generated based on proteins obtained from a protein structure database. Figure 3 A detailed description of step 200 is provided in .

[0048] In step 202, a drug-target interaction (DTI) signature is generated, including a ligand signature and a protein signature. Figure 4 A detailed description of step 202 is provided.

[0049] In step 204, a machine learning model for DTI prediction is generated. Figure 5 A detailed description of step 204 is provided.

[0050] Go to Figure 3 Flowchart describing a method for generating a phantom ligand database. The steps described subsequently are used to exploit homology relationships between proteins to map known ligands to the structures of their homologs (experimental and homology models), called "phantoms".

[0051] In step 300, proteins are obtained from a protein structure database. For each protein, a 3D model can be retrieved.

[0052] In step 302, the obtained proteins are clustered by sequence or domain. In one or more embodiments, protein clustering is defined as any collection of two or more proteins that share similarity in primary sequence or three-dimensional topology (commonly referred to as folding). Protein clustering can be obtained directly from publicly available databases, such as PDB, SCOP, CATH, PFAM or Uniprot. For example, protein sequence alignment tools and clustering tools (such as BLAST, CD-HIT or UCLUST) can be used to manually create protein clusters based on sequence similarity. Alternatively, proteins with three-dimensional structural models can be clustered by grouping unrelated proteins that share a common topology or folding.

[0053] In step 304, one of the clusters is selected for further processing.

[0054] In step 306, a pairwise structural alignment is performed on all proteins in the selected cluster. Fig. 7A An example of a structural alignment of protein 3D structures is shown. Three-dimensional (3D) structural alignment attempts to establish positional equivalence between two proteins. Structural alignment can be performed by applying rotational and / or translational transformations to the coordinates of one protein to minimize the average distance between equivalent residues. For example, structural alignment can be performed on the complete protein structure or on a subselection of residues, such as residues surrounding a single domain or ligand binding site. The choice of ligand binding site residues for 3D structural alignment is the best heuristic for mapping phantom ligands.

[0055] In step 308, phantom ligands are obtained by projecting each ligand to a cluster peer. A confidence score for each projection can be obtained, which is composed of individual scores representing the confidence of different components of the ligand projection. An example of projecting ligands to cluster peers is shown in Fig. 7A The confidence score can be based on any measure that quantifies the uncertainty in the heuristically defined structural representation chosen to model the DTI interactions. At this step, the confidence score can include a homology model quality measure, such as percent sequence identity, sequence similarity, or QMEAN. The confidence score can also include a measure that describes the quality of the structural alignment, such as the root mean square deviation of a local or global alignment. Multiple confidence scores can also be used.

[0056] In step 310, phantom ligands and associated confidence scores are stored in a phantom ligand database.

[0057] In step 312, it is determined whether there are additional clusters to be processed. If there are additional clusters, execution of the method may return to step 304 to select another cluster for processing as described in steps 306-310. If no additional clusters remain, execution of the method may terminate. Once Figure 3 Upon termination of the method, the phantom ligand database may contain a comprehensive collection of phantom ligands and associated confidence scores for all proteins processed as described.

[0058] Go to Figure 4 Flowchart describing a method for generating drug-target interaction (DTI) features. The generated features include features of ligands and features of proteins. These features can then be used to train machine learning models for predicting drug-target interactions. Features for many combinations of proteins and ligands can be generated to ensure the availability of sufficient training samples.

[0059] In step 400, a drug-target interaction for a combination of a ligand and a protein is selected from the DTI database. The subsequent steps are performed for this considered combination of ligand and protein. These steps may be repeated later for other combinations of ligand and protein.

[0060] In step 402, features are generated for the selected ligand. These features may include ligand fingerprints and ligand descriptors. The fingerprint may capture the structure of the ligand in a descriptor format and may be based on a smile representation of the underlying molecule using a fixed length vector. For example, molecular fingerprinting may include atom pairs, extended connection fingerprints, graph-based fingerprints, torsion fingerprints, or pharmacophore fingerprints. For example, molecular weight, number of rotatable bonds, number of hydrogen bond donors, number of hydrogen bond acceptors, hydrophobicity, aromaticity, and functional group composition may be used as ligand descriptors. Molecular shape descriptors, such as ellipticity, geometric descriptors, branching descriptors, or chirality descriptors may also be used.

[0061] In step 404, phantom ligands are retrieved from a phantom ligand database for the selected protein.

[0062] In step 406, each phantom ligand is scored based on its difference from the drug, or more specifically, from the selected ligand. Higher similarity results in higher scores. For example, the similarity between ligands can be scored using a distance metric used to compare molecular fingerprints, such as Tanimoto distance, Dice distance, or cosine distance.

[0063] Steps 404 and 406 may be performed for all protein models (eg, homology models or experimentally derived models) available for the selected protein.

[0064] In step 408, the phantom ligand that is most similar to the selected ligand is selected for further processing.

[0065] In step 410, a confidence vector is generated for the DTI feature, which is composed of individual scores representing the confidence in different components of the DTI feature and its representative phantom ligand. The confidence vector can include confidence scores representing the phantom ligand projections from step 308. The confidence vector can also include confidence scores for fingerprint similarity between the selected ligand and the most similar phantom ligand selected in step 408. In addition, the confidence vector can include a score for the confidence of the selected DTI. The confidence of the selected DTI can be scored based on the source of the DTI data (e.g., different scores can be assigned depending on whether the DTI data was obtained using high-throughput screening or low-throughput screening, etc.).

[0066] In step 412, local features of the selected protein surrounding the most similar phantom ligand are obtained. The local features may include binding site features that exist in concentric shells of increasing radius, such as Figure 7B As shown. For each concentric shell, multiple descriptors are provided, such as atom type descriptors specifying atoms present in the shell. For example, 70 atom type descriptors can be provided for each shell radius. Descriptors may also include, but are not limited to, the flexibility or rigidity of the binding site within the shell region, the contact of residues within the shell region, and / or any other factors representing the biophysical and indirect binding site geometry. However, these features may not specify the exact position or coordinates of the atom. Local features may also include a graphical description of the ligand binding site, which describes the distance between amino acids around the ligand binding site in a network format. Local features may also be defined by the shape of the pocket, corresponding to the gap not occupied by protein residues. The pocket gap can be determined by pocket detection methods, such as overflow filling, concavity or solvent accessibility. Local features defined by the pocket void space may include the shape of the void space, including volume, ellipticity, curvature, branching pattern, or spatial stability based on the dynamics of nearby residues. For example, local features defined by residues adjacent to the pocket void space may include the orientation of the residue, the geometric availability of hydrogen bond donors and acceptor groups, hydrophobicity, aromaticity, or the geometric availability of π stacking interactions. Local features may also include a description of the ligand binding channel, which includes solvent exposed residues near the ligand binding site that are not in direct contact with the ligand in the stable bound state. The ligand binding channel is expected to form transient interactions with the ligand during the dynamic process of binding and dissociation. Ligand binding channel characteristics may include features similar to those defining the pocket, such as the orientation of the residues, amino acid composition, and the availability of hydrogen bond donors and receptors.

[0067] In step 414, global features of the structure and / or sequence of the selected protein are obtained by expanding to an outer shell having a large radius as described in step 412. The shell radius corresponding to the local features may include, for example or The shell radius corresponding to a domain-level description or a global protein description can have, for example, or A greater distance threshold may be used, or there may be no distance threshold. Global features may also include descriptions of domains or folds, and may be derived from publicly available databases such as SCOP, CATH, or PFAM. For example, global features may also include features derived from protein sequences and may include the presence of common sequence motifs. Global features may include descriptions of protein folding states, such as the presence and biophysical properties of intrinsically disordered regions, hinges, loops, ordered regions, or regulatory domains. Global features may also be described based on the distance from the ligand binding site.

[0068] In step 416, functional annotations of the selected proteins are obtained. Functional annotations can be obtained from a protein annotation database. Functional annotations may include, for example, Enzyme Commission (EC) numbers, Gene Ontology (GO) annotations, or Uniprot keywords. Functional annotations may also include the presence or absence of recorded protein position-specific properties, such as catalytic sites, post-translational modifications, disease associations, or genetic variations.

[0069] In step 418, features are generated for the selected proteins. These features may include local features, global features, and / or functional annotations.

[0070] In step 420, it is determined whether there are additional DTIs to be processed. If there are additional DTIs, execution of the method may return to step 400 to select another DTI for processing, as described in steps 402-418. If no additional clusters remain, execution of the method may terminate. Once Figure 4 Once the method is terminated, a comprehensive feature set of ligands and proteins listed in the DTI database can be used.

[0071] Go to Figure 5 Flowchart of FIG. 4 depicts a method for generating a machine learning model for DTI prediction. Figure 4 The obtained DTI features obtain a machine learning model that reflects the compatibility between protein environment and ligand properties.

[0072] In step 500, ligand features and protein features are obtained. The ligand and protein features can be obtained as follows: Figure 4 The steps 402 and 418 are performed as described.

[0073] In step 502, by Figure 4The ligands and proteins are filtered using a function of the confidence vector established in step 408. The filtering can implement a confidence threshold, and only samples above the confidence threshold are considered for further processing. The confidence function can convert the confidence vector into a single score for filtering. For example, the confidence function can convert the confidence scores into probabilities and apply Bayesian statistics to evaluate the combined probabilities. The confidence function can apply a separate cutoff threshold to each element of the confidence vector as a means of selecting which samples are suitable for machine learning. The confidence function threshold or equation can be set by automatically testing different combinations as a hyperparameter of the machine learning algorithm.

[0074] In step 504, ligand and protein features are concatenated to generate positive training samples.

[0075] In step 506, the ligand and protein features are randomly shuffled. The randomly shuffled ligand and protein features are concatenated to generate negative training samples. This step may be repeated multiple times to evaluate different positive and negative training sample ratios, such as 1:1, 1:5, 1:10, 1:19, or 1:20.

[0076] In step 508, a machine learning model for DTI prediction is trained using positive and negative training samples. For example, a back-propagation based learning algorithm may be used. In one or more embodiments, the training samples may be weighted based on their associated confidence vectors. In one or more embodiments, transfer learning is used to more efficiently train the machine learning model. Initially, the machine learning model may be trained by applying an initial confidence threshold in step 502. In a subsequent retraining phase, the confidence threshold may be increased to reduce the number of training instances and improve their quality. Additionally or alternatively, a subsequent retraining phase may limit the training instances to select a class of drug or target. The machine learning model may be a supervised discriminant classification or regression model, such as a random forest, a support vector machine, a single-layer perceptron, or a multi-layer artificial neural network. Artificial neural networks are particularly well suited for this task given the number of training data points (100,000s to 10,000,000s) and the dimensionality of the training data features (1000s to 10,000s). In one embodiment, the artificial neural network representation takes the form of a fully connected network with a feature input layer, two hidden layers, e.g., with 512 and 256 nodes, respectively, and two output nodes corresponding to interacting and non-interacting pairs. In one embodiment, the artificial neural network with multiple hidden layers omits connections between input types to create separate latent spaces representing ligand fingerprints, global protein features, local protein features, and protein functional features.

[0077] Go to Figure 6, describing a method for predicting interactions between a query protein and a query ligand by applying a machine learning model to a set of DTI features corresponding to the query ligand and at least one known binding site of the query protein, as described in reference Figure 2-5 The described trained machine learning model can be used to test the "compatibility" of a query protein and a query compound.

[0078] In step 600, a query protein and a query ligand are obtained. The query protein and the query ligand may be obtained from a user who wants to obtain a prediction of an interaction between the query protein and the query ligand.

[0079] In step 602, as previously described in Figure 4 As described in steps 404-412 of the method, possible binding sites and related local features of the query protein are obtained. Thus, one or more binding sites may be obtained from one or more experiments or homology models. Alternatively, the binding sites and related local features may be obtained from a user, for example, if the user wishes to specify a specific binding site.

[0080] In step 604, the global features and protein annotations of the query protein are obtained. The global features and protein annotations can be obtained as previously described in Figure 4 is obtained as described in steps 414 and 416 of .

[0081] In step 606, features are generated for the query protein. These features may include local features, global features, and / or functional annotations.

[0082] In step 608, as previously described in Figure 4 As described in step 402, a ligand fingerprint and a ligand descriptor are obtained.

[0083] In step 610, features are generated for the query ligand. These features may include ligand fingerprints and ligand descriptors.

[0084] In step 612, the machine learning model for DTI prediction is applied to the features of the query ligand and the features of the query protein to obtain a numerical score for the likelihood of interaction between the query ligand and the query protein.

[0085] The following paragraphs further illustrate the embodiments of the present disclosure based on various examples. It will be understood by those skilled in the art that the present disclosure is not limited to these examples.

[0086] (i) Sample phantom ligand:

[0087] Go to Fig. 7A, shows an example (700) for generating phantom ligands. Three hypothetical protein structures are shown (top row). Two of the three hypothetical protein structures actually interact with the ligand (top row, left and middle columns). The middle row shows various structural arrangements of the three hypothetical proteins. As a result of the structural arrangements, the ligands can be projected onto other proteins. Confidence scores are assigned based on the similarity of the binding sites. The confidence score for the actual ligand-protein pairing is 1.0, but the confidence score for the phantom ligand-protein pairing is lower. The bottom row shows the generated phantom ligand-protein pairings as they can be stored in the phantom ligand database.

[0088] Figure 7B An illustration of a concentric shell model (750) for obtaining binding site features according to one or more embodiments is shown. Concentric shells of increasing radius (r) surround a central chemical structure that is considered to be part of the binding site. The inner shells capture primarily local features near the binding site, while the outer shells capture increasingly global features. Features representing a protein can be based on the concentric shell model, thereby capturing local and global features of the protein without specifying an exact 3D geometry (e.g., at the atomic level).

[0089] (ii) Sample confidence vector:

[0090] Embodiments of the present disclosure rely on a heuristic process to augment drug-target interaction (DTI) data with hypothetical three-dimensional structural representations. These hypothetical DTI representations can provide informative features for machine learning, improving models designed to predict protein-ligand interactions. Obtaining these approximate DTI representations for any given DTI data point requires Figure 2 Several assumptions are made as outlined in steps 200 and 202. For example, the three-dimensional protein structure used to represent the DTI may be derived from a homology model rather than directly from experimental coordinates.

[0091] The confidence vector is composed of multiple metrics that describe the measurable uncertainty in the approximate DTI representation. These metrics are accumulated during the creation of the phantom ligand database (step 200) and the projection of known DTI data onto the phantom ligand database (step 202). In one example, the confidence vector contains four elements, including: (1) the percentage of sequence identity between the homology model representing the DTI protein and its source template, (2) the RMSD of the alignment between the source structure from the phantom ligand and the homology model representing the DTI protein, (3) the Tanimoto similarity between the morgan3 fingerprint of the DTI ligand and the phantom ligand template, and (4) the confidence of the DTI data point.

[0092] In this example, the drug-target interaction (DTI) database indicates that the ligand gefitinib interacts with the protein Aurora kinase A. The DTI database assigns an 85% probability to the interaction based on the accuracy of the source biophysical experiments. The specific interaction between gefitinib and Aurora kinase A was not present in the source three-dimensional structure database. When creating the phantom ligand database, a homology model of the Aurora kinase A protein was created from a close homolog Aurora kinase B that shares 72.5% sequence identity. The closest molecule to gefitinib that was successfully mapped onto the homology model was erlotinib, which shares a Morgan3 fingerprint Tanimoto similarity of 0.372. The erlotinib phantom ligand position was approximated based on a structural alignment between the Aurora kinase A homology model and the erlotinib-EGFR complex crystal structure, with a ligand binding site RMSD of Therefore, the corresponding confidence vector will be: [85%, 72.5%, 0.372, ].

[0093] (iii) Sample training data and negative random permutations:

[0094] The described approach focuses on augmenting drug-target interaction (DTI) pairs from a DTI database with a mixture of relevant features obtained through a series of deterministic mappable relationships and heuristic modeling features (local structural features). Each row in the DTI database can be converted into a feature vector such as Figure 8 The example shown shows the generation of training data (800) with ligand features (columns labeled "Ligand Features") and protein features (columns labeled "Global Features", "Functional Features", and "Local Features") from the corresponding drugs. Global and functional features of proteins can be retrieved from any protein using standard practices for database lookup and protein identifier mapping. Local protein features can be the result of a heuristic definition process outlined in this patent. They are modeled and therefore may not be accurate. Each data row also has a corresponding confidence vector (described above but not shown in the figure), which can be used to imply hard cutoffs or weights for training the machine learning model.

[0095] When a neural network is trained only with real drug-target interaction positive examples extracted from a drug-target interaction dataset, the model may learn to ignore core and obvious patterns of the interactions as they do not provide any signal to the model. In addition, it may be necessary to control for potentially significant biases towards drugs and targets that are highly represented in the drug-target interaction dataset. Therefore, it may be beneficial to sample negative examples using the relative proportions of each drug and target. As a result, the model can learn patterns based on both positive and negative examples. Figure 8In the figure, the randomized negative results randomly permute the ligand component of the feature vector (white) with the protein component of the feature vector (three shades of gray). The resulting negative examples can be used to train the classification engine, balancing the presence of individual ligand or protein features. Using individual ligand and / or protein features equally in both the positive and negative sets avoids the network learning that any single feature is generally particularly relevant to binding.

[0096] Embodiments of the present disclosure use phantom ligands to create local protein features for protein chemometrics (PCM) from protein-ligand datasets. More specifically, drug-target interaction (DTI) data are threaded onto 3D atomic models of protein-ligand complexes to derive local protein features. The hybrid feature datasets for PCM can include local (pocket), regional (domain) and global (whole protein) annotations and / or functional annotations.

[0097] Traditionally, training data for machine learning should be high-confidence 'model quality data'. According to one or more embodiments, using predictions (phantom ligands + connections (threads)) to generate training data for machine learning algorithms seems counterintuitive. Specifically, if heuristic methods are not accurate enough, then the traditional view is that introducing local features derived from the combination of phantom ligands and connections is likely to introduce additional noise, thereby reducing the performance of traditional DTI PMC. However, as Fig. 9 As shown in the performance comparison (900), the performance improvement is achieved by introducing the local features derived by the method described in this patent. Omitting the local features derived by this method is equivalent to the performance achieved by DTI PMC alone.

[0098] Specifically, Fig. 9 A performance comparison of ranking the binding likelihood of small molecule ligands and 8717 proteins is shown. To test the ranking, 100 molecules were randomly removed from the training data and used for testing. The figure plots the predicted ranking of known interactions for these 100 random drugs. For example, without local features, only about 63% of the actual interactions are observed in the predictions for the top 300 proteins in 8717 (top about 3.5%) (dashed line). Including local features estimated by this procedure, the discovery rate increases to about 75% for the same threshold (solid line).

[0099] Various embodiments of the present disclosure have one or more of the following advantages. Embodiments of the present invention are capable of predicting drug-target interactions (DTIs) using machine learning models that reflect compatibility between protein environments and ligand properties. Localized 3D features are created to represent binding sites, even when no 3D information is available for the interaction under consideration.

[0100] Mapping known drug-target interactions to homology models synthetically enhances the rich DTI training data with high-dimensional biophysical information to train deep neural networks. Therefore, this method is able to use a comprehensive DTI database even if the binding position of the drug to the protein is not necessarily known, and can be used for entries in the DTI database.

[0101] The method according to one or more embodiments does not require detailed knowledge of the biophysics of protein-ligand interactions. Therefore, the precise 3D coordinates of atoms are not required, thereby enabling the use of DTI databases and homology models to map drug-target interactions to protein pockets.

[0102] Compared to structure-based deep learning methods that rely on 3D atomic coordinates, embodiments of the present disclosure require a reduced feature space and allow for much larger training data. In addition, embodiments of the present disclosure were found to generalize well. Preliminary performance evaluations indicate that the methods described above execute approximately 1,000,000 times faster than docking simulations. The methods according to one or more embodiments do not require human intervention. Specifically, the most likely protein representations and binding sites are automatically identified. As discussed in Appendices A and B, methods according to one or more embodiments can be used as an accurate in-silico alternative or additional to other in-silico and / or experimental methods for predicting drug-target interactions.

[0103] Embodiments of the present disclosure may have various applications. For example, embodiments may be used for proteomic screening (eg, performing toxicity prediction or phenotypic deconvolution prediction), virtual screening, and general drug discovery and development.

[0104] Embodiments of the present disclosure may be implemented on a computing system. Any combination of mobile, desktop, server, router, switch, embedded device, or other types of hardware may be used. For example, Fig. 10A As shown in FIG. 1 , the computing system (1000) may include one or more computer processors (1002), non-persistent storage devices (1004) (e.g., volatile memory, such as random access memory (RAM), cache memory), persistent storage devices (1006) (e.g., hard disk, optical drive, such as compact disk (CD) drive or digital versatile disk (DVD) drive, flash memory, etc.), communication interface (1012) (e.g., Bluetooth interface, infrared interface, network interface, optical interface, etc.), and many other elements and functions.

[0105] The computer processor(s) (1002) may be an integrated circuit for processing instructions. For example, the computer processor(s) may be one or more cores or micro-cores of a processor. The computing system (1000) may also include one or more input devices (1010), such as a touch screen, keyboard, mouse, microphone, touchpad, electronic pen, or any other type of input device.

[0106] The communication interface (1012) may include an integrated circuit for connecting the computing system (1000) to a network (not shown) (e.g., a local area network (LAN), a wide area network (WAN) such as the Internet, a mobile network, or any other type of network) and / or another device, such as another computing device.

[0107] In addition, the computing system (1000) may include one or more output devices (1008), such as a screen (e.g., a liquid crystal display (LCD), a plasma display, a touch screen, a cathode ray tube (CRT) monitor, a projector, or other display device), a printer, an external storage device, or any other output device. One or more of the output devices may be the same as or different from the input device(s). The input and output device(s) may be connected locally or remotely to the computer processor(s) (1002), the non-persistent storage device(s) (1004), and the persistent storage device(s) (1006). There are many different types of computing systems, and the aforementioned input and output device(s) may take other forms.

[0108] Software instructions in the form of computer-readable program code for executing embodiments of the present disclosure may be stored in whole or in part, temporarily or permanently, on a non-transitory computer-readable medium (such as a CD, DVD, storage device, disk, tape, flash memory, physical memory, or any other computer-readable storage medium). Specifically, the software instructions may correspond to a computer-readable program code that, when executed by (one or more) processors, is configured to execute one or more embodiments of the present disclosure.

[0109] Fig. 10A The computing system (1000) in may be connected to or part of a network. Fig. 10B As shown in FIG. 1 , the network (1020) may include multiple nodes (e.g., node X (1022), node Y (1024)). Each node may correspond to a computing system, such as in Fig. 10A The computing system or a group of nodes shown in the figure may correspond to Fig. 10A. For example, embodiments of the present disclosure may be implemented on a node of a distributed system connected to other nodes. As another example, embodiments of the present disclosure may be implemented on a distributed computing system having multiple nodes, wherein each portion of the present disclosure may be located on a different node within the distributed computing system. In addition, one or more elements of the aforementioned computing system (1000) may be located at a remote location and connected to other elements via a network.

[0110] Despite Fig. 10B Not shown in the figure, the node may correspond to a blade in a server chassis connected to other nodes via a backplane. As another example, the node may correspond to a server in a data center. As another example, a node may correspond to a computer processor or a micro-core of a computer processor with shared memory and / or resources.

[0111] The nodes (e.g., Node X (1022), Node Y (1024)) in the network (1020) can be configured to provide services to client devices (1026). For example, the nodes can be part of a cloud computing system. The nodes can include functionality for receiving requests (1026) from client devices and transmitting responses (1026) to client devices. The client devices (1026) can be computing systems, such as those in Fig. 10A In addition, the client device (1026) may include and / or execute all or part of one or more embodiments of the present disclosure.

[0112] exist Fig. 10A and Fig. 10B The computing system or group of computing systems described in the description may include functionality to perform various operations disclosed herein. For example, (one or more) computing systems may perform communications between processes on the same or different systems. Various mechanisms employing some form of active or passive communication may facilitate data exchange between processes on the same device. Examples representing these inter-process communications include, but are not limited to, implementations of files, signals, sockets, message queues, pipelines, semaphores, shared memory, messaging, and memory mapped files. Additional details related to several of these non-limiting examples are provided below.

[0113] Based on the client-server network model, a socket can be used as an interface or communication channel endpoint to enable bidirectional data transmission between processes on the same device. First, following the client-server network model, a server process (e.g., a process that provides data) can create a first socket object. Next, the server process combines the first socket object, thereby associating the first socket object with a unique name and / or address. After creating and combining the first socket object, the server process then waits and listens for incoming connection requests from one or more client processes (e.g., a process that seeks data). At this point, when the client process wants to obtain data from the server process, the client process starts by creating a second socket object. Then, the client process continues to generate a connection request, which includes at least a second socket object and a unique name and / or address associated with the first socket object. Then, the client process transmits the connection request to the server process. Depending on availability, the server process can accept the connection request and establish a communication channel with the client process, or the server process, which is busy processing other operations, can queue the connection request in a buffer until the server process is ready. The established connection notifies the client process that communication can begin. In response, the client process may generate a data request specifying the data that the client process wishes to obtain. The data request is then transmitted to the server process. Upon receiving the data request, the server process analyzes the request and collects the requested data. Finally, the server process then generates a reply including at least the requested data, and transmits the reply to the client process. More commonly, the data may be transmitted as a datagram or a stream of characters (e.g., bytes).

[0114] Shared memory refers to the allocation of virtual memory space in order to verify that data can be communicated and / or accessed by multiple processes. When implementing shared memory, the initialization process first creates a shareable segment in a persistent or non-persistent storage device. After creation, the initialization process then loads the shareable segment and subsequently maps the shareable segment to the address space associated with the initialization process. After loading, the initialization process continues to identify and grant access rights to one or more authorized processes, which can also write data to and read data from the shareable segment. Changes made to the data in the shareable segment by one process may immediately affect other processes that are also linked to the shareable segment. In addition, when one of the authorized processes accesses the shareable segment, the shareable segment is mapped to the address space of the authorized process. Typically, at any given time, only one authorized process can load the shareable segment in addition to the initialization process.

[0115] Without departing from the scope of the present disclosure, other techniques may be used to share data between processes, such as the various data described in this application. These processes may be part of the same or different applications and may be executed on the same or different computing systems.

[0116] Instead of or in addition to sharing data between processes, the computing system executing one or more embodiments of the present disclosure may include the function of receiving data from a user. For example, in one or more embodiments, a user may submit data via a graphical user interface (GUI) on a user device. Data may be submitted via a graphical user interface by selecting one or more graphical user interface widgets or inserting text and other data into a graphical user interface widget using a touch pad, keyboard, mouse or any other input device. In response to selecting a specific project, information about the specific project may be obtained from a persistent or non-persistent storage device by a computer processor. When a user selects a project, the content of the data obtained about the specific project may be displayed on the user device in response to the user's selection.

[0117] As another example, a request to obtain data about a specific project can be sent to a server that is operably connected to a user device through a network. For example, a user can select a uniform resource locator (URL) link in a network client of a user device, thereby starting to send a hypertext transfer protocol (HTTP) or other protocol request to a network host associated with the URL. In response to the request, the server can extract data about a specific selected project and send the data to the device that initiated the request. Once the user device has received data about a specific project, the content of the data received about the specific project can be displayed on the user device in response to the user's selection. Following the above example, the data received from the server after the URL link is selected can provide a web page of hypertext markup language (HTML) that can be presented by the network client and displayed on the user device.

[0118] Once the data is obtained, such as by using the above-described techniques or from a storage device, the computing system, when executing one or more embodiments of the present disclosure, can extract one or more data items from the obtained data. Fig. 10AThe computing system in performs the extraction as follows. First, an organizational pattern (e.g., syntax, schema, layout) of the data is determined, which can be based on one or more of: position (e.g., bit or column position, Nth token in a data stream, etc.); attribute (where an attribute is associated with one or more values); or hierarchical / tree structure (comprising layers of nodes at different levels of detail, such as in nested package headers or nested document sections). Then, in the context of the organizational pattern, the original, unprocessed stream of data symbols is parsed into a stream (or hierarchical structure) of tokens (where each token can have an associated token "type").

[0119] Next, extraction criteria are used to extract one or more data items from the token stream or structure, where the extraction criteria are processed according to an organizational pattern to extract one or more tokens (or nodes from a hierarchical structure). For position-based data, the token(s) at the position(s) identified by the extraction criteria are extracted. For attribute / value-based data, the token(s) and / or node(s) associated with the attribute(s) that meet the extraction criteria are extracted. For hierarchical / hierarchical data, the token(s) associated with the node(s) that match the extraction criteria are extracted. The extraction criteria can be as simple as an identifier string, or can be a query provided to a structured data repository (where the data repository can be organized according to a database schema or data format (such as XML)).

[0120] The extracted data can be used for further processing by a computing system. For example, Fig. 10AWhen the computing system executes one or more embodiments of the present disclosure, it may perform data comparison. Data comparison can be used to compare two or more data values (e.g., A, B). For example, one or more embodiments may determine whether A > B, A = B, A!= B, A < B, etc. The comparison can be performed by submitting operation codes A, B that specify operations related to the comparison to an arithmetic logic unit (ALU) (i.e., a circuit that performs arithmetic and / or bitwise logical operations on two data values). The ALU outputs a numerical result of the operation and / or one or more status flags related to the numerical result. For example, the status flag can indicate whether the numerical result is positive, negative, zero, etc. By selecting appropriate operation codes and then reading the numerical result and / or status flags, the comparison can be performed. For example, to determine whether A > B, B can be subtracted from A (i.e., A - B), and the status flag can be read to determine whether the result is positive (i.e., if A > B, then A - B > 0). In one or more embodiments, if A = B or if A > B as determined using the ALU, then B can be regarded as a threshold, and A is regarded as meeting the threshold. In one or more embodiments of the present disclosure, A and B can be vectors, and comparing A with B requires comparing the first element of vector A with the first element of vector B, comparing the second element of vector A with the second element of vector B, etc. In one or more embodiments, if A and B are strings, then the binary values of the strings can be compared.

[0121] Fig. 10A The computing system in [description] can implement and / or be connected to a data repository. For example, one type of data repository is a database. A database is a collection of information configured to facilitate data retrieval, modification, reorganization, and deletion. A database management system (DBMS) is a software application that provides an interface for a user to define, create, query, update, or manage a database.

[0122] A user or software application can submit a statement or query to the DBMS. Then the DBMS interprets the statement. The statement can be a select statement, update statement, create statement, delete statement, etc. that requests information. Moreover, the statement can include parameters that specify data or data containers (database, table, record, column, view, etc.), (one or more) identifiers, conditions (comparison operators), functions (e.g., join, full join, count, average, etc.), sorting (such as ascending, descending), or others. The DBMS can execute the statement. For example, the DBMS can access memory buffers, reference, or index files for reading, writing, deleting, or any combination thereof in response to the statement. The DBMS can load data from persistent or non-persistent storage devices and perform calculations in response to a query. The DBMS can return the (one or more) results to the user or software application.

[0123] Fig. 10A The computing system may include functionality to provide raw and / or processed data (such as results of comparisons and other processing). For example, providing data may be accomplished through various presentation methods. Specifically, data may be provided through a user interface provided by the computing device. The user interface may include a GUI that displays information on a display device (a touch screen on a computer monitor or handheld computer device). The GUI may include various GUI widgets that organize what data is shown and how the data is provided to the user. In addition, the GUI may provide data directly to the user, for example, data provided as actual data values ​​via text, or presented by the computing device as a visual representation of the data, such as via a visual data model.

[0124] For example, the GUI may first obtain a notification from a software application requesting that a particular data object be provided within the GUI. Next, the GUI may identify the data object type associated with the particular data object, for example, by obtaining data from a data attribute within the data object that identifies the data object type. The GUI may then determine any rules specified for displaying the data object type, such as rules specified by a software framework for a data object class or rules specified according to any local parameters defined by the GUI for presenting the data object type. Finally, the GUI may obtain a data value from the particular data object and present a visual representation of the data value within a display device according to the specified rules for the data object type.

[0125] The data may also be provided by various audio methods. In particular, the data may be presented in an audio format and provided as sound through one or more speakers operatively connected to the computing device.

[0126] Data can also be provided to the user through tactile methods. For example, tactile methods can include vibrations or other physical signals generated by the computing system. For example, vibrations generated by a handheld computer device with a predetermined duration and vibration intensity can be used to provide data to the user to convey the data.

[0127] The above description of the function only presents the Fig. 10A computing systems and / or Fig. 10B Other functions may be performed using one or more embodiments of the present disclosure.

[0128] While the present disclosure has been described with respect to a limited number of embodiments, those skilled in the art having benefit of this disclosure will appreciate that other embodiments can be devised which do not depart from the scope of the present disclosure as disclosed herein. Accordingly, the scope of the present disclosure should be limited only by the claims appended hereto.

[0129] The embodiments and examples set forth herein are intended to best explain the present invention and its specific applications, thereby enabling those skilled in the art to make and use the present invention. However, those skilled in the art will recognize that the foregoing description and examples are presented for purposes of illustration and example only. The set forth description is not intended to be exhaustive or to limit the present invention to the precise form disclosed.

[0130] While the present invention has been described with respect to a limited number of embodiments, those skilled in the art, having benefit of this disclosure, will appreciate that other embodiments can be devised which do not depart from the scope of the invention as disclosed herein. Accordingly, the scope of the present invention should be limited only by the claims appended hereto.

Claims

1. A computer-implemented method for predicting drug-target binding using synthetic augmentation data, the method comprising: Generate multiple phantom ligands for multiple proteins in the protein structure database; Wherein generating the plurality of phantom ligands comprises: For a protein cluster selected from the plurality of proteins: Performing structural comparison on proteins in the protein cluster; After the structural alignment, obtaining the plurality of phantom ligands by projecting the ligand of one of the proteins in the cluster onto all other proteins in the cluster; obtaining, for each of the plurality of phantom ligands, a confidence score; generating a plurality of DTI features for proteins and ligands in a drug-target interaction (DTI) database using the plurality of phantom ligands; generating a machine learning model using the plurality of DTI features; Wherein generating the machine learning model comprises: obtaining positive training samples based on the plurality of DTI features of proteins and ligands; Obtaining negative training samples based on the multiple DTI features by randomly permuting the multiple DTI features of the protein and the ligand at least once; Using the positive training samples and the negative training samples, training the machine learning model to perform DTI prediction; and The machine learning model is used to predict the likelihood of interaction of a combination of a query protein and a query ligand.

2. The method according to claim 1, wherein the protein clustering is obtained based on one selected from the group consisting of: The similarity of the sequences, Similarity of 3D topological structures, and Existing clusters in the database.

3. The method of claim 1, wherein the confidence score quantifies the uncertainty of the associated phantom ligand.

4. The method of claim 1 , wherein generating the plurality of DTI features comprises: For each of the multiple combinations of ligand and protein in the DTI database: selecting from the plurality of phantom ligands a phantom ligand that is most similar to the ligand under consideration for the combination; and Signatures are generated for proteins under consideration for the combination, wherein the generated signatures characterize the proteins under consideration for the combination.

5. The method of claim 4, wherein the generated features include one selected from the group consisting of: at least one local feature comprising binding site features in concentric shells of increasing radius, at least one global feature other than said binding site feature, and At least one functional annotation. The method of claim 4 , wherein the selection of the most similar phantom ligands is performed based on a distance metric.

7. The method of claim 4, wherein generating the plurality of DTI features further comprises: A confidence vector is obtained that represents confidence in a plurality of components of the DTI feature associated with the most similar phantom ligand.

8. The method of claim 7, wherein the confidences of the plurality of components include at least one selected from the group consisting of: a first confidence score quantifying the uncertainty associated with the most similar phantom ligand, a second confidence score quantifying the fingerprint similarity between the ligand considered for the combination and the most similar phantom ligand, and A third confidence score is dependent on a source from which the DTI feature is obtained.

9. The method according to claim 1, wherein generating the machine learning model comprises, before obtaining the positive training samples and the negative training samples: The plurality of DTI features of proteins and ligands are filtered using a confidence threshold applied to confidence vectors associated with the plurality of DTI features.

10. The method of claim 1, wherein the machine learning model is one selected from a classifier model and a regression model.

11. The method of claim 1, wherein predicting the likelihood of interaction of the combination of the query protein and the query ligand comprises: obtaining possible binding sites and related local features of a query protein based on the plurality of phantom ligands; generating a feature of the query protein, wherein the feature of the query protein includes the local feature; generating a feature of the query ligand, wherein the feature of the query ligand comprises a ligand fingerprint and a ligand descriptor; and The machine learning model is applied to the features of the query protein and the features of the query ligand to obtain the likelihood of interaction between the query ligand and the query protein.

12. The method according to claim 11, wherein the features of the query protein further comprise at least one selected from the group consisting of global features and functional annotations.

13. A non-transitory computer readable medium comprising computer readable program code for predicting drug-target binding using synthetic enhancement data, the computer readable program code causing a computer system to: Generate multiple phantom ligands for multiple proteins in the protein structure database; Wherein generating the plurality of phantom ligands comprises: For a protein cluster selected from the plurality of proteins: Performing structural comparison on proteins in the protein cluster; After the structural alignment, obtaining the plurality of phantom ligands by projecting the ligand of one of the proteins in the cluster onto all other proteins in the cluster; obtaining, for each of the plurality of phantom ligands, a confidence score; generating a plurality of DTI features for proteins and ligands in a drug-target interaction (DTI) database using the plurality of phantom ligands; Generate a machine learning model using the DTI features; Wherein generating the machine learning model comprises: obtaining positive training samples based on the plurality of DTI features of proteins and ligands; Obtaining negative training samples based on the multiple DTI features by randomly permuting the multiple DTI features of the protein and the ligand at least once; Using the positive training samples and the negative training samples, training the machine learning model to perform DTI prediction; and The machine learning model is used to predict the likelihood of interaction of a combination of a query protein and a query ligand.

14. A system for differential drug discovery, the system comprising: Protein structure database; a phantom ligand identification engine configured to generate a plurality of phantom ligands for a plurality of proteins in the protein structure database; Wherein generating the plurality of phantom ligands comprises: For a protein cluster selected from the plurality of proteins: Performing structural comparison on proteins in the protein cluster; After the structural alignment, obtaining the plurality of phantom ligands by projecting the ligand of one of the proteins in the cluster onto all other proteins in the cluster; obtaining, for each of the plurality of phantom ligands, a confidence score; a phantom ligand database storing the plurality of phantom ligands; Stores the drug-target interaction (DTI) database of proteins and ligands; a feature generation engine configured to generate a plurality of DTI features for proteins and ligands in the DTI database using the plurality of phantom ligands in the phantom ligand database; a machine learning model training engine configured to generate a machine learning model using the DTI features; Wherein generating the machine learning model comprises: obtaining positive training samples based on the plurality of DTI features of proteins and ligands; Obtaining negative training samples based on the multiple DTI features by randomly permuting the multiple DTI features of the protein and the ligand at least once; Using the positive training samples and the negative training samples, training the machine learning model to perform DTI prediction; and A DTI prediction engine is configured to predict the likelihood of interaction of a combination of a query protein and a query ligand using the machine learning model.

15. The system of claim 14, wherein generating the plurality of DTI features comprises: For each of the multiple combinations of ligand and protein in the DTI database: selecting from the plurality of phantom ligands a phantom ligand that is most similar to the ligand under consideration for the combination; and Signatures are generated for proteins under consideration for the combination, wherein the generated signatures characterize the proteins under consideration for the combination.

16. The system of claim 15, wherein the generated features include one selected from the group consisting of: at least one local feature comprising binding site features in concentric shells of increasing radius, at least one global feature other than said binding site feature, and At least one functional annotation.

17. The system of claim 15, wherein generating the plurality of DTI features further comprises: obtaining a confidence vector representing confidence in a plurality of components of the DTI feature associated with the most similar phantom ligand, wherein the confidences of the plurality of components include at least one selected from the group consisting of: a first confidence score quantifying the uncertainty associated with the most similar phantom ligand, a second confidence score quantifying the fingerprint similarity between the ligand considered for the combination and the most similar phantom ligand, and A third confidence score is dependent on a source from which the DTI feature is obtained.

18. The system of claim 14, wherein predicting the likelihood of interaction of the combination of the query protein and the query ligand comprises: obtaining possible binding sites and related local features of a query protein based on the plurality of phantom ligands; generating a feature of the query protein, wherein the feature of the query protein includes the local feature; generating a feature of the query ligand, wherein the feature of the query ligand comprises a ligand fingerprint and a ligand descriptor; and The machine learning model is applied to the features of the query protein and the features of the query ligand to obtain the likelihood of interaction between the query ligand and the query protein.

Images