Method and system for predicting drug binding using synthetic data

Abstract

To acquire superior predictions of drug-target interactions.SOLUTION: A method for predicting drug-target binding using synthetically augmented data includes: generating a multitude of ghost ligands for a multitude of proteins in a protein structure database; generating a multitude of drug-target interaction (DTI) features for proteins and ligands in a DTI database, using the multitude of ghost ligands; generating a machine learning model using the multitude of DTI features; and predicting a likelihood of an interaction for a combination of a query protein and a query ligand using the machine learning model.SELECTED DRAWING: Figure 1A

Description

[Background technology]

[0001] [CROSS REFERENCE TO RELATED APPLICATIONS] This application claims priority under 35 U.S.C. § 119(e) to U.S. Provisional Application No. 62 / 788,682, entitled "METHOD AND SYSTEM FOR PREDICTING DRUG BINDING USING SYNTHETIC DATA," filed on January 4, 2019, which has at least one of the same inventors as this application. U.S. Provisional Application No. 62 / 788,682 is incorporated herein by reference.

[0002] Computational methods exist for predicting interactions between ligands and proteins. They are generally classified as "ligand-based" or "structure-based", depending on the type of information used to make the prediction.

[0003] Protein-based predictions may be more universally generalizable due to the potential for learning biophysical compatibility, but are very data-bound. Specifically, protein-based predictions use 3D molecular structures of ligands co-crystallized with proteins to assess or predict interactions. These methods are computationally intensive and tend to be trained only on hundreds to thousands of different proteins. Neural networks trained on these tend to have a very high ratio of feature space to data. As a result, these methods can generate a significant number of false negatives and / or false positives in docking when applied to previously unseen protein systems or drug scaffolds.

[0004] 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. When the ligand-to-protein data ratio is high, one standard approach is to derive many different models for each of the 2,000 proteins. These tend to be successful, often outperforming high-throughput experimental results, but (1) they represent only about 10% of human proteins, (2) they can be weaker when there is not much chemical diversity among individual protein examples, and (3) individual models do not learn the physical attributes of drug-protein compatibility and cannot benefit from the data used to generate other models. Summary of the Invention

[0005] In general, in one aspect, one or more embodiments relate to a method for predicting drug-target binding using synthetically augmented data, the method including: generating a plurality of ghost ligands for a plurality of proteins in a protein structure database; generating a plurality of drug-target interaction (DTI) features for the proteins and ligands in a DTI database using the plurality of ghost ligands; generating a machine learning model using the plurality of DTI features; and predicting a likelihood of an interaction for a combination of a query protein and a query ligand using the machine learning model.

[0006] In general, in one aspect, one or more embodiments relate to a non-transitory computer-readable medium including computer-readable program code for predicting drug-target binding using synthetically augmented data, the computer-readable program code causing a computer system to generate a plurality of ghost ligands for a plurality of proteins in a protein structure database, generate a plurality of drug-target interaction (DTI) features for the proteins and ligands in a DTI database using the plurality of ghost 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.

[0007] In general, in one aspect, one or more embodiments relate to a system for specific drug discovery, the system comprising: a protein structure database; a ghost ligand identification engine configured to generate a plurality of ghost ligands for a plurality of proteins in the protein structure database; a ghost ligand database storing the plurality of ghost ligands; a drug-target interaction (DTI) database storing proteins and ligands; a feature generation engine configured to generate a plurality of DTI features for the proteins and ligands in the DTI database using the plurality of ghost ligands in the ghost 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 predict an interaction probability for a combination of a query protein and a query ligand using the machine learning model. [Brief description of the drawings]

[0008] The present embodiments are illustrated by way of example and are not intended to be limited by the figures of the accompanying drawings. [Figure 1A] FIG. 1 shows a block diagram of a system for predicting drug binding according to one or more embodiments. [Figure 1B] FIG. 1 shows a block diagram of a protein structure database according to one or more embodiments. [Figure 1C] FIG. 1 shows a block diagram of a ghost ligand database according to one or more embodiments. [Figure 1D] FIG. 1 shows a block diagram of a drug-target interaction database according to one or more embodiments. [Figure 1E] FIG. 1 shows a block diagram of a protein annotation database according to one or more embodiments. [Diagram 2] 1 shows a flowchart illustrating a method for training a machine learning model for predicting drug-target interactions, according to one or more embodiments. [Diagram 3] 1 shows a flowchart illustrating a method for generating a ghost ligand database according to one or more embodiments. [Figure 4] 1 shows a flowchart illustrating a method for generating drug-target interaction (DTI) features according to one or more embodiments. [Diagram 5] 1 shows a flowchart illustrating a method for generating a machine learning model for DTI prediction, according to one or more embodiments. [Figure 6] 1 shows a flowchart illustrating a method for predicting an interaction between a query protein and a query ligand, according to one or more embodiments. [Figure 7A] 13 provides an example for generating ghost ligands, according to one or more embodiments. [Figure 7B] FIG. 1 shows a diagram of a concentric shell model used to obtain binding site features, in accordance with one or more embodiments. [Figure 8] 1 illustrates the generation of training data for a machine learning model, in accordance with one or more embodiments. [Figure 9] 1 shows a performance comparison between an embodiment of the present disclosure and a conventional approach. [Figure 10A] 1 illustrates a computing system in accordance with one or more embodiments. [Figure 10B]1 illustrates a computing system in accordance with one or more embodiments. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] Certain embodiments disclosed herein are described in detail below with reference to the accompanying drawings, in which like elements in the various figures may be designated by like reference numbers and / or like names for consistency.

[0010] The following detailed description is merely exemplary in nature and is not intended to limit the embodiments disclosed herein or the application and uses of the embodiments disclosed herein. Furthermore, there is no intention to be bound by any expressed or implied theory presented in the preceding technical field, background, brief summary or the following detailed description.

[0011] In the following detailed description of some embodiments disclosed herein, numerous specific details are set forth in order to provide a more thorough understanding of the various embodiments disclosed herein. However, it will be apparent to those skilled in the art that the embodiments may be practiced without these specific details. In other instances, well-known features have not been described in detail to avoid unnecessarily complicating the description.

[0012] Throughout this application, ordinal numbers (e.g., first, second, third, etc.) may be used as adjectives for elements (i.e., any noun in this application). The use of ordinal numbers does not imply or create any particular order of elements or limit any element to being only a single element, unless expressly disclosed, such as by use of terms such as "before," "after," "single," etc. Rather, the use of ordinal numbers is to distinguish elements. By way of example, a first element is different from a second element, a first element encompasses two or more elements, and may follow (or precede) a second element in the ordering of elements.

[0013] In one or more embodiments of the invention, elements of protein-based 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).

[0014] A wealth of data points exist for protein-ligand interactions in DTI databases. For example, the ChEMBL database contains approximately 15,000,000 records that describe protein-ligand pairs as bound or non-bound (and often provide a measure of affinity or confidence). However, using the content of these DTI databases to predict interactions of new pairs of proteins and ligands has certain limitations, such as limited coverage of the human proteome and weak predictions when the diversity between individual protein examples is limited. Thus, the quality of predictions by machine learning models operating purely on DTI database records may be limited. Alternatively, machine learning models may manipulate the 3D molecular structure of the ligand co-crystallized with the protein to capture the biophysics of the interaction between the protein and the ligand. Although several databases exist that capture protein-ligand interactions (e.g., sc-PDB), these databases contain a relatively small number of data points and tend to be highly redundant. Furthermore, these databases tend to lack the diversity of protein structures and the diversity of chemical species. Thus, training machine learning models on these data points can be challenging due to insufficient data volume. In particular, machine learning models that operate on the 3D molecular structures of ligands co-crystallized with proteins have a high-dimensional feature space that, combined with the limited availability of suitable training samples, makes machine learning models vulnerable to overfitting.

[0015] In one or more embodiments of the present disclosure, the local 3D features and the DTI recordings are used in combination by a machine learning model to enable predictions that are better than the traditional protein-based and ligand-based predictions described above. More specifically, synthetic data is generated by projecting the DTI recordings onto a 3D structural model of a known protein-ligand complex to generate local protein features that are not otherwise available. The 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 about the query protein and query ligand pair.

[0016] Turning to Figure 1A, a system for predicting drug binding using synthetic data is shown, according to one or more embodiments. The system (100) may include a ghost 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 ghost ligand database (170), a drug-target interaction database (180), and a protein annotation database (190). Each of these components is described below.

[0017] The ghost ligand identification engine (110), according to one or more embodiments, includes instructions in the form of computer readable program code for performing at least one of the steps described in Figures 2 and 3 to generate a ghost ligand database (170) of ghost ligands and associated confidence scores from proteins in the protein structure database (160). The ghost ligand identification engine (110) can obtain ghost ligands for proteins by structurally aligning known homologs and projecting the ligands of these known homologs onto the aligned sites of the proteins. Although these ghost ligands may not interact with the proteins, they may function as placeholders indicating structural compatibility between the ghost ligands and the aligned sites of the proteins. The ghost ligand identification engine (110) is operatively connected to the protein structure database (160) and the ghost ligand database (170).

[0018] The feature generation engine (120), according to one or more embodiments, includes instructions in the form of computer readable program code for performing at least one of the steps described in Figures 2 and 4 to generate drug-target interaction (DTI) features used to train a machine learning model for DTI prediction. The feature generation engine (110) can generate features for proteins and ligands using data from the ghost ligand database (170), the drug-target interaction database (180), and the protein annotation database (190). Thus, the feature generation engine (120) is operatively connected to the ghost ligand database (170), the drug-target interaction database (180), and the protein annotation database (190).

[0019] Continuing with FIG. 1A, the machine learning model training engine (130) includes instructions in the form of computer readable program code for performing at least one of the steps described in FIG. 2 and FIG. 5 to train the machine learning model for DTI prediction (140), according to one or more embodiments. The machine learning model training engine (130) may use the DTI features generated by the feature generation engine (120) for training. Thus, the machine learning model training engine (130) is operatively connected to the feature generation engine (120). The resulting machine learning model for DTI prediction (140) may be any type of classifier capable of predicting the interaction between a query drug and a query protein. In one or more embodiments, the machine learning model for DTI prediction (140) may be a deep neural network (DNN) based classifier. It is.

[0020] The drug-target interaction (DTI) prediction engine (150), according to one or more embodiments, includes instructions in the form of computer readable program code for performing at least one of the steps described in FIG. 6 to predict a drug-target interaction for a query drug and a query protein using the machine learning model (140). The DTI prediction engine (150) generates features for the query protein and query ligand associated with the query drug that are compatible with the machine learning model (140), and then calculates the interaction probability based on those features 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.

[0021] Protein structure database (160), according to one or more embodiments, may be any type of storage unit and / or device for storing data, such as a file system, a database, a collection of tables, or any other storage mechanism. Protein structure database (160) is described below with reference to FIG. 1B.

[0022] The ghost ligand database (170), according to one or more embodiments, 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). The ghost ligand database (170) is described below with reference to FIG. 1C.

[0023] The drug-target interaction database (180), according to one or more embodiments, 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). The drug-target interaction database (180) is described below with reference to FIG. 1D.

[0024] Protein annotation database (190), according to one or more embodiments, 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). Protein annotation database (190) is described below with reference to FIG. 1D.

[0025] Turning to FIG. 1B, a protein structure database (160) is shown, according to one or more embodiments. The protein structure database (160) can store 3D models of proteins (162A, 162B, 162N). Each of the 3D models can be associated with a homology model (164A, 164B, 164N) and / or an experimental model (166A, 166B, 166N). Examples of publicly available protein structure databases (160) that can be used include, but are not limited to, the Protein Data Bank (PDB) and SWISS-MODEL.

[0026] Turning to FIG. 1C, a ghost ligand database (170) is shown, according to one or more embodiments. The ghost ligand database (170) may store identified ghost ligands (e.g., 174A, 174B, 174N) for multiple 3D models of a protein (e.g., 172A, 172B, 172N). Additionally, for each identified ghost ligand, a confidence score (e.g., 176A, 176B, 176N) that may be based on similarity is included. The ghost ligand database may be established as described in FIG.

[0027] Turning to FIG. 1D, a drug-target interaction database (180) is shown according to one or more embodiments. 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) that can be used include, but are not limited to, STITCH and ChEMBL. These databases can include a large number of data points (approximately 15,000,000 ChEMBL).

[0028] Turning to FIG. 1E, a protein annotation database (190) is shown, according to one or more embodiments. The protein annotation database (190) can store associated annotations (e.g., 194A, 192B, 194N) for multiple proteins (e.g., 192A, 192B, 192N). The annotations associated with a protein can include available information about the protein, which may have been added manually or computationally to the protein annotation database. For example, the UniProt database may be used.

[0029] Figures 2, 3, 4, 5, and 6 show flow charts according to one or more embodiments. The flow charts of Figures 2, 3, 4, and 5 show a method for training a machine learning model to predict drug-target interactions, and the flow chart of Figure 6 shows a method for using a machine learning model to predict drug-target interactions. One or more of the steps of Figures 2, 3, 4, 5, and 6 may be performed by components of the system (100) described above with reference to Figure 1A. In one or more embodiments, one or more of the steps shown in Figures 2, 3, 4, 5, and 6 may be omitted, repeated, and / or performed in a different order than that shown in Figures 2, 3, 4, 5, and 6. Additional steps may be further performed. Thus, the scope of the present invention should not be considered limited to the particular arrangement of steps shown in Figures 2, 3, 4, 5, and 6.

[0030] Turning to the flowchart of Figure 2, a method for generating a machine learning model for predicting drug-target interactions (DTIs) is shown. Figure 2 is intended to introduce the major steps toward generating a machine learning model, while the flowcharts discussed thereafter provide a more detailed explanation. After completion of the method of Figure 2, the resulting machine learning model can be used to make predictions, as illustrated in Figure 6.

[0031] In step 200, a ghost ligand database is generated from proteins retrieved from a protein structure database. A detailed description of step 200 is provided in FIG.

[0032] In step 202, drug-target interaction (DTI) features, including ligand features and protein features, are generated. A detailed description of step 202 is provided in FIG.

[0033] In step 204, a machine learning model for DTI prediction is generated. A detailed description of step 204 is provided in FIG.

[0034] Turning to the flow chart in Figure 3, a method for generating a ghost ligand database is described. Using the steps described subsequently, homology relationships between proteins are exploited to map known ligands to the structures of their homologues (both experimental and homology models), called "ghosts".

[0035] In step 300, proteins are retrieved from a protein structure database. For each protein, a 3D model can be retrieved.

[0036] In step 302, the obtained proteins are clustered by sequence or domain. In one or more embodiments, a protein cluster is defined as any set of two or more proteins that share similarity in primary sequence or three-dimensional topology, commonly known as a fold. Protein clustering can be obtained directly from publicly available databases, such as PDB, SCOP, CATH, PFAM, or Uniprot. Protein clusters may be created manually based on sequence similarity using protein sequence alignment tools and clustering tools, such as BLAST, CD-HIT, or UCLUST. Alternatively, proteins with three-dimensional structural models may be clustered by grouping unrelated proteins that share a common topology or fold.

[0037] In step 304, one of the clusters is selected for further processing.

[0038] In step 306, pairwise structural alignment is performed for all proteins in the selected cluster. Figure 7A shows an example of structural alignment of 3D structures of proteins. Three-dimensional (3d) structural alignment seeks 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. Structural alignment can be performed on the complete protein structure, or on a subselection of residues, such as a single domain, or residues surrounding, for example, a ligand binding site. The selection of ligand binding site residues for 3d structural alignment is an optimal heuristic approach for mapping ghost ligands.

[0039] In step 308, ghost ligands are obtained by projecting each ligand onto a cluster peer. A confidence score for each projection can be obtained, consisting of individual scores representing the confidence for different components of the ligand projection. An example for projecting ligands onto cluster peers is provided in FIG. 7A. The confidence score can be based on any metric that quantifies the uncertainty of the heuristically defined structural representation selected to model the DTI interaction. In this step, the confidence score can include, for example, percentage sequence identity, sequence similarity, or a homology model quality metric such as QMEAN. The confidence score can also include a metric that describes the quality of the structural alignment, for example, the root mean square deviation of the local or global alignment. Multiple confidence scores can also be used.

[0040] In step 310, the ghost ligands and associated confidence scores are stored in a ghost ligand database.

[0041] In step 312, a determination is made as to whether additional clusters remain for processing. If additional clusters remain, 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. Upon completion of the method of FIG. 3, the ghost ligand database may include a comprehensive collection of ghost ligands and associated confidence scores for all proteins that have been processed as described.

[0042] Turning to the flowchart of FIG. 4, a method for generating drug-target interaction (DTI) features is described. The generated features include features for the ligand and features for the protein. These features can then be used to train a machine learning model for predicting drug-target interactions. Features for multiple combinations of proteins and ligands may be generated to ensure the availability of sufficient training samples.

[0043] In step 400, a drug-target interaction for a ligand-protein combination is selected from the DTI database. Subsequent steps are performed for this considered combination of ligand and protein. This step may be repeated later for other combinations of ligand and protein.

[0044] In step 402, features are generated for the selected ligand. These features can include ligand fingerprints and ligand descriptors. Fingerprints can capture the structure of the ligand in a descriptor format and can be based on a SMILES representation of the underlying molecule using fixed length vectors. Molecular fingerprinting can include, for example, atom pair, extended connectivity fingerprints, graph-based fingerprints, torsion fingerprints, or pharmacophore fingerprints. Molecular weight, number of rotatable bonds, number of hydrogen bond donors, number of hydrogen bond acceptors, hydrophobicity, aromaticity, and functional group composition, for example, may be used as ligand descriptors. Molecular shape descriptors, such as ellipticity, geometrical descriptors, branching descriptors, or chirality descriptors, may also be used.

[0045] In step 404, ghost ligands are searched from a ghost ligand database for the selected protein.

[0046] In step 406, each ghost ligand is scored based on its difference from the drug, or more specifically, the selected ligand. The higher the similarity, the higher the score. The similarity between ligands can be measured using a number of methods, such as the Tanimoto distance, the Delta distance, or the cosine distance. A distance metric for comparing molecular fingerprints can be used for scoring.

[0047] Steps 404 and 406 may be performed for all protein models available for the selected protein (eg, homology models or experimentally derived models).

[0048] In step 408, the ghost ligand that is most similar to the selected ligand is selected for further processing.

[0049] 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 ghost ligand. The confidence vector may include confidence scores representing the ghost ligand projections from step 308. The confidence vector may further include a confidence score for the fingerprint similarity between the selected ligand and the most similar ghost ligand selected in step 408. Additionally, the confidence vector may include a score for the confidence of the selected DTI. The confidence of the selected DTI may be scored based on the source of the DTI data (e.g., different scores may be assigned depending on whether the DTI data was obtained using high-throughput screening or low-throughput screening, etc.).

[0050] In step 412, local features for the selected protein surrounding the most similar ghost ligand are obtained. The local features may include binding site features present in concentric shells of increasing radius, as shown in FIG. 7B. For each of the concentric shells, several descriptors are provided, e.g., atom type descriptors that specify the presence of atoms in the shell. For example, 70 atom type descriptors can be provided per shell radius. The descriptors may further include, but are not limited to, the flexibility or rigidity of the binding site within the shell region, residue contacts within the shell region, and / or biophysics and any other factors that indirectly describe the shape of the binding site. However, the features may not specify the exact position or coordinates of the atoms. The local features may also include a graph description of the ligand binding site, describing the distances between the amino acids surrounding the ligand binding site in a network format. The local features may also be defined by the shape of a pocket, which corresponds to a void not occupied by protein residues. The pocket void can be determined by pocket detection methods such as, for example, flooded filling, concavity, or solvent accessibility. The local features defined by the pocked void space may include the shape of the void space, including, for example, the volume, ellipticity, curvature, branching pattern, or the stability of the space based on the dynamics of nearby residues. The local features defined by the residues flanking the pocket void space may include, for example, the orientation of the residues, the geometric availability of hydrogen bond donor and acceptor groups, hydrophobicity, aromaticity, or the geometric availability of pi-stacking interactions. The local features may also include the description of a ligand-binding channel, including 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 a transient interaction with the ligand in the dynamic process of binding and dissociation. The features of the ligand-binding channel may include similar features to those that define the pocket, such as, for example, the orientation of the residues, amino acid composition, and the availability of hydrogen bond donors and acceptors.

[0051] In step 414, global features for the structure and / or sequence of the selected protein are obtained by expanding to an outer shell with a large radius as described in step 412. Shell radii corresponding to local features may include thresholds of, for example, 5 Å, 10 Å, or 15 Å. Shell radii corresponding to domain level descriptions or global protein descriptions may have larger distance thresholds, for example, 20 Å, 25 Å, or 30 Å, or may have no distance threshold. Global features may further include domain or fold descriptions and may be extracted from publicly available databases, for example, SCOP, CATH, PFAM, etc. Global features may further include features derived from the protein sequence and may include, for example, the presence of common sequence motifs. Global features may include descriptions of protein folding states, such as, for example, the presence and biophysical properties of intrinsically disordered regions, hinges, loops, ordered regions, or regulatory domains. Global features may also be described in terms of distance from the ligand binding site.

[0052] In step 416, functional annotations of the selected proteins are obtained. The functional annotations can be obtained from a protein annotation database. The functional annotations can include, for example, Enzyme Commission (EC) numbers, Gene Ontology (GO) annotations, or Uniprot keywords. The functional annotations can also include whether location-specific properties of the protein, such as catalytic sites, post-translational modifications, disease associations, or genetic mutations, are documented or not.

[0053] In step 418, features of the selected proteins are generated. These features may include local features, global features, and / or functional annotations.

[0054] At step 420, a determination is made as to whether additional DTIs remain for processing. If additional DTIs remain, 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 end. Upon completion of the method of FIG. 4, a determination is made as to whether additional DTIs remain for the ligands and proteins listed in the DTI database. A comprehensive collection of features becomes available.

[0055] Turning to the flowchart in Figure 5, a method for generating a machine learning model for DTI prediction is described. Based on the DTI features obtained as described in Figure 4, a machine learning model is obtained that reflects the compatibility between the protein environment and the ligand attributes.

[0056] In step 500, ligand and protein features are obtained. Obtaining the ligand and protein features can be performed as described in steps 402 and 418 of FIG.

[0057] In step 502, the ligands and proteins are filtered by a function of the confidence vector established in step 408 of FIG. 4. The filtering may implement a confidence threshold, and only samples that exceed the confidence threshold are considered for further processing. The confidence function may convert the confidence vector into a single score for filtering. The confidence function may, for example, convert the confidence scores into probabilities and apply Bayesian statistics to evaluate the composite probability. The confidence function may apply individual cutoff thresholds to each element of the confidence vector as a means of selecting samples suitable for use in machine learning. The thresholds or equations for the confidence function may be set by automatically testing different combinations as hyperparameters of the machine learning algorithm.

[0058] In step 504, the ligand features and protein features are concatenated to generate positive training samples.

[0059] In step 506, the ligand features and protein features are shuffled. The shuffled ligand features and protein features are concatenated to generate negative training samples. This step can be repeated multiple times to evaluate different ratios of positive to negative training samples, for example, 1:1, 1:5, 1:10, 1:19, or 1:20.

[0060] In step 508, a machine learning model for DTI prediction is trained using the positive and negative training samples. For example, a backpropagation based learning algorithm may be used. In one or more embodiments, the training samples may be weighted based on their associated confidence vector. In one or more embodiments, transfer learning is used to more effectively 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 stage, the confidence threshold may be raised to reduce the number and improve the quality of the training examples. Additionally or alternatively, the subsequent retraining stage may limit the training examples for selecting the drug or target class. The machine learning model may be a supervised discriminative classification or regression model, such as, for example, 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 size of the training data points (100,000-10,000,000) and the dimensionality of the training data features (1000-10,000). In one embodiment, the artificial neural network representation takes the form of a fully connected network with a feature input layer, e.g., two hidden layers with 512 and 256 nodes, respectively, and two output nodes corresponding to interacting and non-interacting pairs. In one embodiment, an 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 functional features of proteins.

[0061] Turning to the flow chart of Figure 6, a method for predicting interactions between a query protein and a query ligand is described. Using a machine learning model trained as described with reference to Figures 2-5, the "compatibility" of the query protein and the query compound can be tested by applying the 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.

[0062] In step 600, a query protein and a query ligand are obtained. The query protein and the query ligand can be obtained from a user who wishes to obtain a prediction of an interaction between the query protein and the query ligand.

[0063] In step 602, possible binding sites and associated local features are obtained for the query protein, as previously described in steps 404-412 of Figure 4. Thus, one or more binding sites may be obtained from one or more experimental or homology models. Alternatively, the binding site and associated local features may be obtained from a user, for example if the user wishes to specify a particular binding site.

[0064] In step 604, global feature and protein annotations are obtained for the query protein. Global feature and protein annotations can be obtained as previously described in steps 414 and 416 of FIG.

[0065] In step 606, features are generated for the query protein. These features may include local features, global features, and / or functional annotations.

[0066] In step 608, the ligand fingerprint and ligand descriptor are obtained as previously described in step 402 of FIG.

[0067] In step 610, features are generated for the query ligand. These features may include ligand fingerprints and ligand descriptors.

[0068] 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.

[0069] The following paragraphs further describe embodiments of the present disclosure based on various examples. Those skilled in the art will understand that the present disclosure is not limited to these examples.

[0070] (i) Ghost ligand sample: Turning to FIG. 7A, an example (700) for generating ghost ligands is shown. Three hypothetical protein structures are shown (top row). Two of the three hypothetical protein structures are actually interacting with the ligand (left and center columns, top row). The center row shows various structural alignments of the three hypothetical proteins. As a result of the structural alignments, the ligands can be projected onto other proteins. Based on the similarity of the binding sites, a confidence score is assigned. While the real ligand-protein pairing has a confidence score of 1.0, the ghost ligand-protein pairing has a lower confidence score. The bottom row shows the resulting ghost ligand-protein pairs as they can be stored in the ghost ligand database.

[0071] FIG. 7B shows a diagram of a concentric shell model (750) used to obtain characteristics of a binding site, according to one or more embodiments. Concentric shells of increasing radius (r) represent the structure of the binding site. The shells surround a central chemical structure that is considered to be part of the protein. The inner shells mainly capture local features in close proximity to the binding site, while the outer shells capture increasingly global features. Features representing proteins can be based on concentric shell models, thus capturing local and global features of the protein without specifying the exact 3D geometry (e.g., at the atomic level).

[0072] (ii) Sample confidence vector: The embodiments of the present disclosure rely on a heuristic process to augment drug-target interaction (DTI) data with assumed 3D structural representations. These assumed DTI representations can provide information-rich features for machine learning to improve models aimed at predicting protein-ligand interactions. Obtaining these approximate DTI representations for any given DTI data point requires several assumptions, outlined throughout steps 200 and 202 of FIG. 2. For example, the 3D protein structures used to represent DTI may be sourced from homology models rather than directly from experimental coordinates.

[0073] The confidence vector is composed of multiple metrics that account for the measurable uncertainty of the approximated DTI representation. These metrics arise during the creation of the ghost ligand database (step 200) and the projection of known DTI data onto the ghost ligand database (step 202). In one example, the confidence vector includes four components, including: (1) the percentage of sequence identity between the homology model representing the DTI protein and its source template, (2) the RMSD from the alignment between the source structure of the ghost ligand and the homology model representing the DTI protein, (3) the Tanimoto similarity between the morgan3 fingerprint of the DTI ligand and the ghost ligand template, and (4) the confidence of the DTI data point.

[0074] In this example, the drug-target interaction (DTI) database indicates that the ligand gefitinib interacts with the protein Aurora kinase A. The DTI database assigned a probability of interaction of 85% based on the accuracy of the source biophysical experiments. The specific interaction between gefitinib and Aurora kinase A is not present in the source 3D structure database. In creating the ghost ligand database, a homology model of the Aurora kinase A protein was created from its close homolog, Aurora kinase B, with which it shares 72.5% sequence identity. The closest molecule to gefitinib that successfully mapped to the homology model was erlotinib, which shares a Morgan3 fingerprint Tanimoto similarity of 0.372. The location of the erlotinib ghost ligand was estimated based on a structural alignment between the Aurora kinase A homology model and the erlotinib-EGFR cocomplex crystal structure, with an RMSD of 2.345 Å for the ligand binding site. The corresponding confidence vector was therefore [85%, 72.5%, 0.372, 2.345 Å].

[0075] (iii) Sample training data and negative shuffling: The described methodology focuses on enhancing drug-target interaction (DTI) pairs from a DTI database using a mixture of relevant features obtained through a set of deterministic mappable relationships and heuristically modeled features (local structural features). Each row of the DTI database can be converted into a feature vector as illustrated in FIG. 8, which 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 drug. Protein global and functional features may be obtained from any protein using standard techniques of database lookup and protein identifier mapping. Local protein features may be the result of a heuristically defined process outlined in this patent. Because they are modeled, they may be inaccurate. 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.

[0076] If the neural network is trained only by positive examples of true drug-target interactions extracted from the drug-target interaction dataset, the model may learn to ignore core and obvious patterns of interactions, as they provide no signal to the model. In addition, it may be necessary to control for potentially important biases towards drugs and targets that are highly represented in the drug-target interaction dataset. Therefore, it may be beneficial to sample negative examples in relative proportions of each drug and target. As a result, the model may learn patterns based on the positive and negative examples. In Figure 8, the randomized negatives shuffle the ligand component of the feature vector (white) with the protein component of the feature vector (three shades of grey). The resulting negative examples can be used to train the classification engine and balance the presence of individual ligand or protein features. Equal use of individual ligand and / or protein features in the positive and negative sets avoids the network learning that individual features are particularly relevant for binding in general.

[0077] An embodiment of the present disclosure uses ghost ligands to create local protein features for proteochemometrics (PCM) from protein-ligand datasets. More specifically, drug-target interaction (DTI) data is threaded into 3D atomic models of protein-ligand complexes to derive local protein features. The mixed feature dataset for PCM may include local (pocket), regional (domain), and global (whole protein) annotations and / or functional annotations.

[0078] Traditionally, training data for machine learning needs to be "model quality data" with high confidence. The use of prediction (ghost ligand + threading) according to one or more embodiments to generate training data for machine learning algorithms seems counterintuitive. Specifically, if the heuristic approach is not accurate enough, conventional wisdom suggests that the introduction of local features derived from the combination of ghost ligand and threading may introduce additional noise, thereby resulting in poorer performance than conventional DTI PMC. However, as the performance comparison in FIG. 9 (900) shows, improved performance is achieved by the introduction of local features derived by the approach described in this patent. Omitting the local features derived by this approach is comparable to the performance that would be achieved by DTI PMC alone.

[0079] Specifically, Figure 9 shows a performance comparison for ranking small molecule ligands and their binding potential to 8717 proteins. To test the ranking, 100 molecules are randomly removed from the training data and used for testing. The figure plots the predicted ranks of known interactions of these 100 random drugs. For example, without local features, only about 63% of the actual interactions are observed among the predictions for the top 300 proteins of 8717 (top ~3.5%) (dotted line). Including local features estimated by this procedure increased the detection rate to about 75% at the same threshold (solid line).

[0080] Various embodiments of the present disclosure have one or more of the following advantages: Embodiments of the present invention enable prediction of drug-target interactions (DTI) using machine learning models that reflect the compatibility between the protein environment and ligand attributes. Localized 3D features are created to represent the binding site even when no 3D information is available for the interaction under consideration.

[0081] Mapping known drug-target interactions to homology models synthetically augments the rich DTI training data with high-dimensional biophysical information to train deep neural networks. This approach therefore allows the use of comprehensive DTI databases for DTI database entries even when the location of the drug binding to the protein is not necessarily known.

[0082] The method according to one or more embodiments does not require detailed knowledge of the biophysics of protein-ligand interactions, and therefore does not require precise 3D coordinates of atoms, allowing the drug-target interaction to be mapped into protein pockets using DTI databases and homology models.

[0083] The embodiments of the present disclosure allow for very large training data compared to structure-based deep learning approaches that require a reduced feature space and rely on 3D atomic coordinates. Furthermore, the embodiments of the present disclosure have been found to generalize well. Initial performance evaluations suggest that the described methods run approximately 1,000,000 times faster compared to 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 Exhibits A and B, the methods according to one or more embodiments may be used as an accurate in silico alternative or addition to other in silico and / or experimental methods for predicting drug-target interactions.

[0084] The embodiments of the present disclosure may have a variety of applications. For example, the embodiments may be used for proteomic screening (e.g., to perform toxicity prediction or phenotype deconvolution prediction), virtual screening, and generally for drug discovery and development.

[0085] The 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, as shown in FIG. 10A, a computing system (1000) may include one or more computer processors (1002), non-persistent storage (1004) (e.g., volatile memory such as random access memory (RAM), cache memory), persistent storage (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 numerous other elements and functions.

[0086] The computer processor (1002) may be an integrated circuit for processing instructions. For example, the computer processor 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, a keyboard, a mouse, a microphone, a touch pad, an electronic pen, or any other type of input device.

[0087] The communications interface (1012) may include integrated circuits 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 to another device, such as another computing device.

[0088] Additionally, the computing system (1000) may include one or more output devices (1008), such as a screen (e.g., a liquid crystal display (LCD), plasma display, touch screen, cathode ray tube (CRT) monitor, projector, or other display device), a printer, external storage, or any other output device. The one or more output devices may be the same as or different from the input devices. The input and output devices may be locally or remotely connected to the computer processor (1002), non-persistent storage (1004), and persistent storage (1006). Many different types of computing systems exist, and the aforementioned input and output devices may take other forms.

[0089] Software instructions in the form of computer readable program code for carrying out 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, diskette, tape, flash memory, physical memory, or other computer readable storage medium. In particular, the software instructions may correspond to computer readable program code that, when executed by a processor, is configured to carry out one or more embodiments of the present disclosure.

[0090] The computing system (1000) of FIG. 10A may be connected to or may be part of a network. For example, as shown in FIG. 10B, 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 the computing system shown in FIG. 10A, or a group of nodes combined may correspond to the computing system shown in FIG. 10A. As an example, an embodiment of the present disclosure may be implemented on a node of a distributed system that is connected to other nodes. As another example, an embodiment of the present disclosure may be implemented on a distributed computing system having multiple nodes, and each part of the present disclosure may be located on a different node in the distributed computing system. Furthermore, one or more elements of the computing system (1000) described above may be located at a remote location and connected to other elements over a network.

[0091] Although not shown in Figure 10B, a node may correspond to a blade in a server chassis that connects to other nodes via a backplane. As another example, a node may correspond to a server in a data center. As another example, a node may correspond to a computer processor or micro-core of a computer processor having shared memory and / or resources.

[0092] Nodes in the network (1020) (e.g., node X (1022), node Y (1024)) may be configured to provide services to client devices (1026). For example, the nodes may be part of a cloud computing system. The nodes may include functionality to receive requests from and send responses to the client devices (1026). The client devices (1026) may be computing systems, such as the computing system illustrated in FIG. 10A. Additionally, the client devices (1026) may include and / or execute all or a portion of one or more embodiments of the present disclosure.

[0093] The computing system or groups of computing systems illustrated in Figures 10A and 10B may include functionality to perform various operations disclosed herein. For example, the computing systems may perform communication between processes on the same or different systems. Various mechanisms using 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, files, signals, sockets, message queues, pipelines, semaphores, shared memory, message passing, and memory-mapped file implementations. Further details regarding some of these non-limiting examples are provided below.

[0094] Based on the client-server networking model, a socket may function as an interface or communication channel endpoint to enable bidirectional data transfer between processes on the same device. First and foremost, according to the client-server networking model, a server process (e.g., a process providing data) may create a first socket object. The server process then binds to the first socket object, thereby associating the first socket object with a unique name and / or address. After creating and binding the first socket object, the server process waits and listens for incoming connection requests from one or more client processes (e.g., processes seeking data). At this point, if the client process wishes to obtain data from the server process, the client process begins by creating a second socket object. The client process then proceeds to generate a connection request that includes at least the second socket object and the unique name and / or address associated with the first socket object. The client process then sends the connection request to the server process. Depending on availability, the server process may accept the connection request and establish a communication channel with the client process, or the server process may be busy processing other operations and queue the connection request in a buffer until the server process is ready. An established connection notifies the client process that communication may begin. In response, the client process may generate a data request that specifies the data the client process wishes to obtain. The data request is then sent 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 response that includes at least the requested data and sends the response to the client process. Data may more generally be transferred as datagrams or a stream of characters (e.g., bytes).

[0095] Shared memory refers to the allocation of virtual memory space to demonstrate a mechanism by which data may be communicated and / or accessed by multiple processes. In a shared memory implementation, an initialization process first creates a shareable segment in a persistent or non-persistent storage device. After creation, the initialization process mounts the shareable segment and then maps the shareable segment to an address space associated with the initialization process. Following mounting, the initialization process proceeds to identify and grant access permissions to one or more authorized processes that may write data to or read data from the shareable segment. Changes made to data in the shareable segment by one process may immediately affect other processes that are also linked to the shareable segment. Furthermore, when one of the authorized processes accesses the shareable segment, the shareable segment is mapped to the address space of that authorized process. Often, only one authorized process may mount the shareable segment at any given time, other than the initialization process.

[0096] Other techniques may be used to share data between processes, such as the various data described in this application, without departing from the scope of this disclosure. The processes may be part of the same or different applications and may run on the same or different computing systems.

[0097] Rather than or in addition to sharing data between processes, a computing system implementing one or more embodiments of the present disclosure may include functionality for 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 the graphical user interface by a user selecting one or more graphical user interface widgets or inserting text and other data into the graphical user interface widgets using a touchpad, keyboard, mouse, or other input device. In response to a selection of a particular item, information regarding the particular item may be retrieved by the computer processor from persistent or non-persistent storage. Upon selection of an item by a user, the contents of the retrieved data regarding the particular item may be displayed on the user device in response to the user's selection.

[0098] As another example, a request to obtain data regarding a particular item may be sent to a server operatively connected to the user device over a network. For example, a user may select a Uniform Resource Locator (URL) link in a web client of the user device, thereby initiating a Hypertext Transfer Protocol (HTTP) or other protocol request that is sent to a network host associated with the URL. In response to the request, the server may extract data regarding the particular selected item and transmit the data to the device that initiated the request. Once the user device receives the data regarding the particular item, the contents of the received data regarding the particular item may be displayed on the user device in response to the user's selection. In addition to the above examples, data received from the server after selecting the URL link may provide a Hypertext Markup Language (HTML) web page that may be rendered by the web client and displayed on the user device.

[0099] Once data is retrieved, such as by using the techniques described above, or from a storage device, a computing system, in performing one or more embodiments of the present disclosure, may extract one or more data items from the retrieved data. For example, the extraction may be performed by the computing system of FIG. 10A as follows: First, an organizational pattern (e.g., grammar, schema, layout) of the data is determined, which may be based on one or more of position (e.g., bit or string position, Nth token in a data stream, etc.), attribute (where an attribute is associated with one or more values), or hierarchical / tree structure (including layers of nodes at various levels of detail, such as nested packet headers or nested document sections). Next, a stream of raw, unprocessed data symbols is parsed into a stream (or hierarchical structure) of tokens (each token may have an associated token "type") in the context of the organizational pattern.

[0100] The extraction criteria are then used to extract one or more data items from the token stream or structure, and 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, tokens at positions identified by the extraction criteria are extracted. For attribute / value-based data, tokens and / or nodes associated with attributes that satisfy the extraction criteria are extracted. For hierarchical / layered data, tokens associated with nodes that match the extraction criteria are extracted. The extraction criteria may be as simple as an identifier string, or may be a query provided to a structured data repository (which may be organized according to a database schema or data format such as XML).

[0101] The extracted data may be used for further processing by a computing system. For example, the computing system of FIG. 10A may perform data comparison while executing one or more embodiments of the present disclosure. Data comparison may 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. This comparison may be performed by submitting an operation code that specifies an operation related to the comparison to A, B, and an arithmetic logic unit (ALU) (i.e., a circuit that performs arithmetic operations 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 may indicate whether the numerical result is a positive number, a negative number, zero, etc. The comparison may be performed by selecting an appropriate operation code and then reading the numerical result and / or the status flag. For example, to determine whether A>B, B may be subtracted from A (i.e., A - B), and the status flag may be read to determine whether the result is positive (i.e., if A>B, then A - B>0). In one or more embodiments, B may be regarded as a threshold value, and A may be regarded as satisfying the threshold value if A=B or A>B as determined using the ALU. In one or more embodiments of the present disclosure, A and B may be vectors, and comparing A with B may require 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 character strings, the binary values of the character strings may be compared.

[0102] The computing system of FIG. 10A may implement and / or connect to a data repository. For example, one type of data repository is a database. A database is a collection of information structured 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.

[0103] A user or software application may submit a statement or query to the DBMS. The DBMS then interprets the string. The string may be a select statement, update statement, create statement, delete statement, etc., that requests information. Additionally, the statement may include parameters that specify the data or data container (database, table, record, column, view, etc.), identifiers, conditions (comparison operators), functions (e.g., join, full join, count, average, etc.), sort (e.g., ascending, descending), etc. The DBMS may execute the statement. For example, the DBMS may access memory buffers, reference or index files to read, write, delete, or any combination thereof to respond to the statement. The DBMS may load data from persistent or non-persistent storage and perform calculations to respond to the query. The DBMS may return the results to the user or software application.

[0104] The computing system of FIG. 10A may include functionality for providing raw and / or processed data, such as the results of comparisons and other processing. For example, providing the data may be accomplished through a variety of presentation methods. Specifically, the 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, such as a computer monitor or a touch screen on a handheld computing device. The GUI may include various GUI widgets that organize what data is shown and how the data is presented to the user. Additionally, the GUI may display data, e.g., data provided as actual data values ​​through text, or a visual representation of the data, such as through visualization of a data model, by the computing device. Alternatively, the encoded data may be provided directly to the user.

[0105] For example, the GUI may first obtain a notification from a software application that a particular data object is to be provided within the GUI. The GUI may then determine 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 that data object type, for example rules specified by the software framework for the data object class or according to local parameters defined by the GUI for presenting that data object type. Finally, the GUI may obtain a data value from the particular data object and render a visual representation of the data value within a display device according to the rules specified for that data object type.

[0106] The data may also be provided in various audio manners, in particular the data may be rendered into an audio format and provided as sound via one or more speakers operatively connected to the computing device.

[0107] The data may also be provided to the user by tactile methods, which may include, for example, vibrations or other physical signals generated by the computing system, for example, the data may be provided to the user using vibrations generated by a handheld computing device of a predetermined duration and vibration intensity to communicate the data.

[0108] The above functional descriptions provide only a few examples of functions performed by the computing system of Figure 10A and the nodes and / or client devices of Figure 10B. Other functions may be performed using one or more embodiments of the present disclosure.

[0109] While this disclosure has been described with respect to a limited number of embodiments, those skilled in the art having the benefit of this disclosure will appreciate that other embodiments may be devised that do not depart from the scope of the disclosure disclosed herein. Accordingly, the scope of the disclosure should be limited only by the scope of the appended claims.

[0110] The embodiments and examples described herein are presented to best explain the invention and its particular applications, thereby enabling those skilled in the art to make and use the invention. However, those skilled in the art will recognize that the foregoing description and examples have been presented for the purposes of illustration and example only. The description set forth is not intended to be exhaustive or to limit the invention to the precise form disclosed.

[0111] While the present invention has been described with respect to a limited number of embodiments, those skilled in the art, having the benefit of this disclosure, will appreciate that they may devise other embodiments that do not depart from the scope of the invention disclosed herein. Accordingly, the scope of the present invention should be limited only by the appended claims.

Claims

1. 1. A computer-implemented method for predicting drug-target binding using synthetically augmented data, comprising: generating a plurality of ghost ligands for a plurality of proteins in a protein structural database; generating a plurality of drug-target interaction (DTI) features for proteins and ligands in a DTI database using the plurality of ghost ligands; generating a machine learning model using the plurality of DTI features; and and predicting the interaction probability of a combination of a query protein and a query ligand using the machine learning model.

2. generating said plurality of ghost ligands, In the case of a cluster of proteins selected from said plurality of proteins, performing a structural alignment of the proteins within the cluster of proteins; After the structural alignment, obtaining the ghost ligands by projecting a ligand of one of the proteins in the cluster onto all other proteins in the cluster; and obtaining a confidence score for each of the plurality of ghost ligands.

3. The cluster of proteins is Sequence similarity, Similarity of three-dimensional topology, and The method of claim 2 , wherein the clustering is obtained based on one selected from the group consisting of existing clusterings in a database.

4. The method of claim 2 , wherein the confidence score quantifies the uncertainty of the associated ghost ligand.

5. generating the plurality of DTI features, For each of a plurality of combinations of ligands and proteins in the DTI database, selecting from said plurality of ghost ligands a ghost ligand that is most similar to said ligand considered for said combination; and generating features of the proteins considered for the combination, the generated features characterizing the proteins considered for the combination.

6. The generated features are: at least one local feature comprising binding site features of concentric shells of increasing radius; at least one global feature beyond the binding site features; and The method of claim 5 , comprising at least one selected from the group consisting of functional annotations.

7. The method of claim 5 , wherein the selection of the most similar ghost ligands is performed based on a distance metric.

8. generating the plurality of DTI features, The method of claim 5 , further comprising obtaining a confidence vector representing the confidence of a plurality of components of the DTI feature associated with the most similar ghost ligand.

9. The reliability of the plurality of components is a first confidence score that quantifies the uncertainty associated with the most similar ghost ligand; a second confidence score that quantifies the similarity of the fingerprint between the ligand considered for the combination and the most similar ghost ligand; and The method of claim 8 , further comprising at least one selected from the group consisting of: a third confidence score depending on the source from which the DTI features are obtained.

10. generating the machine learning model, obtaining positive training samples based on the plurality of DTI features of the protein and the ligand; obtaining negative training samples based on the plurality of DTI features by shuffling the plurality of DTI features for the protein and the ligand at least once; and training the machine learning model for DTI prediction using the positive training samples and the negative training samples.

11. generating the machine learning model, before obtaining the positive training samples and the negative training samples, The method of claim 10, comprising filtering the plurality of DTI features of the protein and the ligand using a confidence threshold applied to a confidence vector associated with the plurality of DTI features.

12. The method of claim 1 , wherein the machine learning model is selected from the group consisting of a classifier model and a regression model.

13. predicting the possibility of an interaction for the combination of the query protein and the query ligand, obtaining possible binding sites and associated local features for the query protein based on the plurality of ghost ligands; generating features for the query protein including the local features; generating features for said query ligand, including a ligand fingerprint and a ligand descriptor; and applying the machine learning model to the features for the query protein and the features for the query ligand to obtain a probability of interaction between the query ligand and the query protein.

14. The method of claim 13 , wherein the features for the query protein further comprise at least one selected from the group consisting of global features and functional annotations.

15. 1. A non-transitory computer readable medium comprising computer readable program code for predicting drug-target binding using synthetically augmented data, the computer readable program code being configured to include a computer system comprising: Generate multiple ghost ligands for multiple proteins in a protein structure database; using the plurality of ghost ligands to generate a plurality of drug-target interaction (DTI) features of proteins and ligands in a DTI database; generating a machine learning model using the DTI features; A non-transitory computer readable medium, wherein the machine learning model is used to predict the interaction likelihood of a combination of a query protein and a query ligand.

16. A system for specific drug discovery, comprising: A protein structure database; a ghost ligand identification engine configured to generate a plurality of ghost ligands for a plurality of proteins in the protein structural database; a ghost ligand database storing the plurality of ghost ligands; A drug-target interaction (DTI) database that stores proteins and ligands; a feature generation engine configured to generate a plurality of DTI features for the proteins and the ligands in the DTI database using the ghost ligands in the ghost ligand database; a machine learning model training engine configured to generate a machine learning model using the DTI features; A system for specific drug discovery comprising: a DTI prediction engine configured to predict the interaction probability of a combination of a query protein and a query ligand using the machine learning model.

17. generating the plurality of DTI features, For each of a plurality of combinations of ligands and proteins in the DTI database, selecting from said plurality of ghost ligands a ghost ligand that is most similar to said ligand considered for said combination; and generating features of the proteins considered for the combination, the generated features characterizing the proteins considered for the combination.

18. The generated features are: at least one local feature comprising binding site features of concentric shells of increasing radius; at least one global feature beyond the binding site features; and 20. The method of claim 17, comprising at least one selected from the group consisting of functional annotations.

19. generating the plurality of DTI features, obtaining a confidence vector representing a confidence in a plurality of components of the DTI feature associated with the most similar ghost ligand; The reliability of the plurality of components is a first confidence score that quantifies the uncertainty associated with the most similar ghost ligand; a second confidence score that quantifies the similarity of the fingerprint between the ligand considered for the combination and the most similar ghost ligand; and The system of claim 17 , further comprising at least one selected from the group consisting of: a third confidence score depending on the source from which the DTI features are obtained.

20. predicting the possibility of an interaction for the combination of the query protein and the query ligand, obtaining possible binding sites and associated local features for the query protein based on the plurality of ghost ligands; generating features for the query protein including the local features; generating features for said query ligand, including a ligand fingerprint and a ligand descriptor; and applying the machine learning model to the features for the query protein and the features for the query ligand to obtain a probability of interaction between the query ligand and the query protein.

Images