System and method for determining a biological activity parameter of a ligand

Abstract

A system and method for determining a biological activity parameter of a ligand are disclosed, the system comprising a processor configured to: obtain ligand-protein complex structure data and reference biological activity data; determine ligand-protein complex interaction parameters indicative of intermolecular interaction data of the ligand-protein complex in a docking conformation; determine aggregated ligand-protein complex interaction parameters across a plurality of docking conformations; and determine the biological activity parameter based on the aggregated ligand-protein complex interaction parameters, wherein determining the ligand-protein complex interaction parameters comprises determining non-covalent parameters indicative of intermolecular interaction data of a ligand node and a protein node based on ligand and protein attributes and ligand-protein non-covalent parameters indicative of binding stability of a first non-covalent edge of a dispersion force type.

Description

[0001] Cross-references to related applications

[0002] This patent application claims priority and benefit to Singapore Patent Application No. 10202400124Q, filed on January 16, 2024, with the Intellectual Property Office of Singapore, the disclosure of which is incorporated herein by reference in its entirety. Technical Field

[0003] Various aspects of this disclosure relate to a system and method for determining the bioactivity parameters of a ligand. Background Technology

[0004] The following discussion of the background art is intended only to facilitate understanding of this disclosure. It should be understood that this discussion is not an admission or endorsement that any material cited was published, known, or constitutes part of common general knowledge to those skilled in the art in any jurisdiction prior to the priority date of this disclosure.

[0005] The drug discovery process includes stages such as target identification, lead compound identification, lead compound optimization, preclinical studies, and clinical trials. A key stage is lead compound identification, which screens for potential therapeutic candidates with high biological activity. Bioactivity prediction using deep learning has emerged as a tool for early molecular screening, aiding in the efficient identification of active molecules. This prediction also guides the refinement of lead compounds, enhancing their biological activity and other properties. However, the accuracy of current deep learning-based bioactivity prediction methods is hampered by a lack of high-quality training data.

[0006] To improve the accuracy of bioactivity prediction, emerging deep learning techniques address the challenge of data scarcity. Early methods addressed imbalanced samples by using high-quality negative data to distinguish between active and inactive molecules. Current methods include utilizing untested molecule-target pairs as negative examples, leveraging protein family information as priors for transferable inactive ligands, and employing transfer learning and pre-trained neural network layers. Other techniques involve generative adversarial representations and incorporating cell morphology and chemical structure models. Despite these efforts, many deep learning methods neglect fundamental elements that generate bioactivity, focusing primarily on quantitative structure-activity relationship (QSAR) modeling between ligand structure and bioactivity, involving generative adversarial representations and incorporating cell morphology and chemical structure models.

[0007] For example, current ligand bioactivity prediction models are based on two popular approaches. The first approach includes ligand-based virtual screening (LBVS) models, which predict bioactivity solely based on ligand structure and its corresponding bioactivity. However, ligand bioactivity is influenced by many factors beyond ligand structure. The other approach includes structure-based virtual screening (SBVS) models, which dock candidates to proteins and estimate their binding affinity. However, it is important to note that reaching a certain binding affinity threshold is a prerequisite for bioactivity, but higher binding affinity beyond that threshold does not necessarily guarantee higher bioactivity.

[0008] Bioactivity is determined not only by ligand structure but also by drug-target interactions (e.g., binding conformation and binding affinity), reaction environment, pharmacokinetic properties, and species differences. However, high-quality data on drug-target interactions are much scarcer than molecular structure. This is one reason why current deep learning methods primarily model the molecular structure of drugs.

[0009] Therefore, there is a need for an improved system and / or method for determining the biological activity of a drug (e.g., a ligand), which is designed to address at least one of the aforementioned problems. Summary of the Invention

[0010] The present disclosure aims to provide an improved system and method for determining the bioactivity of ligands (e.g., drugs) by circumventing the aforementioned difficulties through molecular docking to generate ligand-protein complexes and using model self-attention to accurately identify docking pockets and conformations. To this end, a Drug-Target (i.e., Ligand-Protein) Interaction Graph Neural Network (DTIGN) is constructed to determine the non-covalent parameters of the ligand-protein complex. This network considers interatomic distances in Coulomb and London forces to capture force field information reflecting binding stability, a prerequisite for bioactivity. Furthermore, a machine learning algorithm, specifically a multi-head self-attention mechanism, is employed to further identify binding pockets and conformations within the molecular docking complex. To validate the correctness of the model's attention, unsupervised training with crystal structure data demonstrates accurate identification of binding pockets and conformations. Finally, despite the scarcity of crystal structure data, this disclosure utilizes this limited high-quality data for semi-supervised learning within DTIGN to refine the model's attention and bioactivity predictions.

[0011] Furthermore, this disclosure provides a novel benchmark dataset for evaluating bioactivity prediction models in the context of ligand-protein complexes. This novel benchmark dataset is unique in that it contains a large number of ligand-protein complexes generated through molecular docking, distinguishing it from other bioactivity datasets that primarily focus on ligand structure and binding affinity datasets that independently utilize ligand structure, target sequence, target structure, or binding site.

[0012] According to a first aspect of this disclosure, a system is provided for determining bioactivity parameters of a ligand, wherein the ligand binds to a protein to form a ligand-protein complex, the ligand binding to the protein in multiple docking conformations, the system comprising a processor configured to: acquire ligand-protein complex structure data indicating the structure of the ligand-protein complex and reference bioactivity data of the ligand, the protein, and / or the ligand-protein complex; the ligand-protein complex structure data comprising: multiple ligand nodes of the ligand, each ligand node having ligand properties; multiple protein nodes of the protein, each protein node having protein properties; multiple non-covalent edges connecting the ligand nodes to corresponding protein nodes, each non-covalent edge having non-covalent bond properties; and for each of the multiple docking conformations, determining intermolecular interaction data indicating the ligand-protein complex in each docking conformation. The ligand-protein complex interaction parameters are determined by: for each non-covalent edge, determining non-covalent parameters indicating intermolecular interaction data of the first ligand node and the first protein node located nearby and bound to the first ligand node via the first non-covalent edge, based on the first ligand property of the first ligand node, the first protein property of the first protein node, and the binding stability of the first non-covalent edge based on the first ligand-protein non-covalent parameter indicating the type of dispersion force of the first non-covalent edge; determining the ligand-protein complex interaction parameters based on the aggregation of the non-covalent parameters of the plurality of non-covalent edges of the ligand-protein complex; determining aggregated ligand-protein complex interaction parameters indicating the aggregation of the ligand-protein complex in the plurality of docking conformations based on the intermolecular interaction data; and determining the bioactivity parameters based on the aggregated ligand-protein complex interaction parameters.

[0013] In various implementations, the first ligand-protein nonvalent parameter may be based on a predetermined ligand-protein vector, which represents the force field amplitude corresponding to the dispersion force type of the first nonvalent edge, as a function of the nonvalent bond property of the first nonvalent edge.

[0014] In various embodiments, the predetermined ligand-protein vector may include a Coulomb ligand-protein vector comprising a predetermined radial basis function (RBF) parameterized by a Coulomb power value, indicating the magnitude of the force field of the Coulomb dispersion force of the first non-covalent edge. In various other embodiments, the predetermined ligand-protein vector may include a London ligand-protein vector comprising a predetermined RBF parameterized by a predetermined London power value, indicating the magnitude of the force field of the London dispersion force of the first non-covalent edge.

[0015] In various embodiments, the non-covalent property of the first non-covalent edge may include a ligand-protein distance property, which represents a physical distance value between a first ligand node and a first protein node in the ligand-protein complex. The physical distance value may indicate a force field interaction between the first ligand node and the first protein node. In some embodiments, the processor may be further configured to determine the ligand-protein distance property by: measuring the physical distance value between the first ligand node and the first protein node on the ligand-protein complex structural data; comparing the physical distance value with a predetermined distance threshold indicating the presence of the first non-covalent edge; and determining the presence of the first non-covalent edge based on the comparison of the physical distance value with the predetermined distance threshold.

[0016] In various embodiments, the ligand-protein complex structural data may further include a plurality of ligand edges in the ligand that bind the plurality of ligand nodes to each other, each ligand edge having ligand covalent bond chemical properties and ligand covalent bond physical properties, wherein, to determine the ligand-protein complex interaction parameters, the processor is further configured to: for each of the plurality of ligand edges, determine ligand edge parameters indicating intramolecular interaction data of a first ligand node and a second ligand node located adjacent to it and bound to the first ligand node via the first ligand edge, the ligand edge parameters being determined based on: a first ligand property of the first ligand node, a second ligand property of the second ligand node, a ligand covalent bond chemical property of the first ligand edge, and a first ligand covalent parameter indicating the binding stability of the first ligand edge, and further determine the ligand-protein complex interaction parameters based on the aggregation of the ligand edge parameters of the plurality of ligand edges.

[0017] In various implementations, the first ligand covalent parameter may be based on a predetermined ligand vector, which represents the distance property of the intramolecular covalent bond of the first ligand edge, as a function of the physical property of the ligand covalent bond of the first ligand edge.

[0018] In various embodiments, the predetermined ligand vector may include a predetermined RBF parameterized by a first power value, indicating the distance property of the intramolecular covalent bond of the first ligand edge.

[0019] In various implementations, the ligand covalent bond physical properties of the first ligand edge may include a ligand distance property, which represents the physical distance value between the first ligand node and the second ligand node of the ligand.

[0020] In various embodiments, the ligand-protein complex structural data may further include multiple protein edges in the protein that bind the plurality of protein nodes to each other, each protein edge having protein covalent chemical properties and protein covalent physical properties, wherein, to determine the ligand-protein complex interaction parameters, the processor is further configured to: for each of the plurality of protein edges, determine protein edge parameters indicating intramolecular interaction data of a first protein node and a second protein node located adjacent to it and bound to the first protein node through the first protein edge of the plurality of protein edges, the protein edge parameters being determined based on: a first protein property of the first protein node, a second protein property of the second protein node, a protein covalent chemical property of the first protein edge, and a first protein covalent parameter indicating the binding stability of the first protein edge, and further determine the ligand-protein complex interaction parameters based on the aggregation of the protein edge parameters of the plurality of protein edges.

[0021] In various implementations, the first protein covalent parameter may be based on a predetermined protein vector, which represents the distance attribute of the intramolecular covalent bonds of the first protein edge, as a function of the physical property of the protein covalent bonds of the first protein edge.

[0022] In various embodiments, the predetermined protein vector may include a predetermined RBF parameterized by a second power value, indicating the distance property of intramolecular covalent bonds of the first protein edge.

[0023] In various implementations, the protein covalent bond physical properties of the first protein edge may include a protein distance property, which represents the physical distance value between the first protein node and the second protein node of the protein.

[0024] In various embodiments, the processor may further include a first machine learning algorithm, which optionally includes a deep learning algorithm, such as, but not limited to, a neural network-based deep learning algorithm.

[0025] In various embodiments, the processor may further include a second machine learning algorithm for determining the ligand-protein complex interaction parameters, wherein the second machine learning algorithm optionally includes a multilayer perceptron (MLP).

[0026] In various embodiments, the processor may further include a third machine learning algorithm for determining ligand-protein complex interaction parameters of the aggregated protein, wherein the third machine learning algorithm optionally includes a multi-head self-attention algorithm.

[0027] In various embodiments, the bioactivity parameter may include at least one of the following: an EC50 parameter indicating the half-maximal effective concentration of the ligand; and / or an IC50 parameter indicating the half-maximal inhibitory concentration of the ligand.

[0028] According to a second aspect of this disclosure, a drug discovery platform is provided, which includes the system of the first aspect of this disclosure.

[0029] According to a third aspect of this disclosure, a method is provided for determining bioactivity parameters of a ligand, wherein the ligand binds to a protein to form a ligand-protein complex, the ligand binding to the protein in multiple docking conformations, the method comprising providing a processor for: acquiring ligand-protein complex structure data indicating the structure of the ligand-protein complex and reference bioactivity data of the ligand, the protein, and the ligand-protein complex; the ligand-protein complex structure data comprising: multiple ligand nodes of the ligand, each ligand node having ligand properties; multiple protein nodes, each protein node having protein properties; multiple non-covalent edges connecting the ligand nodes to the corresponding protein nodes, each non-covalent edge having non-covalent bond properties; and for each of the multiple docking conformations, determining a ligand-protein complex indicating intermolecular interaction data of the ligand-protein complex in each docking conformation. The complex interaction parameters, wherein determining the ligand-protein complex interaction parameters includes: for each non-covalent edge, determining, based on a first ligand property of a first ligand node, a first protein property of a first protein node, and a first ligand-protein non-covalent parameter indicating the binding stability of the dispersion force type of the first non-covalent edge, a non-covalent parameter indicating intermolecular interaction data of the first ligand node and a first protein node located nearby and bound to the first ligand node via the first non-covalent edge; and determining the ligand-protein complex interaction parameters based on the aggregation of the non-covalent parameters of the plurality of non-covalent edges of the ligand-protein complex; determining aggregated ligand-protein complex interaction parameters indicating the aggregation of the ligand-protein complex in the plurality of docking conformations; and determining the bioactivity parameters based on the aggregated ligand-protein complex interaction parameters.

[0030] In various implementations, the first ligand-protein nonvalent parameter may be based on a predetermined ligand-protein vector, which represents the force field amplitude corresponding to the dispersion force type of the first nonvalent edge, as a function of the nonvalent bond property of the first nonvalent edge.

[0031] In various embodiments, the predetermined ligand-protein vector may include a Coulomb ligand-protein vector comprising a predetermined RBF parameterized by a Coulomb power value, indicating the magnitude of the force field of the Coulomb dispersion force of the first non-covalent edge. In various other embodiments, the predetermined ligand-protein vector may include a London ligand-protein vector comprising a predetermined RBF parameterized by a predetermined London power value, indicating the magnitude of the force field of the London dispersion force of the first non-covalent edge.

[0032] In various embodiments, the non-covalent property of the first non-covalent edge may include a ligand-protein distance property, which represents a physical distance value between a first ligand node and a first protein node in the ligand-protein complex. The physical distance value may indicate a force field interaction between the first ligand node and the first protein node. In some embodiments, determining the ligand-protein distance property may include: measuring the physical distance value between the first ligand node and the first protein node on the ligand-protein complex structural data; comparing the physical distance value with a predetermined distance threshold indicating the presence of the first non-covalent edge; and determining the presence of the first non-covalent edge based on the comparison of the physical distance value with the predetermined distance threshold.

[0033] In various embodiments, the ligand-protein complex structural data may further include a plurality of ligand edges in the ligand that bind the plurality of ligand nodes to each other, each ligand edge having ligand covalent bond chemical properties and ligand covalent bond physical properties, wherein the method further includes: for each of the plurality of ligand edges, determining ligand edge parameters that indicate intramolecular interaction data of a first ligand node and a second ligand node located adjacent to it and bound to the first ligand node through the first ligand edge, the ligand edge parameters being determined based on: a first ligand property of the first ligand node, a second ligand property of the second ligand node, a ligand covalent bond chemical property of the first ligand edge, and a first ligand covalent parameter indicating the binding stability of the first ligand edge; and further determining the ligand-protein complex interaction parameters based on the aggregation of the ligand edge parameters of the plurality of ligand edges.

[0034] In various implementations, the first ligand covalent parameter may be based on a predetermined ligand vector, which represents the distance property of the intramolecular covalent bond of the first ligand edge, as a function of the physical property of the ligand covalent bond of the first ligand edge.

[0035] In various embodiments, the predetermined ligand vector may include a predetermined RBF parameterized by a first power value, indicating the distance property of the intramolecular covalent bond of the first ligand edge.

[0036] In various implementations, the ligand covalent bond physical properties of the first ligand edge may include a ligand distance property, which represents the physical distance value between the first ligand node and the second ligand node of the ligand.

[0037] In various embodiments, the ligand-protein complex structural data may further include a plurality of protein edges in the protein that bind the plurality of protein nodes to each other, each protein edge having protein covalent chemical properties and protein covalent physical properties, wherein the method further includes: for each of the plurality of protein edges, determining protein edge parameters that indicate intramolecular interaction data of a first protein node and a second protein node located adjacent to it and bound to the first protein node through the first protein edge of the plurality of protein edges, the protein edge parameters being determined based on: a first protein property of the first protein node, a second protein property of the second protein node, a protein covalent chemical property of the first protein edge, and a first protein covalent parameter indicating the binding stability of the first protein edge, and further determining the ligand-protein complex interaction parameters based on the aggregation of the protein edge parameters of the plurality of protein edges.

[0038] In various implementations, the first protein covalent parameter may be based on a predetermined protein vector, which represents the distance attribute of the intramolecular covalent bonds of the first protein edge, as a function of the physical property of the protein covalent bonds of the first protein edge.

[0039] In various embodiments, the predetermined protein vector may include a predetermined RBF parameterized by a second power value, indicating the distance property of intramolecular covalent bonds of the first protein edge.

[0040] In various implementations, the protein covalent bond physical properties of the first protein edge may include a protein distance property, which represents the physical distance value between the first protein node and the second protein node of the protein.

[0041] In various implementations, the method may further include a first machine learning algorithm, which optionally includes a deep learning algorithm, such as, but not limited to, a neural network-based deep learning algorithm.

[0042] In various embodiments, the method may further include a second machine learning algorithm for determining the ligand-protein complex interaction parameters, wherein the second machine learning algorithm optionally includes an MLP.

[0043] In various embodiments, the method may further include a third machine learning algorithm for determining ligand-protein complex interaction parameters of the aggregated protein, wherein the third machine learning algorithm optionally includes a multi-head self-attention algorithm.

[0044] In various embodiments, the bioactivity parameter includes at least one of the following: an EC50 parameter indicating the half-maximal effective concentration of the ligand; and / or an IC50 parameter indicating the half-maximal inhibitory concentration of the ligand.

[0045] According to a fourth aspect of this disclosure, a computer-readable medium is provided, comprising instructions that, when executed by a processor, cause the processor to perform the method of the third aspect of this disclosure. Brief description of the attached diagram

[0047] This disclosure will be better understood when considered in conjunction with the detailed description and accompanying drawings, in which:

[0048] Figure 1 An exemplary schematic diagram of system 100 is shown, which is used to determine the bioactivity parameter 180 of ligand 210 when ligand 210 binds to protein 220 to form ligand-protein complex 250, and ligand 210 binds to protein 220 in multiple docking conformations.

[0049] Figure 2 An exemplary schematic diagram of ligand-protein complex structural data 200 is shown, indicating the structure of ligand-protein complex 250 as a 3D diagram;

[0050] Figure 3 Exemplary ligand properties 212a, 214a, 216a, protein properties 222a, 224a, 226a, 228a, 232a, 234a, ligand covalent bond chemical properties 215a, 217a, protein covalent bond chemical properties 221a, 223a, 225a, 227a, 229a, and radial basis functions (RBFs) of the covalent intramolecular interaction forces between ligand nodes 212, 214, and 216 are shown. The radial basis functions (RBFs) of the covalent intramolecular interactions between protein nodes 222, 224, 226, 228, 232, and 234. The two radial basis functions of the non-covalent intermolecular interaction forces between ligand nodes 212, 214, 216 of the ligand-protein complex 250 and the corresponding protein nodes 222, 224, 226, 228, 232, 234. and ;

[0051] Figure 4 An exemplary architecture 400 of a ligand-protein (i.e. drug-target) complex 250 graph neural network (DTIGN) is shown for determining ligand-protein complex parameters 130.

[0052] Figure 5An exemplary flowchart of method 500 is shown, which is used to determine the bioactivity parameters of a ligand when the ligand binds to a protein to form a ligand-protein complex, the ligand binds to the protein in multiple docking conformations.

[0053] Figure 6 An exemplary schematic diagram of the overall DTIGN 600 based on drug-target interaction bioactivity prediction is shown;

[0054] Figure 7 The docking pocket selection 700 of the exemplary ligand CHEMBL202 is shown;

[0055] Figure 8 Table 800 shows a comparison of the predictive performance of bioactivity (pIC50 for CHEMBL202) on different data sources, AutoDock Vina and QVina-W, where the ↑ or ↓ arrows indicate that larger or smaller values ​​are better.

[0056] Figure 9 A table is shown for a benchmark dataset 900 used to predict the bioactivity of ligand-protein complexes;

[0057] Figure 10 A table illustrating the ablation studies 1000 on the I1 task (dihydrofolate reductase) on the benchmark dataset is shown;

[0058] Figure 11 The diagram shows the attention value of 1100 for the test ligand on the protein pocket during training. The natural binding sites depicted above are from the RCSB PDB database. Details and corresponding pockets are given in the legend.

[0059] Figure 12 A table illustrating the cross-union ratio of 1200 between the "natural pocket" and the "attention pocket" is shown;

[0060] Figure 13 Box plot 1300 shows the attention distribution of DTIGN on sorted docking conformations involved in the pIC50 test set (containing 9,975 re-docking conformations on 1,011 native PDB structures). “Avg RMSD” represents the average RMSD between ligand-protein docking conformations and their respective native structures in the PDBBind dataset;

[0061] Figure 14Box plot 1400 shows the attention distribution of DTIGN on the sorted docking conformations involved in the pKd test set (containing 14,411 re-docking conformations on 1,455 native PDB structures). “Avg RMSD” represents the average RMSD between ligand-protein docking conformations and their respective native structures in the PDBBind dataset;

[0062] Figure 15 Box plot 1500 shows the attention distribution of DTIGN on the sorted docking conformations involved in the pKi test set (containing 6,196 re-docking conformations on 626 native PDB structures). “Avg RMSD” represents the average RMSD between ligand-protein docking conformations and their respective native structures in the PDBBind dataset;

[0063] Figure 16 A table showing the best cross-validation Pearson r 1600 on the benchmark dataset is provided;

[0064] Figure 17 A table showing the best cross-validation RMSE of 1700 on the benchmark dataset is provided;

[0065] Figure 18 This demonstrates the best cross-validation on the benchmark dataset. Table 1800;

[0066] Figure 19 The binding conformation 1900 of the exemplary target protein CHEMBL202 with the test ligand CHEMBL7492 is shown in the following cases: (A) the native conformation and (B) the AutoDock Vina docking structure in the pocket of interest (pocket 2) of this model. Non-covalent bonds (polar interactions) between the ligand and protein are marked with dashed lines, and the binding sites are highlighted above the protein sequence. Common binding sites between the native and docking structures are circled in dashes; and

[0067] Figure 20 The results show “CHEMBL7492” in (A) “1u72” and “CHEMBL22” in (B) “2w3a”, indicating the test ligand ID and its PDB ID relative to the protein’s native structure. This demonstrates that DTIGN identifies and learns some native docking conformations, showing an RMSD of less than 5 Å from the native structure. The correlation between RMSD and mean attention value during model training was assessed using the Pearson correlation coefficient (r).

[0068] Detailed description

[0069] The following detailed description refers to the accompanying drawings, which illustrate specific details and embodiments that may enable the practice of this disclosure. These embodiments are described in sufficient detail to enable those skilled in the art to practice this disclosure. Other embodiments may be utilized, and structural and logical changes may be made without departing from the scope of this disclosure. The embodiments are not necessarily mutually exclusive, as some embodiments may be combined with one or more other embodiments to form new embodiments.

[0070] Features described in the context of an implementation scheme may be applied accordingly to the same or similar features in other implementation schemes. Features described in the context of an implementation scheme may be applied accordingly to other implementation schemes, even if not explicitly described in those other implementation schemes. Furthermore, additions and / or combinations and / or alternatives described for a feature in the context of an implementation scheme may be applied accordingly to the same or similar features in other implementation schemes.

[0071] In the context of various implementations, the articles “a,” “an,” and “the” used with respect to a feature or element include references to one or more of the stated features or elements. As used herein, the term “and / or” includes any and all combinations of one or more of the listed related items.

[0072] While terms such as "first," "second," etc., can be used to describe various elements, these terms are used only for clarity to distinguish one element from another and do not define the corresponding elements, such as their order and / or importance. Without departing from the scope of the claims in this specification, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element.

[0073] Throughout this specification, the term "ligand" as used herein may refer to a compound known to be effective against a target (e.g., a protein). The compound may be a small molecule with a molecular weight less than or equal to 900 Da. In various embodiments, a ligand may refer to a drug, such as a pharmaceutical drug. In some embodiments, a ligand may be a ligand antagonist and may bind to a target protein to block the function of the target protein. In some other embodiments, a ligand may be a ligand agonist and may bind to a target protein to activate the function of the target protein. In the context of this disclosure, the term "ligand molecule" may refer to the ligand, which may include a plurality of ligand atoms bonded together by ligand covalent bonds.

[0074] Throughout this specification, the term "protein" as used herein may refer to a target having a specific function or behavior. A ligand may bind to a protein to form a ligand-protein complex to alter (e.g., inhibit or activate) the protein's function or behavior. In various embodiments, a protein may include at least one docking pocket containing at least one binding site for binding a ligand. Binding sites may be categorized into various docking pockets of a protein based on spatial clustering of binding sites. Each binding site may include at least one receptor for binding the ligand. In the context of this disclosure, the term "protein molecule" may refer to the protein comprising a plurality of protein atoms bound together by protein covalent bonds. While the term "protein" is used in this disclosure, targets are also contemplated to include genes, ribonucleic acid (RNA), such as mRNA or ncRNA.

[0075] Throughout this specification, the term "ligand-protein complex" can refer to a complex formed when a ligand binds to at least one binding site (specifically, a receptor) of a protein. In various embodiments, the ligand may bind to at least one binding site of a protein in multiple docking conformations. In the context of this disclosure, the term "ligand-protein complex molecule" can refer to said ligand molecule having ligand atoms that can bind to at least one protein atom of a protein molecule via intermolecular non-covalent bonds.

[0076] Throughout this specification, the term "bioactivity" as used herein can refer to the ability of a ligand molecule to provoke, induce, or modulate a particular biological effect or cellular response in an organism (e.g., a living organism or tissue). The bioactivity of a ligand molecule can be described relative to a target protein molecule and can be induced when the ligand molecule binds to a receptor contained in at least one binding site in the protein molecule. In some embodiments, the bioactivity of the ligand molecule can be inhibitory, for example, through a ligand antagonist. In some other embodiments, the bioactivity of the ligand molecule can activate a cellular response, for example, through a ligand agonist.

[0077] Throughout this specification, the term "proximity," as used herein, can refer to the location of a ligand or protein atom directly adjacent to another ligand or protein atom. The term "proximity" also refers to the location of a ligand atom directly adjacent to a protein atom. That is, there are no intermediate atoms between a ligand atom and a neighboring ligand atom, between a protein atom and a neighboring protein atom, or between a ligand atom and a neighboring protein atom. Therefore, the term "intramolecular interaction data" can refer to the interaction data between a ligand or protein atom and another ligand or protein atom located directly adjacent to it. Furthermore, the term "intermolecular interaction data" can refer to the interaction data between a ligand atom and a protein atom located directly adjacent to it.

[0078] Throughout the specification, the phrase "a function of..." may include a mathematical function describing the relationship between the input and the output. In various embodiments, the input may include: a ligand distance attribute representing the physical distance value between two ligand nodes covalently bound to each other; and a protein distance attribute representing the physical distance value between two protein nodes covalently bound to each other. Therefore, the output may include: a ligand covalent parameter indicating intramolecular interaction data between two covalently bound ligand nodes, and a protein covalent parameter indicating intramolecular interaction data between two covalently bound protein nodes. In various other embodiments, the input may include a ligand-protein distance attribute representing the non-covalent physical distance value binding a ligand node to a corresponding protein node. The non-covalent physical distance value may indicate a force field interaction between the ligand node and the corresponding protein node. The output may include a ligand-protein non-covalent parameter indicating intermolecular interaction data (e.g., force field data) between the ligand node and the corresponding protein node.

[0079] Throughout this specification, the term "data" may be understood to include information in any suitable analog or digital form, such as as a file, part of a file, a set of files, a signal or stream, a portion of a signal or stream, a set of signals or streams, waveforms, etc. However, the term data is not limited to the examples above and may take various forms and represent any information as understood in the art.

[0080] According to various embodiments, a circuit may include analog circuits or components, digital circuits or components, or hybrid circuits or components. Any other type of implementation of the corresponding function, which will be described in more detail below, may also be understood as a “circuit” according to alternative embodiments. A digital circuit may be understood as any type of logical implementation entity, which may be a dedicated circuit or a processor executing software, firmware, or any combination thereof stored in memory. Thus, in various embodiments, a processor may include one or more circuits. A “circuit” may be a digital circuit, such as hardwired logic circuitry or programmable logic circuitry, such as a programmable processor, such as a microprocessor (e.g., a Complex Instruction Set Computer (CISC) processor or a Reduced Instruction Set Computer (RISC) processor). A “circuit” may also include a processor executing software, such as any type of computer program, such as a computer program using virtual machine code, such as Java.

[0081] Figure 1 An exemplary schematic diagram of system 100 is shown, which is used to determine the bioactivity parameter 180 of ligand 210 when ligand 210 binds to protein 220 to form ligand-protein complex 250, and ligand 210 binds to protein 220 in multiple docking conformations.

[0082] See Figure 1The system 100 includes a processor 110 configured to: acquire ligand-protein complex structure data 200 as a 3D diagram of the structure of the ligand-protein complex 250, and reference bioactivity data for ligand 102, protein 104, and ligand-protein complex 106; for each of the plurality of docking conformations, determine ligand-protein complex interaction parameters 130 indicating intermolecular interaction data of the ligand-protein complex 250 in each docking conformation; determine aggregated ligand-protein complex interaction parameters 170 indicating intermolecular interaction data of the aggregation of the ligand-protein complex 250 in the plurality of docking conformations; and determine bioactivity parameters 180 based on the aggregated ligand-protein complex interaction parameters 170.

[0083] In various embodiments, the bioactivity parameter 180 may indicate a predicted or estimated induced biological effect on an organism. In some embodiments, ligand 210 may be a ligand agonist, and the bioactivity parameter may refer to an EC50 parameter indicating the half-maximal effective concentration of ligand 210 after a specific exposure time. The EC50 parameter may include a value representing 50% of the concentration at which the ligand exerts its maximum biological effect after a specific exposure time. In some other embodiments, ligand 210 may be a ligand antagonist, and the bioactivity parameter may refer to an IC50 parameter indicating the half-maximal inhibitory concentration of ligand 210. The IC50 parameter may include a measure of the concentration of ligand 210 required to inhibit a specific biological effect by 50% after a specific exposure time. Other non-limiting examples of the bioactivity parameter 180 may include an inhibition constant indicating a measure of the ligand's ability to inhibit a biological effect. The dissociation constant indicating the ratio of association and dissociation between the ligand and the protein receptor. The association (binding or affinity) constant is a measure of how ligands and proteins bind to each other. The elimination constant is a measure of the rate at which ligands are removed from an organism. Inhibition index reflecting the measure of inhibitory activity; hit fraction representing the success rate of ligands in screening assays; CC50 value indicating the concentration of a compound that causes 50% cytotoxicity; selectivity index (SI) measuring the specificity of a ligand to a target relative to off-target interactions; activity value quantifying the overall potency of a ligand; area under the curve (AUC) representing the integral of the response over a concentration or time range; and AUC ratio comparing AUC values ​​between different conditions or treatments.

[0084] Ligand-protein complex structural data 200

[0085] Figure 2An exemplary schematic diagram of ligand-protein complex structural data 200 is shown, indicating the structure of ligand-protein complex 250, depicted as a 3D diagram. Figure 3 Exemplary ligand properties 212a, 214a, 216a, protein properties 222a, 224a, 226a, 228a, 232a, 234a, ligand covalent bond chemical properties 215a, 217a, protein covalent bond chemical properties 221a, 223a, 225a, 227a, 229a, and radial basis functions (RBFs) of the covalent intramolecular interaction forces between ligand nodes 212, 214, and 216 are shown. The radial basis functions (RBFs) of the covalent intramolecular interactions between protein nodes 222, 224, 226, 228, 232, and 234. The two radial basis functions (RBFs) of the non-covalent intermolecular interaction forces between ligand nodes 212, 214, 216 of the ligand-protein complex 250 and the corresponding protein nodes 222, 224, 226, 228, 232, 234. and ;

[0086] In various embodiments, ligand-protein complex structural data 200, as well as reference bioactivity data for ligand 102, protein 104, and ligand-protein complex 106, can be obtained from memory 120, which communicates data with processor 110. In some embodiments, processor 110 may be configured to: obtain data indicating the structure of ligand-protein complex 250 from memory 120, and represent the data indicating the structure of ligand-protein complex 250 as a 3D graph, as detailed below.

[0087] Processor 110 may include a wireless communication device to acquire ligand-protein complex structure data 200 and / or data indicating the structure of ligand-protein complex 250 from memory 120 based on one or more predefined wireless protocols. Examples of predefined wireless communication protocols include: Global System for Mobile Communications (GSM), Enhanced Data GSM Environment (EDGE), Wideband Code Division Multiple Access (WCDMA), Code Division Multiple Access (CDMA), Time Division Multiple Access (TDMA), Wi-Fi, Voice over Internet Protocol (VoIP), Global Interoperability for Microwave Access (Wi-MAX), Wi-Fi Direct (WFD), Ultra Wideband (UWB), Infrared Data Association (IrDA), Bluetooth, ZigBee, SigFox, LPWan, LoRaWan, GPRS, 3G, 4G, LTE, 5G, and 6G communication systems. Alternatively, processor 110 may acquire ligand-protein complex structure data 200 and / or data indicating the structure of ligand-protein complex 250 via a wired means (e.g., LAN or Ethernet cable).

[0088] See Figure 2 and Figure 3 The ligand-protein complex 250 structure, as a 3D diagram, may include a ligand molecule represented as ligand 210, comprising a first ligand node 212, a second ligand node 214, and a third ligand node 216, each corresponding to a ligand atom of the ligand molecule. The first, second, and third ligand nodes 212, 214, and 216 may be positioned adjacent to each other and covalently bound to each other. Each of the first, second, and third ligand nodes 212, 214, and 216 may also include corresponding first, second, and third ligand properties 212a, 214a, and 216a, such as ligand chemical characteristics, which may be obtained from reference bioactivity data of ligand 102. In various implementations, the first, second, and third ligand properties 212a, 214a, and 216a may include data indicating the following: the type of the corresponding ligand atom; the number of covalent bonds in the ligand molecule; the implicit valence of the ligand atom; the hybridization of the ligand atom; whether the ligand atom is part of an aromatic system; and the number of hydrogen atoms bonded to the ligand atom. The first, second, and third ligand nodes 212, 214, and 216 may also each have unique 3D coordinates indicating their positions on a 3D graph.

[0089] Intramolecular covalent bonds that bind ligand atoms together can be represented as multiple ligand edges 215, 217 that bind first, second, and third ligand nodes 212, 214, 216 together. For example, a first ligand edge 215 can bind a first ligand node 212 to a second ligand node 214, and a second ligand edge 217 can bind a first ligand node 212 to a third ligand node 216. The first and second ligand edges 215, 217 may each include first and second ligand covalent bond chemical properties 215a, 217a, which can be obtained from reference bioactivity data of ligand 102. In various embodiments, the first and second ligand covalent bond chemical properties 215a, 217a may include data indicating: the type of covalent bond; whether the bond is conjugated; whether the bond contains a ring; and the spatial nature of the covalent bond. The first and second ligand edges 215, 217 may also each include first and second ligand covalent bond physical properties 215b, 217b.

[0090] In various embodiments, the first and second ligand covalent bond physical properties 215b and 217b may include ligand distance properties representing the physical distance value between one ligand atom and another ligand atom. For example, the first ligand covalent bond physical property 215b may include the physical distance value between the first ligand node 212 and the second ligand node 214, and the second ligand covalent bond physical property 217b may include the physical distance value between the first ligand node 212 and the third ligand node 216. In some embodiments, the processor 110 may be configured to determine the first and second ligand covalent bond physical properties 215b and 217b by measuring the physical distances between the first ligand node 212 and the second ligand node 214 and between the first ligand node 212 and the third ligand node 216, respectively, on the ligand-protein complex structure data 200.

[0091] See Figure 2 and Figure 3The ligand-protein complex 250 structure, as a 3D diagram, may further include a protein molecule represented as protein 220, comprising a first protein node 222, a second protein node 224, a third protein node 226, a fourth protein node 228, a fifth protein node 232, and a sixth protein node 234, where each of the first to sixth protein nodes 222, 224, 226, 228, 232, and 234 corresponds to a protein atom in the protein molecule. Each of the first to sixth protein nodes 222, 224, 226, 228, 232, and 234 may be located as adjacent to each other and covalently bound to each other. Each of the first to sixth protein nodes 222, 224, 226, 228, 232, and 234 may also include corresponding first to sixth protein properties 222a, 224a, 226a, 228a, 232a, and 234a, such as protein chemical characteristics, which can be obtained from reference bioactivity data of protein 104. In various implementations, the first to sixth protein attributes 222a, 224a, 226a, 228a, 232a, and 234a may include data indicating the following: the corresponding protein atom type; the number of covalent bonds in the protein molecule; the implicit valence of the protein atom; the hybridization of the protein atom; whether the protein atom is part of an aromatic system; and the number of bonded hydrogen atoms in the protein atom. The first to sixth protein nodes 222, 224, 226, 228, 232, and 234 may also each have unique 3D coordinates indicating their respective positions on a 3D graph.

[0092] In various embodiments, the intramolecular covalent bonds that bind protein atoms together can be represented as multiple protein edges 221, 223, 225, 227, and 229 that bind the first to sixth protein nodes 222, 224, 226, 228, 232, and 234 together. For example, the first protein edge 221 can bind the first protein node 222 to the second protein node 224; the second protein edge 223 can bind the second protein node 224 to the third protein node 226; the third protein edge 225 can bind the third protein node 226 to the fourth protein node 228; the fourth protein edge 227 can bind the fourth protein node 228 to the fifth protein node 232; and the fifth protein edge 229 can bind the fifth protein node 232 to the sixth protein node 234. The first to fifth protein edges 221, 223, 225, 227, and 229 may each include covalent bond chemical properties 221a, 223a, 225a, 227a, and 229a, which can be obtained from reference bioactivity data of protein 104. In various embodiments, the covalent bond chemical properties 221a, 223a, 225a, 227a, and 229a may include data indicating the following: the type of covalent bond; whether the bond is conjugated; whether the bond contains a ring; and the spatial nature of the covalent bond.

[0093] The first to fifth protein edges 221, 223, 225, 227, and 229 may each further include the first to fifth protein bond physical properties 221b, 223b, 225b, 227b, and 229b, respectively. In various embodiments, the first to fifth protein covalent bond physical properties 221b, 223b, 225b, 227b, and 229b may include protein distance properties representing the physical distance value between one protein atom and another protein atom. For example, the first protein covalent bond physical property 221b may include the physical distance value between the first protein node 222 and the second protein node 224; the second protein covalent bond physical property 223b may include the physical distance value between the second protein node 224 and the third protein node 226; the third protein covalent bond physical property 225b may include the physical distance value between the third protein node 226 and the fourth protein node 228; the fourth protein covalent bond physical property 227b may include the physical distance value between the fourth protein node 228 and the fifth protein node 232; and the sixth protein covalent bond physical property 229b may include the physical distance value between the fifth protein node 232 and the sixth protein node 234. In some embodiments, processor 110 may be configured to determine the physical properties 221b, 223b, 225b, 227b, 229b of the first to fifth proteins by measuring the physical distances between the first and second protein nodes 222, 224, between the second and third protein nodes 224, 226, between the third and fourth protein nodes 226, 228, between the fourth and fifth protein nodes 227, 232, and between the fifth and sixth protein nodes 232, 234, respectively, on the ligand-protein complex structure data 200.

[0094] See Figure 2 and Figure 3The ligand-protein complex 250 structure, as a 3D diagram, may further include multiple intermolecular non-covalent bonds binding ligand atoms to corresponding protein atoms located adjacent to the ligand atoms, represented as multiple non-covalent edges 252, 254. For example, a first ligand node 212 may be connected to a first protein node 222 via a first non-covalent edge 252 and to a sixth protein node 234 via a second non-covalent edge 254. The first and second non-covalent edges 252, 254 may each include first and second non-covalent bond properties 252b, 254b. In various embodiments, the first non-covalent bond property 252b may include a ligand-protein distance property representing a physical distance value between the first ligand node 212 and the first protein node 222; and the second non-covalent bond property 254b may include a physical distance value between the first ligand node 212 and the sixth protein node 234. Accordingly, the physical distance value of the first non-covalent property 252b can indicate the force field interaction between the first ligand node 212 and the first protein node 222; and the second non-covalent property 254b can indicate the force field interaction between the first ligand node 212 and the sixth protein node 234. The ligand-protein distance property can be multiplied by a predetermined ligand-protein vector that takes into account different types of dispersion forces (specifically, Coulomb and London dispersion forces of intermolecular non-covalent bonds).

[0095] In various embodiments, processor 110 may be configured to determine the first non-covalent property 252b by: measuring a physical distance value between a first ligand node 212 and a first protein node 222 on ligand-protein complex structural data 200; comparing the physical distance value with a predetermined distance threshold indicating the presence of the first non-covalent edge 252; and determining the presence of the first non-covalent edge 252 based on the comparison of the physical distance value with the predetermined distance threshold. In some embodiments, the first and second non-covalent properties 252b, 254b may be obtained from reference bioactivity data of the ligand-protein complex 106.

[0096] In various implementations, the predetermined distance threshold may include a distance of 5 Å, and determining the presence of the first non-covalent edge 252 may include determining whether the physical distance value is less than the predetermined distance threshold of 5 Å, which indicates the presence of the first non-covalent edge 252. In other words, if the physical distance value between the first ligand node 212 and the first protein node 222 is less than the predetermined distance threshold, then the first non-covalent edge 252 exists.

[0097] See Figures 1 to 3 To capture overall ligand-protein complex 250 interaction information, the ligand-protein complex 250 structure as a 3D diagram can be represented as follows: , where graph nodes Corresponding to ligand nodes, for example, nodes 212, 214, and 216 representing the first, second, and third ligand atoms; graph nodes Corresponding to protein nodes, for example, the first to sixth protein nodes representing protein atoms: 222, 224, 226, 228, 232, 234; graph edges Corresponding to ligand edges, such as the first and second ligand edges 215 and 217 representing the intramolecular covalent bond of ligand 210; graph edges Corresponding to protein edges, such as the first to fifth protein edges 221, 223, 225, 227, and 229 representing the intramolecular covalent bonds of protein 220; and graph edges. Corresponding to non-covalent edges, for example, are the first and second non-covalent edges 252 and 254 representing the intermolecular non-covalent bonds that bind ligand 210 to protein 220. Furthermore, Represents the 3D coordinates of ligand nodes (e.g., first, second, and third ligand nodes 212, 214, and 216) in the 3D graph. The 3D coordinates of protein nodes (e.g., the first to sixth protein nodes 222, 224, 226, 228, 232, 234) in the 3D graph.

[0098] The first, second, and third ligand properties 212a, 214a, and 216a, and the first to sixth protein properties 222a, 224a, 226a, 228a, 232a, and 234a, can each be represented as: and their respective 3D coordinates Related. The first and second ligand edges 215, 217 and the first to fifth protein edges 221, 223, 225, 227, 229 can be represented as This represents the presence of ligands or protein atoms. With another corresponding ligand or protein atom The covalent bonds between the first and second ligand edges 215 and 217. The chemical properties 215a and 217a of the first and second ligand covalent bonds, and the chemical properties 221a, 223a, 225a, 227a, and 229a of the first to fifth protein edges 221, 223, 225, 227, and 229a, can be represented as follows: .

[0099] The first and second non-covalent edges 252 and 254 can be represented as The first and second non-covalent bond properties 252b and 254b can be represented as Indicator ligand atoms and the corresponding protein atoms that are non-covalently bound to it. The physical distance between them. When the physical distance is less than 5 Å, the first and second non-covalent bond properties 252b and 254b can exist.

[0100] Ligand-protein complex interaction parameter 130

[0101] Figure 4 An exemplary architecture 400 of a ligand-protein (i.e. drug-target) complex 250 graph neural network (DTIGN) is shown for determining ligand-protein complex parameters 130.

[0102] See Figures 1 to 4 The processor 110 is configured to, for each docking conformation, determine ligand-protein complex interaction parameters 130 that indicate intermolecular interaction data of the ligand-protein complex 250. The ligand-protein complex interaction parameters 130 may indicate overall ligand-protein complex 250 interaction data, including intramolecular bond characteristic data defining intramolecular interactions between atoms of the ligand molecule and between atoms of the protein molecule, and intermolecular bond characteristic data defining intermolecular interactions between atoms of the ligand molecule and atoms of the protein molecule within the ligand-protein complex molecule.

[0103] For each docking conformation, the ligand-protein complex parameter 130 may include: a global ligand edge parameter 132 indicating intramolecular interaction data of the ligand covalent bonds that bind the plurality of ligand atoms together; a global protein edge parameter 134 indicating intramolecular interaction data of the protein covalent bonds that bind the plurality of protein atoms together; and a global non-covalent parameter 136 indicating intermolecular interaction data between ligand atoms and corresponding protein atoms in the ligand-protein complex molecule, to capture overall ligand-protein complex molecular interaction information.

[0104] This can be represented based on a first machine learning algorithm, which optionally includes a deep learning algorithm, such as, but not limited to, a neural network-based deep learning algorithm. In various embodiments, the deep learning algorithm may include a graph neural network and optionally a graph convolutional network (GCN). The graph neural network may include the message passing (MP) algorithm of DTIGN, which can be formed by stacking two sets of blocks: a first module 410 for covalent interactions (including overall ligand side parameters 132 and overall protein side parameters 134) and a second module 420 for non-covalent interactions (including overall non-covalent parameters 136), repeated T times, and may be defined by equation (1).

[0105] Equation (1),

[0106] The second type of machine learning algorithm, such as MLP It aggregates representations from two MP stages. To predict the bioactivity parameter 180. In equation (1), covalent and non-covalent messages on all atoms are summed to distinguish molecular size and normalized in the batch normalization of the MLP.

[0107] The properties of the atoms involved in forming covalent bonds and the distances between them play a crucial role in covalent bond formation. See also Figure 1 and Figure 4 The first module 410 for covalent interactions may include a total ligand edge parameter 132. To determine the total ligand edge parameter 132, the processor 110 may be configured to, for the first ligand edge 215, determine, based on the first and second ligand properties 212a, 214a, the first ligand covalent bond chemical property 215a, and a first ligand covalent parameter 144 indicating the binding stability of the first ligand edge 215, a first ligand edge parameter 142 indicating intramolecular interaction data of the first ligand node 212 and a second ligand node 214 located adjacent to the first ligand node 212 and bound to the first ligand node 212 via the first ligand edge 215.

[0108] The first ligand covalent parameter 144 may be based on a predetermined ligand vector, which represents the distance attribute of the intramolecular covalent bond of the first ligand edge 215 as a function of the first ligand covalent bond physical property 215b. The first ligand covalent bond physical property 215b may include a ligand distance attribute representing the physical distance value between the first ligand node 212 and the second ligand node 215, and the predetermined ligand vector may include a predetermined RBF parameterized by a first power value, indicating the distance attribute of the intramolecular covalent bond of the first ligand edge 215. .

[0109] The predetermined RBF can be a vector from a set of Gaussian RBFs, with distance as its input. d of l The power can be expressed according to equations (2) and (3).

[0110] Equation (2)

[0111] Equation (3),

[0112] In equation (2), the mean value of RBF used in this disclosure is... At minimum distance and maximum distance Sample at equal intervals, then take l Power; variance of RBF It can be determined by the difference between adjacent means. This configuration allows RBF to be determined based on... lThe value effectively captures different distance ranges. For interatomic interactions, It can be set to 1 Å to represent the minimum atomic distance in a stable state. It can be set to 6 Å to capture significant atomic interactions within this radius. Furthermore, Specify the number of RBFs, where options are available. To cover interatomic distances between 1 Å and 6 Å.

[0113] Based on equations (2) and (3), to capture the distance property of the intramolecular covalent bond at the edge 215 of the first ligand, the first power value can include a value equal to 1, i.e. The first ligand covalent parameter 144 can be expressed as: ,in This refers to the ligand distance attribute, which represents the physical distance between the first ligand node 212 and the second ligand node 214. Therefore, the first ligand edge parameter 142 can be defined according to equation (4).

[0114] Equation (4),

[0115] The data on intramolecular interactions between two covalently linked nodes, namely It can be defined as the element-wise multiplication of two components. ): A representation that can be a concatenation and dimensionality reduction of adjacent atomic and bond features, denoted as Another possibility is the distance feature of upsampling between adjacent atoms, denoted as... .vector Can represent atoms In the t The hidden features of the layer, whose initial values ​​are obtained from atomic features through a single-layer neural network. Obtained, denoted as .

[0116] The processor 110 can be configured to further determine, for the second ligand edge 217, based on the first and third ligand properties 212a, 216, the second ligand covalent bond chemical property 217a, and the second ligand covalent parameter 148 indicating the binding stability of the second ligand edge 217, a second ligand edge parameter 146 indicating the intramolecular interaction data of the first ligand node 212 and the third ligand node 216 located adjacent to the first ligand node 212 and bound to the first ligand node 212 via the second ligand edge 217. The second ligand covalent parameter 148 can be determined in the same manner as the first ligand covalent parameter 144, and the second ligand edge parameter 146 can be determined in the same manner as the first ligand edge parameter 142, as described above.

[0117] Therefore, processor 110 can determine the overall ligand side parameter 132 based on the aggregation of ligand side parameters of multiple ligand sides of ligand 210. For example, the overall ligand side parameter 132 can be based on the sum of the first ligand side parameter 142 and the second ligand side parameter 144. The overall ligand side parameter 132, which indicates the intramolecular interaction data of the ligand covalent bonds that bind multiple ligand atoms together, can also be expressed according to equation (5).

[0118] Equation (5).

[0119] See Figure 1 and Figure 4 The first module 410 for covalent interactions may also include a global protein side parameter 134. That is, for each docking conformation, the ligand-protein complex parameter 130 may also include a global protein side parameter 134 indicating intramolecular interaction data of the protein covalent bonds that bind multiple protein atoms together. The global protein side parameter 134 may be determined in a similar manner to the global ligand side parameter 132, as described above.

[0120] In short, the processor 110 can be configured to: for a first protein edge 221, determine a first protein edge parameter 151 indicating intramolecular interaction data between a first protein node 222 and a second protein node 224 located adjacent to the first protein node 222 and bound to the first protein node 222 via the first protein edge 221. The first protein edge parameter 151 can be determined based on a first protein property 222a of the first protein node 222, a second protein property 224a of the second protein node 224, a first protein covalent bond chemical property 221a of the first protein edge 221, and a first protein covalent parameter 152 indicating the binding stability of the first protein edge 221. The first protein edge parameter 151 can be defined according to equation (4).

[0121] In various embodiments, the first protein covalent parameter 152 may be based on a predetermined protein vector, which represents the distance attribute of the intramolecular covalent bonds of the first protein edge 221, as a function of the first protein covalent bond physical attribute 221b. The first protein covalent bond physical attribute 221b may include a protein distance attribute representing the physical distance value between the first protein node 222 and the second protein node 224.

[0122] In various embodiments, the predetermined protein vector may include a predetermined RBF parameterized by a second power value, indicating the magnitude of the distance attribute of the intramolecular covalent bonds of the first protein edge 221. RBF pre-order According to equations (2) and (3), the second power value can include a value equal to 1, that is... The first protein covalent parameter 152 can be expressed as: ,in The protein distance attribute refers to the physical distance value between the first protein node 222 and the second protein node 224.

[0123] The processor 110 may be further configured to: for the second protein edge 223, determine a second protein edge parameter 153 indicating intramolecular interaction data of the second protein node 224 and a third protein node 226 located adjacent to the second protein node 224 and bound to the second protein node 224 via the second protein edge 223. The second protein edge parameter 153 may be determined based on a second protein property 224a of the second protein node 224, a third protein property 226a of the third protein node 226, a second protein covalent bond chemical property 223a of the second protein edge 223, and a second protein covalent parameter 154 indicating the binding stability of the second protein edge 223. The second protein covalent parameter 154 and the second protein edge parameter 153 may be determined in the same manner as the first protein covalent parameter 152 and the first protein edge parameter 151, as described above.

[0124] The processor 110 may be further configured to: for the third protein edge 225, determine a third protein edge parameter 155 indicating intramolecular interaction data of the third protein node 226 and the fourth protein node 228 located adjacent to the third protein node 226 and bound to the third protein node 226 via the third protein edge 225. The third protein edge parameter 155 may be determined based on the third protein property 226a of the third protein node 226, the fourth protein property 228a of the fourth protein node 228, the third protein covalent bond chemical property 225a of the third protein edge 225, and a third protein covalent parameter 156 indicating the binding stability of the third protein edge 225. The third protein covalent parameter 156 and the third protein edge parameter 155 may be determined in the same manner as the first and second protein covalent parameters 152, 154 and the first protein edge parameters 151, 153, as described above.

[0125] The processor 110 may be further configured to: for the fourth protein edge 227, determine a fourth protein edge parameter 157 indicating intramolecular interaction data of the fourth protein node 228 and the fifth protein node 232 located adjacent to the fourth protein node 228 and bound to the fourth protein node 228 via the fourth protein edge 227. The fourth protein edge parameter 157 may be determined based on the fourth protein property 228a of the fourth protein node 228, the fifth protein property 232a of the fifth protein node 232, the fourth protein covalent bond chemical property 227a of the fourth protein edge 227, and the fourth protein covalent parameter 158 indicating the binding stability of the fourth protein edge 227. The fourth protein covalent parameter 158 and the fourth protein edge parameter 157 may be determined in the same manner as the first, second, and third protein covalent parameters 152, 154, and 156 and the first, second, and third protein edge parameters 151, 153, and 155, as described above.

[0126] The processor 110 may be further configured to: for the fifth protein edge 229, determine a fifth protein edge parameter 159 indicating intramolecular interaction data of the fifth protein node 232 and the sixth protein node 234 located adjacent to the fifth protein node 232 and bound to the fifth protein node 232 via the fifth protein edge 229. The fifth protein edge parameter 159 may be determined based on the fifth protein property 232a of the fifth protein node 232, the sixth protein property 234a of the sixth protein node 234, the fifth protein covalent bond chemical property 229a of the fifth protein edge 229, and a fifth protein covalent parameter 160 indicating the binding stability of the fifth protein edge 229. The fifth protein covalent parameter 160 and the fifth protein edge parameter 159 may be determined in the same manner as the first to fourth protein covalent parameters 152, 154, 156, 158 and the first to fourth protein edge parameters 151, 153, 155, 157, as described above.

[0127] To determine the overall protein side parameter 134, each of the first to fifth protein side parameters 151, 153, 155, 157, and 159 can be aggregated. In other words, the overall ligand side parameter 132 can be based on the sum of each protein side parameter of each protein covalent bond of the indicator protein 220 (e.g., the first to fifth protein sides 221, 223, 225, 227, and 229). The overall protein side parameter 134 can be represented according to equation (5) defined above.

[0128] Although non-covalent atoms do not share electrons, they exert forces on neighboring atoms due to the charge of the electrons and the resulting electric field. These forces correspond to different types of dispersion forces unique to non-covalent bonds, including Coulomb-type and London-type dispersion forces. Therefore, to consider these forces, see [reference needed]. Figure 1 and Figure 4 The second module 420 for non-covalent interactions of ligand-protein complex 250 may include a total non-covalent side parameter 136.

[0129] To determine the overall non-covalent edge parameter 136, the processor 110 may be configured to: for the first non-covalent edge 252, based on the first ligand property 212a of the first ligand node 212, the first protein property 222a of the first protein node 222, and the first ligand-protein non-covalent parameter 164 indicating the binding stability of the dispersion force type of the first non-covalent edge 252, determine the first non-covalent parameter 162 indicating the intermolecular interaction data of the first ligand node 212 and the first protein node 222 located adjacent to the first ligand node 212 and bound to the first ligand node 212 through the first non-covalent edge 252.

[0130] In various embodiments, the first ligand-protein nonvalent parameter 164 may be based on a predetermined ligand-protein vector, representing the force field amplitude corresponding to the dispersion force type of the first nonvalent edge 252, as a function of the first nonvalent bond property 252b of the first nonvalent edge 252. The first nonvalent bond property 252b may include a ligand-protein distance property representing the physical distance value between the first ligand node 212 and the first protein node 222, and in some embodiments, the ligand-protein distance property may include a distance less than 5 Å. Therefore, the physical distance value may indicate the force field interaction between the first ligand node 212 and the first protein node 222.

[0131] In various embodiments, the predetermined ligand-protein vector may include a coulombic ligand-protein vector comprising a predetermined RBF parameterized by a coulombic power value, indicating the magnitude of the force field of the coulombic intermolecular dispersion force at the first non-covalent edge 252. RBF pre-order According to equations (2) and (3), the Coulomb power value can include a value equal to -2, i.e. Therefore, the non-covalent parameter 164 of the first ligand-protein bond with coulombic dispersion force can be expressed as: ,in The ligand-protein distance attribute refers to the physical distance value between the first ligand node 212 and the first protein node 222, and can indicate the force field interaction between the first ligand node 212 and the first protein node 222.

[0132] In various other embodiments, the predetermined ligand-protein vector may include a London ligand-protein vector, which includes a predetermined RBF parameterized by a London power value, indicating the magnitude of the force field of the London-type intermolecular dispersion force of the first non-covalent edge 252. RBF pre-order According to equations (2) and (3), the London power value can include values ​​equal to -6, i.e. Therefore, the non-covalent parameter 164 of the first ligand-protein with London-type dispersion force can be expressed as: ,in The ligand-protein distance attribute refers to the physical distance value between the first ligand node 212 and the first protein node 222, and can indicate the force field interaction between the first ligand node 212 and the first protein node 222.

[0133] Therefore, the first non-covalent parameter 162 can be defined according to equation (6).

[0134] Equation (6),

[0135] Among them, intermolecular interaction data, i.e., non-covalent ligand atoms With distance Non-covalent protein atoms within (e.g., 5 Å) Between As a hidden layer feature of neighboring atoms (denoted as...) ) and element-wise multiplication of the spliced, upsampled -2 ​​power distance (Coulomb dispersion force) and -6 power distance (London dispersion force) ( ), can be written as .

[0136] The processor 110 may be further configured to: for the second non-covalent edge 254, based on the first ligand property 212a of the first ligand node 212, the sixth protein property 234a of the sixth protein node 234, and a second ligand-protein non-covalent parameter 168 indicating the binding stability of the dispersion force type of the second non-covalent edge 254, determine a second non-covalent parameter 166 indicating the intermolecular interaction data of the first ligand node 212 and the sixth protein node 234 located adjacent to the first ligand node 212 and bound to the first ligand node 212 via the second non-covalent edge 254. The second ligand-protein non-covalent parameter 168 and the second non-covalent parameter 166 may be determined in the same manner as the first ligand-protein non-covalent parameter 164 and the first non-covalent parameter 162, as described above. For example, the second ligand-protein non-covalent parameter 168 may be based on a Coulomb power value (i.e., ) or London power value (i.e. ) Parameterized predefined RBF Consider the type of dispersion force of the second non-covalent edge 254 (e.g., Coulomb-type dispersion force or London-type dispersion force).

[0137] Therefore, processor 110 can determine the overall non-covalent edge parameter 136 based on the aggregation of non-covalent parameters of multiple non-covalent edges of the ligand-protein complex 250. For example, the overall non-covalent edge parameter 136 can be based on the sum of the first non-covalent parameter 162 and the second non-covalent parameter 166. The overall non-covalent edge parameter 136 can indicate intermolecular interaction data of non-covalent bonds binding ligand atoms to corresponding protein atoms, and can be expressed according to equation (7).

[0138] Equation (7).

[0139] In various embodiments, processor 110 may also include a second machine learning algorithm, such as MLP (see also Equation (1)), for determining ligand-protein complex interaction parameters 130. The second machine learning algorithm can be used to update ligand and protein atom hiding features based on intramolecular interaction data of the first and second ligand edges 215, 217; the intermolecular interaction portions of the first to fifth protein edges 221, 223, 225, 227, 229 (see Equation (5)) and the first and second non-covalent edges 252, 254 (see Equation (7)), along with the combination of atom hiding features, and can be represented according to Equation (8).

[0140] Equation (8)

[0141] Therefore, the representation of the ligand-protein complex 250 in one of the plurality of docking pockets (including at least one binding site) f This can be obtained by summing the hidden features of all atoms after undergoing T-layer updates, i.e. And that statement f It can be used for self-attention learning in the third machine learning algorithm, and for semi-supervised learning in the first machine learning algorithm. In other words, see equation (1), It can be represented as f Therefore, the objective function for determining the predictive power of the bioactivity parameter 180 used to train DTIGN can be expressed according to equation (9).

[0142] Equation (9),

[0143] in and These represent the index and total number of training samples, respectively.

[0144] The interaction parameter of the polymerized ligand-protein complex is 170.

[0145] To determine the ligand-protein complex representation parameter 130 discussed above, candidate pockets containing at least one validated binding site of the protein were selected to gain insight into ligand-protein complex 250 interactions. While these candidate pockets represent possible regions for ligand-protein complex 250 binding, many ligands 210 can interact with target protein 220 without explicitly identifying their binding pockets due to the lack of native structures.

[0146] To solve these problems, see also Figure 1 The processor 110 may include a first machine learning algorithm, such as a deep learning algorithm, including but not limited to a graph neural network, for identifying natural-like binding pockets and docking conformations from pocket-ligand complexes docked to candidate pockets of target protein 220. In various embodiments, the processor 110 may also include a third machine learning algorithm, which may include a multi-head self-attention algorithm (also referred to as one or more self-attention head mechanisms) for determining ligand-protein complex representation parameters 130.

[0147] During model training, docking conformation data of the same ligand 210 docking to multiple candidate pockets can be packaged to facilitate training DTIGN using a third machine learning algorithm. In various implementations, the self-attention layer of the third machine learning algorithm is combined with a residual module to integrate pocket-ligand representations to determine the aggregated ligand-protein complex interaction parameters 170 of the intermolecular interaction data indicating the aggregation of ligand-protein complex 250 under multiple docking conformations, thereby capturing the participation of ligand 210 in the interaction of the associated target protein 220. In some implementations, the residual module may include a dynamic residual module for modeling dynamic residual attention, or a grouped residual layer for replacing the query, key-value linear layer in the self-attention algorithm. In other words, the residual module can be used to add pocket-ligand representations to the output of the multi-head self-attention algorithm before the pocket-ligand representations are aggregated together (see equations (10) and (15) below).

[0148] To determine the polymerized ligand-protein complex interaction parameters 170, processor 110 may be configured to: for each of a plurality of docking conformations, determine ligand-protein complex representation parameters 172 indicating intermolecular interaction data of the updated ligand-protein complex, in each docking conformation, wherein the updated ligand-protein complex is formed when a ligand binds to at least one identified binding site contained in at least one identified docking pocket of a protein to form the updated ligand-protein complex, wherein determining the ligand-protein complex representation parameters 172 includes processing the ligand-protein complex interaction parameters 130 with a third machine learning algorithm; and determine the polymerized ligand-protein complex interaction parameters 170 based on the polymerization of the ligand-protein complex representation parameters 172 for each docking conformation of the updated ligand-protein complex. This can be expressed according to equations (10) to (16) below.

[0149] In this disclosure, Figure It may include the same ligand 210, which can be docked to the target protein. One candidate pocket, configuration For each pocket-ligand conformation n, the corresponding graphical features can be obtained using DTIGN as described above. Then, the query for each conformation in each attention head of the third machine learning algorithm. ,key Sum From based on Different single-layer neural networks are obtained, as represented by equation (10).

[0150] Equation (10),

[0151] in The weight matrix reduces the dimensionality of the input features to achieve effective self-attention in each attention head.

[0152] Subsequently, for each conformation n, all conformations Query s and keys It can be used to calculate the self-attention weights of a third machine learning algorithm, and can be represented by equation (11).

[0153] Equation (11),

[0154] in The dimensions of the key.

[0155] Then, a normalized exponential function, such as the SoftMax function, can be used to assign self-attention weights between conformations n and m to the third machine learning algorithm. After normalization, it can be expressed according to equation (12).

[0156] Equation (12).

[0157] The ligand-protein complex representation parameter 172 may include the output representation of each attention head of a third machine learning algorithm, which can be derived according to equation (13) by adjusting the normalized self-attention weights. Conformation All values We obtain the result by weighting and summing.

[0158] Equation (13).

[0159] Then the representations from different attention heads from the third machine learning algorithm can be concatenated (denoted as...). ), to obtain the output representation of a third machine learning algorithm (e.g., a multi-head self-attention mechanism), as represented by equation (14),

[0160] Equation (14).

[0161] Then, according to equation (15), the representation of all pocket-ligand complexes n of the same ligand 210 can be compared with the corresponding graphical features. Adding and summing the results yields the ligand-protein complex interaction parameter 170, representing the multiconformation polymerization. ,

[0162] Equation (15).

[0163] In various implementation schemes, the representation of multiconformation aggregation The single-conformation representation in the alternative equation (9) The bioactivity parameter 180 was used to determine the bioactivity loss, and the bioactivity loss can be optimized according to equation (16).

[0164] Equation (16),

[0165] The superscript (attn) can be used to distinguish features aggregated by a third machine learning algorithm (i.e., a multi-head self-attention mechanism). The obtained loss function .

[0166] Therefore, in various embodiments, the bioactivity parameter 180 can be based on the polymerized ligand-protein complex interaction parameter 170 (i.e., the representation of multiconformation polymerization). To determine.

[0167] semi-supervised learning

[0168] In the datasets of reference bioactivity data for ligand 102, reference bioactivity data for protein 104, reference bioactivity data for ligand-protein complex 106, and structural data for ligand-protein complex 200, typically only a small fraction of bioactive ligand molecules have publicly available natural structural data (usually obtained by X-ray or nuclear magnetic resonance) with the target protein molecules.

[0169] Utilizing the natural structural data of ligand 210, protein 220, and / or ligand-protein complex 250, processor 110 may include a first machine learning algorithm, i.e., a deep learning algorithm, such as, but not limited to, a graph neural network algorithm, which can provide a semi-supervised learning strategy to guide the learning of a third machine learning algorithm (e.g., a self-attention mechanism). In various embodiments, the first machine algorithm may be the same machine algorithm used to determine ligand-protein complex parameters 130 and / or to identify natural-like binding pockets and docking conformations from candidate pockets docked to target protein 220. The natural structural data of ligand 210, protein 220, and / or ligand-protein complex 250 may be stored in memory 120 or in another processor or server.

[0170] In various embodiments, based on the natural structural data of ligand 210, protein 220, and / or ligand-protein complex 250, protein atoms within residues of protein 220 can be extracted to form at least one local pocket-ligand complex, wherein at least one protein atom is located within 5 Å of any ligand atom. Data regarding the at least one local pocket-ligand complex can be entered into DTIGN to determine natural ligand-protein complex interaction parameters 182 and naturally polymerized ligand-protein complex interaction parameters 184, in a manner similar to that used to determine ligand-protein complex interaction parameters 130 and polymerized ligand-protein complex interaction parameters 170, respectively; details have been omitted for brevity. This can be represented as a natural representation. .

[0171] In various embodiments, for ligand-protein pairs having native structures, processor 110 may be further configured to determine similarity, for example, the interaction parameters 170 of the polymerized ligand-protein complex (i.e., multiconformation polymerization representation). The interaction parameter 184 between the naturally polymerized ligand-protein complex and the naturally expressed ligand-protein complex (i.e., the natural representation) is 184. The cosine similarity between the features is used to ensure feature alignment. Therefore, the loss function can be defined according to equation (17).

[0172] Equation (17),

[0173] in The number of ligand-protein pairs with native structural data in the reference bioactivity data of ligand 102, the reference bioactivity data of protein 104, and the reference bioactivity data of ligand-protein complex 106.

[0174] In various implementations, the interaction parameter 184 of the naturally polymerized ligand-protein complex (i.e., naturally represented) is... This data can be further input into DTIGN and processed by a second machine learning algorithm (e.g., MLP) to optimize the determination of the bioactivity parameter 180, ensuring the actual bioactivity with ligand 210. Alignment. In this respect, the total loss function used to train DTIGN can be defined according to equation (18).

[0175] Equation (18)

[0176] in The coefficient γ equals 1 when the training ligand s has at least one native structure with the target protein 220; otherwise, it equals 0. The coefficient γ balances the contribution of the biological activity parameter 180 to the prediction loss and cosine similarity loss. (Symbol) and These represent the number of training ligands s and the number of native structures of target protein 220, respectively. It is noteworthy that when ligand 210 interacts with target protein 220 through multiple native structures, i.e. Its representation can be summed to obtain the interaction parameters 184 of the naturally aggregated ligand-protein complex, i.e., the natural representation. .

[0177] According to another aspect of this disclosure, a drug discovery platform is provided, which includes a reference... Figures 1 to 4 The system 100 is described above.

[0178] Figure 5 An exemplary flowchart of method 500 is shown, which is used to determine the bioactivity parameters of a ligand when the ligand binds to a protein to form a ligand-protein complex, the ligand binding to the protein in multiple docking conformations. The bioactivity parameters, ligand, protein, and ligand-protein complex may refer to the bioactivity parameters 180, ligand 210, protein 220, and ligand-protein complex 250 of system 100, and references are also provided. Figures 1 to 4 Let's have a discussion.

[0179] Method 500 may include providing a processor for performing the following steps: acquiring ligand-protein complex structure data and reference bioactivity data of the ligand, protein, and ligand-protein complex indicating the structure of the ligand-protein complex; the ligand-protein complex structure data including: a plurality of ligand nodes of the ligand, each ligand node having ligand properties; a plurality of protein nodes, each protein node having protein properties; a plurality of non-covalent edges connecting the ligand nodes and the corresponding protein nodes, each non-covalent edge having non-covalent bond properties (step 502); for each of the plurality of docking conformations, determining ligand-protein complex interaction parameters indicating intermolecular interaction data of the ligand-protein complex under each docking conformation, wherein determining the ligand-protein complex interaction parameters includes, for each non-covalent... The process involves: determining, based on the first ligand properties of the first ligand node, the first protein properties of the first protein node, and the binding stability of the first ligand-protein non-covalent parameters indicating the type of dispersion force of the first non-covalent edge, non-covalent parameters indicating the intermolecular interaction data of the first ligand node and the first protein node located nearby and bound to the first ligand node through the first non-covalent edge; determining, based on the aggregation of non-covalent parameters of multiple non-covalent edges of the ligand-protein complex, ligand-protein complex interaction parameters (step 504); determining, based on the intermolecular interaction data indicating the aggregation of the ligand-protein complex in multiple docking conformations, ligand-protein complex interaction parameters of the aggregation (step 506); and determining, based on the aggregation of ligand-protein complex interaction parameters, bioactivity parameters (step 508).

[0180] In various implementations, the first ligand-protein nonvalent parameter may be based on a predetermined ligand-protein vector, which represents the force field amplitude corresponding to the dispersion force type of the first nonvalent edge, as a function of the nonvalent bond property of the first nonvalent edge.

[0181] In various embodiments, the predetermined ligand-protein vector may include a Coulomb ligand-protein vector comprising a predetermined RBF parameterized by a Coulomb power value, indicating the magnitude of the force field of the Coulomb dispersion force of the first non-covalent edge. In various other embodiments, the predetermined ligand-protein vector may include a London ligand-protein vector comprising a predetermined RBF parameterized by a predetermined London power value, indicating the magnitude of the force field of the London dispersion force of the first non-covalent edge.

[0182] In various embodiments, the non-covalent property of the first non-covalent edge may include a ligand-protein distance property, which represents a physical distance value between a first ligand node and a first protein node in the ligand-protein complex. The physical distance value may indicate a force field interaction between the first ligand node and the first protein node. In some embodiments, to determine the ligand-protein distance property, method 500 may further include: measuring the physical distance value between the first ligand node and the first protein node on the ligand-protein complex structural data; comparing the physical distance value with a predetermined distance threshold indicating the presence of the first non-covalent edge; and determining the presence of the first non-covalent edge based on the comparison of the physical distance value with the predetermined distance threshold.

[0183] In various embodiments, the ligand-protein complex structural data may further include a plurality of ligand edges in the ligand that bind the plurality of ligand nodes to each other, each ligand edge having ligand covalent chemical properties and ligand covalent physical properties, wherein method 500 may further include: for each of the plurality of ligand edges, determining ligand edge parameters indicating intramolecular interaction data of a first ligand node and a second ligand node located adjacent to it and bound to the first ligand node via the first ligand edge, the ligand edge parameters being determined based on: a first ligand property of the first ligand node, a second ligand property of the second ligand node, a ligand covalent chemical property of the first ligand edge, and a first ligand covalent parameter indicating the binding stability of the first ligand edge, and further determining the ligand-protein complex interaction parameters based on the aggregation of the ligand edge parameters of the plurality of ligand edges.

[0184] In various implementations, the first ligand covalent parameter may be based on a predetermined ligand vector, which represents the distance property of the intramolecular covalent bond of the first ligand edge, as a function of the physical property of the ligand covalent bond of the first ligand edge.

[0185] In various embodiments, the predetermined ligand vector may include a predetermined RBF parameterized by a first power value, indicating the distance property of the intramolecular covalent bond of the first ligand edge.

[0186] In various implementations, the ligand covalent bond physical properties of the first ligand edge may include a ligand distance property, which represents the physical distance value between the first ligand node and the second ligand node of the ligand.

[0187] In various embodiments, the ligand-protein complex structural data may further include a plurality of protein edges in the protein that bind the plurality of protein nodes to each other, each protein edge having protein covalent chemical properties and protein covalent physical properties, wherein method 500 may further include: for each of the plurality of protein edges, determining protein edge parameters that indicate intramolecular interaction data of a first protein node and a second protein node located adjacent to it and bound to the first protein node through the first protein edge of the plurality of protein edges, the protein edge parameters being determined based on: a first protein property of the first protein node, a second protein property of the second protein node, a protein covalent chemical property of the first protein edge, and a first protein covalent parameter indicating the binding stability of the first protein edge, and further determining the ligand-protein complex interaction parameters based on the aggregation of the protein edge parameters of the plurality of protein edges.

[0188] In various implementations, the first protein covalent parameter may be based on a predetermined protein vector, which represents the distance attribute of the intramolecular covalent bonds of the first protein edge, as a function of the physical property of the protein covalent bonds of the first protein edge.

[0189] In various embodiments, the predetermined protein vector may include a predetermined RBF parameterized by a second power value, indicating the distance property of intramolecular covalent bonds of the first protein edge.

[0190] In various implementations, the protein covalent bond physical properties of the first protein edge may include a protein distance property, which represents the physical distance value between the first protein node and the second protein node of the protein.

[0191] In various implementations, method 500 may also include a first machine learning algorithm, which optionally includes a deep learning algorithm, such as, but not limited to, a neural network-based deep learning algorithm.

[0192] In various embodiments, method 500 may further include a second machine learning algorithm for determining the ligand-protein complex interaction parameters, wherein the second machine learning algorithm optionally includes an MLP.

[0193] In various embodiments, method 500 may further include a third machine learning algorithm for determining ligand-protein complex interaction parameters of the polymer, wherein the third machine learning algorithm optionally includes a multi-head self-attention algorithm.

[0194] In various embodiments, the bioactivity parameter may include at least one of the following: an EC50 parameter indicating the half-maximal effective concentration of the ligand; and / or an IC50 parameter indicating the half-maximal inhibitory concentration of the ligand.

[0195] According to another aspect of this disclosure, a computer-readable medium is provided, comprising instructions that, when executed by a processor, cause the processor to perform method 500.

[0196] Although the determination of ligand-protein complex interaction parameter 130, polymerized ligand-protein complex interaction parameter 170, and bioactivity parameter 180 has been referenced Figure 2 The exemplary ligand-protein complex 250 shown has been discussed, but embodiments of the present disclosure are not limited thereto, and the system 100 and method 500 of the present disclosure are applicable to other ligand-protein (i.e. drug-target) complexes with different numbers of ligand nodes, ligand edges, protein nodes, protein edges and / or non-covalent edges.

[0197] Example

[0198] The system 100 and method 500 disclosed herein for determining the bioactivity parameter 180 of ligand 210 are further illustrated in the following examples, which are provided by way of illustration and are not intended to limit the scope of this disclosure.

[0199] In the following embodiments, the term "ligand" may refer to ligand 210 of the present disclosure; the term "protein" may refer to protein 220 of the present disclosure; the term "drug-target" may refer to ligand-protein complex 250 of the present disclosure; the term "bioactivity" may refer to bioactivity parameter 180 of the present disclosure; the term "multi-head self-attention; self-attention mechanism" may refer to the third machine learning algorithm of the present disclosure; and the term "semi-supervised component" may refer to the first machine learning algorithm of the present disclosure, such as a deep learning algorithm, including but not limited to graph neural networks.

[0200] Figure 6 An exemplary schematic diagram of the overall DTIGN 600 based on drug-target interaction-based bioactivity prediction is shown. (Reference) Figure 6 The contributions of this disclosure can be summarized as follows:

[0201] By utilizing molecular interaction information generated through molecular docking methods, bioactivity prediction in drug discovery has been extended beyond traditional single-molecule QSAR models. Current bioactivity prediction methods largely ignore this information. In this disclosure, this information is validated using a newly designed network, DTIGN, and a comprehensive set of experiments.

[0202] Based on interatomic interactions, DTIGN was designed. Employing multi-head self-attention within DTIGN helps identify true binding pockets and conformations in docking results. Added semi-supervised components (e.g., first-order machine learning algorithms, including deep learning algorithms such as, but not limited to, graph neural networks) enable DTIGN to refine model attention using scarce ligand-protein natural structures.

[0203] The experimental results presented in this paper confirm that DTIGN outperforms nine leading non-interaction-based bioactivity prediction methods, with an average improvement of 27.03%. Furthermore, naturally occurring docking conformations receive high attention in DTIGN, indicating that its accurate self-attention to naturally occurring conformations is crucial for accurate bioactivity prediction.

[0204] I. Data Preparation

[0205] A. Bioactivity data collection and preprocessing

[0206] Bioactivity data sources: This disclosure utilizes bioactivity data from publicly available datasets, primarily from ChEMBL, one of the leading repositories of bioactive molecules with drug-like properties. ChEMBL contains over 2 million molecules, covering 1.5 million assays for 15,000 protein targets across a wide range of organisms. This disclosure focuses on single proteins (distinguished from protein compounds or families) with over 1,000 bioactive small molecules. Of the 6,778 human proteins in ChEMBL, only 900 meet its criteria. The dataset includes various bioactivity measurements (e.g., IC50, EC50, Kd, ​​Ki, % inhibition, potency, etc.) and molecular representations (e.g., SMILES and ChEMBL identifiers), totaling 6.75 million records.

[0207] Bioactivity Data Preprocessing: The raw data underwent several preprocessing steps. Initially, molecules with “Null” assay values ​​were removed. Assay types with sample sizes less than 6 were excluded because the data was insufficient for robust model training, validation, and testing. Some low-bioactivity assays did not have precise values ​​but were marked as being greater than a certain threshold. For these non-critical samples, the threshold was used as their assay value (e.g., setting assay ">10000" as assay "10000"). Values ​​were standardized to appropriate scales (e.g., from “IC50 = 1 nM” to “pIC50 = -log 10⁻⁹ M = 9”). In cases where the same assay type had different units that could not be converted, the assay type was split into different datasets (e.g., “activity in percentage” and “activity in nM”). Repeated assays with the same units were averaged to maintain data consistency.

[0208] Bioactivity dataset construction: Each distinct assay category is treated as a separate dataset for each specific target, as different assay types have different biological implications. Within each dataset, the data is split in a 5:1 ratio, with 5 parts used as training data and the remainder as test data. The model is trained on every 4 parts of the training data, and its performance is validated on the remaining parts. The model with the best validation performance is then tested on the test set. During this process, molecules are segmented based on their Bemis-Murcko backbones, and similar backbones are grouped together within the same dataset whenever possible. This method amplifies the differences between training, validation, and test data, facilitating the evaluation of the model's generalization ability.

[0209] B. Generation of drug-target interaction data

[0210] Ligands with similar structures can exhibit significantly different biological activities, which can be distinguished by their interactions with proteins. This fact prompts this disclosure to incorporate drug-target interactions into the prediction of biological activity.

[0211] Figure 7 The docking pocket selection 700 of the exemplary ligand CHEMBL202 is shown. Figure 8 Table 800 shows a comparison of the predictive performance of bioactivity (pIC50 for CHEMBL202) on different data sources, AutoDock Vina and QVina-W, where the ↑ or ↓ arrows indicate that larger or smaller values ​​are better. Figure 9 A table is shown for a benchmark dataset 900 used to predict the bioactivity of ligand-protein complexes.

[0212] Binding Sites and Docking Pockets: This disclosure first retrieves functional binding sites for target proteins from the UniProt database using identifiers from the ChEMBL bioactivity dataset (constructed as described in Section IA). Next, this disclosure selects corresponding crystal structures of proteins from the RCSB Protein Database (PDB) that have high resolution and ligand binding at most binding sites. Subsequently, this disclosure labels the binding sites on the crystal structures of each protein and divides pockets based on spatial clustering of the binding sites. Finally, this disclosure incorporates a large pocket containing all identified binding sites. An exemplary pocket selected for the exemplary target protein CHEMBL202 is shown below. Figure 7 As shown.

[0213] Docking Methods: This disclosure compares the predictive performance of data generated by two molecular docking software programs, AutoDock Vina and QVina-W. Using CHEMBL202 as an example, this disclosure docks 957 ligands with known biological activities to seven pockets, generating 6,446 molecules. These were split into training (5,312) and test (1,134) sets using a 5-fold cross-validation strategy. A Geometric Interaction Graph Neural Network (GIGN) was used as the backbone network to optimize the mean squared error (MSE). Despite the noise introduced by multiple pockets, this disclosure evaluates the docking methods used to generate the training data. Graph Convolutional Networks (GCN) and a Universal 3D Molecular Representation Learning Framework (Uni-Mol) were used as controls. Metrics included Pearson correlation coefficient (r), root mean square error (RMSE), and Kendall Tau-b coefficient. The performance of models trained using the docking results of AutoDock-Vina and QVina-W was compared. Figure 8 The results highlight their superior performance compared to ligand-based methods, particularly Vina-Docked, which exhibits a lower RMSE but higher [performance / efficacy]. r and Therefore, AutoDock Vina was chosen for ligand-protein complex generation in this disclosure.

[0214] It is worth noting that with the development of computing power and the emergence of technologies such as cloud computing, large-scale molecular docking between molecules and proteins is no longer a challenge. For example, on a single NVIDIA GeForce RTX 3060 GPU card, the GPU version of AutoDock Vina can dock an average of 16,222 molecules to common proteins per day.

[0215] Benchmark datasets: This disclosure selects target proteins with abundant pIC50 or pEC50 bioactivity measurements from the ChEMBL bioactivity dataset to create diverse benchmark datasets, such as... Figure 9 The datasets listed above differ in characteristics, including the number of unique ligand molecules (from 957 to 9,573), binding sites (from 2 to 24), selected pockets (from 2 to 7), and the first n conformations of each pocket (from 1 to 4). In the dataset IDs, "E" represents pEC50 bioactivity, and "I" represents pIC50 bioactivity. Higher ID values ​​indicate more unique ligand molecules (the number approximately doubles with each additional digit).

[0216] II. Experiment

[0217] A. Experimental Preparation

[0218] Data Introduction: Bioactivity data used to evaluate the proposed method were obtained from ChEMBL. Details of data collection and preprocessing are provided in Section I above. Pocket-ligand complex data were obtained by docking 3D ligand molecular conformations (generated using RDKit based on molecular SMILES strings) with the target protein X-ray structure (obtained from UniProt) using AutoDock Vina. The selection of docking pockets was based on binding sites recorded on UniProt (further details on pocket selection are provided in Section IB above). Figure 7 ).

[0219] Comparison Methods: Deep learning-based ligand bioactivity prediction primarily employs graph neural networks (GNNs) to predict ligand structure. Therefore, this disclosure compares two main classes of methods: 2D and 3D GNNs that construct ligand structures. 2D GNNs include GCNs, graph attention networks (GATs), graph isomorphic networks (GINs) pre-trained using supervised learning and context prediction, message passing neural networks (MPNNs), Weave, neural fingerprinting (Neural FP), and attentive FP. For 3D GNNs, this disclosure considers typical spatial graph convolutional networks (SGCNs) and state-of-the-art Uni-Mol pre-trained models. As an additional baseline, this disclosure also employs a simple MLP to predict bioactivity based on the AutoDock-Vina score of the ligand-protein conformation involved.

[0220] Evaluation metrics: To comprehensively evaluate model performance, this disclosure employs two commonly used metrics for bioactivity prediction: Pearson correlation coefficient (r) and RMSE, as well as a ranking metric. .

[0221] The r-index assesses the linear correlation between two variables and can be defined according to the following equation (19):

[0222] Equation (19),

[0223] in and These represent actual and predicted bioactivity values, respectively. and These are the mean values ​​of the actual and predicted bioactivity values, respectively. n represents the total number of test samples. A higher r value indicates higher overall accuracy of the model's predictions.

[0224] RMSE is a widely used performance evaluation metric for regression models, defined according to equation (20) as the square root of the average of the squared differences between predicted and actual values:

[0225] Equation (20),

[0226] in and These represent the actual and predicted bioactivity values, respectively, and n represents the total number of test samples. A lower RMSE value indicates better regression model performance.

[0227] It is a nonparametric statistic used to measure the consistency between two rankings, and is commonly used in fields such as statistics and computational biology. According to equation (21), it is calculated by counting the number of consistent and inconsistent pairs in the rankings:

[0228] Equation (21),

[0229] Where C represents the number of consistent pairs and D represents the number of inconsistent pairs. Consistent pairs mean that the items in the two rankings are similar in rank, while inconsistent pairs mean that the items are different in rank. The value ranges from -1 to 1, where 1 indicates complete consistency, 0 indicates no consistency, and -1 indicates complete inconsistency between rankings.

[0230] As for 3D GNNs, these include SGCN and Uni-Mol. The source code and usage instructions for both 3D GNNs are available on GitHub. Notably, GIN and Uni-Mol provide models pre-trained on large-scale graph data, enabling this disclosure to fine-tune these models to enhance performance.

[0231] Implementation details: Training data is divided into 5 subsets for 5-fold cross-validation, where RMSE is used to evaluate the model's generalization ability on the validation set. Training stops when the validation RMSE no longer decreases within 100 epochs, and the model with the lowest validation RMSE is evaluated on the test set. The model with the lowest validation RMSE at the 5-fold stopping point is considered the final result. The validation RMSE during the first 40 epochs of training is not considered in the stopping criterion, as this is considered a warm-up phase for the model. The balancing factor is set to 1 in all experiments.

[0232] During batch training, all conformations of a single ligand docking with a protein pocket are treated as a single entity to fully train the self-attention mechanism. Since the number of pockets and docking conformations varies among different proteins, this disclosure sets a uniform batch size of 128, which is divided by the number of pockets and conformations to determine the number of unique ligands sampled per batch. For optimization, this disclosure employs the Adam optimizer with an initial learning rate of 10. -4 The weight decays to 10. -4.3The learning rate decays by 95% every 10 training epochs. The graph convolutional module has 3 layers (i.e., T = 3). The hidden feature dimension of the model is consistently set to 256. The multi-head self-attention layer uses 8 heads with a dropout rate of 0. The MLP consists of three linear layers with LeakyReLU activation and BatchNorm1d.

[0233] Data containing both bioactive records and natural ligand-protein complexes are extremely limited. Therefore, aside from ablation studies, these scarce data instances are reserved specifically for validating the accuracy of the self-attention mechanism, rather than for further model training.

[0234] Experimental setup: To ensure fair comparison, DTIGN should be configured as closely as possible to the recommended model settings of the baseline model.

[0235] B. Ablation Studies

[0236] Figure 10 A table illustrating the ablation studies 1000 on the I1 task (dihydrofolate reductase) on the benchmark dataset is shown.

[0237] To verify the effectiveness of self-attention mechanisms and semi-supervised learning, benchmark datasets (such as...) were used. Figure 9 Ablation experiments were performed on the I1 task (protein name: dihydrofolate reductase) as detailed in the description. Figure 10 The comparison of this method with or without edge features (denoted as "w / o edge features") and with or without multi-head self-attention mechanism (denoted as "w / o self-attention") is shown. The default training uses MSE loss for bioactivity prediction, i.e., in equation (9). However, with the introduction of multi-head self-attention mechanisms, this disclosure switches to using equation (18). To optimize DTIGN based on attention-aggregated features. If semi-supervised learning is enabled, the full loss defined in equation (18) can be used.

[0238] DTIGN outperforms leading GNN methods in predicting molecular biological activity against dihydrofolate reductase: Figure 10 The results indicate that DTIGN and its variants, which utilize pocket-ligand complex data to dock to dihydrofolate reductase proteins, generally outperform state-of-the-art methods that rely solely on ligand data. Specifically, compared to the best-performing method on each metric, DTIGN improves by 44.83% in biological activity prediction relevance (measured by Pearson correlation coefficient r), by 11.02% in accuracy (measured by RMSE), and by ranking (measured by...). In terms of measurement, it improved by 59.82%.

[0239] Ablation studies on edge features and self-attention mechanisms of DTIGN: Figure 10 The first two rows of the DTIGN results show that incorporating edge features improved the accuracy of bioactivity value prediction by 20%, while ranking consistency decreased slightly by 6.4%. Furthermore, as... Figure 10 As shown in the second and third rows of the DTIGN results, integrating the multi-head self-attention mechanism improved the bioactivity prediction relevance and ranking consistency by 16.8% and 17.5%, respectively, while slightly reducing accuracy by 1.4%. Notably, none of these models were trained using natural ligand-protein complex structures.

[0240] Ablation studies on DTIGN semi-supervised learning: such as Figure 10 The last two rows of the DTIGN results show that introducing a limited number (approximately 3% of the training data) of X-ray ligand-protein complexes for additional semi-supervised learning improved all metrics. Specifically, bioactivity prediction relevance, accuracy, and ranking consistency increased by 3.4%, 7.4%, and 8.2%, respectively. In summary, the results highlight the benefits of the DTIGN design, including its edge feature definition, self-attention module design, and the integration of semi-supervised learning, in achieving accurate bioactivity prediction.

[0241] C. Self-attention accuracy verification

[0242] Figure 11 The diagram shows the attention value of 1100 for the test ligand on the protein pocket during training. The natural binding sites depicted above are from the RCSB PDB database. Details and corresponding pockets are given in the legend. Figure 12 A table illustrating the crossover ratio of 1200 between the "natural pocket" and the "attention pocket" is shown. Figure 13 Box plot 1300 shows the attention distribution of DTIGN on sorted docking conformations involved in the pIC50 test set (containing 9,975 re-docking conformations on 1,011 native PDB structures). “Avg RMSD” represents the average RMSD between ligand-protein docking conformations and their respective native structures in the PDBBind dataset. Figure 14 Box plot 1400 shows the attention distribution of DTIGN on the sorted docking conformations involved in the pKd test set (containing 14,411 re-docking conformations on 1,455 native PDB structures). “AvgRMSD” represents the average RMSD between ligand-protein docking conformations and their respective native structures in the PDBBind dataset. Figure 15Box plot 1500 shows the attention distribution of DTIGN on the sorted docking conformations involved in the pKi test set (containing 6,196 re-docking conformations on 626 native PDB structures). “Avg RMSD” represents the average RMSD between ligand-protein docking conformations and their respective native structures in the PDBBind dataset.

[0243] The ligand-protein conformation docked in the actual binding pocket tends to attract DTIGN's attention during training: In this section, this disclosure describes, through dihydrofolate reductase (a protein with numerous pockets, such as...) Figure 9 Unsupervised training on (as shown) demonstrates the ability to combine the self-attention mechanism with pocket and conformation recognition capabilities. This disclosure... Figure 11 This study investigated the alignment between the self-attention-prominent pocket and the natural pocket. First, the RCSB PDB database was queried to identify known binding sites on the protein (CHEMBL202) and binding sites for the test ligand (CHEMBL7492), such as... Figure 11 As indicated in the above footnote. Subsequently, this disclosure is in Figure 11 The model attention values ​​for ligand-protein docking conformations obtained during training on various pockets are plotted. This disclosure observes that ligand-protein conformations docked on pockets 1 and 2 consistently receive a high proportion of attention throughout training, consistent with ligand-protein binding sites (within pockets 1 and 2) recorded in the RCSB PDB database.

[0244] Quantifying the accuracy of self-attention in DTIGN: The model's attention pockets (denoted as "attention pockets") are sorted based on the average attention values ​​during model training, selecting the same number of pockets as the actual binding pockets from high to low. Then, in Figure 12 The cross-union ratio (IoU) between the “attention pocket” and the natural structure pocket (denoted as the “natural pocket”) was calculated. The results showed that 13 of the 17 training ligands with natural structures associated with the target achieved IoU scores exceeding 0.5, indicating successful detection of the binding pocket. Notably, the natural structures associated with the training ligands were not utilized during unsupervised training; the training ligands were trained only on multiple docking pocket-ligand complexes and their biological activities.

[0245] DTIGN's model attention generalizes on unseen test data: the model also demonstrates its ability to correctly identify binding pockets for unseen test ligands. Of the nine instances with natural structures, six achieved IoU scores above 0.5 (e.g., ...). Figure 12(As shown). This further emphasizes the effectiveness of the self-attention mechanism of this disclosure in recognizing binding pockets. Furthermore, the training and testing ligands consistently maintained an average IoU score above 0.5 between the “attention pocket” and the “natural pocket,” demonstrating the generalization ability of this method in pocket recognition.

[0246] The generalization of DTIGN's model attention on PDBBind re-docking data. To validate the model's generalization ability on the PDBBind dataset, this disclosure utilizes QuickVina 2.1 to re-dockle proteins and ligands from the PDBBind v2020 dataset (http: / / www.pdbbind.org.cn / download.php), generating 10 ligand-protein docking conformations for each PDB code (one code represents a unique natural structure), where feasible. These conformations are generated using box centers set to the average coordinates of the natural ligands and docked at a box size of 15 Å. The generated 179,878 docking conformations involve 18,197 PDB codes and three types of bioactivity assays (pIC50, pKd, and pKi). For each type of bioactivity assay, this disclosure, where possible, groups identical backbones within the same subset (i.e., training, validation, and test sets) based on the Bemis-Murcko backbone to evaluate the generalization ability of DTIGN. The three types of bioactivity assays used to train DTIGN (pIC50, pKd, and pKi) comprised 56,725, 53,289, and 39,282 conformations re-connected to 5,756, 5,390, and 3,959 PDB codes, respectively. The three types of bioactivity assays used to test DTIGN comprised 9,975, 14,411, and 6,196 conformations re-connected to 1,011, 1,455, and 626 PDB codes, respectively.

[0247] Subsequently, DTIGN was trained and tested using these re-doped conformations and each type of bioactivity assay from the PDBBind v2020 dataset. To more realistically evaluate the attention generalization ability of DTIGN, this disclosure avoids using native structures to refine the model attention during training, considering the scarcity of native structures with bioactivity assays in practice. This disclosure then analyzes the attention distribution of the re-doped ligand-protein conformations on various RMSD rankings. Figure 13 As shown, external test conformations with high RMSD ranking (indicating class naturalness) typically receive more attention from DTIGN. The natural structures of these test samples were not involved in DTIGN training. Figure 14 and Figure 15Additional experiments showed that consistent results were observed across different bioactivity assays, demonstrating the robust generalization ability of DTIGN attention. However, this disclosure also noted that docking conformations with an average RMSD greater than 9 Å distracted some attention from the DTIGN model. Therefore, this disclosure does not recommend using docking software with generated docking conformations having an average RMSD greater than 9 Å for DTIGN learning.

[0248] D. Benchmarking

[0249] Figure 16 A table showing the best cross-validation Pearson r 1600 on the benchmark dataset is presented. Figure 17 A table showing the best cross-validation RMSE of 1700 on the benchmark dataset is presented. Figure 18 This demonstrates the best cross-validation on the benchmark dataset. The table is 1800.

[0250] In this section, this disclosure validates its method on benchmark datasets (such as...). Figure 9 Overall effectiveness across all tasks (as detailed herein). Follow Figure 9 The experimental setup outlined in the document, along with the performance of "DTIGN unsupervised" testing, are presented in... Figures 16 to 18 Benchmark results show that DTIGN outperforms the comparison methods on most datasets and achieves the best overall performance across all metrics. On all datasets, compared to the best-performing baseline method, r is increased by an average of 32.12%, RMSE is reduced by an average of 5.52%, and τ is increased by an average of 43.44%. Overall, the average improvement calculated by averaging the improvements across all the above metrics is 27.03%.

[0251] E. Attention-based docking conformation review

[0252] Figure 19 The binding conformation 1900 of the exemplary target protein CHEMBL202 with the test ligand CHEMBL7492 is shown in (A) the native conformation and (B) the AutoDock Vina docking structure in the pocket of interest (pocket 2) of this model. Non-covalent bonds (polar interactions) between the ligand and the protein are marked with dashed lines, and the binding sites are highlighted above the protein sequence. Common binding sites between the native and docking structures are circled with dashes. Figure 20The results show “CHEMBL7492” in (A) “1u72” and “CHEMBL22” in (B) “2w3a”, indicating the test ligand ID and its PDB ID relative to the protein’s native structure. This demonstrates that DTIGN identifies and learns some native docking conformations, showing an RMSD of less than 5 Å from the native structure. The correlation between RMSD and mean attention value during model training was assessed using the Pearson correlation coefficient (r).

[0253] The docking conformations that the model tends to focus on share some similarities with their native conformations: the accuracy of self-attention has been validated in Section II-C. Here, this disclosure uses the pocket backtracking molecular docking results of the model's attention. Taking CHEMBL202 and CHEMBL7492 with native structures as examples, this disclosure compares the docking results used for training. The conformations of AutoDock Vina docking in attention-focused pockets 1 and 2 are very similar to each other (one of which is in Figure 19 (Analysis in B). These conformations are located between two pockets and have four of the five binding sites between the ligand and the protein that are associated with its native structure (e.g., ...). Figure 19 (As shown in A) Shares, making their overall conformations similar. Focusing on these accurate docking conformations can improve DTIGN's prediction of biological activity.

[0254] DTIGN learns by identifying several natural-like docking conformations: the docking accuracy within the model's attention pocket was investigated by calculating the RMSD between the natural conformation and the docking conformation within each pocket. To eliminate coordinate bias, this disclosure aligns the docking results with the natural ligand-protein complex before RMSD calculation. Results show that docking conformations with high attention rank higher in the RMSD results (e.g., ...). Figure 20 As shown in the figure, this helps to extract more useful information about ligand-protein interactions.

[0255] III. Summary

[0256] This disclosure pushes the boundaries of bioactivity prediction using deep learning, leveraging molecular docking to create ligand-protein complex data and extract valuable insights from molecular interactions. This disclosure introduces DTIGN, which integrates intermolecular interactions, self-attention, and semi-supervised learning to enhance the accuracy of bioactivity prediction. DTIGN captures force field information by defining intermolecular interactions using higher-order negative powers of the distance in Coulomb and London forces. Its self-attention mechanism identifies several natural-like docking conformations for DTIGN to learn from. Furthermore, DTIGN can enhance its predictive performance using limited natural structure data through semi-supervised learning. Overall, as demonstrated by the benchmarks of this disclosure, DTIGN demonstrates superior performance compared to state-of-the-art bioactivity prediction methods.

[0257] While this disclosure has been shown and described with reference to specific embodiments, those skilled in the art will understand that various changes in form and detail may be made therein without departing from the spirit and scope of this disclosure as defined by the appended claims. Therefore, the scope of this disclosure is indicated by the appended claims, and all changes within the meaning and equivalents of the claims are thus intended to be covered.

Claims

1. A system for determining bioactivity parameters of a ligand, wherein the ligand binds to a protein in multiple docking conformations to form a ligand-protein complex, the system comprising a processor configured to: Obtain ligand-protein complex structure data indicating the structure of the ligand-protein complex and reference bioactivity data of the ligand, protein and / or ligand-protein complex; The ligand-protein complex structural data includes: multiple ligand nodes of the ligand, each ligand node having ligand properties; multiple protein nodes of the protein, each protein node having protein properties; and multiple non-covalent edges connecting the ligand nodes and the corresponding protein nodes, each non-covalent edge having non-covalent bond properties. For each of the plurality of docking conformations, ligand-protein complex interaction parameters are determined that indicate intermolecular interaction data of the ligand-protein complex under each docking conformation. Determining the ligand-protein complex interaction parameters includes: for each non-covalent edge, determining non-covalent parameters indicating intermolecular interaction data of the first ligand node and first protein nodes located adjacent to it and bound to the first ligand node via the first non-covalent edge, based on a first ligand property of the first ligand node, a first protein property of the first protein node, and a first ligand-protein non-covalent parameter indicating the binding stability of the dispersion force type of the first non-covalent edge; and determining the ligand-protein complex interaction parameters based on the aggregation of the non-covalent parameters of the plurality of non-covalent edges of the ligand-protein complex. The ligand-protein complex interaction parameters for polymerization, which indicate intermolecular interaction data of the polymerization of the ligand-protein complex in the plurality of docking conformations, are determined; and The bioactivity parameters are determined based on the interaction parameters of the polymerized ligand-protein complex.

2. The system according to claim 1, wherein the first ligand-protein non-covalent parameter is based on a predetermined ligand-protein vector, the ligand-protein vector representing the force field amplitude corresponding to the dispersion force type of the first non-covalent edge, as a function of the non-covalent bond property of the first non-covalent edge.

3. The system of claim 2, wherein the predetermined ligand-protein vector comprises a Coulomb ligand-protein vector, the Coulomb ligand-protein vector comprising a predetermined radial basis function (RBF) parameterized by a Coulomb power value, indicating the magnitude of the force field of the Coulomb dispersion force of the first non-covalent edge.

4. The system of claim 2, wherein the predetermined ligand-protein vector includes a London ligand-protein vector, the London ligand-protein vector including a predetermined RBF parameterized by a predetermined London power value, indicating the magnitude of the force field of the London dispersion force of the first non-covalent edge.

5. The system according to any one of claims 2 to 4, wherein the non-covalent bond property of the first non-covalent edge includes a ligand-protein distance property, the ligand-protein distance property representing the physical distance value between the first ligand node and the first protein node of the ligand-protein complex.

6. The system of claim 5, wherein the processor is further configured to determine the ligand-protein distance property by: The physical distance between the first ligand node and the first protein node was measured on the ligand-protein complex structural data. The physical distance value is compared with a predetermined distance threshold that indicates the existence of the first non-covalent edge; as well as The existence of the first non-covalent edge is determined by comparing the physical distance value with the predetermined distance threshold.

7. The system according to claim 1 or claim 2, wherein the ligand-protein complex structural data further includes a plurality of ligand edges in the ligand binding the plurality of ligand nodes together, each ligand edge having ligand covalent bond chemical properties and ligand covalent bond physical properties. To determine the ligand-protein complex interaction parameters, the processor is further configured to: For each of the plurality of ligand edges, ligand edge parameters are determined that indicate intramolecular interaction data of a first ligand node and a second ligand node located adjacent to it and bound to the first ligand node via the first ligand edge. These ligand edge parameters are determined based on: a first ligand property of the first ligand node, a second ligand property of the second ligand node, the ligand covalent bond chemical property of the first ligand edge, and a first ligand covalent parameter indicating the binding stability of the first ligand edge. Based on the aggregation of the ligand edge parameters of the multiple ligand edges, the interaction parameters of the ligand-protein complex are further determined.

8. The system according to claim 7, wherein the first ligand covalent parameter is based on a predetermined ligand vector, the ligand vector representing the distance attribute of the intramolecular covalent bond of the first ligand edge as a function of the physical property of the ligand covalent bond of the first ligand edge.

9. The system of claim 8, wherein the predetermined ligand vector includes a predetermined RBF parameterized by a first power value, indicating the distance property of the intramolecular covalent bond of the first ligand edge.

10. The system according to any one of claims 7 to 9, wherein the ligand covalent bond physical property of the first ligand edge includes a ligand distance property, the ligand distance property representing the physical distance value between the first ligand node and the second ligand node of the ligand.

11. The system according to claim 1 or claim 2, wherein the ligand-protein complex structural data further comprises a plurality of protein edges in the protein that bind the plurality of protein nodes to each other, each protein edge having protein covalent chemical properties and protein covalent physical properties. To determine the ligand-protein complex interaction parameters, the processor is further configured to: For each of the plurality of protein edges, protein edge parameters are determined that indicate intramolecular interaction data of a first protein node and a second protein node located adjacent to it and bound to the first protein node via the first protein edge. These protein edge parameters are determined based on: a first protein property of the first protein node, a second protein property of the second protein node, the protein covalent bond chemical property of the first protein edge, and a first protein covalent parameter indicating the binding stability of the first protein edge. Based on the aggregation of protein edge parameters of the multiple protein edges, the ligand-protein complex interaction parameters are further determined.

12. The system of claim 11, wherein the first protein covalent parameter is based on a predetermined protein vector, the protein vector representing the distance attribute of the intramolecular covalent bonds of the first protein edge as a function of the physical property of the protein covalent bonds of the first protein edge.

13. The system of claim 12, wherein the predetermined protein vector includes a predetermined RBF parameterized by a second power value, indicating the distance property of intramolecular covalent bonds of the first protein edge.

14. The system according to any one of claims 11 to 13, wherein the protein covalent bond physical property of the first protein edge includes a protein distance property, the protein distance property representing the physical distance value between the first protein node and the second protein node of the protein.

15. The system according to any one of claims 1 to 14, wherein the processor further comprises a first machine learning algorithm, the first machine learning algorithm optionally comprising a deep learning algorithm.

16. The system according to any one of claims 1 to 15, wherein the processor further comprises a second machine learning algorithm for determining the ligand-protein complex interaction parameters. The second machine learning algorithm optionally includes a multilayer perceptron.

17. The system according to any one of claims 1 to 16, wherein the processor further comprises a third machine learning algorithm for determining the ligand-protein complex interaction parameters of the polymerized protein. The third machine learning algorithm may optionally include a multi-head self-attention algorithm.

18. The system according to any one of claims 1 to 17, wherein the bioactivity parameter includes at least one of the following: The EC50 parameter indicating the half-maximum effective concentration of the ligand; and / or The IC50 parameter indicates the half-maximal inhibitory concentration of the ligand.

19. A drug discovery platform comprising the system according to any one of claims 1 to 18.

20. A method for determining bioactivity parameters of a ligand, wherein the ligand binds to a protein to form a ligand-protein complex, the ligand binding to the protein in a plurality of docking conformations, the method comprising providing a processor for: Obtain ligand-protein complex structural data indicating the structure of the ligand-protein complex and reference bioactivity data of the ligand, protein, and ligand-protein complex; The ligand-protein complex structural data includes: multiple ligand nodes of the ligand, each ligand node having ligand properties; multiple protein nodes, each protein node having protein properties; and multiple non-covalent edges connecting the ligand nodes and the corresponding protein nodes, each non-covalent edge having non-covalent bond properties. For each of the plurality of docking conformations, ligand-protein complex interaction parameters are determined that indicate intermolecular interaction data of the ligand-protein complex under each docking conformation. Determining the ligand-protein complex interaction parameters includes: for each non-covalent edge, determining non-covalent parameters indicating intermolecular interaction data of the first ligand node and first protein nodes located adjacent to it and bound to the first ligand node via the first non-covalent edge, based on a first ligand property of the first ligand node, a first protein property of the first protein node, and a first ligand-protein non-covalent parameter indicating the binding stability of the dispersion force type of the first non-covalent edge; and determining the ligand-protein complex interaction parameters based on the aggregation of the non-covalent parameters of the plurality of non-covalent edges of the ligand-protein complex. The ligand-protein complex interaction parameters for polymerization, which indicate intermolecular interaction data of the polymerization of the ligand-protein complex in the plurality of docking conformations, are determined; and The bioactivity parameters are determined based on the interaction parameters of the polymerized ligand-protein complex.

21. The method of claim 20, wherein the first ligand-protein non-covalent parameter is based on a predetermined ligand-protein vector, the ligand-protein vector representing the force field amplitude corresponding to the dispersion force type of the first non-covalent edge as a function of the non-covalent bond property of the first non-covalent edge.

22. The method of claim 21, wherein the predetermined ligand-protein vector comprises a Coulomb ligand-protein vector, the Coulomb ligand-protein vector comprising a predetermined RBF parameterized by a Coulomb power value indicating the magnitude of the force field of the Coulomb dispersion force of the first non-covalent edge.

23. The method of claim 21, wherein the predetermined ligand-protein vector includes a London ligand-protein vector, the London ligand-protein vector including a predetermined RBF parameterized by a predetermined London power value indicating the magnitude of the force field of the London dispersion force of the first non-covalent edge.

24. The method according to any one of claims 21 to 23, wherein the non-covalent bond property of the first non-covalent edge includes a ligand-protein distance property, the ligand-protein distance property representing the physical distance value between the first ligand node and the first protein node of the ligand-protein complex.

25. A computer-readable medium comprising instructions that, when executed by a processor, cause the processor to perform the method according to any one of claims 20 to 24.

Images