A drug virtual screening method based on hash learning
By mapping proteins and molecules to binary codes using a hash-based learning method, the problem of low efficiency in virtual drug screening in large-scale molecular databases is solved, achieving high-precision and low-storage-overhead virtual drug screening.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV
- Filing Date
- 2024-07-24
- Publication Date
- 2026-07-14
AI Technical Summary
Existing virtual drug screening methods are inefficient in large-scale molecular databases, and cannot simultaneously meet the requirements of high accuracy and low storage overhead.
A hash-based learning approach is used to map proteins and molecules to binary codes. A multimodal hashing strategy is used for virtual drug screening. The hash-based learning method is used to train protein and molecule encoders to generate a binary vector database for fast retrieval.
Significantly reduces storage overhead, improves retrieval speed, and enhances the accuracy of virtual drug screening.
Smart Images
Figure CN118866167B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a drug virtual screening method based on hash learning, which is a method for rapidly screening candidate drug molecules in a molecular database based on protein targets in a computer. It is particularly suitable for scenarios involving efficient virtual drug screening in large-scale molecular databases and belongs to the fields of bioinformatics and computer science. Background Technology
[0002] Drug discovery is a lengthy and costly process because candidate drug molecules often have a high failure rate. Virtual drug screening is a crucial step in computer-aided drug discovery, enabling the identification of molecules that can bind to specific protein targets within a computer. Effective virtual drug screening can uncover more reliable candidate drug molecules, thereby reducing the time and resource costs required for drug discovery.
[0003] An effective virtual drug screening typically requires a large-scale molecular database and a highly accurate virtual screening method. Numerous large-scale molecular databases have been developed for drug discovery, such as ZINC and the Enamine REAL database. These existing molecular databases are already in the billions of entries, so there is no shortage of large-scale databases required for virtual drug screening in real-world applications.
[0004] However, existing virtual drug screening methods cannot meet the needs of practical applications. Traditional virtual drug screening methods, such as molecular docking, are too time-consuming to be used on large-scale databases. Several learning-based methods have been proposed to complete virtual drug screening more quickly; these can be further divided into supervised and unsupervised methods. Supervised methods are limited by the scale of training data, and their accuracy cannot exceed that of traditional methods, resulting in a high number of false positives in the screening results. Unsupervised methods treat virtual drug screening as a process of retrieving the vector representation of a molecule from the vector representation of a protein. Although they have higher accuracy, pre-storing a large number of real-valued vectors incurs huge storage costs, and the calculation of similarity between vectors and the ranking of results also incur significant computational costs. Summary of the Invention
[0005] Objective: To address the problems and shortcomings of existing technologies, this invention provides a drug virtual screening method based on hash learning to solve the problem of performing drug virtual screening in large-scale molecular databases. This invention treats drug virtual screening as a retrieval task using binary vectors, employing a multimodal hashing strategy to simultaneously learn binary vectors for proteins and molecules. This invention not only improves retrieval accuracy but also significantly reduces storage overhead and increases retrieval speed.
[0006] Technical solution: A drug virtual screening method based on hash learning, including the steps of training using hash learning and searching a molecular database based on protein targets.
[0007] The specific steps for training using the hash learning method are as follows:
[0008] Step 100: Input the protein-molecular complex dataset;
[0009] Step 101, define the protein and molecular encoder;
[0010] Step 102: Define the contrastive learning objective function to learn the similarity information between proteins and molecules;
[0011] Step 103: Define the multimodal hash objective function, which is a binary vector learned by proteins and molecules;
[0012] Step 104: Use the contrastive learning and multimodal hashing objective function to form the final loss function to train the model, output and save the model;
[0013] Step 105: Based on the trained molecular encoder, the molecules in the molecular database are represented as binary vectors to obtain a binary vector database.
[0014] The specific steps for searching molecular databases based on protein targets are as follows:
[0015] Step 200: Use the trained protein encoder to represent the protein target as a binary vector;
[0016] Step 201: Calculate the Hamming distance between the binary vector of the protein target and the binary vector of the molecules in the molecular database, or construct an inverted index through the binary vector database to retrieve similar molecules;
[0017] Step 202: Select molecules in a set proportion as candidate drug molecules according to actual needs.
[0018] The protein-molecule complex in step 100 refers to a complex formed by the interaction of proteins and small molecules.
[0019] Step 101 defines the protein and molecular encoder, which refers to the process of constructing and updating initial representations based on the atoms that make up proteins and molecules. For the l-th layer of the encoder, the protein and molecular representations each contain atomic-level representations a. l And the representation q at the atom pair level l The vector representation of the i-th atom is as follows: The representation between atom i and atom j is as follows First, obtain the vector representation of the i-th atom based on the atom type. And obtain the representation between atoms i and j based on the distance between the atom pairs.
[0020] According to the representation of the l-th layer a l and q l For the (l+1)th layer, a is represented as l+1 and q l+1 The update process is as follows
[0021]
[0022] Where H is the number of attention heads, h is used to index the attention heads, d is the hidden layer dimension of K, and l represents the current layer number. Q, K, and V are the Query, Key, and Value in the attention mechanism, respectively. Specifically, Q... l,h K l,h V l,h These refer to a in the l-th layer. l The matrix obtained after different linear mappings under the h-th attention head, a l+1,h This represents the atomic representation of the (l+1)th layer under the h-th attention head.
[0023] Finally, we will analyze a under different attentional perspectives. l+1,h The data is concatenated and then linearly mapped to obtain the input a for the next layer. l+1 The above update process will be performed L times, with the initial representation being a. 0 and q 0 It will be updated to a L and q L For a L After dimensionality reduction and linear mapping, the final vector representation y of proteins and molecules is obtained. p and y m .
[0024] In step 102, the contrastive learning objective function is defined as follows:
[0025]
[0026] Where τ is the temperature coefficient, and These are the representations of the k-th protein and molecule obtained after passing through the protein and molecule encoder, respectively. n represents the number of protein-molecule pairs, and log represents the logarithmic function with the natural constant e as the base.
[0027] In step 103, the multimodal hash objective function is defined as follows:
[0028]
[0029] in sign(·) is the sign function.
[0030] The final loss function in step 104 is:
[0031]
[0032] Where λ is the hyperparameter balancing the two losses.
[0033] In step 105, representing the molecule as a binary vector means using a molecular encoder to encode the molecule into a real-valued vector, and then obtaining the binary vector through the sign(·) function.
[0034] In step 200, representing the protein target as a binary vector means using a protein encoder to encode the protein into a real-valued vector, and then obtaining the binary vector through the sign(·) function.
[0035] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the hash-based virtual drug screening method as described above.
[0036] A computer-readable storage medium having a computer program / instructions thereon that, when executed by a processor, implements the steps of the hash-learning-based virtual drug screening method as described above.
[0037] Beneficial effects: Compared with existing technologies, the drug virtual screening method based on hash learning provided by this invention utilizes hash learning to map proteins and molecules into binary codes, significantly reducing storage overhead and improving retrieval speed. Hash learning also helps reduce model overfitting, thereby improving retrieval accuracy. Attached Figure Description
[0038] Figure 1 This is a flowchart illustrating the training process of the hash-learning-based virtual drug screening method according to an embodiment of the present invention.
[0039] Figure 2 This is a flowchart illustrating the prediction process of the drug virtual screening method based on hash learning, as described in an embodiment of the present invention. Detailed Implementation
[0040] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.
[0041] The training workflow of the hash learning-based virtual drug screening method is as follows: Figure 1As shown in the diagram. First, input the protein-molecule complex dataset (Step 10). During training, sample several protein-molecule complexes from the dataset each time (Step 11). Then, encode the proteins and molecules using protein and molecular encoders respectively (Step 12). Define a contrastive learning objective function to learn the similarity information between proteins and molecules (Step 13). Define a multimodal hash objective function to learn the binary vectors of proteins and molecules (Step 14). Combine the contrastive learning and multimodal hash objective functions to form the final loss function and train the model (Step 15). After each iteration, determine whether the training stopping condition has been met (Step 16). If the stopping condition has not been met, return to Step 11 to sample a new round of data; otherwise, output the training results and save the model (Step 17). Based on the molecular encoder in the trained model, represent the molecules in the molecular database as binary vectors and output the binary vector database (Step 18).
[0042] Protein-molecule complexes are complexes formed by the interaction of proteins and small molecules.
[0043] Encoding proteins and molecules using protein and molecular encoders refers to the process of constructing initial representations based on the atoms that make up proteins and molecules, and then updating these representations. The initial representations of proteins and molecules contain atomic-level representations. 0 And the representation q at the atom pair level 0 Specifically, the vector representation of the i-th atom is obtained based on the atom type. And obtain the representation between atoms i and j based on the distance between the atom pairs.
[0044] The process of updating a and q is as follows
[0045]
[0046] Where H is the number of attention heads, h is used to index the attention heads, d is the hidden layer dimension of K, and l represents the current layer number. Q, K, and V are the Query, Value, and Key in the attention mechanism, respectively. Specifically, Q... l,h K l,h V l,h These refer to a in the l-th layer. l The matrix obtained after different linear mappings under the h-th attention head.
[0047] Finally, we will analyze a under different attentional perspectives. l+1,h The data is concatenated and then linearly mapped to obtain the input a for the next layer. l+1 The above update process will be performed L times, with the initial representation being a. 0 and q 0 It will be updated to a Land q L For a L After dimensionality reduction and linear mapping, the final vector representation y of proteins and molecules is obtained. p and y m .
[0048] Define the contrastive learning objective function as follows:
[0049]
[0050] Where τ is the temperature coefficient, and These are the representations of the k-th protein and molecule after passing through the protein and molecule encoder, respectively, where n represents the number of protein-molecule pairs.
[0051] Define the multimodal hash objective function as follows:
[0052]
[0053] in sign(·) is the sign function.
[0054] The final loss function is
[0055]
[0056] Where λ is the hyperparameter balancing the two losses.
[0057] Representing molecules in a molecular database as binary vectors means using a molecular encoder to encode the molecules into real-valued vectors, and then obtaining the binary vectors through the sign(·) function.
[0058] The workflow for using a trained model to predict molecules in a molecular database most likely to bind to protein targets is as follows: Figure 2 As shown in the diagram, the process begins by reading the trained model and binary vector database (step 20), followed by reading the protein targets (step 21). The protein encoder in the model is then used to represent the protein targets as binary vectors (step 22). The Hamming distance between the binary vectors of the protein targets and those in the molecular database is calculated, or an inverted index is constructed from the binary vector database to retrieve the most likely binding molecules (step 23). Based on actual needs, a certain proportion of molecules are selected as candidate drug molecules (step 24). Finally, the candidate molecules are output as the results of the virtual drug screening.
[0059] Representing a protein target as a binary vector means using a protein encoder to encode the protein into a real-valued vector, and then obtaining the binary vector through the sign(·) function.
[0060] Obviously, those skilled in the art should understand that the steps of the hash-learning-based virtual drug screening method described in the above embodiments of the present invention can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using device-executable program code, thereby storing them in a storage device for execution by the computing device. Furthermore, in some cases, the steps shown or described can be performed in a different order than presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of the present invention are not limited to any particular hardware and software combination.
[0061] This invention was tested on two datasets, comparing the performance of the proposed method with the best existing method. The first dataset is the DUD-E dataset, containing 102 protein targets, 22,886 bindable molecules, and approximately 1 million non-binding molecules. The second dataset is the LIT-PCBA dataset, containing 15 protein targets, 7,844 bindable molecules, and approximately 400,000 non-binding molecules. BEDROC, EF05, and EF1 are commonly used metrics for evaluating the accuracy of virtual drug screening. Experimental results show that the proposed method outperforms the best method on both the DUD-E and LIT-PCBA datasets, particularly on BEDROC, EF05, and EF01. Storage space experiments show that the proposed method reduces storage space by approximately 32 times compared to the current best method. In big data scenarios, the proposed method improves retrieval speed by at least 3.5 times compared to real-valued vector methods.
Claims
1. A drug virtual screening method based on hash learning, characterized in that, This includes the steps of training using hash learning methods and the steps of searching molecular databases based on protein targets; The specific steps for training using the hash learning method are as follows: Step 100: Input the protein-molecular complex dataset; Step 101, define the protein and molecular encoder; Step 102: Define the contrastive learning objective function to learn the similarity information between proteins and molecules; Step 103: Define the multimodal hash objective function, which is a binary vector learned by proteins and molecules; Step 104: Use the contrastive learning and multimodal hashing objective function to form the final loss function to train the model, output and save the model; Step 105: Based on the trained molecular encoder, the molecules in the molecular database are represented as binary vectors to obtain a binary vector database. The specific steps for searching molecular databases based on protein targets are as follows: Step 200: Use the trained protein encoder to represent the protein target as a binary vector; Step 201: Calculate the Hamming distance between the binary vector of the protein target and the binary vector of the molecules in the molecular database, or construct an inverted index through the binary vector database to retrieve similar molecules; Step 202: Select molecules in a set proportion as candidate drug molecules according to actual needs; In step 103, the multimodal hash objective function is defined as follows: in , It is a symbolic function; The final loss function in step 104 is: in To balance the hyperparameters of the two losses.
2. The drug virtual screening method based on hash learning according to claim 1, characterized in that, The protein-molecule complex in step 100 refers to a complex formed by the interaction of proteins and small molecules.
3. The drug virtual screening method based on hash learning according to claim 1, characterized in that, Step 101 defines the protein and molecular encoder, which refers to the process of constructing an initial representation based on the atoms that make up proteins and molecules, and then updating it; for the encoder's... The representations of layers, proteins, and molecules each contain atomic-level representations. and the representation of atomic pairs Among them, the first The vector representation of atoms is as follows ,atom and atoms The representation between them is First, obtain the first [atom type]. Vector representation of atoms And to obtain atoms based on the distance between atomic pairs and Representation between ; According to the Layer representation and For the first Layer representation and The update process is as follows in It's about the number of heads to focus on. Used for indexing attention heads yes The hidden dimension, Represents the current floor number; These refer to the Query, Key, and Value in the attention mechanism. Specifically... They refer to After the first The matrix obtained after different linear mappings under each attention head. Indicates the first The first thing to pay attention to. Atomic representation of layers; Finally, different attention levels will be considered. The inputs are concatenated and then linearly mapped to obtain the input for the next layer. The above update process will be performed. Next, initial representation and It will be updated to and ;right After dimensionality reduction and linear mapping, the final vector representations of proteins and molecules are obtained. and .
4. The drug virtual screening method based on hash learning according to claim 1, characterized in that, The final loss function in step 104 is: in To balance the hyperparameters of the two losses.
5. The drug virtual screening method based on hash learning according to claim 1, characterized in that, In step 105, representing the molecule as a binary vector means using a molecular encoder to encode the molecule into a real-valued vector, and then... The function produces a binary vector.
6. The drug virtual screening method based on hash learning according to claim 1, characterized in that, In step 200, representing the protein target as a binary vector means using a protein encoder to encode the protein into a real-valued vector, and then... The function produces a binary vector.
7. A computer device, characterized in that, The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the drug virtual screening method based on hash learning as described in any one of claims 1-6.
8. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the drug virtual screening method based on hash learning as described in any one of claims 1-6.