Computer-aided drug design system and method based on deep learning and tree search

Abstract

This invention provides a computer-aided drug design system and method based on deep learning and tree search. The system includes: a receptor-ligand feature learning module, employing a hierarchical skip connection Transformer neural network to perform multi-level feature extraction and fusion of protein receptor residue sequences and ligand SMILES sequences. A Hidden Markov Model is established through supervised learning training to obtain a SMILES sequence growth strategy; and a tree search module, which uses the SMILES sequence growth strategy for heuristic searching. The tree search module is also guided by a value evaluation function, ultimately outputting novel SMILES sequences. For a given target, this invention can efficiently search a vast chemical space, generating potential ligand molecules with high affinity.

Description

Technical Field

[0001] This invention relates to the field of computer-aided drug design technology, and more specifically, to a computer-aided drug design system and method based on deep learning and tree search. Background Technology

[0002] The quality and quantity of medicines are crucial factors concerning human health. New drug development is characterized by high investment, long development cycles, and significant risks. Computer-aided drug design can help shorten the discovery cycle, reduce investment costs and risks, and is increasingly gaining attention from academia and industry.

[0003] In recent years, the rapid development of Artificial Intelligence (AI), represented by deep learning, and the accumulation of a large amount of known drug molecule structures and their corresponding pharmacological activities have provided new opportunities for computer-aided drug design. The continuous advancements in life sciences, pharmacy, chemistry, and medicine have led to the generation of abundant data on drug molecules and their corresponding pharmacological activities, genomics, proteomics, structures, and drug-target binding structures. For example, the public compound database PubChem records the chemical structures and biological properties of a large number of compounds, containing information on most drugs and targets, and can be used as an estimate of the distribution of rationally existing molecular structures in nature. The ChEMBL database collects bioactivity data on various targets and compounds from a large amount of literature. The DrugBank database provides detailed drug data and drug target information, including medicinal chemistry, pharmacology, pharmacokinetics, ADME and their interactions, related clinical trials, and drug toxicity data. The PDBBind database systematically collects three-dimensional structural information and affinity experimental data of various types of complexes from the Protein Data Bank (PDB) database, making it the most direct receptor-ligand binding database.

[0004] While the vast amount of drug-related biochemical experimental data provides unprecedented opportunities for training intelligent drug design models, managing, modeling, analyzing, and applying the storage, management, and heterogeneous unstructured chemical and biological data to drug design remains a challenging task. Artificial intelligence for big data learning, especially deep learning with its powerful data fitting and automatic feature extraction capabilities, is increasingly becoming a powerful tool for various applications, data analysis, and modeling, helping to efficiently mine and extract relevant patterns and information from drug target biochemical data. Recent work has focused on ab initio chemical simulations, predicting the physicochemical properties of drugs, and predicting the strength of drug-target binding. These works are mainly applied to virtual drug screening scenarios and are therefore limited by the high computational cost and limited complexity of compound libraries. Furthermore, some works use deep generative network models for molecular generation, such as differential autoencoders and generative adversarial networks, to generate molecular SMILES strings or molecular graphs. However, these methods neglect information about target proteins, making it difficult to generate candidate drug molecules with high target affinity. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the purpose of this invention is to provide a computer-aided drug design system and method based on deep learning and tree search.

[0006] This invention is achieved through the following technical solution:

[0007] According to one aspect of the present invention, a computer-aided drug design system based on deep learning and tree search is provided, the system comprising:

[0008] The receptor-ligand feature learning module is used to extract and fuse multi-level features of protein receptor residue sequences and ligand SMILES sequences using a hierarchical skip connection Transformer network. In a supervised learning manner, it is modeled as a Hidden Markov Model to obtain a preliminary SMILES sequence growth strategy.

[0009] The tree search module is used to take the growth strategy of the initial SMILES sequence as the initial information for adding candidate nodes, and then perform a heuristic search. The tree search module is also guided by the value evaluation function, and finally obtains a robust growth strategy for the SMILES sequence.

[0010] According to another aspect of the present invention, a computer-aided drug design method based on deep learning and tree search is provided, implemented based on the aforementioned computer-aided drug design system based on deep learning and tree search, the method comprising:

[0011] Obtain a protein receptor-small molecule ligand pairing dataset and divide the dataset into a training set, a validation set, and a test set;

[0012] The receptor-ligand feature learning module is trained based on the training set;

[0013] The receptor-ligand feature learning module is validated based on the validation set.

[0014] Protein receptor residue sequences are extracted from the test set and input into a trained receptor-ligand feature learning module. Combined with a tree search module, heuristic searches are continuously performed to ultimately generate novel SMILES sequences.

[0015] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0016] The computer-aided drug design system and method based on deep learning and tree search of the present invention improves the efficiency of neural networks in characterizing drug molecule-protein target interactions by using a hierarchical skip connection Transformer network; and by adding a value evaluation function to score, it effectively assists the tree search algorithm, which can efficiently generate potential ligand molecules with high affinity for a given target, greatly reducing the computational complexity caused by the huge chemical space, thereby accelerating the drug design and development process. Attached Figure Description

[0017] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0018] Figure 1 This is a schematic diagram of the receptor-ligand feature learning module in an embodiment of the present invention;

[0019] Figure 2 This is a schematic diagram of the internal encoding and decoding layers of the receptor-ligand feature learning module in an embodiment of the present invention;

[0020] Figure 3 This is a schematic diagram of the tree search module in an embodiment of the present invention;

[0021] Figure 4 This is a schematic diagram of the testing phase of the computer-aided drug design method based on deep learning and tree search in an embodiment of the present invention. Detailed Implementation

[0022] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0023] Current drug design technologies are limited by the high cost of virtual screening computations, the limited complexity of compound libraries, and the difficulty in generating candidate drug molecules with high target affinity. It is worth noting that artificial intelligence for search optimization, especially heuristic search methods supplemented by deep learning models, has demonstrated outstanding performance in many problems requiring large-scale sequential decision-making and complex planning decisions. Therefore, using deep learning network models to extract structural information of target proteins and drug molecules, combined with heuristic search algorithm models, holds great potential as a powerful tool for solving design optimization problems in drug molecule space.

[0024] To this end, embodiments of the present invention provide a computer-aided drug design system based on deep learning and tree search. The system includes: a receptor-ligand feature learning module, used to perform multi-level feature extraction and fusion of protein receptor residue sequences and existing ligand SMILES sequences using a hierarchical skip-connection Transformer network, aiming to learn the interaction patterns between small molecule ligands and protein receptor residue sequences and predict a preliminary SMILES sequence growth strategy; during the training phase, based on the cross-entropy loss function, the neural network parameters of the hierarchical skip-connection Transformer network in the receptor-ligand feature learning module are updated using backpropagation; a tree search module, used to take the preliminary SMILES sequence growth strategy as initial information for new candidate nodes, and then perform heuristic search to improve or correct the probability selection of the next atom generation. The tree search module is also guided by a value evaluation function such as molecular docking software scoring to improve the binding affinity of the generated small molecules to the protein receptor, ultimately obtaining a robust SMILES sequence growth strategy and outputting a novel ligand SMILES sequence.

[0025] To better extract the interaction features between protein receptors and small molecule ligands, the receptor-ligand feature learning module takes the protein receptor residue sequence and the known small molecule ligand (SMILES) sequence as input. It employs a hierarchical skip connection Transformer, aiming to implicitly extract the interaction features between the protein receptor and small molecule ligand within the network. Through supervised learning training, a Hidden Markov Model is established to obtain a preliminary SMILES sequence growth strategy. In some preferred embodiments, such as... Figure 1As shown, the hierarchical skip connection Transformer network includes a protein embedding representation layer, a protein position encoding layer, four encoding layers, a SMILES embedding representation layer, a SMILES position encoding layer, four decoding layers, a linear layer, and a softmax layer. The protein embedding representation layer includes a word vector layer for initial encoding of the protein receptor residue sequence, obtaining a preliminary protein residue semantic vector. The protein position encoding layer adds position encoding to the protein residue semantic vector. Since the multi-head attention mechanism in the Transformer does not record the positional information of the protein sequence, a position encoding layer is introduced to enable the Transformer to capture the positional information of the protein residues. The four encoding layers have identical structures and are stacked. The protein residue semantic vector with added position encoding is input to the bottommost encoding layer. Each encoding layer receives protein information vectors from the layers below as input, and then processes them. The protein information vector is then passed as output to the upper coding layer and the decoding layer at the same level. The SMILES embedding representation layer includes a word vector layer, which performs preliminary encoding on the ligand SMILES sequence to obtain a preliminary SMILES sequence semantic vector; a SMILES position encoding layer, which adds position encoding to the SMILES sequence semantic vector; four decoding layers have identical structures and are stacked on top of each other, with each decoding layer corresponding to one of the four coding layers; the SMILES sequence semantic vector with added position encoding is input to the bottom decoding layer. Each decoding layer receives the SMILES information vector from the lower layer and the protein information vector from the coding layer at the same level as input, and then passes the processed information vector as output to the upper decoding layer. The processed information vector serves as the Transformer for understanding the interaction between proteins and molecules through hierarchical skip connections, fusing information from the protein sequence and the SMILES sequence; for example... Figure 1 As shown, the first encoding layer is at the bottom, and then the information vector will pass through the previous encoding layer in turn, and finally reach the softmax layer. The linear layer and the softmax layer are used together to make appropriate adjustments to the information vector. The softmax layer outputs the information vector as the probability distribution of the SMILES sequence.

[0026] The input to the hierarchical skip connection Transformer network is the protein receptor residue sequence and a partial SMILES sequence. The output is the growth strategy of the next molecule in the SMILES sequence. That is, given the partially generated molecule, the output is the set of atoms that can be grown next.

[0027] Specifically, the embedding representation layer is a lookup table storing a fixed-size dictionary of embedding vectors. That is, given a number, the embedding layer returns the embedding vector corresponding to that number, and the embedding vector reflects the semantic relationships between the symbols represented by each number. Therefore, for protein receptor residue sequences and ligand SMILES sequences, preliminary encoding is performed using the protein embedding representation layer and the SMILES embedding representation layer, respectively, to obtain preliminary protein residue semantic vectors and SMILES sequence semantic vectors. Since the encoding and decoding layers themselves lack access to sequence order information, positional encoding needs to be added to the protein residue semantic vectors and SMILES sequence semantic vectors. Both the protein positional encoding layer and the SMILES positional encoding layer include a trigonometric cosine function calculation layer, using the following trigonometric function: where pos represents the position, i is the dimension, and d... model The dimension of the embedding vector:

[0028]

[0029]

[0030] The protein sequence semantic vector, with added positional information, is input to the bottommost coding layer. The four coding layers have identical structures and are stacked on top of each other. Each coding layer receives the information vector from the layer below as input, and then passes the processed information vector as output to the coding layer above and the decoding layer at the same level. Therefore, coding layers at different levels can extract protein sequence semantic information at different levels and pass it to decoding layers at different levels. The decoding layers thus receive more details about the receptor-ligand interaction pattern at different feature levels, which helps the model better learn the receptor-ligand interaction pattern at different levels. Figure 2 As shown, each encoding layer includes a multi-head self-attention layer, a first summation and normalization layer, and a first feedforward neural network layer. The multi-head self-attention layer in the encoding layer captures the global interaction forces of proteins. The first summation and normalization layer and the first feedforward neural network layer are commonly used layers in deep neural networks. Normalization accelerates model convergence, while the summation and first feedforward neural network layers increase the model's expressive power, enabling it to better fit the target distribution. Specifically, the self-attention block outputs a weighted sum of the query vector Q and the key-value vector pair (K,V) mapped to values.

[0031]

[0032] Multi-head self-attention layers employ multiple self-attention blocks to perform different linear mappings on multiple query vectors Q and key-value pairs (K,V), enabling the model to focus on key information about the interactions between symbols in the sequence from different chemical feature perspectives. The computational mechanism is as follows:

[0033] MultiHead(Q,K,V)=Concat(H1,…,H h W O

[0034]

[0035] The SMILES sequence semantic vector, with positional information appended, is input to the bottommost decoding layer. The four decoding layers have identical structures and are stacked on top of each other. Each decoding layer accepts the SMILES information vector from the layer below and the protein sequence information vector from the coding layer at the same level as input, and then passes the processed information vector as output to the decoding layer above. For example... Figure 2 As shown, each decoding layer includes a masked multi-head self-attention layer, a second summation and normalization layer, a multi-head cross-attention layer, and a second feedforward neural network layer. The masked multi-head self-attention layer captures the interactions between positions within the unmasked SMILES sequence, and the multi-head cross-attention layer captures the interactions between the protein sequence and the SMILES sequence. The remaining layers function identically to their coding layers; that is, the second summation and normalization layer functions the same as the first summation and normalization layer, and the second feedforward neural network layer functions the same as the first feedforward neural network layer. Each of the four decoding layers receives the outputs of the four coding layers at its corresponding level as input to the multi-head cross-attention layer. To prevent label leakage, i.e., to prevent symbolic information in later sequences from affecting earlier predictions, the multi-head self-attention layer of the decoding layer needs to mask the information of later sequences, calculated in the same way as the multi-head self-attention layer of the coding layer. The cross-attention layer of the decoding layer receives the protein sequence vector output from the coding layer at the same level as a key-value pair (K,V) and calculates the attention sum, calculated in the same way as the self-attention layer of the coding layer. The information output from the highest decoding layer passes through a linear layer and a softmax layer to obtain the probability distribution of the growth strategy for the next bit of the SMILES sequence.

[0036] To improve or correct the probability selection of the next atom generation through better look-ahead searching, in some preferred embodiments, the tree search module establishes connections between nodes in a tree structure, where each node is used to store the following information:

[0037] The initial node priority value, which is the growth probability of the corresponding sequence, represents the priority of selecting the corresponding symbol node and is output by the receptor-ligand feature learning module.

[0038] The number of visits represents the number of times the corresponding node has been visited during a search.

[0039] The average result information is obtained from the cumulative average value of binding affinity calculated by value assessment functions such as molecular docking software.

[0040] In some preferred embodiments, the tree search module repeats the following steps:

[0041] Selection: The optimal priority principle is used to visit child nodes, and the leaf node reached in the end is the selected node; specifically, the tree simulation adopts the optimal priority principle, that is, each level visits child data nodes according to the principle of high initial value of node, low number of visits, and high action reward value, and the final terminal node reached in the end is the selected data node.

[0042] Expand: Based on the calculation of the receptor-ligand feature learning module, initialize all legal symbol nodes under the existing SMILES sequence terminal symbol nodes, initialize the node priority to the probability prediction value output by the receptor-ligand feature learning module, and initialize the number of visits and the average returned information approximation value to 0;

[0043] Fast search: When a leaf node in the tree structure is accessed, a greedy strategy is used to quickly generate SMILES sequences until a complete SMILES sequence is generated. Then, a value evaluation function is used to evaluate the value of the complete SMILES sequence. Specifically, in each step of the fast search process, the receptor-ligand feature learning module is used to calculate the initial predicted probability value. Then, the selection is carried out according to the greedy method. The selection is continued until a terminal node is generated or the maximum tree depth is reached. At this point, the symbol string path from the starting root node to the terminal node constitutes a complete SMILES sequence. Then, a value evaluation function, such as the molecular docking software SMINA, is called to score the binding affinity as the value evaluation.

[0044] Feedback: Update data node information layer by layer upwards until the initial data node, i.e., the root node; specifically, the access count increments by 1, and the average value is taken after the cumulative evaluation of the average molecular docking score returned.

[0045] After repeating the above steps multiple times, the access frequency of each node is calculated and used as the final priority of the corresponding node; specifically, the final priority of selecting each symbol growth is calculated by dividing the access count of each child data node by the sum of the access counts of all child data nodes.

[0046] This invention also provides a computer-aided drug design method based on deep learning and tree search, implemented based on the computer-aided drug design system based on deep learning and tree search described in the above embodiments. The method includes:

[0047] S1. Obtain the protein receptor-small molecule ligand pairing dataset and divide the dataset into training set, validation set and test set;

[0048] S2. Training the receptor-ligand feature learning module based on the training set;

[0049] S3. Validating receptor-ligand feature learning module based on validation set;

[0050] S4. Extract protein receptor residue sequences from the test set and input them into the trained receptor-ligand feature learning module. Combined with the tree search module, continuously perform heuristic searches to generate a robust SMILES sequence growth strategy, i.e., a novel SMILES sequence.

[0051] In some preferred embodiments, the receptor-ligand feature learning module is trained based on a training set, including:

[0052] S21. Input the protein receptor residue sequences and SMILES sequences in the training set into the receptor-ligand feature learning module. Through supervised learning training, establish a hidden Markov model and learn the initial growth strategy of SMILES sequences.

[0053] S22. Based on the cross-entropy loss function, the model loss is calculated and backpropagated. Then, the neural network updates the parameters of the Transformer with hierarchical skip connections in the receptor-ligand feature learning module using gradient descent. Specifically, the neural network is trained using supervised learning. The cross-entropy loss is calculated based on the probability prediction of the SMILES sequence output by the module and the true label values. This loss is backpropagated, and then the neural network updates the parameters of the Transformer with hierarchical skip connections in the receptor-ligand feature learning module using gradient descent.

[0054] In some preferred embodiments, protein receptor residue sequences are extracted from the test set and input into a trained receptor-ligand feature learning module. Combined with a tree search module, a lookahead search is continuously performed, outputting the generated SMILES sequences, including:

[0055] S41. Input the protein receptor sequence and the existing SMILES sequence fragment (i.e., the generated partial ligand sequence fragment) into the receptor-ligand feature learning module. Each time the receptor-ligand feature learning module calculates, it can obtain the growth strategy for the next position of the SMILES sequence fragment.

[0056] S42. Assign the initial probability of the next chemical symbol prediction to the node priority initial value of the tree search module. Each growth strategy will create a new tree node and use the probability of its selection as the node priority initial value.

[0057] S43. The iterative execution tree search module performs a preset number of iterations. Preferably, the iterative execution tree search module performs 50 iterations, using the access frequency of the node as the growth strategy for the next node in the final SMILES sequence. At the same time, it obtains the node with the highest access frequency and adds its corresponding SMILES symbol to the end of the existing SMILES sequence.

[0058] S44. Add a SMILES symbol corresponding to the optimal growth strategy to the end of the existing SMILES sequence fragment; the hierarchical jump connection Transformer model adopts an autoregressive generation method to realize the continuous growth of the SMILES sequence and finally generate a complete new SMILES sequence.

[0059] First, let's explain the basic settings of this embodiment: The dataset is a protein receptor-small molecule ligand pair set, where each protein receptor-small molecule ligand pair is {τ,m}, and m is a SMILES string m1,…,m t , construct s t ={τ,m1,…,m t As input to the hierarchical skip connection Transformer network, m t+1 The next symbolic true label value is the true label value that the neural network should fit. The goal of this embodiment is to train a hierarchical skip connection Transformer that can learn the receptor-ligand interaction pattern, and combine it with a value evaluation function such as affinity scoring from molecular docking software, auxiliary tree search, and autoregressive generation of SMILES sequences, thereby generating potential ligand molecules with high affinity.

[0060] In this embodiment, the training phase of the model generally includes several iterations, which will be described in detail with one iteration as an example.

[0061] like Figure 3 As shown, taking Monte Carlo tree search as an example, the single iteration process of model training in this embodiment includes:

[0062] The protein receptor residue sequence τ is input into the protein embedding representation layer of the receptor-ligand feature learning module. Through information processing of the protein position coding layer and four coding layers, protein semantic representation encoding at different levels is obtained.

[0063] Given the existing molecular partial sequence m = {m1, ..., m} t The molecular SMILES embedding representation layer of the input receptor-ligand feature learning module is passed through the SMILES position encoding layer and four decoding layers. At the same time, the four decoding layers receive the protein semantic representation encoding of the corresponding level of the encoding layer. Finally, after passing through the linear layer and the softmax layer, the symbol probability distribution of the next position of the SMILES sequence is obtained.

[0064] Based on the predicted probability distribution of the next SMILES symbol and the actual m t+1 Calculate the cross-entropy loss function.

[0065] The obtained cross-entropy loss is used to update the neural network parameters based on the backpropagation algorithm.

[0066] like Figure 4 As shown, the testing phase of the model in this embodiment includes: inputting the protein receptor sequence and the already generated partial SMILES ligand sequence (initially the SMILES sequence is empty) into the receptor-ligand feature learning module to obtain the preliminary probability distribution of the next SMILES symbol; simultaneously initializing the corresponding nodes in the tree and using the corresponding probabilities as the initial priority values ​​of the nodes; after iteratively executing the tree search module a certain number of times, using the access frequency of each node as the growth strategy for the next SMILES symbol in the final SMILES sequence; adding the SMILES symbol with the highest access frequency to the end of the already generated SMILES ligand sequence; each time only one SMILES symbol is added, and by continuously repeating, that is, by autoregression, the entire SMILES sequence is generated until a complete SMILES sequence is generated.

[0067] The computer-aided drug design system and method based on deep learning and tree search in the above embodiments of the present invention improves the efficiency of neural networks in characterizing drug molecule-protein target interactions by using hierarchical skip connection Transformer networks; and by adding value evaluation functions such as affinity scoring from molecular docking software, it effectively assists the tree search algorithm. For a given target, it can efficiently generate potential ligand molecules with high affinity, greatly reducing the computational complexity caused by the huge chemical space, thereby accelerating the drug design and development process.

[0068] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention. The above preferred features can be used in any combination without conflict.

Claims

1. A computer-aided drug design system based on deep learning and tree search, characterized by, include: The receptor-ligand feature learning module is used to extract and fuse multi-level features of protein receptor residue sequences and ligand SMILES sequences using a hierarchical skip connection Transformer network. In a supervised learning manner, it is modeled as a Hidden Markov Model to obtain a preliminary SMILES sequence growth strategy. The tree search module is used to take the growth strategy of the initial SMILES sequence as the initial information for adding candidate nodes, and then perform a heuristic search. The tree search module is also guided by the value evaluation function, and finally obtains a robust growth strategy for the SMILES sequence. The hierarchical jump connection Transformer network includes: The protein embedding representation layer, including the word vector layer, is used to initially encode the protein receptor residue sequence to obtain a preliminary protein residue semantic vector. The protein position encoding layer is used to add position encoding to the semantic vector of protein residues; Four coding layers, all of which have the same structure and are stacked on top of each other; plus a position-encoded protein residue semantic vector as input to the bottom coding layer, each coding layer receives the protein information vector from the layer below as input, and then passes the processed protein information vector as output to the coding layer above and the decoding layer in the same layer. The SMILES embedding representation layer, including the word vector layer, is used to initially encode the ligand SMILES sequence to obtain the initial SMILES sequence semantic vector. The SMILES position encoding layer is used to add position encoding to the semantic vector of the SMILES sequence. There are four decoding layers, all of which have the same structure and are stacked on top of each other. Each of the four decoding layers corresponds one-to-one with the four coding layers. The bottom decoding layer is input with a position-encoded SMILES sequence semantic vector. Each decoding layer accepts the SMILES information vector from the lower layer and the protein information vector from the coding layer of the same layer as input, and then passes the processed information vector as output to the upper decoding layer. The linear layer and the softmax layer are used to adjust the information vector, and the softmax layer outputs the information vector as a probability distribution of the SMILES sequence.

2. The computer-aided drug design system based on deep learning and tree search according to claim 1, characterized in that, The encoding layer includes a multi-head self-attention layer, a first summation and normalization layer, and a first feedforward neural network layer. The multi-head self-attention layer is used to capture the global interaction forces of proteins. The normalization of the first summation and normalization layer is used to accelerate model convergence. The summation of the first summation and normalization layer and the first feedforward neural network layer are used to increase the expressive power of the model.

3. The computer-aided drug design system based on deep learning and tree search according to claim 1, characterized in that, The decoding layer includes a masked multi-head self-attention layer, a second summation and normalization layer, a multi-head cross-attention layer, and a second feedforward neural network layer; the four decoding layers respectively receive the output of the corresponding level of the encoding layer as the input of the multi-head cross-attention layer; The masked multi-head self-attention layer is used to capture the interaction forces between various positions within the unmasked SMILES sequence. The multi-head cross-attention layer is used to capture the interaction forces between the protein sequence and the SMILES sequence. The normalization of the second summation and normalization layer is used to accelerate model convergence. The summation of the second summation and normalization layer and the second feedforward neural network layer are used to increase the model's expressive power.

4. The computer-aided drug design system based on deep learning and tree search according to claim 1, characterized in that, Both the protein position coding layer and the SMILES position coding layer include a trigonometric cosine function calculation layer.

5. The computer-aided drug design system based on deep learning and tree search according to claim 1, characterized in that, The tree search module establishes connections between nodes in a tree structure, where each node stores the following information: The initial node priority value represents the priority of selecting the corresponding node and is obtained from the output of the receptor-ligand feature learning module. The number of visits represents the number of times the corresponding node has been visited during a search. The average result information is obtained by calculating the cumulative average of binding affinity using the value assessment function.

6. The computer-aided drug design system based on deep learning and tree search according to claim 5, characterized in that, The tree search module repeats the following four steps: Selection: The best-priority principle is used to select child nodes, and the final leaf node reached is the selected node; Expand: Based on the calculation of the receptor-ligand feature learning module, initialize all legal symbol nodes under the terminal symbol node of the SMILES sequence, initialize the node priority to the probability prediction value of the growth strategy output by the receptor-ligand feature learning module, and initialize the number of visits and the average return information approximation value to 0. Fast search: When a leaf node in the tree is visited, a greedy strategy is used to quickly generate the SMILES sequence until a complete SMILES sequence is generated. Then, the value evaluation function is used to evaluate the value of the complete SMILES sequence. Postback: Update data node information layer by layer upwards until the root node; After repeating the above steps multiple times, calculate the access frequency of each node and use it as the final priority of the corresponding node.

7. A computer-aided drug design method based on deep learning and tree search, implemented based on the computer-aided drug design system based on deep learning and tree search as described in any one of claims 1-6, characterized in that, include: Obtain a protein receptor-small molecule ligand pairing dataset and divide the dataset into a training set, a validation set, and a test set; The receptor-ligand feature learning module is trained based on the training set; The receptor-ligand feature learning module is validated based on the validation set. Protein receptor residue sequences are extracted from the test set and input into a trained receptor-ligand feature learning module. Combined with a tree search module, heuristic searches are continuously performed to ultimately generate novel SMILES sequences.

8. The computer-aided drug design method based on deep learning and tree search according to claim 7, characterized in that, The process of training the receptor-ligand feature learning module based on the training set includes: The protein receptor residue sequences and SMILES sequences in the training set are input into the receptor-ligand feature learning module. A hidden Markov model is established through supervised learning training to learn the initial growth strategy of SMILES sequences. Based on the cross-entropy loss function, the model loss is calculated and backpropagated. Then, the neural network parameters of the Transformer in the hierarchical skip connections of the receptor-ligand feature learning module are updated by the neural network through gradient descent.

9. The computer-aided drug design method based on deep learning and tree search according to claim 7, characterized in that, The process involves extracting protein receptor residue sequences from the test set, inputting them into a trained receptor-ligand feature learning module, and combining this with a tree search module to continuously perform heuristic searches, ultimately generating novel SMILES sequences, including: The protein receptor sequence and the generated partial SMILES ligand sequence fragments are input into the receptor-ligand feature learning module. Each time the receptor-ligand feature learning module performs a calculation, it can obtain the growth strategy for the next position of the SMILES sequence fragment. Each growth strategy creates a new tree node, and the probability of its selection is used as the initial priority value for that node. The tree search module is iterated a preset number of times to obtain the growth strategy for the next position in the final SMILES sequence; A SMILES symbol corresponding to the optimal growth strategy is added to the end of the existing SMILES sequence fragment; the hierarchical jump connection Transformer model adopts an autoregressive generation method to achieve continuous growth of the SMILES sequence, and finally generate a complete new SMILES sequence.

Images