Molecular generation optimization method and system for fusion protein pocket sequence and structural information

Abstract

The invention discloses a fusion protein pocket sequence and structure information molecule generation optimization method and system. The method comprises the following steps: coding a protein pocket sequence based on a Transform encoder; representing a three-dimensional structure of a protein pocket by using a graph, introducing a random walk position code and a graph attention mechanism into a structure encoder, and carrying out protein pocket structure encoding; characterizing the small molecules using a self-referenced embedded character string (SELFIES); the sequence output and the structure output of a protein pocket are combined with small molecule SELFIES representation, after decoding is conducted through a Transform decoder, fusion output is obtained through an attention mechanism, and finally SELFIES characters are generated according to word list probability distribution of the fusion output; and optimizing the physicochemical properties of the generated molecules by using a strategy gradient method of reinforcement learning. Drug molecules with high binding affinity, rich diversity and excellent physicochemical properties can be generated, and the drug design performance is improved.

Description

Technical Field

[0001] The present invention relates to a molecular generation optimization method and system that integrates protein pocket sequence and structure information, and belongs to the technical fields of drug design and artificial intelligence. It is used to generate candidate drug molecules with high affinity for specific protein pockets and optimize the physicochemical properties of the generated molecules through reinforcement learning. Background Art

[0002] In the process of new drug design and discovery, screening or designing candidate small molecule drugs that bind to drug targets is crucial. Traditional computer-aided drug design (CADD) methods, such as high-throughput screening and virtual screening, although they can accelerate the drug discovery process to a certain extent, the chemical space they search is limited, and the discovered molecules lack novelty. In recent years, deep generative models have been widely used in the fields of computer vision and natural language processing, and artificial intelligence-assisted drug design (AIDD) has gradually become a research hotspot in the field of drug design.

[0003] Molecular generation models based on deep learning can be divided into two categories: ligand-based and protein pocket-based models. Ligand-based models generate new molecules different from existing molecules by learning the hidden chemical rules formed by the structures of existing small molecules, but they cannot guarantee that the generated molecules have good binding affinity for specific protein targets. Protein pocket-based models use various characterization methods to extract the sequence or structure information of protein pockets and can generate molecules with high affinity for specific protein targets. However, most of the existing protein pocket-based molecular generation models are only based on single information (sequence or structure) of protein pockets and cannot take into account both the sequence and structure information of protein pockets at the same time, resulting in poor performance of the generated molecules in terms of binding affinity and diversity.

[0004] Therefore, the present invention proposes a molecular generation optimization method that integrates protein pocket sequence and structure information. By combining the sequence information, structure information of protein pockets, and the self-referencing embedded string (SELFIES) characterization of small molecules, it generates candidate drug molecules with high affinity for specific protein pockets and optimizes the physicochemical properties of the generated molecules through reinforcement learning. Summary of the Invention

[0005] Aiming at the deficiency that existing molecular generation models cannot take into account both the sequence and structure information of protein pockets at the same time, the purpose of the present invention is to provide a molecular generation optimization method and system that integrates protein pocket sequence and structure information. By combining the sequence information, structure information of protein pockets, and the self-referencing embedded string (SELFIES) characterization of small molecules, it generates candidate drug molecules with high affinity for specific protein pockets and optimizes the physicochemical properties of the generated molecules through reinforcement learning.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A molecular generation optimization method integrating protein pocket sequence and structure information, comprising the following steps:

[0008] (1) Based on the Transformer encoder, perform protein pocket sequence encoding;

[0009] (2) Use a graph to represent the three-dimensional structure of the protein pocket, introduce random walk position encoding and graph attention mechanism into the structure encoder, and perform protein pocket structure encoding;

[0010] (3) Use self-referencing embedding strings (SELFIES) to represent small molecules;

[0011] (4) Combine the sequence output and structure output encoded by the protein pocket encoder with the small molecule SELFIES representation, decode through the Transformer decoder to obtain the sequence decoding output and structure decoding output, then use the attention mechanism to obtain the fusion output, and finally generate SELFIES characters according to the word list probability distribution of the fusion output;

[0012] (5) Use the policy gradient method of reinforcement learning to optimize the physicochemical properties of the generated molecules.

[0013] Further, in step (1), first perform embedding encoding on the input sequence to obtain an embedding encoding matrix, then add it to the position encoding matrix as the input of the Transformer encoder, and then, through 6 layers of Transformer encoder layers, obtain the protein pocket sequence encoding matrix.

[0014] Further, in step (2), represent the three-dimensional structure of the protein pocket p as a graph where represents the set of nodes, n i represents a single amino acid residue node, represents the three-dimensional coordinates of its α-carbon atom, and n is the number of nodes; represents the set of edges between nodes; in the graph the node n i is represented as a one-hot encoded vector x according to 20 amino acid types i , and the edge e ij is constructed by the k-nearest neighbor (KNN) algorithm; the k-nearest neighbor algorithm selects k nearest neighbor nodes of node i according to the Euclidean distance d ij =||p i -p j ||2 to construct the edge connection relationship, and represent it in the form of an adjacency matrix as the edge representation;

[0015] When performing structure encoding, random walk positional encoding is introduced; for any node i, the random walk positional encoding vector is defined by a random walk of k steps:

[0016]

[0017] where RW ii = AD -1 represents the random walk operator, A is the adjacency matrix of graph G, and D is the degree matrix; the degree matrix is a diagonal matrix, and the element value on the diagonal is the degree of the node, that is, the number of edges associated with the node; this is a low-complexity random walk matrix algorithm that only considers the transition probability from node i to itself;

[0018] In the 6-layer graph Transformer encoder layer, a graph attention mechanism is introduced; the next-layer representation of node i is calculated by aggregating the attention scores between it and all neighbor nodes j in the current layer; the attention matrix is expressed as:

[0019]

[0020] where, represents the query vector of node i, represents the key and value vectors of neighbor node j;

[0021] The structure encoder encodes the protein pocket structure representation to obtain a structure encoding matrix.

[0022] Furthermore, in step (3), self-referencing embedding strings (SELFIES) are used to represent small molecules to ensure that the generated molecules conform to chemical rules; after obtaining the SELFIES representation of the molecules, the SELFIES representation of the small molecules is encoded by constructing a vocabulary and using embedding encoding to obtain a small molecule SELFIES encoding matrix.

[0023] Furthermore, in step (4), the small molecule SELFIES encoding matrix is added to the small molecule sine-cosine positional encoding matrix. First, the addition result and the sequence encoding output are sent to the sequence decoder for decoding to obtain a sequence decoding output; then the addition result and the structure encoding output are sent to the structure decoder for decoding to obtain a structure decoding output; then the attention mechanism is used to fuse the sequence decoding output and the structure decoding output to obtain a normalized SELFIES character generation probability distribution of the fused output, and finally SELFIES characters are generated.

[0024] Furthermore, in step (5), the physicochemical properties of the generated molecules, including drug-likeness and synthesizability, are optimized using the policy gradient method; the reinforcement learning optimization objective is set as the sum of the drug-likeness score and the synthesizability score.

[0025] A system for implementing the molecular generation optimization method that fuses protein pocket sequence and structure information, comprising:

[0026] A protein pocket sequence encoding module that obtains a protein pocket sequence encoding based on a Transformer encoder;

[0027] A protein pocket structure encoding module that uses a graph to represent the three-dimensional structure of the protein pocket, introduces random walk position encoding and graph attention mechanism into the structure encoder, and obtains the encoding of the protein pocket structure;

[0028] A small molecule SELFIES encoding module that represents small molecules using self-referencing embedding strings (SELFIES) to obtain a small molecule SELFIES encoding matrix;

[0029] A fusion decoding module that combines the sequence output and structure output encoded by the protein pocket encoder with the small molecule SELFIES representation, decodes through a Transformer decoder to obtain a sequence decoding output and a structure decoding output, then uses an attention mechanism to obtain a fusion output, and finally generates SELFIES characters according to the vocabulary probability distribution of the fusion output;

[0030] A reinforcement learning module that optimizes the physicochemical properties of the generated molecules using the policy gradient method of reinforcement learning.

[0031] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0032] (1) The method proposed by the present invention can simultaneously fuse the sequence and structure information of the protein pocket, generate candidate drug molecules with high affinity for a specific protein pocket, and can improve the performance of drug design;

[0033] (2) By constructing a reinforcement learning module, this method can further optimize the physicochemical properties of the generated molecules, including drug-likeness (QED) and synthesizability (SA), to ensure that the generated molecules have excellent drug properties;

[0034] (3) This method can not only generate molecules with high drug-likeness and synthesizability, but also is superior to existing models in terms of the diversity of the generated molecules and the docking affinity score, and is applicable to a wide range of drug design scenarios. Description of the Drawings

[0035] Figure 1 The molecular generation model architecture diagram constructed for the embodiment;

[0036] Figure 2 The generated result diagram for the embodiment. Detailed implementation manners

[0037] The present invention will be further explained below in conjunction with the accompanying drawings.

[0038] A molecular generation optimization method that integrates protein pocket sequence and structure information of the present invention includes the following steps:

[0039] (1) Based on the standard Transformer encoder, perform protein pocket sequence encoding.

[0040] In some embodiments, first, the normalized representation vector s of the amino acid sequence of the protein pocket with length L norm is subjected to embedding encoding, transformed into an embedding encoding matrix of L×N, where N is the embedding encoding dimension. The obtained embedding representation is added to the position encoding matrix as the input of the Transformer encoder. Then, through 6 layers of Transformer encoder layers, a sequence encoding matrix with a final dimension of L×N is obtained:

[0041]

[0042] (2) Use a graph to represent the three-dimensional structure of the protein pocket, introduce random walk position encoding and graph attention mechanism into the structure encoder, and perform protein pocket structure encoding.

[0043] In some embodiments, the three-dimensional structure of the protein pocket p is represented as a graph represents the set of nodes, where n i represents a single amino acid residue node, represents the three-dimensional coordinates of its alpha carbon atom, and n is the number of nodes (n = L). ε = {e ij , i, j = 1, 2,..., n & i ≠ j} represents the set of edges between nodes. In the graph , the node n i is represented as a one-hot encoding vector x according to 20 amino acid types i , and the dimension of the node encoding matrix is L×20; the edge e ij is constructed by the k-nearest neighbor (KNN) algorithm. The k-nearest neighbor algorithm selects k nearest neighbor nodes of node i according to the Euclidean distance d ij = ||p i - p j ||2 to construct the edge connection relationship, and represents it in the form of an adjacency matrix as the edge characterization. The above steps construct the structural graph representation of the protein pocket p:

[0044] When performing structure encoding, random walk positional encoding is introduced. For any node i, the random walk positional encoding vector is defined by a random walk of k steps:

[0045]

[0046] where RW ii = AD -1 represents the random walk operator, A is the adjacency matrix of graph G, and D is the degree matrix. The degree matrix is a diagonal matrix, and the elements on the diagonal are the degrees of the nodes, that is, the number of edges associated with the node. This is a low-complexity random walk matrix algorithm that only considers the transition probability from node i to itself.

[0047] After linearly transforming the node encoding matrix and the positional encoding matrix, the unified dimension is L×N, and then the two matrices are added as the input of the graph Transformer encoder layer.

[0048] In the 6-layer graph Transformer encoder layer, a graph attention mechanism is introduced. The next-layer representation of node i is calculated by aggregating the attention scores between it and all neighbor nodes j in the current layer. The attention matrix is expressed as:

[0049]

[0050] where, represents the query vector of node i, represents the key and value vectors of neighbor node j. The graph Transformer encoder encodes the protein pocket structure representation to obtain a protein pocket structure encoding matrix with dimension L×N:

[0051]

[0052] (3) Represent small molecules using self-referencing embedding strings (SELFIES) to obtain a small molecule SELFIES encoding module. The SELFIES representation avoids generating molecules that do not conform to chemical rules by mapping each chemical symbol to a specific bracket structure, thus effectively solving the problem of generating unbalanced brackets or circular identifiers during the generation process.

[0053] In some embodiments, after converting the small molecule Mol object into a SMILES string, it is further converted into the SELFIES format, and then a vocabulary is constructed and the SELFIES string of length l is encoded using embedding encoding to obtain a small molecule SELFIES encoding matrix with dimension l×N:

[0054]

[0055] (4) Combine the sequence output and structure output encoded by the protein pocket encoder with the small molecule SELFIES representation, decode through the Transformer decoder to obtain the sequence decoding output and structure decoding output, then use the attention mechanism to obtain the fusion output, and finally generate SELFIES characters according to the vocabulary probability distribution of the fusion output.

[0056] In some embodiments, add the small molecule SELFIES encoding matrix and the position encoding matrix to obtain the small molecule input:

[0057]

[0058] where pe represents the Transformer sine-cosine position encoding matrix of the small molecule.

[0059] Send the small molecule input and the sequence encoding output into the sequence decoder for decoding to obtain the sequence decoding output:

[0060]

[0061] Send the small molecule input and the structure encoding output into the structure decoder for decoding to obtain the structure decoding output:

[0062]

[0063] The Transformer decoder layers of both the sequence decoder and the structure decoder are 6 layers.

[0064] Then use the attention mechanism to obtain the fusion representation of the sequence decoding output and the graph decoding output:

[0065]

[0066] Thus, obtain the fusion decoding output probability distribution z. The essence of the attention mechanism fusion is weighted summation. The advantage of doing this is to allow the model to give different weights to the sequence output and the graph output at different positions, so as to reflect the relative importance of a certain type of output. Finally, after linear layer transformation to output the dimension, and then use the softmax function to obtain the normalized SELFIES character generation probability distribution, that is, the distribution probability of the characters that may be generated at each position of the sequence.

[0067] output_prob = Softmax(Linear(z))

[0068] (5) Use the policy gradient method of reinforcement learning to optimize the physicochemical properties of the generated molecules.

[0069] In some embodiments, the policy gradient method is used to optimize the generated molecular properties. The optimization objective of reinforcement learning is set as the sum of the drug-likeness (QED) and the synthetic accessibility score (SA).

[0070] In the molecular generation task, the decoder can be regarded as the executor of the stochastic policy π θ , and the generation process of a molecule is a sampling process of a policy trajectory: S=(s0, s1, …, s T )~π θ , where T is the preset maximum generation length. The probability value of generating a small molecule S can be expressed as the negative log-likelihood of the generation conditional probability:

[0071]

[0072] where NLL represents the negative log-likelihood, and G(x t+1 |x 1:t ) represents the probability of sampling a character at time step t.

[0073] The loss function is expressed as:

[0074] loss = [NLL(S) Prior -R(S)-NLL(S) Agent 2

[0075] where the reward function R(S) is used to calculate the final reward after a trajectory, that is, a molecule S is completed, including the drug-likeness and synthetic accessibility scores of the sampled molecule. Prior represents the initial model without reinforcement learning training. Agent represents the agent model, which is the current optimal model obtained by training the initial model Prior through reinforcement learning.

[0076] The present invention also provides a system for implementing the above-mentioned molecular generation optimization method that fuses protein pocket sequence and structure information, including:

[0077] A protein pocket sequence encoding module, based on a standard Transformer encoder, to obtain the protein pocket sequence encoding;

[0078] A protein pocket structure encoding module, using a graph to represent the three-dimensional structure of the protein pocket, introducing random walk position encoding and graph attention mechanism into the structure encoder, to obtain the protein pocket structure encoding;

[0079] A small molecule SELFIES encoding module, using self-referential embedding strings (SELFIES) to represent small molecules;

[0080] ​The fusion decoding module combines the sequence output and structure output encoded by the protein pocket encoder with the small molecule SELFIES representation, decodes them through the Transformer decoder to obtain the sequence decoding output and structure decoding output, then uses the attention mechanism to obtain the fusion output, and finally generates SELFIES characters according to the vocabulary probability distribution of the fusion output;

[0081] The reinforcement learning module uses the policy gradient method of reinforcement learning to optimize the physicochemical properties of the generated molecules.

[0082] The present invention will be further described below in conjunction with embodiments.

[0083] Embodiment

[0084] (1) Collect data of protein pockets and corresponding drug ligand small molecules: The dataset used for training this model is Crossdocked2020, which is split by MMseqs2 with 30% sequence identity to obtain a training set consisting of 100,000 protein-ligand pairs and a test set consisting of 100 protein-ligand pairs. In the training set, through random splitting during training, 99,900 pairs of data are used as training pairs and 100 pairs of data are used as the validation set. In the test set, 100 proteins are used as test proteins for molecular generation experiments and the performance of the generated molecules is evaluated.

[0085] (2) Data preprocessing: Represent the amino acid sequence of the protein pocket as a one-dimensional real vector, and represent the three-dimensional structure of the protein pocket as a graph structure.

[0086] Taking the 20 types of amino acid residue types that make up the protein pocket as the vocabulary, represent the amino acid sequence of the protein pocket as a one-dimensional real vector. To normalize the length of all data, set the maximum length to the length of the longest sequence, and use the number 0 to fill in the positions at the end of the shorter sequences, so as to establish the sequence structure representation of the protein pocket:

[0087] s norm =(a1,a2,…,a i ,…,a n ,0,0,…,0),a i ∈{1,2,3,…,20)

[0088] Represent the three-dimensional structure of the protein pocket p as a graph represents the set of nodes, where n i represents a single amino acid residue node, represents the three-dimensional coordinates of its alpha carbon atom, and n is the number of nodes. ε={e ij , i,j=1,2,…,n&i≠j} represents the set of edges between nodes. In the graph nodes n iRepresented as a one-hot encoded vector x according to 20 amino acid types i , edge e ij is constructed by the k-nearest neighbor (KNN) algorithm. The k-nearest neighbor algorithm selects the k nearest neighbor nodes of node i according to the Euclidean distance d ij = ||p i - p j ||2 to construct the edge connection relationship, and represents it in the form of an adjacency matrix as the edge representation. The above steps construct the structural graph representation of the protein pocket p:

[0089] (3) Construct a molecular generation model according to the five modules included in the system of the present invention: Write the model code using the Pytorch framework; the constructed model includes a sequence encoding module, a structure encoding module, a small molecule encoding module, a fusion decoding module, and a reinforcement learning module. The model architecture is as Figure 1 shown.

[0090] (4) Model training, and fine-tune the model using the reinforcement learning method: When training the model, use the cross-entropy loss function to measure the error between the prediction result and the true result. And reasonable hyperparameter settings are made for the ordinary training process and the reinforcement learning fine-tuning process.

[0091] (5) Case study: Select the p21-activated kinase protein pocket (PID: 5I0B) in the test set as the protein pocket for the case study. After running the molecular generation model, select the 5 molecules with the highest pocket molecule affinity score (Vina Score) generated by it as the model generation results. Figure 2 is the graph of the model generation results for the case study. The test set reference molecule is located in the leftmost column, and the 5 generated molecules are arranged in the remaining columns from high to low according to the Vina docking score. To comprehensively evaluate the properties of the generated molecules, the druggability score (QED), synthesizability score (SA), Lipinski-like drug comprehensive score (Lin), and docking binding affinity score (Vina) of each molecule are also marked in the figure.

[0092] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can still be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A molecular generation optimization method that integrates protein pocket sequence and structural information, characterized in that: It includes the following steps: (1) Based on the Transformer encoder, perform protein pocket sequence encoding; (2) Use a graph to represent the three-dimensional structure of the protein pocket, introduce random walk position encoding and graph attention mechanism into the structure encoder, and perform protein pocket structure encoding; (3) Use self-referencing embedding strings (SELFIES) to represent small molecules; (4) Combine the sequence output and structure output encoded by the protein pocket encoder with the small molecule SELFIES representation, decode through the Transformer decoder to obtain the sequence decoding output and structure decoding output, then use the attention mechanism to obtain the fusion output, and finally generate SELFIES characters according to the vocabulary probability distribution of the fusion output; (5) Use the policy gradient method of reinforcement learning to optimize the physicochemical properties of the generated molecules.

2. The molecular generation optimization method integrating the pocket sequence and structure information of the fusion protein, characterized in that: In step (1), first perform embedding encoding on the input sequence to obtain an embedding encoding matrix, then add it to the position encoding matrix as the input of the Transformer encoder, and then, through 6 layers of Transformer encoder layers, obtain the protein pocket sequence encoding matrix.

3. The molecular generation optimization method integrating the pocket sequence and structural information of a fusion protein, characterized in that: In step (2), the three-dimensional structure of the protein pocket p is represented as Figure where represents a set of nodes, n i represents a single amino acid residue node, represents the three-dimensional coordinates of its alpha carbon atom, and n is the number of nodes; represents a set of edges between nodes; in the graph nodes n i are represented as a one-hot encoded vector x according to 20 amino acid types i , and edges e ij are constructed by the k-nearest neighbor (KNN) algorithm; the KNN algorithm selects k nearest neighbor nodes of node i according to the Euclidean distance d ij = ||p i - p j ||2 to construct edge connection relationships and represent them in the form of an adjacency matrix as edge representations; When performing structural encoding, random walk positional encoding is introduced; for any node i, the random walk positional encoding vector is defined by a random walk of k steps: where RW ii = AD -1 represents the random walk operator, A is the adjacency matrix of graph G, and D is the degree matrix; In the 6-layer graph Transformer encoder layer, a graph attention mechanism is introduced; the next-layer representation of node i is calculated by aggregating the attention scores between it and all neighbor nodes j in the current layer; the attention matrix is represented as: Among them, represents the query vector of node i, represents the key and value vectors of neighbor node j; The structure encoder encodes the protein pocket structure representation to obtain a structure encoding matrix.

4. The molecular generation optimization method integrating the pocket sequence and structural information of a fusion protein, characterized in that: In step (3), use self-referencing embedding strings (SELFIES) to represent small molecules to ensure that the generated molecules conform to chemical rules; after obtaining the SELFIES representation of the molecules, encode the SELFIES representation of the small molecules by constructing a vocabulary and using the embedding encoding method to obtain the small molecule SELFIES encoding matrix.

5. The molecular generation optimization method integrating protein pocket sequence and structure information according to claim 1, characterized in that: In step (4), add the small molecule SELFIES encoding matrix to the small molecule sine-cosine position encoding matrix. First, send the addition result and the sequence encoding output to the sequence decoder for decoding to obtain the sequence decoding output; Then send the addition result and the structure encoding output to the structure decoder for decoding to obtain the structure decoding output; Then use the attention mechanism to fuse the sequence decoding output and the structure decoding output to obtain the normalized SELFIES character generation probability distribution of the fusion output, and finally generate SELFIES characters.

6. The molecular generation optimization method integrating protein pocket sequence and structure information according to claim 1, characterized in that: In step (5), use the policy gradient method to optimize the physicochemical properties of the generated molecules, including drug-likeness and synthesizability; The reinforcement learning optimization goal is set as the sum of the drug-likeness score and the synthesizability score.

7. A system for implementing the molecular generation optimization method of fusing the pocket sequence and structural information of the fusion protein described in any one of claims 1 to 6, characterized in that: It includes: A protein pocket sequence encoding module, based on the Transformer encoder, to obtain the protein pocket sequence encoding; A protein pocket structure encoding module, using a graph to represent the three-dimensional structure of the protein pocket, introducing random walk position encoding and graph attention mechanism into the structure encoder, to obtain the encoding of the protein pocket structure; A small molecule SELFIES encoding module, using SELFIES to represent small molecules, to obtain the small molecule SELFIES encoding; The fusion decoding module combines the sequence output and the structure output encoded by the protein pocket encoder with the small molecule SELFIES representation, decodes them through the Transformer decoder to obtain the sequence decoding output and the structure decoding output, then uses the attention mechanism to obtain the fusion output, and finally generates SELFIES characters according to the vocabulary probability distribution of the fusion output; The reinforcement learning module uses the policy gradient method of reinforcement learning to optimize the physicochemical properties of the generated molecules.

Images