Method, device, equipment and storage medium for generating active polypeptide sequence

Abstract

The present invention relates to the technical field of polypeptide design, and discloses a method, device, equipment and storage medium for generating an active polypeptide sequence, the method comprising: inputting the sequence information of a polypeptide receptor and the generated polypeptide word-meta information into a generative deep learning model to generate new polypeptide word-meta information; splicing the polypeptide word-meta information generated by the generative deep learning model in a generation order to obtain the sequence information of a target polypeptide; obtaining the first geometric property of the target polypeptide and the second geometric property of the polypeptide receptor; inputting the first geometric property and the second geometric property, as well as the sequence information of the polypeptide receptor and the sequence information of the target polypeptide into a polypeptide activity prediction model to predict the binding activity of the target polypeptide and the polypeptide receptor. The present invention can significantly improve the efficiency and quality of active polypeptide design, thereby realizing high-throughput active polypeptide design.

Description

Technical Field

[0001] The present invention relates to the technical field of polypeptide design, and in particular to a method, device, equipment and storage medium for generating an active polypeptide sequence. Background Art

[0002] Peptides are chain molecules formed by amino acids connected by peptide bonds, and the number of amino acids in a peptide chain is usually less than about 50. Peptide drugs have some unique advantages. For example, peptides can target specific targets, such as receptors, enzymes, etc., and are highly selective; compared with large molecular antibodies, peptides are easier to chemically synthesize and modify; and the immunogenicity of peptides can be reduced through appropriate modification. Therefore, peptide drugs are a hot topic in drug research and development.

[0003] The process of peptide drug development is generally as follows: identify disease-related protein targets (such as receptors); design peptide sequences that can bind to the targets based on their structural information; prepare the designed peptides using solid-phase or multi-liquid-phase synthesis techniques; verify the activity of the peptides through experiments, and optimize their structures to improve their stability, specificity, and efficacy.

[0004] However, this peptide drug design method of verifying the activity of the peptide through experiments requires a lot of time to do experimental verification, resulting in low efficiency of peptide drug design. Summary of the invention

[0005] In view of this, the present invention provides a method, device, equipment and storage medium for generating an active polypeptide sequence to solve the problem of low efficiency in polypeptide drug design.

[0006] In a first aspect, the present invention provides a method for generating an active polypeptide sequence, the method comprising:

[0007] Input the sequence information of the peptide receptor and the generated peptide word information into the generative deep learning model to generate new peptide word information;

[0008] splicing the polypeptide word information generated by the generative deep learning model in a generation order to obtain sequence information of the target polypeptide;

[0009] Acquiring a first geometric property of the target polypeptide and a second geometric property of the polypeptide receptor;

[0010] The first geometric property and the second geometric property, as well as the sequence information of the polypeptide receptor and the sequence information of the target polypeptide, are input into a polypeptide activity prediction model to predict the binding activity of the target polypeptide and the polypeptide receptor; the polypeptide activity prediction model includes a first feature learning model, a second feature learning model and a classification model, the first feature learning model is used to learn the first feature information of the first geometric property and the second geometric property, the second feature learning model is used to learn the second feature information of the sequence information of the polypeptide receptor and the sequence information of the target polypeptide, and the classification model is used to predict the binding activity of the target polypeptide and the polypeptide receptor based on the first feature information and the second feature information.

[0011] In an optional embodiment, the obtaining of the first geometric property of the target polypeptide and the second geometric property of the polypeptide receptor comprises:

[0012] Using a deep learning model to predict the first three-dimensional structure corresponding to the sequence information of the target polypeptide and the second three-dimensional structure of the polypeptide receptor;

[0013] A first geometric property of the target polypeptide is calculated based on the first three-dimensional structure, and a second geometric property of the polypeptide receptor is calculated based on the second three-dimensional structure.

[0014] In an optional embodiment, the polypeptide activity prediction model also includes a Transformer Encoder model, which is used to align and fuse the first feature information and the second feature information to obtain third feature information, and the classification model predicts the binding activity of the target polypeptide and the polypeptide receptor based on the third feature information.

[0015] In an optional embodiment, the classification model predicts the binding activity between the target polypeptide and the polypeptide receptor based on the third feature information and the second feature information.

[0016] In an optional implementation, the input layer of the generative deep learning model is an embedding layer, and the embedding layer is used for word embedding and position embedding;

[0017] The decoder of the generative deep learning model includes a plurality of network modules connected in sequence, each of which includes a first normalization layer, a masked multi-head self-attention layer, a second normalization layer and a feedforward fully connected layer connected in sequence; the masked multi-head self-attention layer is a Flash Attention mechanism; and a Gaussian error linear unit activation function is used in the feedforward fully connected layer;

[0018] The decoder further includes a third normalization layer, a linear layer, and a Softmax layer, and the embedding layer and the linear layer share weights.

[0019] In an optional embodiment, before inputting the sequence information of the polypeptide receptor and the generated polypeptide word information into the generative deep learning model, the method further includes:

[0020] Acquire a word unit table, wherein one amino acid residue is one word unit, one modification information is one word unit, and the word unit table further includes a first special word unit, a second special word unit, and an unknown word unit, wherein the first special word unit is used to complete the sequence information of the polypeptide receptor to a preset length, the second special word unit is used to separate the sequence information of the polypeptide receptor and the polypeptide word unit information, and the unknown word unit is used to represent other words units other than the amino acid residue, the modification information, the first special word unit, and the second special word unit;

[0021] Performing word segmentation on the polypeptide receptor sequence according to the words in the word table;

[0022] According to the numerical value corresponding to each word in the word table, the segmented polypeptide receptor sequence is digitized to obtain the sequence information of the polypeptide receptor.

[0023] In an optional embodiment, the generative deep learning model is trained using a Fused AdamW optimizer.

[0024] In an optional embodiment, the generative deep learning model is trained using a half-precision floating point format.

[0025] In a second aspect, the present invention provides a device for generating an active polypeptide sequence, the device comprising:

[0026] A generation module, used to input the sequence information of the polypeptide receptor and the generated polypeptide word information into the generative deep learning model to generate new polypeptide word information;

[0027] A splicing module, used to splice the polypeptide word information generated by the generative deep learning model according to the generation order to obtain the sequence information of the target polypeptide;

[0028] An acquisition module, used to acquire a first geometric property of the target polypeptide and a second geometric property of the polypeptide receptor;

[0029] An activity prediction module is used to input the first geometric property and the second geometric property, as well as the sequence information of the polypeptide receptor and the sequence information of the target polypeptide, into a polypeptide activity prediction model to predict the binding activity of the target polypeptide and the polypeptide receptor; the polypeptide activity prediction model includes a first feature learning model, a second feature learning model and a classification model, the first feature learning model is used to learn the first feature information of the first geometric property and the second geometric property, the second feature learning model is used to learn the second feature information of the sequence information of the polypeptide receptor and the sequence information of the target polypeptide, and the classification model is used to predict the binding activity of the target polypeptide and the polypeptide receptor based on the first feature information and the second feature information.

[0030] In a third aspect, the present invention provides a computer device comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the method for generating an active polypeptide sequence of the first aspect or any corresponding embodiment thereof by executing the computer instructions.

[0031] In a fourth aspect, the present invention provides a computer-readable storage medium having computer instructions stored thereon, the computer instructions being used to enable a computer to execute the method for generating an active polypeptide sequence of the first aspect or any corresponding embodiment thereof.

[0032] In a fifth aspect, the present invention provides a computer program product, comprising computer instructions for causing a computer to execute the method for generating an active polypeptide sequence according to the first aspect or any corresponding embodiment thereof.

[0033] The method, device, equipment and storage medium for generating active polypeptide sequences provided in this embodiment can not only use the generative deep learning model to quickly generate a variety of different receptor sequences for different polypeptide receptor sequences. It also combines the polypeptide generative deep learning model with the polypeptide activity screening deep learning model (i.e., polypeptide activity prediction model). That is, the polypeptide designed by the generative deep learning model is input into the polypeptide activity prediction model to screen out active polypeptides, thereby significantly improving the efficiency and quality of active polypeptide design, thereby realizing high-throughput active polypeptide design.

[0034] In addition, when predicting the binding activity of the generated target polypeptide and the polypeptide receptor, the embodiment of the present invention not only predicts based on the sequence information of the two, but also combines the geometric properties of the two. Specifically, the first geometric property and the second geometric property specifically include the relative accessible area (RASA), depth index (DPX), protrusion index (CX), and hydrophobicity of the residue. RASA is the ratio of each residue between the calculated accessible surface area and the standard accessible surface area of ​​a specific residue. DPX is used to calculate the volume occupied by the polypeptide and the receptor in the spatial dimension, and the CX of the atom is obtained by calculating the ratio of the occupied volume and the unoccupied volume in the spatial dimension. Hydrophobicity is conducive to the formation of secondary structure and alpha helix between the receptor and the polypeptide, affecting its stability. By introducing the geometric properties of the receptor and the polypeptide, the relationship between the polypeptide and the receptor can be analyzed from the spatial dimension level, which can improve the accuracy of the prediction of the binding activity of the polypeptide and the receptor. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the related technologies, the drawings required for use in the specific embodiments or the related technical descriptions will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0036] Figure 1 is a schematic diagram of a process for generating an active polypeptide sequence according to an embodiment of the present invention;

[0037] Figure 2 is a schematic diagram of the structure of a polypeptide activity prediction model according to an embodiment of the present invention;

[0038] Figure 3 is a schematic diagram of the structure of a generative deep learning model according to an embodiment of the present invention;

[0039] Figure 4 is a schematic diagram of the input of a generative deep learning model according to an embodiment of the present invention;

[0040] Figure 5 is a schematic diagram of a word meta-table according to an embodiment of the present invention;

[0041] Figure 6 is a schematic diagram of the content of data collected for training a model according to an embodiment of the present invention;

[0042] Figure 7 is a schematic diagram of segmentation and digitization of a polypeptide receptor sequence according to an embodiment of the present invention;

[0043] Figure 8is a schematic diagram of a process for generating a polypeptide sequence according to an embodiment of the present invention;

[0044] Fig. 9 is a schematic diagram of a three-dimensional structure of a polypeptide sequence generated according to an embodiment of the present invention;

[0045] Fig.10 is a schematic diagram of the three-dimensional structure of another polypeptide sequence generated according to an embodiment of the present invention;

[0046] Fig.11 is a structural block diagram of a device for generating an active polypeptide sequence according to an embodiment of the present invention;

[0047] Fig.12 Schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0049] According to an embodiment of the present invention, an embodiment of a method for generating an active polypeptide sequence is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of executable computer instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0050] In this embodiment, a method for generating an active polypeptide sequence is provided, which can be used in various computer devices, including terminals and servers. Figure 1 is a flow chart of a method for generating an active polypeptide sequence according to an embodiment of the present invention, such as Figure 1 As shown, the process includes the following steps:

[0051] Step S101, the sequence information of the polypeptide receptor and the generated polypeptide word-meta information are input into the generative deep learning model to generate new polypeptide word-meta information. The sequence information here can be the information after the sequence is digitized.

[0052] When the polypeptide word unit information has not yet been generated, a fixed preset starting word unit information can be input. The preset starting word unit information is not used as the generated polypeptide word unit information, and is not subsequently involved in splicing to obtain the sequence information of the target polypeptide.

[0053] The generative deep learning model in the embodiment of the present invention can be a GPT model built from scratch.

[0054] Step S102, splicing the polypeptide word information generated by the generative deep learning model according to the generation order to obtain the sequence information of the target polypeptide.

[0055] Step S103, obtaining the first geometric property of the target polypeptide and the second geometric property of the polypeptide receptor.

[0056] In some specific embodiments, step S103, i.e., obtaining the first geometric property of the target polypeptide and the second geometric property of the polypeptide receptor, comprises:

[0057] Step S1031, using a deep learning model to predict the first three-dimensional structure corresponding to the sequence information of the target polypeptide and the second three-dimensional structure of the polypeptide receptor. Specifically, the HelixFold-Single tool (tool reference: Fang X, Wang F, Liu L, et al. A method for multiple-sequence-alignment-free protein structure prediction using a protein language model[J]. Nature Machine Intelligence, 2023, 5(10): 1087-1096.) can be used to predict the three-dimensional structure corresponding to the sequence information of the target polypeptide and the three-dimensional structure of the polypeptide receptor. If the three-dimensional structure of the polypeptide receptor is known, the three-dimensional structure of the polypeptide receptor can be directly obtained without using a deep learning model to predict it.

[0058] Step S1032, calculating the first geometric property of the target polypeptide according to the first three-dimensional structure, and calculating the second geometric property of the polypeptide receptor according to the second three-dimensional structure. Specifically, the geometric property can be calculated using the PSAIA tool (tool reference: Mihel J, Šikić M, Tomić S, et al. PSAIA-protein structure and interaction analyzer[J]. BMC structural biology, 2008, 8: 1-11.).

[0059] Step S104, input the first geometric property and the second geometric property, as well as the sequence information of the polypeptide receptor and the sequence information of the target polypeptide, into a polypeptide activity prediction model to predict the binding activity of the target polypeptide and the polypeptide receptor; the polypeptide activity prediction model includes a first feature learning model, a second feature learning model and a classification model, the first feature learning model is used to learn the first feature information of the first geometric property and the second geometric property, the second feature learning model is used to learn the second feature information of the sequence information of the polypeptide receptor and the sequence information of the target polypeptide, and the classification model is used to predict the binding activity of the target polypeptide and the polypeptide receptor based on the first feature information and the second feature information.

[0060] Specifically, the predicted result is active or inactive. Finally, active peptides can be screened out according to the activity prediction results of the model, that is, a peptide sequence with binding activity to the receptor is designed. The first feature learning model can be a Vision Transformer model, and the second feature learning model can be a Star-Transformer model. The Vision Transformer model is a deep learning model based on the Transformer architecture, which breaks the traditional convolutional architecture and adopts a sequence-based processing method, so that the model can better capture global information. Star-Transformer is an improved Transformer architecture that solves the problems of the standard Transformer when processing long sequences, especially computational efficiency and memory burden. Unlike the Full Self-Attention (FSA) mechanism used by the standard Transformer, Star-Transformer adopts a new mechanism called "Star-Attention", which reduces the number of parameters and reduces computational complexity.

[0061] The method for generating active polypeptide sequences provided in this embodiment can not only use the generative deep learning model to quickly generate a variety of different receptor sequences for different polypeptide receptor sequences. It also combines the polypeptide generative deep learning model with the polypeptide activity screening deep learning model (i.e., polypeptide activity prediction model). That is, the polypeptide designed by the generative deep learning model is input into the polypeptide activity prediction model to screen out active polypeptides, thereby significantly improving the efficiency and quality of active polypeptide design, thereby realizing high-throughput active polypeptide design.

[0062] In addition, when predicting the binding activity of the generated target polypeptide and the polypeptide receptor, the embodiment of the present invention not only predicts based on the sequence information of the two, but also combines the geometric properties of the two. Specifically, the first geometric property and the second geometric property specifically include the relative accessible area (RASA), depth index (DPX), protrusion index (CX), and hydrophobicity of the residue. RASA is the ratio of each residue between the calculated accessible surface area and the standard accessible surface area of ​​a specific residue. DPX is used to calculate the volume occupied by the polypeptide and the receptor in the spatial dimension, and the CX of the atom is obtained by calculating the ratio of the occupied volume and the unoccupied volume in the spatial dimension. Hydrophobicity is conducive to the formation of secondary structure and alpha helix between the receptor and the polypeptide, affecting its stability. By introducing the geometric properties of the receptor and the polypeptide, the relationship between the polypeptide and the receptor can be analyzed from the spatial dimension level, which can improve the accuracy of the prediction of the binding activity of the polypeptide and the receptor.

[0063] In some optional specific implementation methods, the polypeptide activity prediction model also includes a Transformer Encoder model, which is used to align and fuse the first feature information and the second feature information to obtain third feature information, and the classification model predicts the binding activity of the target polypeptide and the polypeptide receptor based on the third feature information.

[0064] In some other optional embodiments, the classification model predicts the binding activity between the target polypeptide and the polypeptide receptor based on the third feature information and the second feature information. Figure 2 As shown, after the output of the VisionTransformer model and the output of the Star-Transformer model enter the Transformer Encoder model, the Transformer Encoder output is spliced ​​with the output of the Star-Transformer model and used as input to enter the classification model to predict the activity of the polypeptide. The embodiment of the present invention uses Transformer Encoder and feature splicing to fuse different modal features (sequence features and geometric property features) to improve the accuracy of the model's binding activity prediction.

[0065] In the embodiment of the present invention, sequence information and geometric properties are combined to predict the binding activity of the generated target polypeptide with the polypeptide receptor. After using two feature learning models to obtain the feature information of the sequence information and the feature information of the geometric properties respectively, the Transformer Encoder model is also used to align and fuse the two feature information, thereby improving the accuracy of the polypeptide activity prediction model in predicting the binding activity of the generated target polypeptide with the polypeptide receptor based on the two feature information.

[0066] In some optional specific implementations, such as Figure 3 Shown and Figure 4 As shown, the input layer of the generative deep learning model is an embedding layer, and the embedding layer is used for word unit embedding and position embedding;

[0067] like Figure 3 As shown, the decoder of the generative deep learning model includes a plurality of network modules connected in sequence, each of which includes a first normalization layer, a masked multi-head self-attention layer, a second normalization layer, and a feedforward fully connected layer connected in sequence; the masked multi-head self-attention layer is a Flash Attention mechanism; a Gaussian error linear unit (GELU) activation function is used in the feedforward fully connected layer;

[0068] The decoder further includes a third normalization layer, a linear layer, and a Softmax layer, and the embedding layer and the linear layer share weights.

[0069] In the decoder of the embodiment of the present invention, a normalization layer (also called a normalization layer) is provided before the masked multi-head self-attention layer, and a normalization layer is also provided before the feed-forward fully connected layer. In addition, an additional normalization layer is added after the decoder to improve the stability of model training.

[0070] The Gaussian Error Linear Unit (GELU) activation function used in the feed-forward fully connected layer can have more variations than the rectified linear unit (ReLU) activation function. In addition, the GELU activation function can be approximated using the Tanh function. GELU is a nonlinear activation function used to accelerate model training, and it combines the advantages of the ReLU function (such as sparse activation) and the advantages of the traditional Sigmoid function (such as smooth transition), so it can provide better performance. The calculation of GELU is relatively complex, so the Tanh function is used to simplify the calculation of GELU.

[0071] In some optional specific embodiments, before inputting the sequence information of the polypeptide receptor and the generated polypeptide word information into the generative deep learning model, the method further includes:

[0072] A word unit table is obtained, wherein one amino acid residue is one word unit, one modification information is one word unit, and the word unit table further includes a first special word unit, a second special word unit, and an unknown word unit, wherein the first special word unit is used to complete the sequence information of the polypeptide receptor to a preset length, the second special word unit is used to separate the sequence information of the polypeptide receptor and the polypeptide word unit information, and the unknown word unit is used to represent other words units other than the amino acid residue, the modification information, the first special word unit, and the second special word unit; specifically, Figure 5 As shown in the figure, the word list contains 20 amino acid residues (Y, A, T, V, L, D, E, G, R, H, I, W, Q, K, M, F, N, S, P, C), 2 modifier words (modification 1: ~, modification 2: !), 2 special words ( <pad> 、 <sep>), 1 unknown word ( <unknown>). Among them, special words <pad>Used to fill the input to the preset length, special word <sep>Used to separate receptor words and peptide words, unknown words <unknown>It is used to represent all words except 20 amino acid residue sequences, 2 modifier words, and 2 special words, to prevent unknown words from appearing in the input sequence or modification, which would cause word segmentation errors.

[0073] Performing word segmentation on the polypeptide receptor sequence according to the words in the word table;

[0074] According to the numerical value corresponding to each word in the word table, the segmented polypeptide receptor sequence is digitized to obtain the sequence information of the polypeptide receptor.

[0075] In an embodiment of the present invention, the polypeptide receptor sequence is directly segmented into words according to each amino acid residue as a word. In other embodiments, k-mer can also be used to segment the polypeptide receptor sequence into words. Using k-mer for word segmentation also does not involve the segmentation of modification information, and the modification information is directly used as a special word. K-mer refers to a continuous amino acid sequence of length k. Using k-mer to segment the polypeptide receptor sequence into words means using a sliding window of length k to move on the polypeptide sequence, moving one amino acid position each time, and extracting the corresponding k-mer word.

[0076] In some optional specific implementations, the generative deep learning model is trained using a Fused AdamW optimizer. The AdamW optimizer is an improved version of the Adam (Adaptive Moment Estimation) optimizer used to improve the performance and efficiency of the training model. Specifically, the AdamW optimizer adds a weight decay processing method based on the original Adam optimizer to more effectively control the size of the model weight and prevent overfitting. Fused usually means fusion, specifically, it means merging multiple independent computing operations into a single operation to improve computing efficiency and reduce the number of memory accesses. This technology can significantly improve the model training speed and resource utilization. Fused specifically achieves the following ways to improve the model training speed and resource utilization: merging multiple operations into one operation, reducing the number of computing calls, thereby improving computing efficiency; optimizing memory access patterns, reducing memory bandwidth bottlenecks; using the parallel computing capabilities of GPUs (Graphics Processing Units) or TPUs (Tensor Processing Units) to speed up the optimizer's computing speed.

[0077] In some optional specific implementations, the generative deep learning model is trained using a half-precision floating point format.

[0078] In summary, the embodiments of the present invention also adopt a variety of model training acceleration strategies (including sharing weights, changing activation functions, and using Flash Attention, Fused AdamW optimizer, half precision, etc.), which further improves the efficiency of peptide design.

[0079] In a specific embodiment, the specific training process of the generative deep learning model and the polypeptide activity prediction model is:

[0080] Step 1, collect data. Collect polypeptide receptor sequences, polypeptide sequences (including modifications), and corresponding activity values. In this embodiment, a total of 1,657 data (after deduplication) were obtained from channels such as polypeptide data from the AVPdb database (http: / / crdd.osdd.net / servers / avpdb / ) and polypeptide data from the DRAVP database (http: / / dravp.cpu-bioinfor.org / ). Then, active data are screened out according to the set threshold. Specifically, Activity IC50 / EC50 10uM can be used as the threshold, and Activity IC50 / EC50>10uM is marked as 0, that is, inactive. Activity IC50 / EC50≤10uM is marked as 1, that is, active.

[0081] like Figure 6 As shown in the figure, the 1657 data include ID (peptide identification number), PeptideSeq (peptide sequence), Target_Organism (target virus), Activity (activity), Binding_Target (binding target), Mechanism (peptide mechanism), Activity (uM) (activity value after unified unit), Label (label), ReceptorSeq (receptor sequence), Receptor_Sequence_Ref (reference sequence information of receptor sequence), Source (data source). Finally, only the data with label 1 are retained, and the data with label 0 are discarded.

[0082] Step 2: Create a word unit table. Here, each amino acid residue is directly used as a word unit. See above for details, which will not be repeated here.

[0083] Step 3: Process the input information. A polypeptide sequence and a corresponding polypeptide receptor sequence are combined into a set of input data. A special word element is used between the polypeptide sequence and the polypeptide receptor sequence. <sep>Separate, with the polypeptide receptor sequence in front and the polypeptide sequence in the back. Then use the above-mentioned word table to segment the word and convert it into a corresponding digital sequence, that is, digitization. For example, the HR1 sequence of the respiratory syncytial virus (RSV) receptor F protein "AVSKVLHLEGEVNKIKSALLSTNKAVVSLSNGVSVLTSKVLDLKNYIDKQLLPIVNKQSCRI" (NCBI identification number FJ614815), and the polypeptide sequence FDASISQVNEKINQSLAFIRKSDELLHNVNAGKST bound to it (from the AVPdb database: identification number AVP0143), constitute a set of input data, in which the HR1 sequence of the RSV receptor F protein is in front, followed by a special word. <sep>The final result is a peptide sequence. The input data is segmented according to single amino acid residues, and peptide modifications are not segmented and are directly used as modification terms. Figure 5 The lemma table shown in the figure converts the segmented lemmas into numerical values. The numerical values ​​are embedded as lemmas and combined with position embedding, which can be used as the input of the model.

[0084] Step 4: Train the generative deep learning model. Obtain the GPT model built from scratch, input the peptide receptor sequence and peptide sequence word embedding plus position embedding into the model for training, and save the optimal model. Specifically, Figure 3 As shown in the figure, the input of the GPT model built from scratch is the word embedding plus the position embedding, and the input enters the decoder part of the GPT model. The decoder contains multiple identical layers stacked together, including a masked multi-head self-attention layer and a feedforward fully connected layer. The masked multi-head self-attention layer is preceded by a normalization layer, and the feedforward fully connected layer is also preceded by a normalization layer. The masked multi-head self-attention layer masks the word information after the current position to prevent the attention mechanism from obtaining the word to be predicted in the future. Since the GPT model generates the word information of the target peptide sequence one by one, the word information to be predicted in the future cannot be obtained in advance. In addition, a fast and memory-efficient attention mechanism-Flash Attention is used to replace the attention mechanism in the GPT model to speed up the training of the GPT model. The Gaussian Error Linear Unit (GELU) activation function is used in the feedforward fully connected layer, and the Tanh function is used for approximation, which can have more variations than the rectified linear unit (ReLU) activation function.

[0085] An additional normalization layer is required after the decoder, followed by a linear layer. The linear layer and the input embedding layer share weights to reduce the amount of parameter training and make GPT model training more efficient. Finally, a Softmax layer is added to obtain the probability distribution of the word table size.

[0086] During the GPT model training process, the AdamW optimizer is used to fix the vulnerability of the Adam optimizer, and fused is used to accelerate training. In addition, the GPT model is trained with half precision (BFLOAT16) to better utilize the GPU's video memory resources, thereby speeding up the GPT model training.

[0087] Step 5: Model evaluation and peptide generation. The optimal GPT model saved in step 4 needs to be tested on the test set. Specifically, the perplexity value of the model on the test set is 2.1316.

[0088] By inputting the sequence information of the peptide receptor into the trained optimal GPT model, a peptide sequence (including modifications) that specifically binds to the peptide receptor can be generated. Figure 7 As shown, the receptor sequence of RSV polypeptide AVSKVLHLEGEVNKIKSALLSTNKAVVSLSNGVSVLTSKVLDLKNYIDKQLLPIVNKQSCRI (derived from NCBI identifier FJ614815) was input and the input sequence was converted into a numerical sequence according to the word list (6, 25, 21, 17, 25, 16, 14, 16, 12, 13, 12, 25,8, 17, 15, 17, 21, 6, 16, 16, 21, 22, 8, 17, 6, 25, 25, 21, 16, 21, 8, 13,25, 21, 25, 16, 22, 21, 17, 25, 16, 9, 16, 17, 8, 24, 15, 9, 17, 11, 16, 16,20, 15, 25, 8, 17, 11, 21, 10, 7, 15), and use special words <pad>(The corresponding number is 0) to fill the preset input receptor sequence length, convert the numerical sequence into a tensor. Finally, transfer the numerical sequence to the GPU device, which is Figure 7 device='cuda:0' in .

[0089] like Figure 8 As shown in the figure, after the receptor sequence of RSV polypeptide AVSKVLHLEGEVNKIKSALLSTNKAVVSLSNGVSVLTSKVLDLKNYIDKQLLPIVNKQSCRI is converted into a numerical sequence and input into the trained GPT model, a word (E) will be generated first. The generated word (E) continues to be input into the GPT model, and the next word (L) will continue to be generated. This process continues until the set maximum output length is reached. Finally, all the generated words are spliced ​​(EL...), which is the final generated polypeptide sequence (including modifications).

[0090] like Fig. 9 and Fig.10 As shown, by inputting the receptor sequence AVSKVLHLEGEVNKIKSALLSTNKAVVSLSNGVSVLTSKVLDLKNYIDKQLLPIVNKQSCRI of the RSV polypeptide and the receptor sequence YAADQKSTQNAINGITNKVNTVIEKMNIQFTAVGKEFNKLEKRMENLNKKVDD of the INFVA polypeptide (derived from the PDB database, identification number 5VLI), a polypeptide sequence (including modifications) binding to the corresponding receptor can be generated. Fig. 9 and Fig.10 The peptide sequences generated in the experiment all contain α-helical features.

[0091] Step 6: Train the peptide activity prediction model. Figure 2 As shown in the figure, the peptide activity prediction model is built based on Transformer, and the model specifically includes the Vision Transformer model and the Star-Transformer model. The input of the Vision Transformer model is the geometric properties of the receptor and the geometric properties of the peptide. The input of the Star-Transformer model is the receptor sequence and the peptide sequence. After the output of the Vision Transformer model and the output of the Star-Transformer model enter the Transformer Encoder model, the Transformer Encoder output is spliced ​​with the output of the Star-Transformer model and enters the classification model as input to predict the activity of the peptide.

[0092] In the present embodiment, a generation device of an active polypeptide sequence is also provided, and the device is used to implement the above method embodiment and preferred implementation mode, and the descriptions that have been made are not repeated here. As used below, the term "module" can implement a combination of software and / or hardware of a predetermined function. Although the device described in the following embodiments is preferably implemented with software, the implementation of hardware, or a combination of software and hardware is also possible and conceived.

[0093] This embodiment provides a device for generating an active polypeptide sequence, such as Fig.11 As shown, including:

[0094] A generation module 1101 is used to input the sequence information of the polypeptide receptor and the generated polypeptide word-meta information into the generative deep learning model to generate new polypeptide word-meta information;

[0095] A splicing module 1102 is used to splice the polypeptide word information generated by the generative deep learning model according to the generation order to obtain the sequence information of the target polypeptide;

[0096] An acquisition module 1103 is used to acquire a first geometric property of the target polypeptide and a second geometric property of the polypeptide receptor;

[0097] The activity prediction module 1104 is used to input the first geometric property and the second geometric property, as well as the sequence information of the polypeptide receptor and the sequence information of the target polypeptide, into a polypeptide activity prediction model to predict the binding activity of the target polypeptide and the polypeptide receptor; the polypeptide activity prediction model includes a first feature learning model, a second feature learning model and a classification model, the first feature learning model is used to learn the first feature information of the first geometric property and the second geometric property, the second feature learning model is used to learn the second feature information of the sequence information of the polypeptide receptor and the sequence information of the target polypeptide, and the classification model is used to predict the binding activity of the target polypeptide and the polypeptide receptor based on the first feature information and the second feature information.

[0098] In some optional implementations, the acquisition module 1103 includes:

[0099] A three-dimensional structure prediction unit, used to predict the first three-dimensional structure corresponding to the sequence information of the target polypeptide and the second three-dimensional structure of the polypeptide receptor respectively by using a deep learning model;

[0100] The geometric property calculation unit is used to calculate the first geometric property of the target polypeptide according to the first three-dimensional structure, and calculate the second geometric property of the polypeptide receptor according to the second three-dimensional structure.

[0101] In some optional embodiments, the polypeptide activity prediction model also includes a Transformer Encoder model, which is used to align and fuse the first feature information and the second feature information to obtain third feature information, and the classification model predicts the binding activity of the target polypeptide and the polypeptide receptor based on the third feature information.

[0102] In some optional embodiments, the classification model predicts the binding activity between the target polypeptide and the polypeptide receptor based on the third feature information and the second feature information.

[0103] In some optional embodiments, the input layer of the generative deep learning model is an embedding layer, and the embedding layer is used for word embedding and position embedding;

[0104] The decoder of the generative deep learning model includes a plurality of network modules connected in sequence, each of which includes a first normalization layer, a masked multi-head self-attention layer, a second normalization layer and a feedforward fully connected layer connected in sequence; the masked multi-head self-attention layer is a Flash Attention mechanism; and a Gaussian error linear unit activation function is used in the feedforward fully connected layer;

[0105] The decoder further includes a third normalization layer, a linear layer, and a Softmax layer, and the embedding layer and the linear layer share weights.

[0106] In some optional embodiments, the device for generating an active polypeptide sequence further comprises:

[0107] A word-unit table acquisition module, used to acquire a word-unit table, wherein one amino acid residue in the word-unit table is one word-unit, one modification information is one word-unit, and the word-unit table also includes a first special word-unit, a second special word-unit and an unknown word-unit, wherein the first special word-unit is used to complete the sequence information of the polypeptide receptor to a preset length, the second special word-unit is used to divide the sequence information of the polypeptide receptor and the polypeptide word-unit information, and the unknown word-unit is used to represent other words other than the amino acid residue, the modification information, the first special word-unit and the second special word-unit;

[0108] A word unit segmentation module, used for performing word unit segmentation on the polypeptide receptor sequence according to the word units in the word unit table;

[0109] The digitization module is used to digitize the segmented polypeptide receptor sequence according to the numerical value corresponding to each word in the word table to obtain the sequence information of the polypeptide receptor.

[0110] In some optional embodiments, the generative deep learning model is trained using a Fused AdamW optimizer.

[0111] In some optional embodiments, the generative deep learning model is trained using a half-precision floating point format.

[0112] The further functional description of each of the above modules and units is the same as that of the above corresponding embodiments and will not be repeated here.

[0113] The active polypeptide sequence generation device in this embodiment is presented in the form of a functional unit, where the unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that executes one or more software or fixed programs, and / or other devices that can provide the above functions.

[0114] The embodiment of the present invention also provides a computer device having the above Fig.11 The device for generating the active polypeptide sequence shown.

[0115] See also Fig.12 , Fig.12 is a schematic diagram of the structure of a computer device provided by an optional embodiment of the present invention, such as Fig.12 As shown, the computer device includes: one or more processors 10, a memory 20, and interfaces for connecting various components, including high-speed interfaces and low-speed interfaces. Various components are connected to each other using different buses for communication, and can be installed on a common mainboard or installed in other ways as needed. The processor can process the instructions executed in the computer device, including instructions stored in or on the memory to display the graphical information of the GUI on an external input / output device (such as, a display device coupled to the interface). In some optional embodiments, if necessary, multiple processors and / or multiple buses can be used together with multiple memories and multiple memories. Similarly, multiple computer devices can be connected, and each device provides some necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system). Fig.12 A processor 10 is taken as an example.

[0116] The processor 10 may be a central processing unit, a network processor or a combination thereof. The processor 10 may further include a hardware chip. The hardware chip may be a dedicated integrated circuit, a programmable logic device or a combination thereof. The programmable logic device may be a complex programmable logic device, a field programmable gate array, a general purpose array logic or any combination thereof.

[0117] The memory 20 stores instructions executable by at least one processor 10, so that the at least one processor 10 executes the method shown in the above embodiment.

[0118] The memory 20 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 20 may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some optional embodiments, the memory 20 may optionally include a memory remotely arranged relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0119] The memory 20 may include a volatile memory, such as a random access memory; the memory may also include a non-volatile memory, such as a flash memory, a hard disk or a solid state drive; the memory 20 may also include a combination of the above types of memory.

[0120] The computer device also includes an input device 30 and an output device 40. The processor 10, the memory 20, the input device 30 and the output device 40 may be connected via a bus or other means. Fig.12 The example of connecting through bus is taken in the following.

[0121] The input device 30 can receive input digital or character information, and generate key signal input related to the user settings and function control of the computer device, such as a touch screen, a keypad, a mouse, a track pad, a touch pad, an indicator bar, one or more mouse buttons, a trackball, a joystick, etc. The output device 40 may include a display device, an auxiliary lighting device (e.g., an LED) and a tactile feedback device (e.g., a vibration motor), etc. The above-mentioned display device includes but is not limited to a liquid crystal display, a light emitting diode, a display and a plasma display. In some optional embodiments, the display device may be a touch screen.

[0122] The computer device also includes a communication interface, which is used for the computer device to communicate with other devices or a communication network.

[0123] The embodiment of the present invention also provides a computer-readable storage medium. The method according to the embodiment of the present invention can be implemented in hardware, firmware, or can be implemented as a computer code that can be recorded in a storage medium, or can be implemented as a computer code that is originally stored in a remote storage medium or a non-temporary machine-readable storage medium and will be stored in a local storage medium through a network download, so that the method described herein can be stored in such software processing on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only storage memory, a random access memory, a flash memory, a hard disk or a solid-state hard disk, etc.; further, the storage medium can also include a combination of the above types of memories. It can be understood that a computer, a processor, a microprocessor controller, or programmable hardware includes a storage component that can store or receive software or computer code. When the software or computer code is accessed and executed by a computer, a processor, or hardware, the method shown in the above embodiment is implemented.

[0124] A part of the present invention may be applied as a computer program product, such as a computer program instruction, which, when executed by a computer, can call or provide the method and / or technical solution according to the present invention through the operation of the computer. Those skilled in the art should understand that the existence of the computer program instruction in a computer-readable medium includes, but is not limited to, a source file, an executable file, an installation package file, etc., and accordingly, the way in which the computer program instruction is executed by the computer includes, but is not limited to: the computer directly executes the instruction, or the computer compiles the instruction and then executes the corresponding compiled program, or the computer reads and executes the instruction, or the computer reads and installs the instruction and then executes the corresponding installed program. Here, the computer-readable medium may be any available computer-readable storage medium or communication medium accessible to the computer.

[0125] Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations are all within the scope defined by the appended claims.< / pad> < / sep> < / sep> < / unknown> < / sep> < / pad> < / unknown> < / sep> < / pad>

Claims

1. A method for generating an active polypeptide sequence, characterized in that: The method comprises: Input the sequence information of the peptide receptor and the generated peptide word information into the generative deep learning model to generate new peptide word information; splicing the polypeptide word information generated by the generative deep learning model in a generation order to obtain sequence information of the target polypeptide; Acquiring a first geometric property of the target polypeptide and a second geometric property of the polypeptide receptor; The first geometric property and the second geometric property, as well as the sequence information of the polypeptide receptor and the sequence information of the target polypeptide, are input into a polypeptide activity prediction model to predict the binding activity of the target polypeptide and the polypeptide receptor; the polypeptide activity prediction model comprises a first feature learning model, a second feature learning model and a classification model, the first feature learning model is used to learn the first feature information of the first geometric property and the second geometric property, the second feature learning model is used to learn the second feature information of the sequence information of the polypeptide receptor and the sequence information of the target polypeptide, and the classification model is used to predict the binding activity of the target polypeptide and the polypeptide receptor based on the first feature information and the second feature information; Among them, the first geometric property and the second geometric property specifically include the relative accessible area, depth index, protrusion index, and hydrophobicity of the residues; the relative accessible area of ​​the residues is the ratio of the calculated accessible surface area for each residue to the standard accessible surface area of ​​the specific residue; the depth index is used to calculate the volume occupied by the polypeptide and the receptor in the spatial dimension, and the protrusion index of the atom is obtained by calculating the ratio of the occupied volume and the unoccupied volume in the spatial dimension.

2. The method according to claim 1, characterized in that The step of obtaining the first geometric property of the target polypeptide and the second geometric property of the polypeptide receptor comprises: Using a deep learning model to predict the first three-dimensional structure corresponding to the sequence information of the target polypeptide and the second three-dimensional structure of the polypeptide receptor; A first geometric property of the target polypeptide is calculated based on the first three-dimensional structure, and a second geometric property of the polypeptide receptor is calculated based on the second three-dimensional structure.

3. The method according to claim 1 or 2, characterized in that The polypeptide activity prediction model also includes a Transformer Encoder model, which is used to align and fuse the first feature information and the second feature information to obtain third feature information, and the classification model predicts the binding activity of the target polypeptide and the polypeptide receptor based on the third feature information.

4. The method according to claim 3, characterized in that The classification model predicts the binding activity between the target polypeptide and the polypeptide receptor based on the third feature information and the second feature information.

5. The method according to claim 1, characterized in that The input layer of the generative deep learning model is an embedding layer, and the embedding layer is used for word unit embedding and position embedding; The decoder of the generative deep learning model includes a plurality of network modules connected in sequence, each of which includes a first normalization layer, a masked multi-head self-attention layer, a second normalization layer and a feedforward fully connected layer connected in sequence; the masked multi-head self-attention layer is a Flash Attention mechanism; and a Gaussian error linear unit activation function is used in the feedforward fully connected layer; The decoder further includes a third normalization layer, a linear layer, and a Softmax layer, and the embedding layer and the linear layer share weights.

6. The method according to claim 1, characterized in that Before inputting the sequence information of the polypeptide receptor and the generated polypeptide word information into the generative deep learning model, the method further includes: Acquire a word unit table, wherein one amino acid residue is one word unit, one modification information is one word unit, and the word unit table further includes a first special word unit, a second special word unit, and an unknown word unit, wherein the first special word unit is used to complete the sequence information of the polypeptide receptor to a preset length, the second special word unit is used to separate the sequence information of the polypeptide receptor and the polypeptide word unit information, and the unknown word unit is used to represent other words units other than the amino acid residue, the modification information, the first special word unit, and the second special word unit; Performing word segmentation on the polypeptide receptor sequence according to the words in the word table; According to the numerical value corresponding to each word in the word table, the segmented polypeptide receptor sequence is digitized to obtain the sequence information of the polypeptide receptor.

7. The method according to claim 1, characterized in that The generative deep learning model is trained using a FusedAdamW optimizer; and / or, The generative deep learning model is trained using half-precision floating point format.

8. A device for generating an active polypeptide sequence generated by the generation method according to any one of claims 1 to 7, characterized in that: The device comprises: A generation module, used to input the sequence information of the polypeptide receptor and the generated polypeptide word information into the generative deep learning model to generate new polypeptide word information; A splicing module, used to splice the polypeptide word information generated by the generative deep learning model according to the generation order to obtain the sequence information of the target polypeptide; An acquisition module, used to acquire a first geometric property of the target polypeptide and a second geometric property of the polypeptide receptor; An activity prediction module is used to input the first geometric property and the second geometric property, as well as the sequence information of the polypeptide receptor and the sequence information of the target polypeptide, into a polypeptide activity prediction model to predict the binding activity of the target polypeptide and the polypeptide receptor; the polypeptide activity prediction model includes a first feature learning model, a second feature learning model and a classification model, the first feature learning model is used to learn the first feature information of the first geometric property and the second geometric property, the second feature learning model is used to learn the second feature information of the sequence information of the polypeptide receptor and the sequence information of the target polypeptide, and the classification model is used to predict the binding activity of the target polypeptide and the polypeptide receptor based on the first feature information and the second feature information.

9. A computer device, characterized in that: include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the method for generating an active polypeptide sequence according to any one of claims 1 to 7 by executing the computer instructions.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a computer to execute the method for generating an active polypeptide sequence according to any one of claims 1 to 7.

Images