Protein design method based on deep learning

Abstract

The invention discloses a protein design method based on deep learning. The method comprises the following steps: calculating amino acid distribution in a protein sequence from five pre-trained models; respectively calculating the distribution of amino acid residues of proteins in the five pre-trained models; combining amino acid distribution in the five pre-trained models to a new latitude, inputting the combined distribution into an integrated network based on self-attention, and obtaining attention output Z; the attention output Z is processed through a multi-layer feedforward neural network, and a result H is obtained; projection is carried out on the linear layer, and a non-normalized score is obtained; and normalizing the non-normalized scores to obtain final protein sequence distribution. According to the protein design method based on deep learning provided by the invention, by integrating the structures of a plurality of pre-training models, the corresponding relationship of protein structure sequence modals is comprehensively learned, and more diversified protein sequences can be designed.

Description

Technical Field

[0001] The present invention relates to the field of protein design, and in particular to a protein design method based on deep learning. Background Art

[0002] Protein reverse folding is an efficient method for protein design. It redesigns proteins by generating protein sequences that can fold into a given skeleton. Currently, protein reverse folding models mainly include ProteinMPNN, ESM-IF, Frame2seq, PiFold and other models. However, proteins designed by existing models have the following disadvantages:

[0003] 1. Protein sequence diversity is low. Existing models are often limited by the model structure and feature extraction process during training, resulting in low diversity and lack of diversity in the generated protein sequences, making it difficult to expand the boundaries of the protein sequence space.

[0004] 2. Existing models often have erroneous biases in the frequency of use of certain types of amino acids during training. For example, ProteinMPNN tends to prefer glutamate, while ESM-IF uses less glutamate and alanine, and Frame2seq uses less leucine for design. These erroneous preferences often lead to failures in protein design.

[0005] 3. The designed protein cannot be folded. The first condition for a protein to perform its biological function is that it can be correctly folded into the corresponding conformation. However, the sequences designed by many current models often cannot be correctly folded due to unreasonable design of internal amino acid residues.

[0006] 4. The ability to design large proteins is poor. Existing models are often unable to design longer sequences due to autoregressive generation. Summary of the invention

[0007] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is that the existing protein design methods have low diversity, some erroneous preferences that lead to protein design failure, the designed proteins cannot be folded, or the large protein design ability is poor. Therefore, the present invention provides a protein design method based on deep learning, an ensemble learning model based on attention, which integrates the structures of multiple pre-trained models, comprehensively learns the modal correspondence between protein structure sequences, and can design more diverse protein sequences.

[0008] To achieve the above object, the present invention provides a protein design method based on deep learning, comprising the following steps:

[0009] Compute amino acid distributions in protein sequences from several pre-trained models;

[0010] Then, the distribution of each amino acid residue of the protein in the five pre-trained models was calculated respectively;

[0011] Then, the amino acid distributions in several pre-trained models are merged into a new dimension, and the merged distribution is input into the integrated network based on self-attention, and the attention output Z is obtained;

[0012] Process the attention output Z through a multi-layer feedforward neural network and obtain the result H;

[0013] Then the result H obtained by the multi-layer feedforward neural network is projected to through the output linear layer and the unnormalized score is obtained;

[0014] The unnormalized scores are then normalized by the function to obtain the final protein sequence distribution.

[0015] Furthermore, several pre-trained models are integrated into a heterogeneous pre-trained inverse folding model, including ProteinMPNN-vanilla, ProteinMPNN-soluble, ESM-IF, Frame2seq, and PiFold; the pre-trained model is a protein inverse folding model, the original input is the three-dimensional coordinates of the protein amino acid residue skeleton, and the output is the protein amino acid residue sequence distribution

[0016] Furthermore, the input of the Frame2seq pre-trained model requires four protein amino acid residues and a virtual Cβ atomic coordinate X Cβ .

[0017] Furthermore, the amino acid distributions in several pre-trained models are then merged into a new dimension, and the merged distribution is input into the integrated network based on self-attention. Specifically, the amino acid distributions S of each base model are merged to obtain Among them, M is the number of base models, L is the number of amino acid residues, C is the type of amino acid, and the merged S concat Enter self-attention based ensemble networks.

[0018] Furthermore, the combined S concat Expand along the last dimension and get S flat , then S flat Through the linear layer W inproj Projected into the 128-dimensional latent variable space, we get S hidden Then S hidden Perform sine and cosine position encoding, and then convert the position encoded S hidden Input into the Transformer encoder based on the multi-head attention mechanism to obtain the encoded attention output Z.

[0019] Furthermore, the multi-layer feedforward neural network includes two fully connected layers FC and ReLU activation function.

[0020] Furthermore, the attention output Z is processed by a multi-layer feedforward neural network to obtain the result H, and then the attention output Z is processed by a multi-layer feedforward neural network, and then the result H obtained by the multi-layer feedforward neural network is projected to through the output linear layer and obtains the unnormalized score logits, and then the unnormalized score is normalized by the Softmax function to first obtain the amino acids that conform to the given structure, and finally obtain the final protein sequence distribution.

[0021] Furthermore, it also includes comparing the probability distribution of the protein sequence normalized by the Softmax function with the original sequence, and calculating the loss using the cross entropy loss function.

[0022] Furthermore, the cross entropy loss function is:

[0023]

[0024] Where S ij is the amino acid index at position i in the target sequence, The index of the sequence predicted by the model.

[0025] Furthermore, it also includes using a back-propagation algorithm to calculate the gradient of the loss function with respect to the model parameters, and updating the model parameters according to the gradient to minimize the loss function value until the loss function converges.

[0026] Technical Effects

[0027] A protein design method based on deep learning of the present invention uses an attention-based ensemble learning model to comprehensively learn the modal correspondence between protein structure sequences by integrating the structures of multiple pre-trained models, and can design more diverse protein sequences. A protein design method based on deep learning of the present invention, the designed amino acids are more consistent with the distribution of natural amino acids, with less amino acid error deviation; the recovery rate of buried internal amino acids reaches more than 80%, showing better correct folding ability; at the same time, the present invention adopts a one-time generation method, which can design long protein sequences of more than 1,000 amino acids.

[0028] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a schematic diagram of a framework of a protein design method based on deep learning according to a preferred embodiment of the present invention;

[0030] Figure 2 It is a schematic diagram of sequence recovery and perplexity of a protein design method based on deep learning in a preferred embodiment of the present invention;

[0031] Figure 3 It is a schematic diagram of buried amino acid recall rate of a protein design method based on deep learning according to a preferred embodiment of the present invention;

[0032] Figure 4 It is a schematic diagram of the amino acid residue error preference of a protein design method based on deep learning in a preferred embodiment of the present invention;

[0033] Figure 5 It is a schematic diagram of an index evaluation (short protein) of sequence structure and function designed by different models of a protein design method based on deep learning in a preferred embodiment of the present invention;

[0034] Figure 6 It is a schematic diagram of an index evaluation of sequence structure and function (medium-length protein) of different model designs of a protein design method based on deep learning in a preferred embodiment of the present invention;

[0035] Figure 7 It is a schematic diagram of an index evaluation (long protein) of sequence structure and function designed by different models of a protein design method based on deep learning in a preferred embodiment of the present invention;

[0036] Figure 8 It is a schematic diagram of an index evaluation of sequence structure and function (proteins of different lengths) of different model designs of a protein design method based on deep learning in a preferred embodiment of the present invention;

[0037] Fig. 9 This is a schematic diagram of the functional prediction capability of deep learning models for mutants using different models of a deep learning-based protein design method according to a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0038] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0039] In the following description, specific details such as specific internal procedures and techniques are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present invention. However, it should be clear to those skilled in the art that the present invention may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present invention.

[0040] like Figure 1 As shown, an embodiment of the present invention provides a protein design method based on deep learning, comprising the following steps:

[0041] The amino acid distribution in the protein sequence is calculated from several pre-trained models; wherein, several pre-trained models are integrated into a heterogeneous pre-trained inverse folding model, and in this embodiment, there are five pre-trained models, including ProteinMPNN-vanilla, ProteinMPNN-soluble, ESM-IF, Frame2seq, and PiFold; the pre-trained model is a protein inverse folding model, the original input is the three-dimensional coordinates of the protein amino acid residue skeleton, and the output is the protein amino acid residue sequence distribution; wherein, the original input is the three-dimensional coordinates X=[x1, x2, ...x n ],in n is the number of amino acid residues.

[0042] Then calculate the distribution of each amino acid residue of the protein in the five pre-trained models respectively The 21-dimensional code includes the one-hot encoding of the twenty common amino acids and the default value X. The Frame2seq input requires the above four amino acid residues and a virtual Cβ atomic coordinate X Cβ , the calculation formula is as follows:

[0043] b=X Cα -X N

[0044] c=X C -X Cα

[0045] a=b·c

[0046] X Cβ =-0.58273431*a+0.56802827*b-0.54067466*c

[0047] Then, the amino acid distributions of the five pre-trained models are merged into a new dimension, and the merged distribution is input into the integrated network based on self-attention, and the attention output Z is obtained; specifically, the amino acid distributions S of each base model are merged to obtain Among them, M is the number of base models, L is the number of amino acid residues, C is the type of amino acid, and the merged S concat Input the integrated network based on self-attention. concat Expand along the amino acid length L dimension, in this embodiment, expand along the last dimension, and obtain S flat , then S flat Through the linear layer W inproj Projected into the 128-dimensional latent variable space, we get S hidden Then S hidden Perform sine and cosine position encoding, and then convert the position encoded S hidden Input into the Transformer encoder based on the multi-head attention mechanism to obtain the encoded attention output Z.

[0048] Among them, the integrated network used in the embodiment of the present invention is mainly based on the multi-head attention mechanism. First, S concat The amino acid one-hot encoding dimension along the last dimension is flattened to S flat , and then pass through the input linear layer W inproj Mapped to a 128-dimensional latent variable space, namely:

[0049] S hidden =S flat W inproj

[0050] Since the multi-head attention mechanism lacks the recognition of amino acid residue position information, the latent variables are positionally encoded. Specifically, the latent variables are positionally encoded by sine and cosine, and the encoding method is as follows:

[0051]

[0052] S hidden =PE+S hidden

[0053] Where pos is the position index, i is the dimension index, d is the encoding dimension, S hidden is a hidden variable.

[0054] The position-encoded latent variables are then passed through three layers of Transformer decoders based on a multi-head attention mechanism to capture the global dependencies between input features. The attention mechanism calculation formula is as follows:

[0055]

[0056] Among them, Q, K, and V are query matrix, key matrix, and value matrix respectively. k is the scaling factor. hidden =[s1,s2,...,s n ]The query matrix maps to query, key, and value spaces,

[0057] Q=S hidden W Q

[0058] K=S hidden W k

[0059] V=S hidden W V

[0060] Where W Q , W K , W V is a trainable weight parameter. The multi-head self-attention mechanism is to output multiple attention heads i The attention output Z is concatenated to enhance the expressive power of the model. The formula is as follows:

[0061] Z=MultiHeadAttention(Q,K,V)=Concat(head1,head2,...head h )W o

[0062]

[0063] Next, the attention output Z is processed through a multi-layer feedforward neural network and the result H is obtained; then the attention output Z is processed through a multi-layer feedforward neural network, which includes two layers of fully connected layers FC and ReLU activation function, namely:

[0064] H=FeedForward(Z)=FC2(ReLU(FC1(Z)))

[0065] Then the result G obtained by the feedforward neural network is projected to the output linear layer and the unnormalized score logits is obtained, that is,

[0066] logits=HW outproj ;

[0067] Then the unnormalized scores are normalized by the Softmax function to obtain the final protein sequence distribution:

[0068] Probs = Softmax(logits)

[0069]

[0070] Among them, logits i Indicates the corresponding amino acid type index. The obtained Probs are the final designed amino acids that meet the given structure.

[0071] In the embodiment of the present invention, the probability distribution of the protein sequence normalized by the Softmax function is compared with the original sequence, and the loss is calculated using the cross entropy loss function. The specific formula is:

[0072]

[0073] Where S ij is the amino acid index at position i in the target sequence, The index of the sequence predicted by the model.

[0074] The back propagation algorithm is further used to calculate the gradient of the loss function with respect to the model parameters. Then, the Adam optimizer is used to update the model parameters at a learning rate of 0.000001. First, the gradient of the loss function with respect to the model parameters is calculated. The model parameters are updated according to the gradient to minimize the loss function value until the loss function converges.

[0075] A protein design method based on deep learning in the above embodiment of the present invention is implemented by relying on the CHIEF model. The protein design method based on deep learning and the training method based on the CHIEF model are as follows: First, the CATH4.2 data set is split into a training set, a validation set, and a test set in a ratio of 8:1:1. The default values ​​of the extracted amino acid residues are filled with 0, and the mask tensor is selected to record the default position. The three-dimensional coordinates of the N atom, the Cα atom, the O atom, and the carboxyl C atom are obtained. and sequence Contains 20 amino acids and a default value X. Further, all sequences were input into the ProteinMPNN-vanilla, ProteinMPNN-soluble, ESM-IF, Frame2seq, and PiFold pre-trained models to obtain the sequence probability distribution given by each pre-trained model. For the output of each model, the common 20 amino acids and default value symbols were extracted, totaling 21 dimensions, and the probability distributions of each pre-trained model were merged to obtain S concat , Where B is the training batch size, L is the longest amino acid length in the batch, and M is the number of pre-trained models.

[0076] S concat Expanding along the last dimension, we get Then pass it through the linear layer W inproj Projected into the 128-dimensional latent variable space, we get S hidden Perform sine and cosine position encoding, and then convert the position encoded S hidden Input into the Transformer encoder based on the multi-head attention mechanism to obtain the encoded representation

[0077] Input the Transformer encoder representation Z to the linear layer W outproj , and get the unnormalized score The normalized protein sequence probability distribution is further obtained through the Softmax function.

[0078] Calculate the loss function: Compare the probability distribution of the protein sequence normalized by the Softmax function with the original sequence, and use the cross-entropy loss function to calculate the loss. The specific formula is:

[0079]

[0080] Where S ij is the amino acid index at position i in the target sequence, The index of the sequence predicted by the model.

[0081] The back propagation algorithm is further used to calculate the gradient of the loss function with respect to the model parameters. Then, the Adam optimizer is used to update the model parameters at a learning rate of 0.000001. First, the gradient of the loss function with respect to the model parameters is calculated. The model parameters are updated according to the gradient to minimize the loss function value until the loss function converges.

[0082] Model validation: Evaluate the performance of the model on the validation set, calculate the sequence recovery and sequence perplexity, and save the model parameters after training for subsequent use. Load the saved model for prediction or further fine-tuning.

[0083] A protein design method based on deep learning of the present invention uses deep learning to promote the transformation of protein design from energy function to artificial intelligence paradigm. However, due to the weak interpretability of deep neural networks, there are large differences between the sequences generated by various models, and it is often difficult to select a suitable model in protein design practice. The pre-trained base models that have been developed at present often have problems such as low sequence recovery, inability to correctly fit the distribution of natural protein sequences, poor design ability of longer proteins, inability to fold the generated proteins, and large amino acid residue preferences. The CHIEF model used in the present invention uses voting networks as technical motivations and provides an emerging heterogeneous integrated learning network based on attention mechanisms. The CHIEF model of the present invention integrates the results of different pre-trained models as voting network features to generate new protein sequences. By adaptively integrating the advantages of different models, this model has significant improvements in sequence recovery, model perplexity, buried amino acid accuracy, long protein design ability, and reduction of amino acid error preferences compared to other pre-trained models. The present invention is based on the protein inverse folding problem as a design principle. Protein inverse folding design is an optimization problem. The optimization goal is to enable the model to generate a protein sequence that can fold into a given protein skeleton structure, that is, a given protein skeleton structure X minimizes the generated sequence P θ (S|X) and the original sequence S native The distribution difference of , the optimization objective function is as follows:

[0084]

[0085] First, in order to evaluate the performance of the method of the present invention, the sequence recall rate and perplexity, two golden indicators of protein design, were selected for evaluation on the most commonly used CATH4.2, TS50, and TS500 validation sets. The sequence recall rate is used to measure the ability of the model to restore the original sequence for a given protein structure, and the perplexity is used to describe the difference between the protein sequence designed by the model and the natural amino acid distribution. Figure 2 As shown, the recall rate of the CHIEF model of the present invention on CATH reached 62%, while the models such as ProteinMPNN, ESMIF, Frame2seq, pifold, etc. were only around 50%, which is far better than the 50% of the most advanced five models at present, and reached 71% on TS50 and TS500. The perplexity performance also reached 3.45 on the three data sets, which was 11% lower than the average of the base model. At the same time, in the long protein design task, the sequence recovery reached 57%, which quickly dropped to 40% compared with ESM-IF, suggesting that the CHIEF model has a more robust performance for the design of long protein sequences. This fully proves the good performance of the protein design of the CHIEF model of the present invention.

[0086] Secondly, at each position of the protein, whether inside or outside the protein, the amino acid residue recall rate of the CHIEF model of the present invention is higher than that of other models, especially for buried amino acids located inside the protein, which play an extremely important role in protein folding. The index of amino acid burial degree is as follows: Figure 3 As shown in the figure, the recall rates of the other five models for buried amino acids are all around 80%, while the recall rate of the CHIEF model of the present invention for internal amino acid residues reaches 87%, and the present invention has a higher recall rate at each position of the protein. This experiment means that the CHIEF model of the present invention has better folding ability for the designed protein.

[0087] The proteins designed by previous models often have strong preferences for the wrong amino acids. For example, ProteinMPNN tends to prefer glutamate, while ESM-IF uses less glutamate and alanine, and Frame2seq uses less leucine for design. These wrong preferences are also an important reason for the failure of protein design. By evaluating the wrong preferences of amino acid residues of each model on the validation set, Figure 4 As shown, the five models previously studied all have high deviations, while the overall deviation of the CHIEF model of the present invention is greatly reduced, which is more consistent with the distribution of amino acids in nature.

[0088] Sequence- and structure-based indicators are further used to evaluate the performance of different models and the CHIEF model of the present invention. First, different models are used to design a variety of proteins, and the generated sequences are further evaluated based on sequence, structure, and function. In terms of sequence, protein solubility, DEDAL score (used to describe the homology between the designed sequence and the original sequence), BLOSUM62 substitution matrix score (BLAST amino acid substitution penalty), and ESM2 protein language model score (used to describe the possibility of the original sequence evolving to the designed sequence) are selected. In terms of structural indicators, root mean square distance error RMSD (used to evaluate the overall similarity of the structure), TM score (used to evaluate the alignment of protein templates), pLDDT (ESMfold gives the confidence for evaluating the structure) and secondary structure recovery are selected.

[0089] like Figure 5 As shown in the figure, when designing shorter (200-300 amino acid residues) proteins, the CHIEF model of the present invention outperforms the other five models in five sequence indicators, namely solubility, BLOSUM62, sequence recovery, DEDAL homology score, and wild-type edge probability. In terms of structural indicators, TM score, structure confidence score, and distance root mean square error are all optimal. Secondary structure sequence recovery ranks second only to ESM-IF.

[0090] like Figure 6 As shown in the figure, when designing proteins of medium length (500-600 amino acid residues), the CHIEF model of the present invention outperforms the other five models in five of the six indicators, and MMP ranks second only to ESM-IF. In the structural indicators, the root mean square error of distance, the secondary structure recovery, and the pLDDT performance are better than other models, and the TM score ranks second.

[0091] For designing long proteins (greater than 900 amino acid residues), such as Figure 7 As shown, the CHIEF model of the present invention also exhibits stable performance, with the best performance in five sequence indicators far superior to the second place, and the performance in four structural indicators is worse than other models. The ESM-IF, which performed well before, performed worse than other models in all aspects, indicating that the ESM-IF is insufficient in its ability to design long proteins, while the CHIEF model has better design capabilities for long proteins.

[0092] Comprehensive protein design of different lengths, such as Figure 8 As shown, the proteins involved in the CHIEF model of the present invention have stable performance in both sequence indicators and structural indicators, and also have the ability to design long proteins.

[0093] Further, if Fig. 9 As shown, in order to evaluate the ability of the CHIEF model of the present invention to design functional proteins, the CHIEF model of the present invention evaluates the predictive ability of mutant functions. The mutant effect prediction task is closely related to the protein design task for the inverse folding model, and the model that can accurately predict the mutant effect theoretically has a higher protein design success rate. Deep mutation scanning (DMS) is a data set associated with protein mutants and their biological functions. By comparing the model's scoring of mutant sequences with the experimentally obtained mutant function scores, the CHIEF model of the present invention ranks second among the various models, better than ProteinMPNN-soluble and ProteinMPNNvanilla, as well as PiFold, and Frame2seq. A correlation of 0.7 was reached. ESM-IF, which ranks first, performs better because it already includes mutant function information in training. Even if the CHIEF model of the present invention does not train mutant data, this also reflects the generalization ability of the CHIEF model of the present invention for data modalities that have not been seen.

[0094] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A protein design method based on deep learning, characterized in that: The following steps are involved: Compute amino acid distributions in protein sequences from several pre-trained models; Then, the distribution of each amino acid residue of the protein in several pre-trained models is calculated respectively; Then, the amino acid residue distribution in the pre-trained model is merged into the new dimension, and the merged distribution is input into the integrated network based on self-attention to obtain the attention output Z; Processing the attention output Z through a multi-layer feedforward neural network to obtain a result H; Then the result H obtained by the multi-layer feedforward neural network is projected to through the output linear layer and the unnormalized score is obtained; The unnormalized scores are then normalized by the function to obtain the final protein sequence distribution.

2. A protein design method based on deep learning as claimed in claim 1, characterized in that: Several pre-trained models are integrated into a heterogeneous pre-trained inverse folding model, including ProteinMPNN-vanilla, ProteinMPNN-soluble, ESM-IF, Frame2seq, and PiFold; the pre-trained model is a protein inverse folding model, the original input is the three-dimensional coordinates of the protein amino acid residue skeleton, and the output is the protein amino acid residue sequence distribution.

3. A protein design method based on deep learning as claimed in claim 2, characterized in that: Then, the amino acid distributions of several pre-trained models are merged into a new dimension, and the merged distribution is input into the integrated network based on self-attention. Specifically, the amino acid distribution S of each base model is merged to obtain Among them, M is the number of base models, L is the number of amino acid residues, C is the type of amino acid, and the merged S concat Enter self-attention based ensemble networks.

4. A protein design method based on deep learning as claimed in claim 3, characterized in that: The merged S concat Expand along the amino acid length L dimension to get S flat , then S flat Through the linear layer W inproj Projecting to the latent variable space, we get S hidden Then S hidden Perform sine and cosine position encoding, and then convert the position encoded S hidden Input into the Transformer encoder based on the multi-head attention mechanism to obtain the encoded attention output Z.

5. A protein design method based on deep learning as claimed in claim 1, characterized in that: The multi-layer feedforward neural network includes two fully connected layers FC and ReLU activation function.

6. A protein design method based on deep learning as claimed in claim 5, characterized in that: The attention output Z is processed by a multi-layer feedforward neural network to obtain the result H, and then the attention output Z is processed by a multi-layer feedforward neural network, and then the result H obtained by the multi-layer feedforward neural network is projected to through the output linear layer and obtains the unnormalized score logits, and then the unnormalized score is normalized by the Softmax function to first obtain the amino acids that meet the given structure, and finally obtain the final protein sequence distribution.

7. A protein design method based on deep learning as claimed in claim 6, characterized in that: It also includes comparing the probability distribution of the protein sequence normalized by the Softmax function with the original sequence, and calculating the loss using the cross entropy loss function.

8. A protein design method based on deep learning as claimed in claim 7, characterized in that: The cross entropy loss function is: Where S ij is the amino acid index at position i in the target sequence, The index of the sequence predicted by the model.

9. A protein design method based on deep learning as claimed in claim 7, characterized in that: It also includes using the back-propagation algorithm to calculate the gradient of the loss function to the model parameters, updating the model parameters according to the gradient to minimize the loss function value until the loss function converges.

Images