A fixed-backbone protein sequence design method based on masked node classification in graph networks

Abstract

A fixed-skeleton protein sequence design method based on graph network mask node classification, in the offline stage, the graph neural network is trained by constructing the protein nearest neighbor graph of the sample as a training set; in the online stage, the nearest neighbor graph of the protein structure of the missing sequence is processed by the trained graph neural network to obtain the missing amino acid category probability, and then the category probability is sampled to obtain the predicted missing amino acid sequence. The present invention adds more structural constraints, adds distance and relative angle features to the connected amino acid pairs in the protein graph and the dihedral angle features of each amino acid, and then establishes a neighbor graph, realizing a more efficient graph network based on the Transformer multi-head attention mechanism and exploring the optimal amino acid mask ratio.

Description

Technical Field

[0001] The present invention relates to a technology in the field of bioengineering, specifically a fixed-skeleton protein sequence design method based on graph network mask node classification. Background Art

[0002] Protein design is an important method to study the relationship between protein sequence and protein structure. It designs new proteins by designing amino acid sequences with specific structures. The current technology of protein design using deep learning generally divides protein design into two steps. First, the protein skeleton is generated according to the required function, and then the protein design of the fixed skeleton is carried out, striving for a specific and stable three-dimensional structure.

[0003] ProteinSolver (https: / / github.com / ostrokach / proteinsolver) is a fixed-skeleton protein design method. This technology models protein design as a constraint satisfaction problem (CSP). Specifically, the structural features of the protein are used as constraints, and some amino acid types are masked, the types of masked amino acids are predicted, and the predicted new sequences are used to fill the specific target structure. However, the structural features used by this technology are not sufficiently constrained. Only the distance matrix is ​​used for constraints, and a more accurate protein 3D structure representation cannot be extracted. The same distance matrix will correspond to too many different homologous sequences, resulting in a large difference between the generated sequence folded structure and the required structure. At the same time, the network structure used is not efficient enough. Summary of the invention

[0004] In response to the problems of insufficient structural feature constraints and inefficient network models in the existing ProteinSolver, the present invention proposes a fixed-skeleton protein sequence design method based on graph network mask node classification, which adds more structural constraints, adds distance and relative angle features to the connected amino acid pairs in the protein graph and the dihedral angle features of each amino acid, and then establishes a neighbor graph, realizes a more efficient graph network based on the Transformer multi-head attention mechanism, and explores the optimal amino acid mask ratio.

[0005] The present invention is achieved through the following technical solutions:

[0006] The present invention relates to a fixed-skeleton protein sequence design method based on graph network mask node classification. In the offline stage, the graph neural network is trained by constructing the protein nearest neighbor graph of the sample as a training set; in the online stage, the nearest neighbor graph of the protein structure of the missing sequence is processed by the trained graph neural network to obtain the missing amino acid category probability, and then the category probability is sampled to obtain the predicted missing amino acid sequence.

[0007] The protein nearest neighbor map is obtained by:

[0008] Step 1) Obtain the C in the 3D structure corresponding to the protein sequence α , C, N, O coordinates;

[0009] Step 2) Use the C α The coordinates of The closest neighbor graph is established based on the distance between the two amino acids. α Coordinate distance is less than ② When the number of neighboring nodes of any amino acid node is greater than 32, select the node with the closest distance to C α The nearest 32 amino acids are used as neighbor nodes.

[0010] Step 3) further constructs the node features and edge features of the nearest neighbor graph in step 2): the node features use the three dihedral angles of the protein backbone One-hot encoding (21 categories) of amino acid types (1-20 categories); edge features use the C of paired amino acids α The distance between atoms ij and the relative direction α ij , specifically: Where: C i ,C j Two amino acids C α The coordinates of the atoms.

[0011] Step 4) According to the mask ratio, the amino acid type information of some amino acid nodes in the nearest graph obtained in step 3) is masked, that is, replaced by one-hot encoding of the 21st category, and a data set with missing sequences is constructed to predict the missing sequences based on the structural features and the features of the non-missing sequences.

[0012] The mask ratio is preferably 30-60%.

[0013] The training set is specifically obtained by obtaining a CATH 40% non-redundant data set, selecting sequences with a length of 50-500, and then dividing them into a training set, a validation set, and a test set in a ratio of 80:15:5.

[0014] The graph neural network includes: an input layer, an encoding layer and an output layer, wherein: the input layer transforms the edge and node features in the input protein nearest neighbor graph to unify the feature dimensions of the edges and nodes, the encoding layer further extracts the features of the nodes and edges through a Transformer-based graph network, and the output layer uses a fully connected layer to convert the node features of the amino acid nodes to 20 dimensions, and then calculates the predicted amino acid category probability through softmax.

[0015] The input layer refers to: aligning the feature dimensions of edges and nodes, specifically: h_x = LayerNorm(Linear(σ(Linear(x)))), h_edge = LayerNorm(Linear(σ(Linear(edge)))), where: LayerNorm is the layer normalization layer, Linear is the fully connected layer, σ is the activation function, x is the amino acid node feature, and edge is the edge feature.

[0016] The encoding layer is composed of n identical transformer-based graph network modules, each of which includes a multi-head attention layer and a forward connection layer, wherein: the multi-head attention layer calculates the amino acid node features, specifically: Q = MultiHead(Q, K, V) + Q, MultiHead(Q, K, V) = Concat(head1, ..., head h )W o , Where: Q is a certain amino acid node feature, K, V is the sum of the node features of its neighboring nodes and the features of the edges; the forward connection layer calculates the amino acid node features, specifically: x = Linear(Linear(Dropout(α(Linear(LayerNorm(x)))))+x), where: x is the amino acid node feature, LayerNorm is the layer regularization layer, Linear is the fully connected layer, σ is the activation function, and Dropout is the dropout layer.

[0017] The amino acid category probability logits=Softmax(Linear(x)), wherein x is the amino acid node feature, Linear is the fully connected layer, and softmax is a normalized exponential function.

[0018] The graph neural network is trained to predict the average accuracy of the masked amino acid types, that is, the recovery rate evaluation index.

[0019] Preferably, the trained graph network is input according to the nearest neighbor graph of the protein of the sequence to be tested to obtain the corresponding amino acid category probability, and then sampled one hundred times and compared with the true label of the amino acid, and the average prediction accuracy of the masked amino acid type is calculated as the recovery rate, and the final test recovery rate is calculated after five-fold cross validation.

[0020] The present invention relates to a system for implementing the above method, comprising: a nearest neighbor graph generation unit and a graph network unit, wherein: the nearest neighbor graph generation unit constructs edge and node features and builds a graph according to the protein 3D structure, and the graph network unit predicts missing sequences according to structural features and non-missing sequence features.

[0021] The graph network unit includes: an input layer, an encoding layer, and an output layer, wherein: the input layer transforms the edge and node features in the input protein nearest neighbor graph to unify the feature dimensions of the edges and nodes; the encoding layer further extracts the features of the nodes and edges through a Transformer-based graph network; the output layer uses a fully connected layer to convert the node features of the amino acid nodes to 20 dimensions, then calculates the predicted amino acid category probability through softmax, and then samples the category probability to obtain the predicted missing amino acid sequence. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a flow chart of the present invention;

[0023] Figure 2 The proportion of different types of amino acids in the training, validation and test set of the embodiment;

[0024] Figure 3 This is a graph showing the recovery rate of amino acids in the example as a function of mask ratio. DETAILED DESCRIPTION

[0025] like Figure 1 As shown in the figure, this embodiment relates to a fixed-skeleton protein sequence classification method based on graph network mask node classification, using the CATH 40% non-redundant data set, selecting sequences with a length of 50-500, and dividing them into a training set, a validation set, and a test set in a ratio of 80:15:5. Then, the C α The coordinates of The nearest neighbor graph is established as the distance between two amino acids C α The coordinates are less than When the two amino acids are neighbors, and when the number of neighbor nodes of a node is greater than 32, the distance C is selected. α The nearest 32 amino acids are used as neighbor nodes. The feature dimension of the node feature of the constructed graph is 24, and the feature dimension of the edge feature is 4.

[0026] like Figure 1 As shown, the graph network in this embodiment includes an input layer, an encoding layer, and an output layer, wherein: the input layer transforms the feature dimension of edges and nodes to 128. The encoding layer uses 6 identical transformer-based graph network blocks, the hidden layer feature dimension of the graph network is 128, the number of heads of the multi-head attention mechanism is 4, and the dropout ratio is 0. The output layer transforms the feature dimension of the node to 20.

[0027] The graph network uses Adam as the optimizer, with an initial learning rate of 0.001, a weight decay of 1e-6, and a learning rate adjustment method of CosineAnnealing. It is trained for a total of 100 epochs, of which the warmup epochs is 5. The batch size of the dataset is 64.

[0028] The evaluation indicator used in the training is the recovery rate: on the test set, the output logits are obtained using the graph network, and then the logits are sampled one hundred times and compared with the true labels of the amino acids, and the average prediction accuracy of the amino acid type of the masked node is calculated as the recovery rate.

[0029] The prediction results of the experiment are as follows:

[0030]

[0031] This embodiment improves the recovery rate of ProteinSolver by about 2.7% when the mask rate is 30%, 50%, and 80%, and according to Figure 2 , the predicted results are much higher than the random results, eliminating the influence of different types of amino acid distributions in the data set.

[0032] like Figure 3 As shown in the figure, as the mask rate increases, the recovery rate increases first and then decreases. At the beginning, the mask ratio is relatively low, the noise is relatively large, the learned features are not clean enough, and the feature expression ability is weak. As the mask ratio increases, the noise gradually decreases and the feature expression ability gradually increases, until the mask ratio is too large and the appropriate feature expression cannot be learned from the effective features. It can be seen that the mask ratio is preferably 30%-60% to obtain better results.

[0033] Compared with the prior art, the recovery rate of deleted sequences in this method is about 2.7% higher than that of ProteinSolver.

[0034] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principle and purpose of the present invention. The protection scope of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. Each implementation scheme within its scope shall be subject to the constraints of the present invention.

Claims

1. A method for designing protein sequences with a fixed skeleton based on mask node classification in a graph network, characterized in that: In the offline stage, the graph neural network is trained by constructing the nearest graph of the sample protein as a training set; in the online stage, the nearest graph of the protein structure of the missing sequence is processed by the trained graph neural network to obtain the missing amino acid category probability, and then the category probability is sampled to obtain the predicted missing amino acid sequence; The protein nearest neighbor map is obtained by: Step 1) Obtain the C in the 3D structure corresponding to the protein sequence α , C, N, O coordinates; Step 2) Use the C α The coordinates of The closest neighbor graph is established based on the distance between the two amino acids. α Coordinate distance is less than ② When the number of neighboring nodes of any amino acid node is greater than 32, select the node with the closest distance to C α The nearest 32 amino acids are used as neighbor nodes; Step 3) further constructs the node features and edge features of the nearest adjacent graph in step 2): Node features use the three dihedral angles of the protein backbone One-hot coding of 1-20 amino acid types; The edge features use the C of the paired amino acids α The distance between atoms ij and the relative direction α ij , specifically: Where: C i ,C j Two amino acids C α Coordinates of atoms; Step 4) According to the mask ratio, the amino acid type information of some amino acid nodes in the nearest graph obtained in step 3) is masked, that is, replaced by one-hot encoding of the 21st category, and a data set with missing sequences is constructed to predict the missing sequences based on the structural features and the features of the sequences that are not missing; The mask ratio is 30-60%; The graph neural network includes: an input layer, an encoding layer and an output layer, wherein: the input layer transforms the edge and node features in the input protein nearest neighbor graph to unify the feature dimensions of the edges and nodes, the encoding layer further extracts the features of the nodes and edges through a Transformer-based graph network, and the output layer uses a fully connected layer to convert the node features of the amino acid nodes to 20 dimensions, and then calculates the predicted amino acid category probability through softmax.

2. The protein sequence design method according to claim 1, characterized in that: The training set is specifically obtained by obtaining a CATH 40% non-redundant data set, selecting sequences with a length of 50-500, and then dividing them into a training set, a validation set, and a test set in a ratio of 80:15:

5.

3. The protein sequence design method according to claim 1, characterized in that: The input layer refers to: aligning the feature dimensions of edges and nodes, specifically: h_x = LayerNorm(Linear((Linear(x)))), h_edge = LayerNorm(Linear(σ(Linear(edge)))), where: LayerNorm is the layer normalization layer, Linear is the fully connected layer, σ is the activation function, x is the amino acid node feature, and edge is the edge feature.

4. The protein sequence design method according to claim 1, characterized in that: The encoding layer is composed of n identical transformer-based graph network modules, each of which includes a multi-head attention layer and a forward connection layer, wherein: the multi-head attention layer calculates the amino acid node features, specifically: Q = MultiHead(Q, K, V) + Q, MultiHead(Q, K, V) = Concat(head1, ..., head h )W o , Where: Q is a certain amino acid node feature, K, V is the sum of the node features of its neighboring nodes and the features of the edges; the forward connection layer calculates the amino acid node features, specifically: x = Linear(Linear(Dropout(α(Linear(LayerNorm(x)))))+x), where: x is the amino acid node feature, LayerNorm is the layer regularization layer, Linear is the fully connected layer, σ is the activation function, and Dropout is the dropout layer.

5. The protein sequence design method according to claim 1, characterized in that: The amino acid category probability logits=Softmax(Linear(x)), wherein x is the amino acid node feature, Linear is the fully connected layer, and softmax is a normalized exponential function.

6. The protein sequence design method according to claim 1, characterized in that: The graph neural network is trained to predict the average accuracy of the masked amino acid types, i.e., the recovery rate evaluation index, specifically: the trained graph network is input according to the nearest graph of the protein of the sequence to be tested to obtain the corresponding amino acid category probability, and then sampled one hundred times and compared with the true label of the amino acid, and the average prediction accuracy of the masked amino acid type is calculated as the recovery rate, and the final test recovery rate is calculated after five-fold cross validation.

7. A system for implementing the protein sequence design method according to any one of claims 1 to 6, characterized in that: include: The nearest graph generation unit and the graph network unit, wherein: the nearest graph generation unit constructs edge and node features and builds a graph according to the protein 3D structure, and the graph network unit predicts the missing sequence according to the structural features and the non-missing sequence features.

8. The system according to claim 7, characterized in that: The graph network unit includes: an input layer, an encoding layer, and an output layer, wherein: the input layer transforms the edge and node features in the input protein nearest neighbor graph to unify the feature dimensions of the edges and nodes; the encoding layer further extracts the features of the nodes and edges through a Transformer-based graph network; the output layer uses a fully connected layer to convert the node features of the amino acid nodes to 20 dimensions, then calculates the predicted amino acid category probability through softmax, and then samples the category probability to obtain the predicted missing amino acid sequence.

Images