A neuropeptide prediction method and system based on multimodal features and twin networks

Abstract

The present invention discloses a neuropeptide prediction method and system based on multimodal features and twin networks, belonging to the field of neuropeptide prediction technology, comprising the following steps: S1: constructing a benchmark dataset; S2: encoding NPs; S3: constructing a twin network module; S4: constructing a protein language module; S5: constructing a prediction model; and S6: using the model for prediction. The present invention constructs a balanced dataset and, based on this dataset, utilizes four encoding methods (token embedding encoding, word2vec embedding encoding, manual feature encoding, and protein language embedding encoding) to obtain feature vectors of neuropeptide sequences. The prediction model is constructed by combining the twin network module and the protein language module, enabling accurate identification of neuropeptides (NPs) with an accuracy exceeding that of currently available prediction methods.

Claims

1. A neuropeptide prediction method based on multimodal features and twin networks, characterized in that: The following steps are involved: S1: Building a benchmark dataset Establish a benchmark dataset and divide it into a training set and a test set; In step S1, the generation process of the benchmark dataset is as follows: S11: Neuropeptide sequence information was obtained through NeuroPep 2.0 and UniProt databases; S12: The screening retained neuropeptides consisting of 5 to 100 residues; S13: CD-HIT was used to eliminate redundant sequences, and finally 4463 neuropeptides and 4463 non-neuropeptides were obtained; S14: The dataset is divided into training set and test set in a ratio of 9:1; S15: Divide the training set into 10 10-fold cross validations to obtain the benchmark dataset. S2: NPs encoding The neuropeptide sequences of the benchmark dataset are encoded into feature vectors using four encoding methods: token embedding encoding, word2vec embedding encoding, manual feature encoding, and protein language embedding encoding; In step S2, the specific processing process of protein language embedding coding is as follows: S231: Select the pre-trained ProtT5 protein language model, namely ProtT5-XL-UniRef50, to extract the contextual semantic information of neuropeptides to generate embedding representations of peptide sequences and secondary structure sequences; S232: Select the average embedding feature of the last layer in the ProtT5 protein language model as the final embedding feature; In the step S231, the secondary structure sequence of the neuropeptide is obtained through the PHAT web server; In step S2, in order to extract comprehensive physicochemical information from the original sequence, the following 12 manual coding strategies were used for manual feature coding: AAC, DPC, CTriad, GAAC, GDPC, GTPC, SE, CTDC, CTDT, CTDD, Moran, and SOCN; S3: Twin network module construction Construct a twin network module through transformer encoder, MS-CNN, and Bi-GRU, take the vector results of token embedding and word2vec embedding as input, and input the result into the FC layer; S4: Protein language module construction A protein language module is constructed through another Bi-GRU. The obtained manual features and protein language embedding features are fused and input into the Bi-GRU unit of the protein language module to obtain the feature representation of neuropeptides, and the result is input into the FC layer. S5: Build a predictive model Combine the twin network module and the protein language module, train based on the feature vector in step S2, and establish a prediction model; S6: Using the model to make predictions The prediction model in step S5 is used to predict the neuropeptide sequences of the test set.

2. A neuropeptide prediction method based on multimodal features and twin networks according to claim 1, characterized in that: In step S2, the specific processing process of token embedding encoding is as follows: S211: Build a vocabulary using each type of residue as a label; S212: Map each tag to an integer; S213: pad the sequences with zeros so that the length of each sequence is a set value, where the set value is the number of residues in the longest sample in the benchmark data; S214: Convert the numeric vector into a matrix through the token embedding layer.

3. A neuropeptide prediction method based on multimodal features and twin networks according to claim 2, characterized in that: In step S2, the specific processing process of word2vec embedding coding is as follows: S221: Use samples from the training set and test set to train the CBOW model in Word2vec; S222: Divide peptide sequence into multiple words x 1, x2,x3,x4,…x n-1 ,x n , where each word is a k-mer, and a k-mer is a series of k consecutive amino acids. The CBOW model is used to predict the target word based on its context words. The objective function of the CBOW model is as follows: Among them, X j As the central word, X j-t ,…,X j-1 ,X j+1 ,X t+j represents the context word, P(X j |X j-t ,…,X j-1 ,X j+1 ,X t+j ) is the conditional probability, defined as follows: Among them, h j is the value of the input word vector, and the matrix M is the weight matrix of the hidden layer; S223: The k value of k-mer is set to 1 and 4 respectively, and two peptide sequence segmentation methods are obtained, thereby obtaining two CBOW models, and then obtaining two two-dimensional feature matrices, and then connecting the two two-dimensional feature matrices to obtain a connected matrix.

4. The neuropeptide prediction method based on multimodal features and twin networks according to claim 1, characterized in that: In step S3, the twin network module includes two identical sub-networks, each of which includes a transformer encoder, an MS-CNN, and a Bi-GRU connected in sequence; the specific processing process of the twin network module is as follows: S31: The result vector after token embedding encoding and word2vec embedding encoding is used as input; S32: The transformer encoder layer is used to capture the local and global information of the neuropeptide sequence. The specific process is as follows: S321:X i,embed is fed into an MSA block and an MLP block, X i,embed is the output of the transformer encoder embedding layer, represented as follows: X i,embed =Embedding(X i )+PosEmbedding(X i ); S322: layer normalization before each block, residual connection after each block, first encoder layer The output is: S323: Input the second encoder layer and generate the representation The output of the previous encoder layer is input into the next encoder layer. The transformer encoder consists of an embedding layer and n encoder layers. The output of the last encoder layer is S324: After the above operations, an embedded representation related to the context is obtained; S33: Use MS-CNN to capture information at different scales. The specific process is as follows: S331: Obtain feature vectors through convolution and use the maximum pooling layer for dimensionality reduction; S332: Connect the results of each convolutional layer to obtain the global feature representation of the peptide sequence; S34: Using Bi-GRU to capture long-range feature representations; In step S34, in the Bi-GRU, for each element in the input sequence, the output is calculated as follows: U t =σ(W U ·[h t-1 ,S t ]) Q t =σ(W Q ·[h t-1 ,S t ]) Among them, S t is the input; h t is the hidden state at time step t, h t-1 is the hidden state at time step t-1 or the initial hidden state at time 0; σ is the sigmoid activation function; tanh is the activation function; U t and is the update and reset gate vector; W U , W Q and W represent weight parameters; ° is the Hadamard product.

5. The neuropeptide prediction method based on multimodal features and twin networks according to claim 4, characterized in that: In the Bi-GRU unit of step S34, the input of the first layer is the hidden state of the previous layer After passing through K layers, the hidden state output of K layers is obtained Among them, l≥2.

6. A neuropeptide prediction system based on multimodal features and twin networks, characterized by: Identifying and predicting neuropeptide sequences using the neuropeptide prediction method based on multimodal features and a Siamese network as described in any one of claims 1 to 5, comprising: The dataset construction module is used to establish a benchmark dataset and divide the benchmark dataset into a training set and a test set; The feature representation module is used to encode the neuropeptide sequences of the benchmark dataset into feature vectors using four encoding methods: token embedding encoding, word2vec embedding encoding, manual feature encoding, and protein language embedding encoding; The model building module is used to train the training set using the twin network module and the protein language module to build a prediction model; The prediction module is used to predict new neuropeptide sequences using a prediction model.

Images