Protein function prediction method and system based on deep learning

Abstract

The invention relates to the field of bioinformatics, and provides a protein function prediction method and system based on deep learning. The method comprises the following steps: constructing a multi-modal data set of protein; performing feature extraction on the multi-modal data set through an encoder to obtain multi-modal features; training a neural network through the multi-modal features to obtain a prediction model; and predicting the function of the to-be-detected protein through the prediction model to obtain a prediction result. The protein function prediction accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics technology, and in particular to a method and system for predicting protein function based on deep learning. Background Technology

[0002] Protein function prediction is one of the core problems in structural biology and computational biology. Traditional protein function annotation mainly relies on methods such as sequence homology alignment, evolutionary analysis, and experimental verification. In recent years, with the rapid development of deep learning technology, functional prediction methods based on protein language models have made significant progress. Pre-trained protein language models such as ESM (Evolutionary Scale Modeling), ProtBERT, and ProtT5, through unsupervised learning on large-scale protein sequence data, can effectively capture evolutionary information and local structural patterns in amino acid sequences, and perform excellently in tasks such as protein secondary structure prediction, contact map prediction, and functional classification.

[0003] However, the function of proteins is highly dependent on their three-dimensional spatial structure. Although traditional protein language models have achieved certain results in structure and function prediction, they rely only on sequence information and cannot explicitly model the spatial relationships and geometric constraints between residues, making it difficult to accurately predict functional regions such as binding sites and catalytic residues. Summary of the Invention

[0004] This invention provides a protein function prediction method and system based on deep learning to overcome the shortcomings of existing technologies.

[0005] This invention provides a deep learning-based method for predicting protein function, comprising: S1: Construct a multimodal dataset of proteins; S2: Extract features from the multimodal dataset using an encoder to obtain multimodal features; S3: Train the neural network using the multimodal features to obtain the prediction model; S4: The prediction model is used to predict the function of the protein to be tested, and the prediction results are obtained.

[0006] According to the protein function prediction method based on deep learning provided by the present invention, the multimodal dataset in step S1 includes: image modality data and language modality data.

[0007] According to the protein function prediction method based on deep learning provided by the present invention, step S1 further includes: S11: Standardize and render the PDB structure file of the protein to obtain image modality data including multi-view images; S12: Standardize the amino acid sequence to obtain language modality data.

[0008] According to the protein function prediction method based on deep learning provided by the present invention, step S2 further includes: S21: Spatial feature extraction is performed on the image modal data in the multimodal dataset using the Vision Transformer encoder to obtain structural image feature vectors; S22: The language modality data in the multimodal dataset is semantically encoded using a Decoder-only Transformer encoder to obtain a sequence semantic feature vector; S23: Perform cross-modal feature fusion of the structural image feature vector and the sequence semantic features to obtain multimodal features.

[0009] According to the protein function prediction method based on deep learning provided by the present invention, step S3 further includes: S31: The multimodal feature representation is linearly transformed using a fully connected classifier of a neural network to obtain the predicted probability distribution; S32: Perform joint optimization on the loss values ​​corresponding to the predicted probability distribution to obtain the total loss function; S33: Based on the loss function, the model parameters of the neural network are updated by gradient to obtain the prediction model.

[0010] According to the protein function prediction method based on deep learning provided by the present invention, step S32 further includes: S321: By applying a contrastive learning loss function to the structural image feature vector and the sequence semantic feature vector in the multimodal features, a consistency constraint is applied to obtain the modality alignment loss; S322: The cross-entropy loss and the modal alignment loss are jointly optimized by weighted summation to obtain the total loss function.

[0011] According to the protein function prediction method based on deep learning provided by the present invention, step S4 further includes: S41: Collect the PDB structure file of the protein to be tested and the corresponding amino acid sequence to obtain input data; S42: Preprocess the input data to obtain preprocessed data; S43: Input the preprocessed data into the prediction model to obtain the prediction result.

[0012] According to the protein function prediction method based on deep learning provided by the present invention, the prediction result in step S43 includes: protein function and confidence score.

[0013] According to the deep learning-based protein function prediction method provided by the present invention, the protein function includes: residue-level function and overall protein function, wherein the residue-level function includes: catalytic residues, binding residues and structural residues.

[0014] This invention also provides a deep learning-based protein function prediction system, comprising: Modules: Used to build multimodal datasets of proteins; Extraction module: used to extract features from the multimodal dataset to obtain multimodal features; Training module: used to train a neural network using the multimodal features to obtain a prediction model; The prediction module is configured with the prediction model trained by the training module, and is used to predict the function of the protein to be tested and obtain the prediction result.

[0015] This invention provides a deep learning-based protein function prediction method and system. First, it standardizes and renders protein PDB structure files to generate six-view structural images, enabling comprehensive capture of the protein's three-dimensional geometric information from multiple spatial dimensions. Compared to traditional single-view or abstract diagram representation methods, this significantly enhances the completeness and expressive power of structural features, allowing the model to identify key functional regions that cannot be observed from a single angle, such as catalytic active centers hidden inside the protein or binding pockets located on the dorsal side of the molecule. This improves the utilization rate of spatial structural information and lays a solid data foundation for subsequent functional prediction. Subsequently, an encoder extracts features. For image features, it automatically learns the geometric correspondence and spatial consistency between different viewpoints, improving feature extraction accuracy, particularly in identifying functional protrusions, depressions, and channel structures on the protein surface. The accuracy of directly converting these features into functional predictions is significantly improved. For semantically encoded features, a unidirectional autoregressive mechanism better simulates sequence dependencies during protein folding, offering advantages over bidirectional encoders in processing directional features of protein sequences, thus improving the quality of sequence semantic representation, especially in capturing functional synergistic effects between distant residues. In supervised learning for protein function prediction, a joint optimization mechanism combining modality consistency constraints not only ensures the accuracy of residue-level function prediction, but also strengthens the semantic consistency of different modality feature representations through contrastive learning loss, avoiding modality bias problems. This allows the model to maintain stable prediction performance even when faced with missing structural information or incomplete sequence information. Furthermore, it enables the model to achieve efficient training with limited computational resources, shortening training time, while ensuring the stability of model convergence and the superiority of final performance.

[0016] Overall, this invention improves the prediction accuracy in protein function prediction tasks and demonstrates significant advantages in specific applications such as catalytic residue identification, binding pocket localization, and functional region segmentation. It provides strong technical support for important application areas such as enzyme engineering, drug target discovery, and protein design, and has significant scientific value and broad application prospects. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 A schematic diagram of a protein function prediction method based on deep learning provided by the present invention; Figure 2 This is a schematic diagram of the structure of a protein function prediction system based on deep learning provided by the present invention.

[0019] Figure labels: 100, construction module; 200, extraction module; 300, training module; 400, prediction module. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0021] The embodiments of the present invention are described below with reference to the figures.

[0022] like Figure 1 As shown, the present invention provides a protein function prediction method based on deep learning, comprising: S1: Construct a multimodal dataset of proteins.

[0023] Step S1 further includes: S11: Standardize and render the protein's PDB structure file to obtain image modal data including multi-view images.

[0024] Furthermore, the PDB structure file contains information such as the spatial coordinates, residue types, chain identifiers, and temperature factors of each atom in the protein. This invention uses PyMOL software for normalized rendering, and its built-in molecular rendering engine parses and processes the PDB file. First, the file header information is read to obtain the basic properties of the protein. Then, the ATOM and HETATM records are parsed line by line to extract the X, Y, and Z three-dimensional coordinates of each atom. In the subsequent normalized rendering process, this invention uses PyMOL to convert the atomic coordinate data into an internal molecular object data structure, establishes the chemical bond connections between atoms, and calculates the van der Waals radius and charge distribution of each atom. The multi-view image includes six viewing angles: front, back, left, right, top, and bottom. The rendering algorithm uses ray tracing. The final image data output format is a three-dimensional array, where the first dimension represents the six viewing angles, the second and third dimensions represent the height and width of the image in pixels, and the fourth dimension represents the RGB color channels.

[0025] S12: Standardize the amino acid sequence to obtain language modality data.

[0026] Furthermore, in step S12, the original sequence data is first cleaned by removing non-character data such as spaces, newlines, and tabs, and checking for non-standard amino acid characters. A sequence verification process is then performed, using regular expression pattern matching to check if each character belongs to the standard amino acid set. Non-standard residues such as pyrrolidone and hydroxyproline are replaced with the closest standard amino acid. In sequence length standardization, excessively short sequences (less than 30 residues) are padded with repeating sequences, while excessively long sequences (more than 1000 residues) are segmented into multiple subsequences using a sliding window method. Finally, a single-letter format conversion is performed, uniformly converting three-letter amino acid codes to single-letter codes. A mapping dictionary is established for batch replacement operations, resulting in a language modality data format of a string list, i.e., a standardized string format of single-letter format.

[0027] The multimodal dataset in step S1 includes image modal data and language modal data.

[0028] Furthermore, after obtaining the image modality data and language modality data in step S1, alignment processing was also performed, that is, matching was performed through protein identifiers to ensure that the six-view image corresponding to each PDB structure file is established in a one-to-one correspondence with its amino acid sequence, and a data index was established to create a dictionary data structure to store the mapping relationship between protein ID, image file path and sequence string.

[0029] S2: Extract features from the multimodal dataset using an encoder to obtain multimodal features.

[0030] Step S2 further includes: S21: Spatial feature extraction is performed on the image modal data in the multimodal dataset using the Vision Transformer encoder to obtain structural image feature vectors.

[0031] Furthermore, the Vision Transformer encoder described herein is an image processing architecture based on a self-attention mechanism. This invention uses the Vision Transformer encoder to segment images into fixed-size image blocks and perform serialization processing. Specifically, firstly, six-view image data is input, and the data is standardized through an image preprocessing module, mapping pixel values ​​from the 0-255 range to the 0-1 range. Subsequently, each image is divided into multiple patches. Then, for each patch, it is converted into an embedding vector through a linear projection layer. A patch embedding vector is obtained through matrix multiplication. Position encoding uses learnable position embedding vectors, assigning a unique position identifier to each patch position to obtain a position embedding matrix. This matrix is ​​added to the patch embedding vector to obtain the final input representation. Then, multi-head attention merges the above input representations into a feature representation through a concatenation operation. After processing by a feedforward neural network and global average pooling, the final structural image feature vector is obtained.

[0032] S22: The language modality data in the multimodal dataset is semantically encoded using a Decoder-only Transformer encoder to obtain a sequence semantic feature vector.

[0033] Furthermore, in step S22, this invention processes the serialized amino acid data using a Decoder-only Transformer encoder. The input language modality data is a standardized amino acid sequence. First, each amino acid character is converted into a corresponding integer index through a vocabulary mapping. Then, the amino acid embedding layer converts the integer indices into dense vector representations. The embedding vector for each amino acid is obtained through an index lookup operation. Position encoding uses fixed-position encoding generated by sine and cosine functions. The input representation is obtained by adding the amino acid embedding vector and the position encoding vector. The Decoder-only Transformer contains 12 decoder layers, each containing a masked multi-head self-attention sublayer and a feedforward neural network sublayer. Residual connections and layer normalization operations are added after each sublayer. The feedforward neural network uses two linear transformation layers with GELU activation function. Finally, the Decoder-only Transformer encoder outputs a sequence semantic feature vector.

[0034] S23: Perform cross-modal feature fusion of the structural image feature vector and the sequence semantic features to obtain multimodal features.

[0035] The cross-modal feature fusion in this invention employs a Cross-Attention mechanism to achieve deep interaction between image features and sequence features. The structural image feature vector and the sequence semantic feature vector are directly used as inputs to the fusion module. In the specific Cross-Attention calculation, the structural image feature vector serves as the query input, while the sequence semantic feature vector serves as both the key and value inputs. The attention score is calculated by matrix multiplication of the transposes of Q and K to obtain the attention score matrix. Subsequently, softmax normalization is used to calculate the attention weight for each sequence position. The final fused feature is calculated by weighted summation of the attention weights and value vectors. The multimodal feature generation process involves concatenating the fused feature with the original sequence features, i.e., adding the fused feature with the sequence semantic feature vector element-wise. The resulting feature vector at each position of the final multimodal feature representation dimension simultaneously contains the sequence semantic information of the corresponding amino acid and the overall structural spatial information of the protein.

[0036] S3: Train the neural network using the multimodal features to obtain a prediction model.

[0037] Step S3 further includes: S31: The multimodal feature representation is linearly transformed using a fully connected classifier of a neural network to obtain the predicted probability distribution.

[0038] Further, in step S31, the fully connected classifier receives the multimodal feature vector output from the cross-modal feature fusion layer. The fully connected classifier uses three linear transformation layers. The first layer maps the input feature vector to the hidden layer dimension and performs a non-linear transformation using the ReLU activation function. The second layer continues the feature transformation, mapping the hidden layer dimension features output from the first layer to another hidden layer dimension, also using the ReLU activation function. The third layer is the output layer, mapping the hidden layer features to the dimension of the number of categories C, where C corresponds to the total number of protein functional categories. During the linear transformation, the multimodal feature vector at each residue position is transformed through a weight matrix and a bias vector. The output layer uses the Softmax function to convert the linear transformation result into a probability distribution. The Softmax function converts the original score of each category into a probability value between 0 and 1, forming a prediction probability distribution matrix with dimensions N×C. Each row represents the predicted probability of a residue position for all functional categories.

[0039] S32: Perform joint optimization on the loss values ​​corresponding to the predicted probability distribution to obtain the total loss function.

[0040] Step S32 further includes: S321: By applying a contrastive learning loss function to the structural image feature vector and the sequence semantic feature vector in the multimodal features, a consistency constraint is applied to obtain the modality alignment loss.

[0041] Furthermore, the objective of the contrastive learning loss function is to enhance the similarity between different modal features of the same protein while reducing the similarity between features of different proteins. Specifically, firstly, a linear projection layer maps the feature vectors of the two input modalities to the same representation space. Then, the cosine similarity between the image features and sequence features of corresponding residue positions within the same protein is calculated. Contrastive learning loss is then applied based on this cosine similarity. The denominator of the loss function includes the similarity between positive and negative sample pairs. Positive sample pairs refer to different modal features at the same residue position, while negative sample pairs refer to feature pairs at different residue positions or different proteins. The final modality alignment loss is obtained by averaging the contrastive losses of all residue positions within a batch.

[0042] S322: The cross-entropy loss and the modal alignment loss are jointly optimized by weighted summation to obtain the total loss function.

[0043] Furthermore, the cross-entropy loss function calculates the difference between the predicted probability distribution and the true label. The cross-entropy loss of all residue positions within a batch is averaged to obtain the total cross-entropy loss. The final total loss function combines the cross-entropy loss and modality alignment loss through a weighted summation method. The total loss function takes into account both the accuracy of function prediction and the consistency of cross-modal feature alignment, thus achieving multi-objective joint optimization.

[0044] S33: Based on the loss function, the model parameters of the neural network are updated by gradient to obtain the prediction model.

[0045] Furthermore, the model parameter gradient update of this invention adopts the AdamW optimizer, the training process adopts gradient accumulation technology, and the training termination condition is that the validation set loss has no improvement for 10 consecutive epochs or reaches the preset maximum number of training epochs. Finally, the model parameters with the best performance on the validation set are saved as the prediction model.

[0046] S4: The prediction model is used to predict the function of the protein to be tested, and the prediction results are obtained.

[0047] Step S4 further includes: S41: Collect the PDB structure file of the protein to be tested and the corresponding amino acid sequence to obtain input data.

[0048] Furthermore, the input data includes the PDB structure file of the protein to be tested and the corresponding amino acid sequence. The PDB structure file is a standard protein three-dimensional structure data file obtained from the Protein Data Bank, which contains atomic coordinates, residue information and secondary structure data. The amino acid sequence is a single-letter amino acid sequence string parsed from the PDB file. The PDB file and the amino acid sequence are associated and matched through the protein ID to ensure data consistency.

[0049] S42: Preprocess the input data to obtain preprocessed data.

[0050] Furthermore, the preprocessing is divided into structural data preprocessing and sequence data preprocessing. For structural data preprocessing, PyMOL software is used to load the PDB file, execute the "remove solvent" command to remove water molecules, the "remove hetero" command to remove heteroatoms, and the "h_add" command to add hydrogen atoms. Uniform rendering parameters are set to generate RGB images from six orthogonal perspectives (front, back, left, right, top, and bottom). For sequence data preprocessing, the amino acid sequence is converted into a single-letter encoded string, truncated or padded to a fixed length of 512 characters. Insufficient parts are padded with the padding character "X", and excessive parts are truncated. The final preprocessed data includes six perspective images and corresponding numerical sequence vectors.

[0051] S43: Input the preprocessed data into the prediction model to obtain the prediction result.

[0052] Furthermore, after obtaining the preprocessed data in step S42, the preprocessed data is input into the prediction model for forward propagation calculation. After the prediction model extracts the features of the preprocessed data, the final features are input into the classification head: residue-level function prediction uses a fully connected layer (768-dimensional input, 3-dimensional output corresponding to catalytic / binding / structural residues) with softmax activation; overall protein function prediction uses a fully connected layer (768-dimensional input, number of functional categories output) with softmax activation; the confidence score is determined by the maximum probability value output by softmax.

[0053] The prediction results in step S43 include: protein function and confidence score. The protein function includes: residue-level function and overall protein function. The residue-level function includes: catalytic residues, binding residues, and structural residues.

[0054] Furthermore, the final prediction results are output in JSON format, containing three levels of data: residue-level prediction results are an array, with each element corresponding to an amino acid position, including residue type (catalytic / binding / structural) and confidence score (0-1); overall protein prediction results are a dictionary, including predicted functional category (such as ANTIMICROBIAL PROTEIN) and confidence score; raw data information includes the PDB ID used and amino acid sequence fragment.

[0055] like Figure 2 As shown, the present invention also provides a protein function prediction system based on deep learning, comprising: Module 100: Used to construct a multimodal dataset of proteins; Extraction module 200: used to extract features from the multimodal dataset to obtain multimodal features; Training module 300: used to train a neural network using the multimodal features to obtain a prediction model; The prediction module 400 is configured to use the prediction model trained by the training module 300 to predict the function of the protein to be tested and obtain the prediction result.

[0056] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0057] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0058] This invention discloses a deep learning-based protein function prediction method and system, aiming to address the problem of insufficient utilization of structural information in protein function prediction. By introducing structural images as a visual modality and inputting them together with sequence information into a multimodal large model, such as DeepSeek Janus, this invention achieves cross-modal joint modeling, thereby significantly improving the accuracy of protein function prediction. This invention can be used for residue-level function prediction (such as identifying catalytic residues and binding pockets) or overall protein function tag prediction, and is suitable for tasks such as protein function annotation, variant effect analysis, and protein engineering.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A protein function prediction method based on deep learning, characterized in that, include: S1: Construct a multimodal dataset of proteins; S2: Extract features from the multimodal dataset using an encoder to obtain multimodal features; S3: Train the neural network using the multimodal features to obtain the prediction model; S4: The prediction model is used to predict the function of the protein to be tested, and the prediction results are obtained.

2. The protein function prediction method based on deep learning according to claim 1, characterized in that, The multimodal dataset in step S1 includes: image modal data and language modal data.

3. The protein function prediction method based on deep learning according to claim 2, characterized in that, Step S1 further includes: S11: Standardize and render the PDB structure file of the protein to obtain image modality data including multi-view images; S12: Standardize the amino acid sequence to obtain language modality data.

4. The protein function prediction method based on deep learning according to claim 1, characterized in that, Step S2 further includes: S21: Spatial feature extraction is performed on the image modal data in the multimodal dataset using the Vision Transformer encoder to obtain structural image feature vectors; S22: The language modality data in the multimodal dataset is semantically encoded using a Decoder-only Transformer encoder to obtain a sequence semantic feature vector; S23: Perform cross-modal feature fusion of the structural image feature vector and the sequence semantic features to obtain multimodal features.

5. The protein function prediction method based on deep learning according to claim 1, characterized in that, Step S3 further includes: S31: The multimodal feature representation is linearly transformed using a fully connected classifier of a neural network to obtain the predicted probability distribution; S32: Perform joint optimization on the loss values ​​corresponding to the predicted probability distribution to obtain the total loss function; S33: Based on the loss function, the model parameters of the neural network are updated by gradient to obtain the prediction model.

6. The protein function prediction method based on deep learning according to claim 5, characterized in that, Step S32 further includes: S321: By applying a contrastive learning loss function to the structural image feature vector and the sequence semantic feature vector in the multimodal features, a consistency constraint is applied to obtain the modality alignment loss; S322: The cross-entropy loss and the modal alignment loss are jointly optimized by weighted summation to obtain the total loss function.

7. The protein function prediction method based on deep learning according to claim 1, characterized in that, Step S4 further includes: S41: Collect the PDB structure file of the protein to be tested and the corresponding amino acid sequence to obtain input data; S42: Preprocess the input data to obtain preprocessed data; S43: Input the preprocessed data into the prediction model to obtain the prediction result.

8. The protein function prediction method based on deep learning according to claim 7, characterized in that, The prediction results in step S43 include: protein function and confidence score.

9. The protein function prediction method based on deep learning according to claim 8, characterized in that, The protein functions include: residue-level functions and overall protein functions. The residue-level functions include: catalytic residues, binding residues, and structural residues.

10. A protein function prediction system based on deep learning, characterized in that, include: Modules: Used to build multimodal datasets of proteins; Extraction module: used to extract features from the multimodal dataset to obtain multimodal features; Training module: used to train a neural network using the multimodal features to obtain a prediction model; The prediction module is configured with the prediction model trained by the training module, and is used to predict the function of the protein to be tested and obtain the prediction result.

Images