A method for predicting protein post-translational modifications based on multimodal deep learning

Abstract

This invention belongs to the field of bioinformatics and relates to a method for predicting protein post-translational modifications based on multimodal deep learning. First, multimodal feature extraction is performed on the input protein sequence and 3D structural data to obtain sequence feature vectors and structural feature vectors. Second, a cross-modal attention mechanism and an adaptive gating network are used for feature fusion. Then, the fused features are combined with disease type information, and the predicted probabilities are fine-tuned using a disease-specific coding network. Next, a multi-task learning framework is used to predict the site probabilities of various protein post-translational modification types in parallel. Finally, gradient backpropagation is used to calculate feature importance, and a comprehensive report is output based on variation impact analysis. This invention achieves high-precision, interpretable, and disease-aware protein post-translational modification prediction, providing an important computational analysis tool for revealing the molecular mechanisms of diseases and discovering precise drug targets.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics, specifically relating to a method for predicting protein post-translational modifications based on multimodal deep learning. This method can automatically integrate protein sequence and three-dimensional structural information, and through adaptive fusion and disease perception mechanisms, achieve high-precision, interpretable PTM (Post-Translational Modification) site prediction. Background Technology

[0002] Post-translational modifications (PTMs) are a core mechanism regulating cellular function and signal transduction, and their abnormalities are closely related to various diseases such as cancer and neurodegenerative diseases. Existing prediction methods, such as MultisiteDeep and DeepPhos, mainly rely on protein sequence information and have the following shortcomings: ① Single feature: They only use sequence information, ignoring the three-dimensional structural information that determines PTM site accessibility and enzyme specificity; ② Lack of context awareness: They do not consider the regulatory influence of disease type on PTM patterns and cannot provide disease-specific predictions; ③ Simple fusion mechanism: The multimodal feature fusion method is relatively rudimentary and fails to fully explore the deep correlation between sequence and structure; ④ Strong structure dependence: They rely too heavily on external databases such as AlphaFold2, resulting in insufficient prediction stability when structural data is missing; ⑤ Poor interpretability: They fail to provide feature importance analysis and variation impact assessment, making it difficult to support precision medicine applications. Therefore, there is an urgent need in this field for a PTM prediction method that can fuse multi-source information, has disease-awareness capabilities, is structurally robust, and is interpretable. Summary of the Invention

[0003] This invention proposes a protein post-translational modification prediction method based on multimodal deep learning, aiming to address the problems of insufficient feature utilization, lack of disease context awareness, and poor interpretability of results in existing technologies. The specific technical solution includes the following five steps:

[0004] Step 1: Multimodal Feature Extraction: The input consists of the protein's amino acid sequence data and its corresponding three-dimensional structural data. Parallel sequence feature extraction and structural feature extraction branches are used to extract deep features containing local patterns and long-range dependencies from the sequence, and multidimensional features including residue confidence, spatial conformation, solvent accessibility, and structural uncertainty from the structural data. The output is a unified sequence feature vector and structural feature vector, laying the foundation for subsequent fusion.

[0005] Step 2, Adaptive Feature Fusion: Using the sequence feature vector and structural feature vector output from Step 1 as input, a cross-modal attention mechanism is employed to establish a dynamic correlation mapping between sequence positions and structural positions. Then, a quality-aware adaptive gating network dynamically calculates and assigns fusion weights to each modality based on the confidence level of the structural data. The output is a deeply fused feature representation that reflects data quality, effectively improving the robustness and information richness of feature representation.

[0006] Step 3, Disease Conditional Prediction: The fused features obtained in Step 2 are combined with specific disease type information. A disease-specific encoding network maps disease categories to adjustment factors, which are then used to fine-tune the base prediction probabilities. The output is a prediction result modulated by the disease context, making the prediction more closely aligned with the biological background of the specific pathological state, thus achieving personalized PTM prediction.

[0007] Step 4, Multi-task PTM Prediction: Based on the conditional feature representation output from Step 3, the input is used. A multi-task learning framework is employed, using multiple parallel, weight-independent prediction heads to simultaneously predict the site probabilities of eight common PTM types, including serine / threonine phosphorylation, tyrosine phosphorylation, and lysine acetylation. The output is the probability of occurrence of each PTM type at each site, achieving collaborative inference of multiple modifications by a single model.

[0008] Step 5: Result Output and Interpretation: A deep interpretability analysis is performed on the predicted probabilities generated in Step 4. Importance scores of the input features are calculated using gradient backpropagation technology, and a detailed variation impact analysis module is used to quantify the perturbation effect of genetic variations on PTM potential. Finally, a comprehensive prediction report is output, including not only site confidence and disease adjustment factors, but also feature attribution and variation impact assessments, greatly enhancing the credibility and practical value of the prediction results.

[0009] A method for predicting protein post-translational modifications based on multimodal deep learning is proposed. Step 1 is implemented as follows: The system receives the protein sequence in FASTA format and UniProt ID, and obtains structural data via the AlphaFold2 API. Sequence feature extraction employs a CNN-BiGRU hybrid network: three parallel one-dimensional convolutional layers with kernel sizes of 3, 5, and 7, and filter numbers of 64, 128, and 256, respectively, followed by batch normalization, ReLU activation, and a dropout layer with a dropout rate of 0.2. The data is then input into a bidirectional GRU with 128 hidden units, outputting 256-dimensional sequence features. Structural feature extraction includes: pLDDT confidence score, normalized to 0-1; secondary structure using one-hot encoding; solvent accessibility score; and prediction alignment error matrix.

[0010] A protein post-translational modification prediction method based on multimodal deep learning is proposed. Step 2 is implemented as follows: a cross-modal attention and adaptive gating fusion mechanism is employed. Structural features are projected onto the sequence feature space via a linear layer. .in, This represents the input structural feature matrix. For sequence length, The original dimension of the structural features; Here is the weight matrix of the projection layer. For bias terms, The target dimension is consistent with the sequence feature dimension; The projected structural features are shown. The attention weight matrix is ​​calculated as follows: .in, For querying the matrix, The key matrix, This is the scaling factor, and also a key dimension for each attention head; The function is normalized along the last dimension to ensure that the sum of each row is 1. The resulting attention weight matrix... Each of its elements Quantified the first The position of the sequence is related to the first... The correlation strength between structural locations. The gating weights are generated by the fully connected layer and the sigmoid function: .in, This indicates a concatenation operation along the feature dimension; Weights and biases for the gating layer; The sigmoid function compresses the output to the (0,1) interval. Gating weights. Dynamic adjustment based on the average pLDDT confidence level of the structure: when pLDDT is below 70, for Applying an upper bound constraint to make it approach 1, the system mainly relies on sequence features; when pLDDT is higher than 90, applying a lower bound constraint to make it approach 0, the system mainly relies on structural features; in the intermediate range, a smooth transition occurs. The fused features are: .in, This indicates element-wise multiplication. This is the feature representation after adaptive fusion, which will be used as the input for step 3.

[0011] A multimodal deep learning-based method for predicting protein post-translational modifications, step 3 of which is implemented as follows: The system predefines eight disease categories, including cancer, neurodegenerative diseases, cardiovascular diseases, metabolic diseases, autoimmune diseases, infectious diseases, genetic diseases, and other diseases. Disease types are mapped to 32-dimensional vectors through an embedding layer. The adjustment factor is output through a two-layer fully connected network. : , .in, These are the parameters of the first layer. It is an intermediate activation vector; These are the parameters of the second layer; The function limits the output to Interval. Adjustment factor for the final output. It is an 8-dimensional vector, with each dimension corresponding to a type of adjustment intensity of PTM; the adjusted probability is: .in, The strength coefficient was determined through a grid search to adjust it; For the pruning function, ensure that the adjusted probability falls within the specified range. Within the effective range This is the final predicted probability after disease context modulation.

[0012] A multimodal deep learning-based method for predicting protein post-translational modifications (PTMs) is presented. Step 4 is implemented as follows: Eight PTM classes are predicted in parallel using a multi-task learning framework, including serine / threonine phosphorylation, tyrosine phosphorylation, lysine acetylation, lysine methylation, arginine methylation, lysine ubiquitination, lysine SUMOylation, and asparagine glycosylation. Each PTM class corresponds to a prediction head. The loss function is a weighted binary cross-entropy, with weights set according to the training set distribution as follows: serine / threonyl phosphorylation 0.25, tyrosine phosphorylation 0.25, lysine acetylation 0.125, lysine methylation 0.125, promethylation 0.125, lysine ubiquitination 0.0625, lysine SUMOylation 0.0625, and asparagine glycosylation 0.0625.

[0013] A protein post-translational modification prediction method based on multimodal deep learning is described, with step 5 implemented as follows: The system outputs the PTM type, confidence score, disease adjustment factor, and feature importance score. Feature importance is calculated via gradient backpropagation. The variation impact analysis includes probability change values, joint probability to log ratio ratio, and calculates empirical p-values ​​based on the background distribution of random variation. Attached Figure Description

[0014] Figure 1 This is a diagram illustrating the overall architecture of a protein post-translational modification prediction method based on multimodal deep learning.

[0015] Figure 2 This is a flowchart of the multimodal feature extraction and fusion module.

[0016] Figure 3 This is a flowchart of the disease type conditional perception module.

[0017] Figure 4 This is a flowchart for the results interpretation and variation impact analysis module. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and examples.

[0019] Multimodal feature extraction implementation. For example... Figure 1 and Figure 2 As shown, this step includes two parallel processing branches: sequence feature extraction and structural feature extraction. During system initialization, data processing parameters are configured as follows: the maximum sequence length is set to 1000 amino acid residues; excessively long sequences are truncated from the N-terminus; sequences shorter than 10 residues are considered invalid input. The technical details of sequence feature extraction are as follows: The input sequence is represented by a 22-dimensional one-hot encoding matrix, covering 20 standard amino acids and the special characters U (selenocysteine) and B (asparagine or aspartic acid). The encoding dictionary is specifically defined as {'A':0, 'R':1, 'N':2, 'D':3, 'C':4, 'Q':5, 'E':6, 'G':7, 'H':8, 'I':9, 'L':10, 'K':11, 'M':12, 'F':13, 'P':14, 'S':15, 'T':16, 'W':17, 'Y':18, 'V':19, 'U':20,} 'B':21}, for the unknown amino acid X, is encoded as an all-zero vector; undefined characters are encoded as uniformly distributed vectors with values ​​of 0.5 in each dimension. The sequence feature extraction neural network employs a three-layer parallel one-dimensional convolutional architecture. The first convolutional layer is configured with a kernel size of 3 residues, 64 filters, a stride of 1, and weights initialized using a He normal distribution. The second convolutional layer is configured with a kernel size of 5 residues and 128 filters. The third convolutional layer is configured with a kernel size of 7 residues and 256 filters. Each convolutional layer is followed by a rigorous processing pipeline: first, batch normalization is performed, with the momentum parameter set to 0.9 and the epsilon value set to 1×10⁻⁶. -5Then, the ReLU activation function is applied with a negative slope parameter of 0.01; finally, a dropout layer is applied with a dropout rate strictly set to 0.2. The convolutional feature outputs are concatenated along the sequence dimension and input into a bidirectional GRU layer with 128 hidden units, 1 network layer, a dropout rate of 0.2, and an output dimension of 256. The technical details of structural feature extraction are as follows: AlphaFold2 API requests are implemented using the Python requests library, with a timeout of 30 seconds, 3 retries, and a request URL format of https: / / alphafold.ebi.ac.uk / api / prediction / {UniProt_ID}. The structural feature analysis includes four dimensions: pLDDT confidence is extracted from the pLDDT field as a floating-point array and normalized by dividing by 100.0; secondary structure is extracted from the secondary_structure field as a string and one-hot encoded according to the mapping rules 'H'→[1,0,0] (α-spiral), 'E'→[0,1,0] (β-fold), and other characters→[0,0,1] (irregular curl); solvent accessibility is directly read from the solvent_accessibility field as a floating-point value; and the prediction alignment error matrix is ​​extracted from the pae field as a two-dimensional array, maintaining the original numerical accuracy. When AlphaFold2 data is unavailable, the system employs a degradation processing mechanism, generating synthetic structural features based on the sequence's physicochemical properties: the pLDDT score is calculated using the formula 70 + 0.1 × sequence complexity - 0.05 × sequence length, where sequence complexity is calculated using Shannon entropy; secondary structure is predicted based on amino acid tendency, with consecutive E, K, or R sequences predicted as helices, and consecutive V, I, or L sequences predicted as folds; solvent accessibility is estimated based on the basic accessibility score of amino acid types. All feature extraction processes are processed in parallel using multiple threads, with a maximum thread count set to 8 to ensure processing efficiency.

[0020] Adaptive feature fusion implementation. For example... Figure 2 As shown, the feature fusion module employs a dual mechanism of cross-modal attention and adaptive gating. The specific implementation process is as follows: First, feature space alignment is performed, projecting the structural features onto the sequence feature space through a linear transformation layer. The weight matrix of this linear transformation layer has a dimension of 256×23, and the bias term has a dimension of 256. The weights are initialized using a uniform Xavier distribution, and the bias is initialized to zero. Second, the projection formula is... ,in The input dimension corresponds to 23 structural features (3-dimensional secondary structure + 1-dimensional pLDDT + 1-dimensional solvent accessibility + 18-dimensional PAE matrix after PCA dimensionality reduction), and the output dimension is 256, consistent with the sequence feature dimension. Next, cross-modal attention calculation is performed using an 8-head attention mechanism, with each head having a key dimension of 64. The query vector is derived from the sequence features. The key and value vectors are derived from the projected structural features. The attention weight matrix is ​​calculated as follows: First, the dot product of the query vector and the key vector is calculated. Then, it is scaled by dividing by the square root of the key dimension (64), and finally, the softmax function is applied to normalize along the last dimension, generating the attention weight matrix A = softmax( / √64), each element A of the matrix [ i , j ] indicates the first i The position of the sequence is related to the first... j The correlation strength of each structural location is determined; then, attention-weighted structural features are applied. We obtain a weighted structural feature representation; finally, we perform adaptive gating fusion: first, we combine the sequence features... and weighted structural features Concatenating along the feature dimensions yields a 512-dimensional fused feature, which is then input into a fully connected layer to calculate the gating weights. This fully connected layer's weight matrix has a dimension of 512×256 and a bias dimension of 256. The gating weights are then output using a sigmoid activation function. Among them, the gating weight Based on the dynamic adjustment of structural characteristics, the specific adjustment strategy is as follows: when the average pLDDT score of the protein is below 70, [the following is applied to] the protein. Apply the correction factor min(1.0, ×1.2) to make it approach 1; when the pLDDT score is higher than 90, apply the correction factor max(0.0, ×0.8) to bring it close to 0; keep the original weights in the 70-90 range, and finally calculate the fused features as follows: , where ⊙ denotes element-wise multiplication. The entire fusion process uses batch processing, with a maximum supported batch size of 32, and all matrix operations are accelerated using the BLAS optimization library.

[0021] Implementation of disease conditional prediction. For example... Figure 3As shown, the disease perception module incorporates disease-specific information into the prediction process through a conditionalization mechanism. The system predefines eight disease types and establishes a complete mapping dictionary: {'Cancer': 0, 'Neurodegenerative Diseases': 1, 'Cardiovascular Diseases': 2, 'Metabolic Diseases': 3, 'Autoimmune Diseases': 4, 'Infectious Diseases': 5, 'Genetic Diseases': 6, 'Other Diseases': 7}. The disease coding process is as follows: First, the disease identifier is mapped to a 32-dimensional dense vector through an embedding layer. The embedding layer weight matrix has a dimension of 8×32, initialized using a uniform distribution, and the input is a one-hot encoding of the disease index. The disease vector is encoded through two fully connected network layers. The first layer is configured with an input dimension of 32, an output dimension of 64, using the ReLU activation function, a weight matrix dimension of 32×64, and a bias dimension of 64. The second layer is configured with an input dimension of 64, an output dimension of 8, using the tanh activation function, a weight matrix dimension of 64×8, and a bias dimension of 8. The network output is an 8-dimensional adjustment factor. These correspond to the adjustment intensities for eight different PTM types. The prediction probability adjustment mechanism uses a weighted combination approach: The intensity coefficient was adjusted to 0.2 and determined through a grid search. The adjusted probability was then limited to the range [0,1] by a value clipping operation, with the clipping threshold set to 1×10⁻⁶. -7 and 1-×10 -7 The automatic disease type detection function includes two approaches: First, it queries the DisGeNET database via the requests library using the UniProt ID, parsing the `disease_associations` field in the JSON response. Second, when prior knowledge is lacking, it infers the disease type based on sequence features, calculating features such as the proportion of charged amino acids (K, R, D, E), the proportion of hydrophobicity (A, V, L, I, P, W, F, M), and the proportion of cysteine ​​(C). A rule engine then makes classification decisions: a charged amino acid proportion > 0.25 and a sequence length > 500 indicate cardiovascular disease; a hydrophobic proportion > 0.4 indicates cancer; a cysteine ​​proportion > 0.05 indicates autoimmune disease; a simple repeat sequence (such as PolyQ) length > 40 indicates neurodegenerative disease; and other conditions are classified as other diseases. All network parameters are trained using the Adam optimizer with a learning rate of 0.001 and a weight decay coefficient of 1×10⁻⁶. -5 .

[0022] To verify the effectiveness of this invention, two known pathogenic variants, R1441C and R1441H, at the R1441 site of the LRRK2 protein were selected as test cases on the sample test set provided by the DeepMVP project. For the prediction of the same methylation site under two different variant scenarios, the confidence level of this invention was improved from 0.609 to 0.903 compared with the baseline method, an improvement of 48.3%, which proves the effectiveness and superiority of this invention.

[0023] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for predicting post-translational modification of proteins based on multi-modal deep learning, characterized in that, Comprising the following steps: Step 1, multi-modal feature extraction: the input is the amino acid sequence data of the protein and its corresponding three-dimensional structure data; through parallel sequence feature extraction branches and structure feature extraction branches, deep features containing local patterns and long-range dependencies are extracted from the sequence, and multi-dimensional features including residue confidence, spatial conformation, solvent accessibility and structural uncertainty are extracted from the structure data; The output is a unified representation of the sequence feature vector and the structure feature vector; Step 2, adaptive feature fusion: taking the sequence feature vector and the structure feature vector output in step 1 as input; A cross-modal attention mechanism is used to establish a dynamic association mapping between sequence positions and structure positions, and then a quality-aware adaptive gating network is used to dynamically calculate and assign the fusion weights of each modality according to the confidence level of the structure data; The output is a deep fused feature representation that can reflect the data quality; Step 3, disease-conditioned prediction: combine the fusion features obtained in step 2 with specific disease type information; through a disease-specific encoding network, map the disease category to an adjustment factor, and fine-tune the base prediction probability with it; the output is the prediction result modulated by the disease context; Step 4, multi-task PTM prediction: PTM stands for Post-Translational Modification; based on the conditioned feature representation output in step 3 as input; Using a multi-task learning framework, through multiple parallel and weight-independent prediction heads, the site probability of 8 common PTM types is predicted simultaneously; Step 5, result output and explanation: perform in-depth analysis of the interpretability of the prediction probability generated in step 4; calculate the importance score of the input features through gradient backpropagation technology, and combine a detailed variant impact analysis module to quantify the disturbance effect of genetic variation on PTM potential; finally output a comprehensive prediction report.

2. The method of claim 1, wherein, In step 1, sequence feature extraction uses a CNN-BiGRU hybrid network: three parallel one-dimensional convolution layers are used, with convolution kernel sizes of 3, 5, and 7, respectively, and filter numbers of 64, 128, and 256, respectively, followed by batch normalization, ReLU activation, and a dropout layer with a dropout rate of 0.2; Then input into a bidirectional GRU with 128 hidden units, output 256-dimensional sequence features; Structure feature extraction includes: pLDDT confidence score, normalized to 0-1; secondary structure encoded by one-hot; solvent accessibility score; predicted alignment error matrix.

3. The method of claim 1, wherein, In step 2, cross-modal attention and adaptive gating fusion mechanism is used; structure features are projected to sequence feature space through a linear layer; Calculate the attention weight matrix; the gating weight is generated by a fully connected layer and a sigmoid function; the gating weight is dynamically adjusted according to the average pLDDT confidence of the structure: when pLDDT is lower than 70, the system mainly relies on sequence features; When pLDDT is higher than 90, the system mainly relies on structure features; in the middle interval, it is smoothly transitioned.

4. The method of claim 1, wherein, In step 3, the system predefines eight categories of diseases, including cancer, neurodegenerative diseases, cardiovascular diseases, metabolic diseases, autoimmune diseases, infectious diseases, genetic diseases, and other diseases; the disease type is mapped to a 32-dimensional vector through an embedding layer, and an adjustment factor is output through two fully connected networks; and the adjusted probability is: .

5. The method of claim 1, wherein, In step 4, each type of PTM corresponds to a prediction head; the loss function is weighted binary cross-entropy, and the weights are set according to the distribution of the training set: 0.25 for serine / threonine phosphorylation, 0.25 for tyrosine phosphorylation, 0.125 for lysine acetylation, 0.125 for lysine methylation, 0.125 for arginine methylation, 0.0625 for lysine ubiquitination, 0.0625 for lysine SUMOylation, and 0.0625 for asparagine glycosylation.

6. The method of claim 1, wherein, In step 5, feature importance is calculated by gradient backpropagation; variation impact analysis includes probability change value, joint probability and log ratio value ratio, and empirical p-value is calculated based on random variation background distribution; finally, a comprehensive prediction report is output, which not only includes site confidence and disease adjustment factor, but also provides feature attribution and variation impact evaluation.

Images