A protein classification method based on multi-modal feature fusion and an interpretable network

Abstract

The application provides a protein classification method based on multi-modal feature fusion and an interpretable network. The method fuses multi-modal features of protein sequences and structures, and adopts an interpretable neural network model for classification. A pre-trained protein language model is used to extract sequence features, and a high-level protein structure prediction model is used to obtain three-dimensional structure information of the protein. Sequence and structure features are fused through an improved attention mechanism or a graph neural network to form a high-dimensional comprehensive representation. Finally, an interpretable classification network is used to output a prediction result and give an explanation of the contribution of key features. The method can efficiently and accurately identify DNA-binding proteins, significantly improve prediction accuracy, has good cross-species generalization ability, and enhances model decision transparency.

Description

Technical Field

[0001] This invention relates to bioinformatics, artificial intelligence, biomedicine, and proteomics, and particularly to a protein classification method based on multimodal feature fusion and interpretable networks, which can be used for the automatic identification of DNA-binding proteins. Background Technology

[0002] As the core executors of life activities, the precise analysis of protein functions is the cornerstone of modern life science research. In particular, the identification of specific functional proteins such as DNA-binding proteins (DBPs) is crucial for understanding gene regulatory networks, elucidating the molecular mechanisms of diseases, and developing novel targeted drugs. With the rapid development of high-throughput sequencing technology, the size of protein sequence databases has grown exponentially, and traditional methods relying on biological experiments for functional identification are far from meeting the demands of analyzing massive amounts of data. Therefore, developing rapid, accurate, and reliable computational prediction methods has become an urgent need and a research frontier in the field of bioinformatics.

[0003] Existing methods for protein function prediction can be broadly categorized into several types. Early methods relied on sequence homology alignment, but their predictive performance drops sharply when the similarity between the target protein and known functional proteins is low. Subsequently, machine learning-based methods were widely adopted, training classifiers by extracting various hand-designed features from protein sequences, such as amino acid composition, physicochemical properties, and evolutionary information. In recent years, deep learning has achieved significant success in protein function prediction tasks due to its powerful nonlinear representation learning capabilities.

[0004] Among them, the multimodal feature fusion strategy is widely recognized as an effective way to improve model performance. The core idea of ​​this strategy is to construct a more comprehensive and robust digital representation of proteins by integrating complementary information from different biological dimensions—for example, evolutionary conservation features that reflect evolutionary pressure, physicochemical features that summarize the overall properties of the sequence, sequence semantic features that capture deep contextual relationships, and even predicted three-dimensional structural features.

[0005] However, despite significant progress, current deep learning methods based on multimodal fusion still face two core challenges that urgently need to be addressed:

[0006] Limitations of Feature Fusion Mechanisms: Most existing methods employ simple feature concatenation before feeding into fully connected layers, or use standard self-attention mechanisms for fusion. These methods tend to treat features from all modalities equally, failing to effectively model and utilize the inherent affinity and complementarity between different information modalities. For example, in specific function prediction tasks, structural and evolutionary information may exhibit stronger synergistic effects than they do with global physicochemical properties. Standard attention mechanisms struggle to explicitly and adaptively capture and reinforce such modal preference relationships, thus limiting the efficiency of information fusion and the final prediction accuracy.

[0007] The "black box" problem of model decision-making: Deep learning models, especially their classification decision-making parts composed of multiple layers of nonlinear transformations, have complex weights and activation functions that make the model's decision-making process highly opaque. Researchers cannot intuitively understand which input feature dimensions the model uses and through what response patterns to make specific predictions. This lack of interpretability severely hinders the credibility and widespread application of prediction results in high-risk decision-making scenarios such as biological research and drug development. Biologists not only need to know "what" the prediction results are, but also urgently need to know "why" they are, in order to conduct subsequent experimental verification and mechanistic exploration.

[0008] Therefore, there is an urgent need in this field for a new technical solution that can not only intelligently integrate multi-source heterogeneous biological features, but also provide a transparent and verifiable decision-making path, thereby achieving high prediction accuracy while endowing the model with high interpretability. Summary of the Invention

[0009] This invention aims to overcome the shortcomings of existing computational identification methods for DNA-binding proteins (DBPs) in terms of characterization ability, structure utilization, model interpretability, and cross-species generalization, and focuses on solving the following key technical problems:

[0010] Constructing robust DBP sequence representations that do not rely on multiple sequence alignments. Existing methods generally rely on position-specific scoring matrices (PSSMs) generated by tools such as PSI-BLAST, which are computationally expensive and show significant performance degradation in novel species or low-homology sequences. How to fully utilize the evolutionary and semantic information contained in large-scale pre-trained protein language models (such as ProtT5 and ESM series) to obtain high-quality sequence representations without explicit multiple sequence alignments is an urgent problem to be solved.

[0011] Effectively integrating predicted 3D structural information to improve discrimination ability. Although structural prediction models such as AlphaFold3 and ESMFold can generate protein 3D structures with high accuracy, existing DBP recognition methods either ignore structural information or only use coarse-grained features such as secondary structure ratios, lacking fine-grained modeling of key structural clues such as spatial proximity relationships and potential DNA binding interfaces, making it difficult to achieve coordinated characterization of sequence and structural modalities at the residue level.

[0012] Achieving endogenous interpretability in the classification decision-making process is crucial. While traditional deep neural networks possess strong fitting capabilities, their "black box" nature makes it impossible to trace key discriminative criteria (such as critical residues, structural regions, or physicochemical features). Even with posterior interpretation methods like SHAP and Grad-CAM, their stability and biological verifiability remain insufficient. Therefore, a classification architecture with interpretability at the model structure level is urgently needed, enabling biologists to intuitively understand and verify the decision-making logic.

[0013] Balancing model complexity with cross-species generalization ability is crucial. DBP identification needs to be applicable to high-throughput screening scenarios involving multiple species and large-scale protein libraries. However, introducing multimodal features and complex networks can easily lead to parameter inflation and overfitting. A key challenge in practical applications is how to control computational costs while ensuring the model's generalization performance on unseen species and supporting efficient deployment.

[0014] In summary, this invention aims to provide a protein classification method based on multimodal feature fusion and interpretable networks. This method can deeply fuse sequence and predicted structural information without the need for multiple sequence alignments, and simultaneously achieve high-precision prediction and transparent decision-making through an endogenous interpretable classification mechanism. This enhances the reliability and practicality of the method in basic research and drug development.

[0015] To address the aforementioned technical problems, this invention proposes a protein classification method based on multimodal feature fusion and interpretable networks. This algorithm uses the protein amino acid sequence as the sole input, jointly utilizing a pre-trained protein language model and a high-precision structure prediction model to characterize proteins from both sequence and structural modal perspectives. It then achieves deep fusion of multimodal features through a graph neural network, and finally uses an interpretable neural network classifier to output the DBP probability and its decision-making basis.

[0016] The overall process can be formally described as follows: given an amino acid sequence (1)

[0017] in For sequence length, For amino acid alphabets; corresponding supervision tags (2)

[0018] This invention learns a mapping from sequence to probability. (3)

[0019] in The predicted probability of it being a DNA-binding protein. These are all the trainable parameters for the algorithm. The following sections explain each step.

[0020] 1. Sequence Modal Feature Extraction

[0021] This invention first utilizes a pre-trained protein language model to construct protein representations at the sequence level. Let... For large-scale pre-trained protein language models such as ProtT5 and ESM, the input sequence is processed. Encoding is performed to obtain a residue-level embedding matrix: (4)

[0022] in For the first A sequence embedding vector of residues.

[0023] To obtain the overall sequence semantic information, pooling can be performed on the residue embeddings to obtain a sequence-level representation. For example, simple average: (5)

[0024] During subsequent mapping, Used as sequence modal features for nodes. It can be used as a global auxiliary feature.

[0025] By using equations (4)–(5), this invention extracts sequence modality features rich in evolutionary and semantic information from a single sequence without relying on multiple sequence alignments.

[0026] 2. Structural modality features and protein map construction

[0027] To utilize three-dimensional structural information, this invention further introduces structural modes and transforms them into a graph structure. Let... For protein structure prediction models (such as ESMFold or AlphaFold3), for the same sequence Predict its residue coordinates: (6)

[0028] The distance matrix between residues is defined based on coordinates: (7)

[0029] Subsequently, a protein structure diagram was constructed. Node set Each residue in the corresponding sequence; the adjacency relationship is determined by both sequence adjacency and spatial proximity, and an adjacency matrix is ​​defined. (8)

[0030] in A preset distance threshold (e.g., 8 Å) is used. Regarding structural modal characteristics, this invention constructs a vector for each residue. This can include secondary structure type, relative solvent-accessible area, local geometric statistics, charge / hydrophobicity, etc., and is denoted as: (9)

[0031] Finally, the initial multimodal features of each node are constructed as follows: (10)

[0032] in This represents vector concatenation. Simultaneously encodes sequence and structural information.

[0033] By using equations (6)–(10), the original linear sequence is transformed into a protein graph representation that explicitly models the structural neighborhood at the node and edge levels.

[0034] 3. Multimodal feature fusion of graph neural networks

[0035] After constructing the protein map And specify the initial characteristics of the node. Subsequently, this invention employs a graph neural network (such as the graph attention network GAT) to perform multi-layer message passing on the graph, thereby achieving deep fusion of sequence-structure multimodal approaches.

[0036] For the graph neural network's... A layer can be represented as follows: first, a linear transformation is performed on the node features. (11)

[0037] in This is a trainable weight matrix. Then, for each edge... Calculate the unnormalized attention score: (12)

[0038] in For trainable vectors, This is a distance-based edge feature mapping. In the neighbor set... The attention weights are obtained by performing softmax normalization within the function. (13)

[0039] Finally, the neighbor features are weighted, aggregated, and updated: (14)

[0040] in For non-linear activation functions, such as ReLU. Through multiple iterations ( Each node ultimately represents It has already integrated multi-round interaction information from its sequence environment and three-dimensional structural neighborhood.

[0041] To obtain a protein-level global representation, this invention employs a graph-level readout (pooling) operation to aggregate the final representations of all nodes into a fixed-dimensional feature vector: (15)

[0042] in It can be a simple average. Vectors can also be learned through graph-level attention mechanisms. This is the multimodal fusion feature representation of the algorithm of this invention.

[0043] 4. Interpretable Classifier Design

[0044] After obtaining multimodal fusion features Subsequently, the present invention employs an interpretable neural network classifier to discriminate the DBP probability and provides interpretable feature contribution information.

[0045] In a preferred implementation, the present invention employs an interpretable classifier in the form of a Kolmogorov-Arnold network (KAN). The idea is to integrate each dimension of the input features... This is mapped to a visual univariate response curve, and then these responses are linearly combined. For simplicity, it can be written as:

[0046] Apply a nonlinear mapping to the fused features (which can be understood as a combination of all unary responses): (16)

[0047] Among them each The interpretable response is composed of a combination of univariate basis functions. The response vector is linearly mapped to a scalar score (logit). (17)

[0048] in , These are trainable parameters. The final predicted probability is obtained through the Sigmoid function: (18)

[0049] During the training phase, the loss function uses binary cross-entropy with a regularization term: (19)

[0050] in It can include weights Regular expressions are used to control model complexity, improve generalization, and enhance sparsity and interpretability.

[0051] In other implementations, interpretable classifiers may also employ differentiable neural decision tree models, which essentially involve... Mapping to a set of piecewise linear or gated paths forms a rule path similar to "if feature A is greater than the threshold and feature B falls into the interval, then it is determined to be DBP", which can also output a probability form similar to equation (17) – (18).

[0052] 5. Overview of Training and Reasoning Process

[0053] Based on the above modules, this invention provides the following training and inference process:

[0054] Input: with tags A collection of protein sequences.

[0055] For each sequence, execute equations (4)–(10) sequentially to obtain the initial node features and protein map;

[0056] Executing equations (11)–(15) yields the multimodal fusion features. ;

[0057] Executing equations (16)–(18) yields the predicted probabilities. ;

[0058] Calculate the loss according to formula (19) All parameters are updated using stochastic gradient descent or optimizers such as Adam. until convergence on the validation set.

[0059] For any unknown protein sequence The predicted probabilities can be obtained simply by following the same forward computation process as in the training phase. ;

[0060] By analyzing the response function or rule structure within an interpretable classifier, the contribution of each feature dimension or each class of residues / structural regions to the prediction results can be analyzed, thereby generating user-oriented interpretable reports.

[0061] Through the above technical solutions, this invention forms a complete algorithm flow from sequence input to "multimodal graph modeling + interpretable classification", which not only improves the accuracy of DNA binding protein identification and cross-dataset generalization ability, but also significantly enhances the transparency and traceability of the model decision-making process through endogenous interpretable structure.

[0062] This invention addresses the problems of existing DNA-binding protein identification methods, such as limited feature representation, insufficient utilization of structural information, heavy reliance on multiple sequence alignment, and the "black box" nature of the model decision-making process. It proposes a multimodal graphical neural network fusion framework based on a pre-trained sequence model and a high-precision structure prediction model, and introduces an endogenously interpretable neural network classifier. Compared with existing technologies, this invention has at least the following advantages:

[0063] High-quality DBP representations can be constructed without the need for multiple sequence alignments, significantly improving prediction accuracy and robustness.

[0064] This invention utilizes large-scale pre-trained protein language models such as ProtT5 and ESM to directly extract sequence embeddings rich in evolutionary tendencies, physicochemical properties, and long-range dependencies from single amino acid sequences, without relying on multiple sequence alignment tools such as PSI-BLAST to generate traditional features like PSSM. On the one hand, this avoids the high computational cost and instability of multiple sequence alignment in large-scale data and new species scenarios; on the other hand, it enables the model to maintain high discriminative ability in datasets lacking homologous sequences or with low sequence diversity. Based on the sequence modality representation proposed in this invention, in the embodiments, higher accuracy, recall, and Matthews correlation coefficients than traditional sequence feature methods can be achieved on multiple benchmark datasets.

[0065] The system incorporates predicted structural information and performs graph structure modeling to more fully capture DNA binding interface features.

[0066] This invention utilizes state-of-the-art structure prediction models such as AlphaFold3 and ESMFold to map the amino acid sequence of the target protein into high-precision three-dimensional structural information. Furthermore, it constructs a protein structure graph with residues as nodes and sequence adjacency and spatial proximity as edges. Compared to existing methods that only use coarse-grained descriptions such as secondary structure ratios, this invention explicitly models the spatial topological relationships at the residue level and potential DNA-binding interfaces (such as positively charged residue clusters and groove structures) at the graph structure level. Combined with the message passing mechanism of graph neural networks, this facilitates the identification of key regions that determine DNA affinity, thus achieving significant advantages in distinguishing between positive and negative samples and adapting to complex folding topologies.

[0067] By using graph neural networks to achieve deep fusion of sequence-structure multimodal features, we can improve feature representation and generalization capabilities.

[0068] This invention does not simply concatenate sequence and structural features; instead, it embeds both into a unified protein graph. Graph convolutional networks, graph attention networks, or graph transformers perform multiple rounds of message passing and updates, achieving deep interaction between sequence and structural modal information at the residue level. This design explicitly captures high-order associations of "which sequence patterns are more likely to form DNA-binding functions in what structural environments," thus significantly enhancing feature representation capabilities. Thanks to this multimodal fusion mechanism, this invention demonstrates superior generalization performance compared to pure sequence models on cross-species test sets, making it suitable for DNA-binding protein prediction tasks from different species and database sources.

[0069] An endogenous interpretable neural network classifier is used to achieve a "white-box" presentation of the decision-making process.

[0070] This invention introduces interpretable neural network architectures such as the Kolmogorov-Arnold Network (KAN) or differentiable neural decision tree models during the classification stage. This explicitly decomposes the complex nonlinear mapping between multimodal fusion features and output probabilities into a set of univariate response functions or class decision tree rules. Compared to ex-post interpretation methods, the interpretability of this invention stems from the model structure itself.

[0071] When using KAN, the response curve corresponding to each input feature can be obtained directly, clearly showing the promoting or inhibiting effect of the feature on "DNA binding" or "non-DNA binding" under different value ranges;

[0072] When using a differentiable neural decision tree model, a discrimination logic similar to a "rule path" can be extracted to form decision conditions that are close to human-readable.

[0073] This inherent interpretability significantly improves the understandability and credibility of predictions in high-risk scenarios such as biological research and drug target discovery.

[0074] The engineering feasibility of achieving large-scale screening and cross-species application is achieved while balancing model complexity and computational cost.

[0075] This invention controls the overall computational complexity and storage overhead of the model through the reuse of pre-trained models, the pipelined invocation of structure prediction models, and the rational parameter design of graph neural networks and interpretable classifiers. Compared with traditional methods that heavily rely on multiple sequence alignments and large-scale feature engineering, this invention can perform parallel inference on a large number of candidate proteins on GPUs or high-performance computing nodes, making it suitable for integration into high-throughput in silico screening workflows. Furthermore, since the core features originate from universal protein patterns learned by the pre-trained and structure prediction models, this invention exhibits good transferability and reusability in cross-species and cross-dataset applications. Attached Figure Description

[0076] Figure 1 This diagram illustrates the overall technical principle of a protein classification method based on multimodal feature fusion and interpretable networks, as provided by this invention. The diagram macroscopically shows the overall technical route of the data flow from sequence feature extraction and structure prediction to multimodal fusion and final classification.

[0077] Figure 2 This is a schematic diagram of the interpretable multimodal feature fusion model described in this invention. It includes a sequence feature extraction module, a structural feature extraction module, a fusion module, and an interpretable classification module. These modules work collaboratively to analyze and classify proteins.

[0078] Figure 3 This is a flowchart illustrating the specific implementation steps of the method described in Embodiment 1 of the present invention. The flowchart details the specific execution logic steps, including data preprocessing, model building, feature fusion calculation, and classification output. Detailed Implementation

[0079] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments. Those skilled in the art should understand that the following embodiments are only for illustrating the present invention and are not intended to limit the scope of protection of the present invention; various equivalent substitutions or modifications made to the embodiments without departing from the spirit and substance of the present invention should fall within the scope of protection of the present invention.

[0080] Unless otherwise stated, the terms "sequence modality feature", "structural modality feature", "protein structure map", "graph neural network fusion module", "multimodal fusion feature representation", "interpretable neural network classifier", etc., used in the embodiments of this invention are consistent with the meanings of the foregoing technical solutions and claims.

[0081] Example 1: Implementing a multimodal interpretable DBP recognition algorithm on a public benchmark dataset

[0082] This embodiment provides a complete implementation flow of the method of the present invention on a publicly available DNA-binding protein (DBP) benchmark dataset. Through this embodiment, those skilled in the art can intuitively understand how to start from the original sequence, sequentially complete data preparation, feature generation, model building, training and testing, and how to obtain interpretable prediction results.

[0083] For ease of description, let's call the training dataset: in For the first A protein sequence, A tag indicating whether it is a DNA-binding protein.

[0084] S101 Dataset Construction and Preprocessing

[0085] Select data source:

[0086] Select a commonly used DNA-binding protein benchmark dataset from a public database or literature, ensuring that each protein has a clear "DBP / Non-DBP" label. Download the raw data in FASTA or a similar format, ensuring that each sequence has a unique ID.

[0087] Cleaning and length filtration:

[0088] Delete samples containing non-standard amino acid characters (such as "B", "Z", "X"), or replace them with the closest standard residues before use. Filter out sequences that are too short (e.g., length < 50) and too long (e.g., length > 1024) to avoid extreme samples negatively impacting model training.

[0089] Redundancy removal and partitioning of training / validation / test sets:

[0090] Cluster the data using tools such as CD-HIT, setting a sequence identity threshold (e.g., 25% or 30%), and retaining one representative sequence in each cluster to reduce homology bias. Randomly partition the deredundant data into training, validation, and test sets, for example, in an 8:1:1 ratio. During partitioning, ensure that sequences from the same cluster are not split into different subsets. After these processes, a clean dataset is obtained for model training and evaluation. , , .

[0091] S102 Sequence Modal Feature Generation

[0092] This step maps the original amino acid sequence to a high-dimensional vector space to represent the semantic and evolutionary information of the sequence.

[0093] Load the pre-trained model:

[0094] Load a pre-trained ProtT5 or ESM protein language model into a deep learning framework (such as PyTorch). Set the model to inference mode (evaluation mode) and do not update its internal parameters.

[0095] Encode each sequence:

[0096] For any sequence in the dataset Convert the sequence into the input format required by the model (such as a tokenID sequence). Input it into the pre-trained model to obtain a vector representation of each residue. in For the first An embedding vector of residues (e.g., 1024 dimensions).

[0097] Preserving sequence embeddings:

[0098] The residue embeddings of each sequence are saved as independent files (e.g., .npy or .pt), indexed by sequence ID, for direct reading during subsequent mapping and training. At this point, each sequence has obtained a set of residue-level sequence modal features.

[0099] S103 Structural Modal Features and Protein Mapping

[0100] This step characterizes proteins at the structural level, predicting the 3D structure of each sequence and constructing a protein map.

[0101] Structural prediction:

[0102] For each sequence Using structural prediction models such as ESMFold, the three-dimensional coordinates of each residue were obtained. The model output is typically in PDB format or a similar data structure, containing information such as residue numbers, coordinates, and confidence scores.

[0103] Calculate the distance between residues and their adjacency:

[0104] For each sample, calculate the distance matrix based on the residue coordinates: Set distance threshold (e.g., 8 Å) If two residues are adjacent in sequence or the spatial distance is less than a threshold, an edge is established between them. This yields the node set corresponding to each protein. Sum of edges The protein structure diagram .

[0105] Structural modal feature construction:

[0106] For each residue, secondary structure tags (such as H / E / C), relative solvent access area (RSA), and local confidence scores are extracted from the structure prediction results. These scalars or small vectors are then concatenated into a structural feature vector. Typically, the dimensions are small (e.g., 16–32 dimensions are sufficient).

[0107] Initial multimodal features of nodes:

[0108] Embedding sequences With structural features By concatenating the features, we obtain the initial characteristics of the nodes: Will The first layer of node features serves as the input to the graph neural network. At this point, each sample has been transformed from a "sequence" into a "protein graph" with multimodal node features.

[0109] S104 Graph Neural Network Model Construction and Hyperparameter Settings

[0110] This embodiment uses a multi-layer graph neural network to perform feature fusion on the protein map.

[0111] Graph Neural Network Structure:

[0112] Network layers: 3-layer graph neural network (e.g., Graph Attention Network GAT); Hidden dimension: 256 dimensions per layer; Number of attention heads: 4 heads per layer, output concatenation; Activation function: ReLU; Dropout: 0.2–0.3, used to prevent overfitting. The graph neural network contains three stacked graph convolutional layers. Each layer updates the representation of the current residue by aggregating features from neighboring nodes, thereby progressively integrating its local structure and sequence context information. After propagation through three layers, the embedding features of each residue have been fused with neighborhood information with a radius of three hops, effectively characterizing its potential DNA binding properties in the three-dimensional structure of the protein.

[0113] Graph-level readout:

[0114] For all node features obtained in the last layer, average pooling or attention pooling is used to aggregate the node-level information into a fixed-length protein-level vector. This vector can be understood as "a holistic representation containing the multimodal information of the protein sequence-structure".

[0115] S105 can explain the classification head and training process.

[0116] Explainable classification head design:

[0117] Fusion features Input a shallow, interpretable classification head: the first layer can be viewed as a combination of several unary response functions (implemented similarly to KAN, or a set of interpretable activation units); the second layer is a simple linear layer that weights and sums the response values ​​to obtain a scalar score. Use the Sigmoid function to convert the scores into DBP probabilities. Due to its simple structure, by analyzing the weights and the output of intermediate units, we can directly obtain "which feature dimensions are the most important and within what range they have positive or negative contributions".

[0118] Loss function and optimization strategy:

[0119] For a training batch, standard binary cross-entropy loss is used to compare the predicted probabilities with the true labels; a suitable additional factor is added to the loss. The regularization term encourages the model to avoid excessively large weights, with some weights approaching 0, thus facilitating the interpretation of "which features are truly useful". The Adam optimizer is used, with common configurations such as: initial learning rate: 0.001; batch size: 16 or 32; number of training epochs: up to 100 epochs, stopping early when the validation set metrics no longer improve.

[0120] Training process:

[0121] Each epoch: A batch of sequences is read from the training set; ProtT5 embeddings and AlphaFold3 structures are generated from these sequences, and a protein map is constructed; Forward computation: Graph neural network → Graph-level readout → Interpretable classification head → Obtain predicted probabilities; Gradients are calculated based on the loss function, backpropagation is performed, and model parameters are updated; Metrics such as ACC, SN, SP, and MCC are evaluated on the validation set to monitor training progress and select the optimal model. After training, the set of model parameters that performed best on the validation set is saved for final testing and deployment.

[0122] S106 Test Visible "Usage Methods"

[0123] Evaluate on the test set:

[0124] Using the trained model, perform a complete forward pass for each protein in the test set: input sequence; read pre-calculated ProtT5 embeddings and structure predictions to construct a graph; obtain predicted probabilities using a graph neural network and an interpretable classification head. Based on a set threshold (e.g., 0.5), it is classified as "DBP / Non-DBP", and the accuracy, recall, specificity, and MCC are statistically analyzed.

[0125] The actual steps for using a single new sequence are as follows:

[0126] If the user now has a completely new protein sequence To find out if it is a DNA-binding protein, you can follow these steps:

[0127] Save the sequence as a FASTA text file;

[0128] Call the already deployed ProtT5 module to perform forward inference on the sequence and obtain the embedding vector of each residue;

[0129] Call the structure prediction module (ESMFold or AlphaFold3) to generate the predicted three-dimensional structure of the protein;

[0130] Construct a protein graph based on the three-dimensional structure: residues are used as nodes, and edges are added between residue pairs whose spatial distance is less than a threshold.

[0131] Input the graph into the trained graph neural network to obtain the fused feature vector;

[0132] The fused features are fed into an interpretable classification head, which outputs a probability value between 0 and 1, such as 0.87.

[0133] If the probability is higher than the threshold (such as 0.5 or 0.7), it indicates that "this protein has a high probability of being a DNA-binding protein".

[0134] Also view the "List of Key Features" and "Response Trends" output by the interpretable module, for example:

[0135] The main structural regions of interest in the model are the α-helices at positions 40–60 and 120–140;

[0136] A high proportion of positively charged residues on the surface is considered one of the main reasons driving the DBP determination.

[0137] As can be seen from the description of this embodiment, as long as a conventional deep learning environment (such as a server with a GPU) is available, and the data and pre-trained model are prepared step by step, those skilled in the art can directly reproduce the protein classification method based on multimodal feature fusion and interpretable network proposed in this invention, and perform prediction and interpretation on new real samples.

[0138] Without departing from the overall concept of this invention, any adjustments to parameters such as the specific pre-trained model type, the number of graph neural network layers, the hidden dimension, and the learning rate are variations that are easily conceived by those skilled in the art and should be considered to fall within the protection scope of this invention.

[0139] Example 2: DNA-binding protein recognition for cross-species generalization and practical applications

[0140] In Example 1, this invention demonstrates how to implement and evaluate a protein classification method based on multimodal feature fusion and interpretable networks proposed in this invention on a single benchmark dataset. This example further illustrates how to construct data, train and evaluate the model in cross-species scenarios, and demonstrates how the algorithm can be used in new species and real-world tasks through practical applications.

[0141] This embodiment emphasizes two key points:

[0142] The model was trained only on the "source species", but still had good recognition ability on the "target species" (species not involved in the training);

[0143] The model output not only indicates whether it is a DNA-binding protein, but also provides explanatory information that is valuable to biologists and can assist in the design of subsequent experiments.

[0144] S201 Cross-Species Dataset Construction

[0145] Select source species:

[0146] Several species are selected from model organisms as "source species", such as Escherichia coli and Saccharomyces cerevisiae.

[0147] The source species dataset is formed by collecting known DNA-binding proteins (positive samples) and non-DNA-binding proteins (negative samples) from public databases.

[0148] Select the target species:

[0149] Select species that are evolutionarily distant from the source species as "target species", such as humans, mice, and plants;

[0150] Similarly, DNA-binding proteins and non-DNA-binding proteins with clearly defined functions from these species were collected to form a test set for the target species.

[0151] Cleaning and Redundancy Removal:

[0152] The data for the source species and the target species were cleaned separately:

[0153] Remove sequences containing a large number of non-standard amino acids;

[0154] Samples are screened by length range (e.g., 50–1024 amino acids).

[0155] Use tools such as CD-HIT to perform redundancy filtering on the source species data (e.g., 25% or 30% identity threshold) to avoid too many highly similar sequences affecting training.

[0156] Special note: No redundancy is removed between the source species and the target species to simulate a real "cross-species generalization" scenario, i.e., training sequences that do not include the target species.

[0157] Data partitioning:

[0158] The source species data is divided into training and validation sets (e.g., 8:2), and only the source species data is used for training and parameter tuning.

[0159] The target species data is only used in the final evaluation phase and is not used in any training or hyperparameter selection.

[0160] After the above steps, we obtain:

[0161] Source species training and validation sets (used for model training and selection);

[0162] Target species test set (used to assess cross-species generalization ability).

[0163] S202 Model Structure and Training Strategy (Cross-Species Setup)

[0164] This embodiment uses the same model structure as Embodiment 1 (pre-trained ProtT5 + structure prediction + graph neural network + interpretable classification head), but the training and evaluation strategies are designed for cross-species use.

[0165] Feature generation and mapping:

[0166] For all sequences in the source species training set, validation set, and target species test set, the same procedure as in Example 1 is used:

[0167] Use ProtT5 (or ESM) to generate residue-level sequence embeddings;

[0168] Predict 3D structures using ESMFold or AlphaFold3;

[0169] Construct a protein map based on spatial distance and sequence adjacency;

[0170] Structural modal features are constructed for each residue, and node features are a concatenation of "sequence embedding + structural features".

[0171] Model structure reuse:

[0172] The hyperparameters of the graph neural network, such as the number of layers, hidden dimensions, number of attention heads, and Dropout, can all use the typical settings given in Example 1, such as:

[0173] A 3-layer GNN, each layer with 256 dimensions, and 4 attention heads;

[0174] ReLU activation;

[0175] Dropout 0.2–0.3.

[0176] Interpretable classification heads remain simple and interpretable, for example: shallow nonlinear transformation + linear output layer + sigmoid probability.

[0177] Training process (using only the source species):

[0178] During training, only samples from the source species training set are used:

[0179] Read several source species sequences in batches from the training set;

[0180] Load these sequences with pre-generated multimodal graph inputs (sequence embeddings + structure + graph);

[0181] The predicted probability is obtained through a graph neural network and an interpretable classification head;

[0182] Calculate the cross-entropy loss and regularization term, and update the model parameters once.

[0183] Indicators such as ACC and MCC are evaluated on the source species validation set for early termination and model selection.

[0184] No data from the target species is used during training to ensure objectivity when assessing cross-species capabilities.

[0185] The model after training:

[0186] After training, the set of model parameters that performs best on the source species validation set is selected and fixed as the "cross-species model" for subsequent target species prediction and practical applications.

[0187] S203 Cross-species assessment in the target species

[0188] Reasoning process:

[0189] For each protein sequence in the target species test set, the exact same forward pass procedure as in the source species was used:

[0190] Input sequence, read the corresponding ProtT5 embedding and structure prediction results;

[0191] Constructing protein graphs and node multimodal features;

[0192] After training the GNN and using an interpretable classification head, the predicted probability that the sequence is a DNA-binding protein is obtained.

[0193] Indicator Calculation:

[0194] Based on the real labels of the target species and the model prediction results, TP, FP, TN, and FN are statistically analyzed, and accuracy, recall, specificity, and MCC are calculated.

[0195] If needed, metrics such as the area under the ROC curve (AUC) and precision-recall curve can also be calculated under different probability thresholds to comprehensively evaluate the model's performance in cross-species recognition tasks.

[0196] Interpretation of Results:

[0197] If the model can still maintain a high MCC and a reasonable recall rate for the target species, it indicates that the multimodal features and interpretable classification head of this invention have not "remembered" the pattern of a certain species, but have learned the structure-sequence rules that are independent of species and related to DNA binding function.

[0198] Compared with control experiments using only sequence features or only PSSM + traditional classifiers, the present invention can significantly improve performance on target species, indicating that the introduction of prediction structure and multimodal fusion is effective.

[0199] Practical application process of S204 in new species or non-model organisms

[0200] This invention can not only verify its generalization ability on manually segmented source / target species datasets, but is also suitable for integration into practical research workflows for batch screening of new species or non-model organisms using DBP. A practical application example is given below.

[0201] Suppose researchers have just obtained the complete proteome sequence of a non-model organism (such as a newly discovered fungus or marine microorganism) and want to quickly screen for potential DNA-binding proteins as candidates for subsequent experiments.

[0202] Prepare the input sequence:

[0203] All protein sequences of this species were compiled into a FASTA file, ensuring that each sequence has a unique identifier (such as gene name or ID).

[0204] Batch generation of multimodal features:

[0205] Using the pre-deployed ProtT5 module, batch generate residue embeddings for each sequence and save them;

[0206] Use ESMFold or AlphaFold3 to predict the structure of each sequence. For large proteomes, parallel processing can be performed in a cluster or multi-GPU environment.

[0207] Based on the predicted structure, a protein map and node structure features are constructed for each sequence.

[0208] Use the pre-trained cross-species model to make predictions:

[0209] Input the above multimodal graph into the "cross-species model" trained in Example 2;

[0210] For each sequence, obtain a DBP probability between 0 and 1;

[0211] Sort the proteins by probability from high to low to obtain a "list of candidate DNA-binding proteins", while retaining the explanatory information for each protein (such as which domains and residue regions the model considers most important).

[0212] Results screening and manual review:

[0213] Multiple probability levels can be set, for example:

[0214] High-confidence candidates are given priority for experimental verification.

[0215] Candidates with medium confidence can be screened together with other information (such as transcription factor annotation and subcellular localization prediction).

[0216] Low confidence level, generally not considered as the first choice, but can still be retained in the context of high throughput.

[0217] Researchers can focus on the structural explanations of high-confidence candidates, for example:

[0218] The model focuses on the presence of certain positively charged α-helices on the surface, concentrated in specific regions;

[0219] The model assumes that a predicted domain is very close to a known DNA binding domain in the multimodal feature space.

[0220] In conjunction with experimental design:

[0221] Researchers can design site-directed mutagenesis experiments (such as mutating certain Lys / Arg residues) based on the key residues or structural fragments highlighted in the model to verify whether DNA binding ability is weakened, thus directly using the model's interpretation results to guide experiments.

[0222] S205 Analogy and Annotation to Known DBP

[0223] In practical applications, this invention can also combine known DBP information in the database to provide "analogical explanations" for the prediction results and help with functional annotations.

[0224] Multimodal feature space analogy:

[0225] During the training or additional processing phase, the multimodal fusion features of known DBPs of the source species can be clustered or visualized (e.g., reduced to 2 or 3 dimensions).

[0226] For a protein predicted as a DBP in a target species or a new species, examine which known DBP its fusion characteristics are most similar to in this space.

[0227] Generate comment hints:

[0228] If the characteristics of a new protein are very similar to those of a known class of transcription factors, a description can be added to the model output: "This protein is similar to a certain class of transcription factors in the multimodal feature space and may have a similar DNA binding pattern."

[0229] This analogy does not constitute a rigorous functional prediction, but it can provide biologists with intuitive "similar objects" as clues for further literature searches and experimental design.

[0230] As can be seen from this embodiment, the protein classification method based on multimodal feature fusion and interpretable networks proposed in this invention not only achieves high accuracy on single-species benchmark datasets but also maintains good generalization ability in cross-species scenarios. Furthermore, it can be naturally embedded into practical research processes such as proteomics analysis of new species, candidate target screening, and site-directed mutagenesis design. Without altering the overall framework of this invention, adjustments to the selection of source and target species, data volume, pre-trained model type, graph neural network structure, and specific settings of training hyperparameters are all reasonable modifications that can be made by those skilled in the art and should be considered within the scope of protection of this invention.

Claims

1. A protein classification method based on multimodal feature fusion and interpretable networks, characterized in that... The following processing steps are included: Obtain the amino acid sequence of the protein to be tested; encode the amino acid sequence based on a pre-trained protein sequence language model to obtain a first feature representation characterizing the local and global contextual information of the sequence, which serves as the sequence modality feature; The amino acid sequence is input into the protein structure prediction model to obtain the predicted three-dimensional structure information of the protein to be tested, and a second feature representation is extracted based on the three-dimensional structure information as a structural modality feature. Constructing a protein structure graph using the three-dimensional structural information and / or the structural modality features specifically includes: treating each amino acid residue in the protein to be tested as a graph node; establishing graph edges between node pairs that meet a preset distance threshold or contact threshold based on the adjacency relationship of residues in the amino acid sequence or the spatial distance relationship between the three-dimensional coordinates of residues; using the sequence modality features as the initial node features, and using the structural modality features or features extracted based on the geometric relationship between residues as the initial edge features, thereby forming a protein graph representation that integrates sequence and structural information; The sequence modality features and the structural modality features are input into the graph neural network fusion module. By transferring and updating the features of nodes and edges in the protein structure graph, the sequence modality and the structural modality are interactively fused to obtain a multimodal fusion feature representation that characterizes the overall properties of the protein under test. The multimodal fusion feature representation input is configured as an interpretable neural network classifier that can provide decision-making basis, to obtain a classification result for characterizing whether the protein to be tested is a DNA-binding protein.

2. The method according to claim 1, characterized in that, The method also includes an interpretability analysis step: Based on the function structure and parameters inside the interpretable neural network classifier, the contribution of each dimension or each amino acid residue of the multimodal fusion feature representation is estimated, and feature importance scores or response curves are generated to explain the decision basis of the classification results.

3. The method according to claim 1, characterized in that, The pre-trained protein sequence language model is configured as follows: Extract sequence embedding features that include residue physicochemical properties, evolutionary conservation, and long-range dependencies without the need for explicit multiple sequence alignment; The pre-trained protein sequence language model includes at least one of the ProtT5 model, the ESM series of protein language models, or a deep pre-trained protein language model based on the Transformer architecture.

4. The method according to claim 1, characterized in that, The protein structure prediction model is configured as follows: The output includes one or more of the following three-dimensional structural information: residue three-dimensional coordinates, residue distance matrix, contact diagram, or secondary structure annotation; the protein structure prediction model includes at least one of the AlphaFold series models, ESMFold models, or deep learning protein structure prediction models based on Transformer architecture or graph neural networks.

5. The method according to claim 1, characterized in that, The graph neural network fusion module includes a graph convolutional neural network or a graph attention neural network, and updates the multimodal fusion features through at least two layers of message passing and aggregation operations, wherein: Each layer of the graph neural network uses weighted weights to aggregate the sequential modal features and structural modal features of the node's neighborhood to obtain the updated node representation; The weighting is adaptively learned by an attention mechanism based on node features, edge features, or modality type to distinguish the contributions of different neighboring nodes and different modal information.

6. The method according to any one of claims 1 to 5, characterized in that, The interpretable neural network classifier is a Kolmogorov–Arnold network or a variant thereof, and the Kolmogorov–Arnold network is configured as follows: It contains at least one nonlinear transformation layer based on a learnable activation function; In the nonlinear transformation layer, each dimension of the input feature is mapped to a set of basis functions, and the basis functions are weighted by trainable control coefficients to construct a univariate nonlinear activation function on the edge. The outputs of the univariate nonlinear activation functions corresponding to each dimension are aggregated or linearly combined to obtain the output features or final classification score of that layer.

7. The method according to any one of claims 1 to 5, characterized in that, The interpretable neural network classifier employs a differentiable neural decision tree model; the differentiable neural decision tree model is configured to generate a discrimination path approximating a decision tree by integrating a segmented decision structure or soft routing mechanism in a feedforward network, for outputting a visualized rule-level explanation result.

Images