A protein ligand affinity prediction method based on multi-source feature fusion
By using a multi-source feature fusion method, a protein-ligand binding affinity prediction model is constructed using dilated convolution and cross-transformer neural networks. This solves the problems of high prediction accuracy and high computational cost in existing technologies, and achieves efficient and accurate protein-ligand binding affinity prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ANQING NORMAL UNIV
- Filing Date
- 2026-04-09
- Publication Date
- 2026-07-03
AI Technical Summary
Existing technologies suffer from insufficient accuracy and generalization ability in predicting protein-ligand binding affinity. In particular, traditional methods are prone to introducing structural biases when dealing with protein flexibility changes and solvent molecule interactions, and they are computationally expensive, making them difficult to apply widely in large-scale compound screening.
A multi-source feature fusion method is adopted, which extracts global protein features through dilated convolutional neural networks and combines one-dimensional convolutional neural networks and cross-transformer neural networks to capture the interaction between proteins and ligands. A protein-ligand binding affinity prediction model is constructed, and ligand features are fused using TTC networks to reduce computational complexity and improve prediction accuracy.
It improves the accuracy and stability of protein-ligand binding affinity prediction, reduces experimental validation costs, increases screening efficiency, and enhances the model's generalization ability and computational efficiency.
Smart Images
Figure CN122337346A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of protein and ligand binding affinity prediction technology, and specifically to a protein ligand affinity prediction method based on multi-source feature fusion. Background Technology
[0002] Protein-ligand binding affinity prediction is a core task in computational drug design, aiming to accurately assess the binding strength between small molecule drug candidates and target proteins, thereby guiding drug screening, chemical structure optimization, and subsequent experimental validation. In traditional drug development, researchers primarily rely on molecular docking, empirical scoring functions, and force-field-based molecular dynamics simulations to predict protein-ligand binding affinity. These methods have played a crucial role in early drug screening, providing feasible candidate sets for high-throughput experiments; however, their predictive accuracy and generalization ability are often limited. For example, molecular docking methods are sensitive to changes in protein flexibility and solvent molecule interactions, easily introducing structural biases; empirical scoring functions have limited applicability across different types of compounds, and their scoring models often rely on specific training data or parameter sets; while molecular dynamics simulations can provide relatively fine kinetic information and high-precision energy assessments, their extremely high computational cost makes them difficult to widely apply in large-scale compound screening.
[0003] Therefore, there is an urgent need for a high-precision and reliable method to predict the previously undiscovered affinity between proteins and ligands. With the continuous advancement of neural network technology, deep learning is increasingly being applied to the prediction of protein-ligand binding affinity. Neural network-based protein-ligand binding affinity prediction methods can be mainly divided into two categories: one relies solely on protein and ligand sequence information, while the other combines protein sequence, protein structure information, and the chemical structural features of the ligand. Methods relying solely on sequence information model by extracting features from the protein amino acid sequence and the linear representation of the ligand. These methods do not fully consider the three-dimensional structure of the protein and the spatial information of its binding pocket region, thus limiting the model's ability to characterize the true interaction mechanism to some extent. In contrast, methods combining protein sequence features, protein structure features, and ligand structure features can not only characterize the protein's sequence and spatial conformation information but also depict the chemical structural characteristics of the ligand, thereby achieving a more comprehensive and refined modeling of protein-ligand interactions. This type of method, to some extent, makes up for the shortcomings of relying solely on sequence information and improves the feature expression ability. However, it also increases the computational complexity and implementation difficulty of the model due to the relatively complex extraction and representation process of protein structural features and ligand structural features.
[0004] In summary, protein-ligand binding affinity prediction, a crucial step in computational drug design, is shifting from traditional physical and empirical methods to intelligent methods based on deep learning. Researchers both domestically and internationally are continuously improving the prediction accuracy, generalization ability, and interpretability of models by introducing advanced technologies such as convolutional neural networks, graph neural networks, attention mechanisms, and multimodal feature fusion. With the accumulation of high-quality structural and experimental data, and the development of more efficient and interpretable deep learning models, protein-ligand binding affinity prediction is expected to play a greater role in new drug development, precision medicine, and personalized treatment, providing more reliable computational tools for drug design. Summary of the Invention
[0005] The technical problem to be solved by this invention is: how to extract key features that can effectively characterize the properties of protein and ligand molecules, and on the basis of obtaining the above key features, to achieve accurate and efficient prediction of protein-ligand affinity relationships; at the same time, completing the affinity prediction process through machine automation helps to reduce experimental verification costs and improve screening efficiency; and a protein ligand affinity prediction method based on multi-source feature fusion is provided.
[0006] The present invention solves the above-mentioned technical problems through the following technical solution, and the present invention includes the following steps:
[0007] S1: Preprocessing of protein-ligand data
[0008] The protein-ligand data were preprocessed to obtain standardized training, validation, and test sets.
[0009] S2: Features of extracting and fusing proteins and ligands
[0010] The basic features of the protein are extracted and split to obtain global protein features and local pocket features. At the same time, the SMILES representation features and SELFIES representation features of the ligands are extracted. The features are then fused through the TTC network to obtain the ligand fusion features.
[0011] S3: Constructing a dilated convolutional neural network
[0012] A dilated convolutional neural network is constructed to extract global protein features, resulting in multi-scale global protein feature representations.
[0013] S4: Construct a one-dimensional convolutional neural network
[0014] A one-dimensional convolutional neural network is constructed to extract features from local pocket features, resulting in an encoded local pocket embedding vector.
[0015] S5: Constructing a cross-Transformer neural network
[0016] A cross-transformer neural network is constructed to jointly model ligand fusion features and local pocket embedding vectors, capturing the interaction between the two and obtaining interactive feature representations.
[0017] S6: Combining affinity prediction
[0018] The global protein feature representation, local pocket embedding vector, and interaction feature representation are concatenated, input into the MLP for regression prediction, and the protein-ligand binding affinity prediction result is output.
[0019] Furthermore, in step S1, the specific processing procedure is as follows:
[0020] S11: Obtain the general set, refined set, and core set of the PDBbind database 2016 version, and standardize each dataset according to a fixed length. Set 1000 characters as the fixed length for protein sequences, 150 characters as the fixed length for SMILES strings, and 63 characters as the fixed length for pocket sequences. Truncate sequences longer than the fixed length and pad sequences shorter than the fixed length with 0s.
[0021] S12: Use the processed core set as the test set, mix the general set and the refined set, randomly select 1000 complexes as the validation set, and use the remaining complexes as the training set.
[0022] Furthermore, in step S2, the basic features of the global protein features and the local pocket features are the same protein feature vector. The process of obtaining the global protein features and the local pocket features is as follows:
[0023] S21: The secondary structure state of each residue of the protein was predicted using SSPro software, and the residue structure features were obtained by 8-dimensional one-hot vector encoding.
[0024] S22: Using a 21-dimensional one-hot vector to encode 21 residues in a protein sequence, residue sequence features are obtained;
[0025] S23: The chemical properties of each residue side chain structure of a protein are represented by an 11-dimensional vector to obtain the residue chemical characteristics, which include polarity, acidity, basicity, dipole and volume characteristics;
[0026] S24: The residue structure features, residue sequence features, and residue chemical features are spliced and fused to construct a 40-dimensional protein residue feature vector. Based on this, the protein feature vector is split according to the positional relationship between the overall protein structure and the binding pocket region to obtain global protein features and local pocket features.
[0027] Furthermore, in step S2, the process of obtaining the ligand fusion features is as follows:
[0028] S25: Convert the original SDF file of the ligand into a SMILES string using Open Babel, split the SMILES characters into single character sequences and encode them into integers to obtain the SMILES representation features; convert the SMILES string into a SELFIES representation, split the SELFIES symbol units and encode them into integers to obtain the SELFIES representation features;
[0029] S26: Input the SMILES representation features and SELFIES representation features into the TTC network for feature interaction and fusion to obtain ligand fusion features.
[0030] Furthermore, in step S26, the specific fusion process is as follows:
[0031] S261: First, input the SMILES representation features into two layers of Transformer Encoder to extract contextual dependencies, and then align them with the dimensions of the SELFIES representation features through linear projection;
[0032] S262: Then input the SELFIES representation features into 1D convolution to extract local patterns, and capture global dependencies through Transformer Encoder to obtain the dense feature representation of SELFIES;
[0033] S263: The dimension-aligned SMILES representation and the SELFIES dense representation are fused using a cross-attention mechanism to obtain ligand fusion features.
[0034] Furthermore, in step S3, a dilated convolutional neural network containing a dilated convolutional module is constructed. The dilated convolutional module consists of 5 layers of dilated convolutions with a kernel size of 3×3. The dilation rates of each layer of dilated convolutions are set to 1, 2, 4, 8, and 16, respectively. The dilated convolutions with different dilation rates are used to extract global protein features at multiple scales. The different receptive field features output by each layer of dilated convolutions are spliced together through hierarchical fusion to obtain a multi-scale global protein feature representation.
[0035] Furthermore, a residual module is introduced into the dilated convolution module. A residual connection is set in each dilated convolution module to directly add the input features and output features of the dilated convolution to form a skip connection, thereby extracting features from global protein features while retaining the original feature information.
[0036] Furthermore, in step S4, the processing procedure for the local pocket features is as follows:
[0037] S41: Construct a one-dimensional convolutional neural network consisting of three Conv1D layers, with each Conv1D layer having a window size of 3.
[0038] S42: The local pocket features are sequentially input into the Conv1D layer, batch normalization layer, and activation function layer for processing;
[0039] S43: The processed features are compressed to a fixed size using adaptive max pooling, and then dimension compression is performed using the Squeeze operation to obtain the encoded local pocket embedding vector.
[0040] Furthermore, in step S5, the specific processing procedure for the interaction between ligand fusion features and local pocket features is as follows:
[0041] S51: The ligand fusion features and local pocket embedding vectors are input into a Transformer neural network to achieve mutual interaction and information fusion between the ligand fusion features and local pocket features;
[0042] S52: Output the interaction feature representation obtained across the attention fusion module for subsequent affinity prediction.
[0043] Furthermore, the TTC network, one-dimensional convolutional neural network, cross-transformer neural network, and MLP are combined to form a protein-ligand binding affinity prediction model. During the model training phase, the Adam optimizer is used to iteratively update all parameters of the model, with an initial learning rate set to 0.0003, 20 training epochs, and mean squared error (MSE) as the loss function. During training, the model performance in each epoch is evaluated using a validation set, and the model with the smallest MSE in the validation set is selected as the final protein-ligand binding affinity prediction model. Finally, the test set is input into the final protein-ligand binding affinity prediction model to obtain the protein-ligand binding affinity prediction results.
[0044] Compared with existing technologies, this invention has the following advantages: This protein ligand affinity prediction method based on multi-source feature fusion fuses the SMILES and SELFIES representations of the ligand, thereby enhancing the ability to characterize the chemical structure information of the ligand; it extracts global protein features through dilated convolution and sets different dilation rates to capture multi-scale contextual information, avoiding the computational complexity and information redundancy problems caused by multi-layer convolution stacking; it uses cross-transformer to fuse local pocket features and ligand features, realizing the interaction and alignment of key information between proteins and ligands; and it uses one-dimensional convolution to extract local pocket features, making the model more focused on the binding site region, thereby improving the ability to characterize protein-ligand binding patterns and the stability of prediction results. Attached Figure Description
[0045] Figure 1 This is a flowchart illustrating the protein ligand affinity prediction method based on multi-source feature fusion in an embodiment of the present invention.
[0046] Figure 2 This is a schematic diagram of the TTC network structure in the protein-ligand binding affinity prediction model in this embodiment of the invention;
[0047] Figure 3 This is a schematic diagram of the structure of the dilated convolutional network in the protein-ligand binding affinity prediction model in this embodiment of the invention;
[0048] Figure 4 This is a schematic diagram of the structure of a one-dimensional convolutional network in the protein-ligand binding affinity prediction model in an embodiment of the present invention;
[0049] Figure 5 This is a schematic diagram of the overall structure of the protein-ligand binding affinity prediction model in an example of the present invention. Detailed Implementation
[0050] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.
[0051] like Figure 1 As shown, this embodiment provides a technical solution: a protein ligand affinity prediction method based on multi-source feature fusion, comprising the following steps:
[0052] S1: Preprocessing of protein-ligand data.
[0053] Step S1 specifically includes the following:
[0054] S11: Obtain the three datasets from the PDBbind database version 2016 (as shown in Table 1): the General Set, the Refined Set, and the Core Set (Core 2016). Since the protein sequences, pocket sequences, and ligand strings have different lengths, a fixed length must be ensured to create a valid representation. The datasets are padded sequentially. We defined a fixed length of 1000 protein sequence characters, 150 characters for the SMILES string, and 63 characters for the pocket sequence to cover approximately 90% of the proteins, 90% of the ligands, and 90% of the pockets in these datasets. Sequences longer than the fixed length were truncated, and sequences shorter than the fixed length were padded with 0.
[0055] Table 1. Three datasets in the 2016 version of the PDBbind database.
[0056]
[0057] It should be noted that in this experiment, the general set and the refined set were mixed and 1000 complexes were randomly selected as the validation set, and the remaining 11,906 complexes were used as the training set; the core set (Core Set 2016) was used for model testing.
[0058] S2: Features of extracting and fusing proteins and ligands
[0059] Step S2 specifically includes the following:
[0060] S21: Extracting protein structure information
[0061] In step S21, the secondary structure state of each residue in the protein sequence is predicted. This prediction is performed using the software SSPro, outputting eight classes of secondary structures. -spiral, spiral, -spiral, -fold, -Bridges, turns, bends and irregular curls), and the structural features of each residue are encoded using an 8-dimensional one-hot vector;
[0062] S22: Extract protein sequence information
[0063] In step S22, 20 different types of amino acids are encoded based on the protein sequence information; in addition, non-standard residues are also included in some proteins. We use 21-dimensional one-hot encoding to encode 21 different types of residues in the protein sequence;
[0064] S23: Extracting information on the chemical properties of proteins
[0065] In step S23, the chemical properties of the residue side chain structure, polarity, acidity, basicity information, as well as dipole and volume characteristics, are used to characterize the physicochemical properties of each residue in a total of 11 dimensions.
[0066] S24: Fusion protein characteristics
[0067] In step S24, a 40-dimensional feature vector of the protein is constructed by combining the structural features, sequence features, and chemical properties of the residue side chain structures. Based on this, the protein feature vector is split according to the positional relationship between the overall protein structure and the binding pocket region to obtain global protein features and local pocket features.
[0068] S25: Extraction of ligand SMLIES representation and SMLIES representation
[0069] In step S25, based on the ligand structure representation, the original SDF file is converted into a SMILES string using Open Babel. Each SMILES character is then split into a single-character sequence and encoded using integers to form the SMILES representation features. Simultaneously, the SMILES string is converted into a SELFIES representation, and the SELFIES symbol units are split and encoded using integers to obtain the SELFIES representation features.
[0070] S26: Fusion ligand characteristic information
[0071] In step S26, after obtaining the independent representations of SMILES and SELFIES, this invention designs a TTC network to fuse the two, obtaining the fused features of SMLIES and SELFIES. Specifically, the SMILES representation features are first processed by two Transformer Encoder layers to extract contextual dependencies, and then aligned with the dimensions of the SELFIES representation features through linear projection to ensure fusionability. The SELFIES branch processes the SELFIES representation features through 1D convolution to extract local patterns, and then captures global dependencies through a Transformer Encoder to obtain a dense feature representation. Subsequently, a cross-attention fusion mechanism is used to fuse the two, obtaining the fused features of SMLIES and SELFIES.
[0072] S3: Constructing a dilated convolutional neural network
[0073] In step S3, convolutions with different dilation rates are used to extract global protein features, and hierarchical fusion is used to fuse features from different receptive fields to cover multi-scale information from local to global.
[0074] It should be noted that the above convolution adds a residual module, which introduces residual connections into each dilated convolution module, allowing input features to be directly added to output features to form "skip connections". This alleviates the gradient vanishing problem in deep networks, while preserving the original feature information and enhancing the model's expressive power and training stability.
[0075] S4: Construct a one-dimensional convolutional neural network
[0076] The local pocket features of the protein are processed using three-layer one-dimensional convolution (Conv1D), with a convolution window size of 3 for each layer. This branch focuses more on the features of local residues.
[0077] S5: Constructing a cross-Transformer neural network
[0078] A cross-transformer neural network is constructed, which uses ligand fusion features and local pocket features as inputs to further capture the potential interaction between ligands and pockets, enabling feature representations from different sources to be interdependent and dynamically aggregated.
[0079] It should be noted that the described cross-Transformer neural network is built based on the Transformer attention mechanism, which is a commonly used attention network structure in this field. This network models the associations of features from different sources through a cross-attention mechanism. In this embodiment, the number of attention heads in the cross-Transformer is set to 4, enabling the model to model the relationships between global protein features, local pocket features, and ligand fusion features from four different subspaces, improving expressive power and stability while maintaining reasonable computational efficiency.
[0080] S6: Prediction of protein-ligand binding affinity
[0081] The outputs of the three networks are concatenated and then used to predict affinity using an MLP.
[0082] Specifically, in step S6, the specific training scheme is as follows:
[0083] S61: During the model training phase, the initial learning rate is set to 0.0003, the mean squared error loss function is used as the objective function, and the Adam optimizer is used to iteratively update the model parameters. The entire training process consists of 20 training epochs.
[0084] S62: After completing model training, the performance of the model corresponding to each training round is evaluated on the validation dataset. The model with the smallest validation set error is selected as the final prediction model to improve the generalization ability and stability of the protein-ligand binding affinity prediction model.
[0085] S63: During model training, to verify the contribution of the SELFIES feature to the overall model performance, an experiment was conducted by removing this feature from the model structure. At the same time, the TTC network with fused features was also removed, and the results were verified on the test dataset (Test).
[0086] Table 2. Performance of SELFIES tested on Test.
[0087]
[0088] S64: Input the test set data (Test) into the protein-ligand binding affinity prediction model for prediction and obtain the prediction results, as shown in Table 3. Our model (our) significantly outperforms other methods on the benchmark dataset Test in five metrics.
[0089] Table 3 Performance comparison with other methods on Test.
[0090]
[0091] In summary, the protein ligand affinity prediction method based on multi-source feature fusion described in the above embodiments first utilizes three datasets from the widely used and publicly available PDBbind database 2016, containing protein sequences, pocket structures, and ligand information. The datasets undergo multi-step preprocessing, including screening for complete protein structures and standardizing ligand representations, to ensure high-quality and consistent training data. During model construction, global protein features, local pocket features, and ligand SMILES and SELFIES representations are used as inputs. Ligand features are fused using a TTC network, global protein features are extracted using a dilated convolutional network, local protein pocket features are processed using one-dimensional convolution, and finally, a cross-transformer is used to capture the potential interactions between local protein pocket features and ligand features. During training, a mean squared error loss function is used to accurately measure the prediction error, and the model parameters are iteratively updated using an Adam optimizer, enabling the model to achieve higher accuracy and stability in affinity prediction tasks. On the test set, by comparing the performance of different models, experimental results show that the multi-source fusion method proposed in this invention can effectively improve the model convergence speed and robustness, significantly enhance the prediction accuracy of protein-ligand binding affinity, and demonstrate good generalization ability. In the future, this method is expected to be deployed on edge computing devices to achieve efficient protein-ligand binding affinity prediction and widespread application.
[0092] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.
Claims
1. A method for predicting protein ligand affinity based on multi-source feature fusion, characterized in that, Includes the following steps: S1: Preprocessing of protein-ligand data The protein-ligand data were preprocessed to obtain standardized training, validation, and test sets. S2: Features of extracting and fusing proteins and ligands The basic features of the protein are extracted and split to obtain global protein features and local pocket features. At the same time, the SMILES representation features and SELFIES representation features of the ligands are extracted. The features are then fused through the TTC network to obtain the ligand fusion features. S3: Constructing a dilated convolutional neural network A dilated convolutional neural network is constructed to extract global protein features, resulting in multi-scale global protein feature representations. S4: Construct a one-dimensional convolutional neural network A one-dimensional convolutional neural network is constructed to extract features from local pocket features, resulting in an encoded local pocket embedding vector. S5: Constructing a cross-Transformer neural network A cross-transformer neural network is constructed to jointly model ligand fusion features and local pocket embedding vectors, capturing the interaction between the two and obtaining interactive feature representations. S6: Combining affinity prediction The global protein feature representation, local pocket embedding vector, and interaction feature representation are concatenated, input into the MLP for regression prediction, and the protein-ligand binding affinity prediction result is output.
2. The protein ligand affinity prediction method based on multi-source feature fusion according to claim 1, characterized in that, In step S1, the specific processing procedure is as follows: S11: Obtain the general set, refined set, and core set of the PDBbind database 2016 version, and standardize each dataset according to a fixed length. Set 1000 characters as the fixed length for protein sequences, 150 characters as the fixed length for SMILES strings, and 63 characters as the fixed length for pocket sequences. Truncate sequences longer than the fixed length and pad sequences shorter than the fixed length with 0s. S12: Use the processed core set as the test set, mix the general set and the refined set, randomly select 1000 complexes as the validation set, and use the remaining complexes as the training set.
3. The protein ligand affinity prediction method based on multi-source feature fusion according to claim 1, characterized in that, In step S2, the basic features of the global protein features and the local pocket features are the same protein feature vector. The process of obtaining the global protein features and the local pocket features is as follows: S21: The secondary structure state of each residue of the protein was predicted using SSPro software, and the residue structure features were obtained by 8-dimensional one-hot vector encoding. S22: Using a 21-dimensional one-hot vector to encode 21 residues in a protein sequence, residue sequence features are obtained; S23: The chemical properties of each residue side chain structure of a protein are represented by an 11-dimensional vector to obtain the residue chemical characteristics, which include polarity, acidity, basicity, dipole and volume characteristics; S24: The residue structure features, residue sequence features, and residue chemical features are spliced and fused to construct a 40-dimensional protein residue feature vector. Based on this, the protein feature vector is split according to the positional relationship between the overall protein structure and the binding pocket region to obtain global protein features and local pocket features.
4. The protein ligand affinity prediction method based on multi-source feature fusion according to claim 3, characterized in that, In step S2, the process of obtaining ligand fusion features is as follows: S25: The original SDF file of the ligand is converted into a SMILES string using Open Babel. The SMILES characters are split into single-character sequences and then encoded into integers to obtain the SMILES representation features. The SMILES string is converted into a SELFIES representation, and after splitting the SELFIES symbol units, integer encoding is performed to obtain the SELFIES representation features; S26: Input the SMILES representation features and SELFIES representation features into the TTC network for feature interaction and fusion to obtain ligand fusion features.
5. The protein ligand affinity prediction method based on multi-source feature fusion according to claim 4, characterized in that, In step S26, the specific fusion process is as follows: S261: First, input the SMILES representation features into two layers of Transformer Encoder to extract contextual dependencies, and then align them with the dimensions of the SELFIES representation features through linear projection; S262: Then input the SELFIES representation features into 1D convolution to extract local patterns, and capture global dependencies through Transformer Encoder to obtain the dense feature representation of SELFIES; S263: The dimension-aligned SMILES representation and the SELFIES dense representation are fused using a cross-attention mechanism to obtain ligand fusion features.
6. The protein ligand affinity prediction method based on multi-source feature fusion according to claim 4, characterized in that, In step S3, a dilated convolutional neural network containing a dilated convolutional module is constructed. The dilated convolutional module consists of 5 layers of dilated convolutions with a kernel size of 3×3. The dilation rates of each layer of dilated convolutions are set to 1, 2, 4, 8, and 16, respectively. Global protein features are extracted at multiple scales using dilated convolutions with different dilation rates. The different receptive field features output by each layer of dilated convolutions are spliced together through hierarchical fusion to obtain a multi-scale global protein feature representation.
7. The protein ligand affinity prediction method based on multi-source feature fusion according to claim 6, characterized in that, The dilated convolution module introduces a residual module, and a residual connection is set in each dilated convolution module. The input features and output features of the dilated convolution are directly added to form a skip connection, which extracts features of global protein features while retaining the original feature information.
8. The protein ligand affinity prediction method based on multi-source feature fusion according to claim 7, characterized in that, In step S4, the processing procedure for the local pocket features is as follows: S41: Construct a one-dimensional convolutional neural network consisting of three Conv1D layers, with each Conv1D layer having a window size of 3. S42: The local pocket features are sequentially input into the Conv1D layer, batch normalization layer, and activation function layer for processing; S43: The processed features are compressed to a fixed size using adaptive max pooling, and then dimension compression is performed using the Squeeze operation to obtain the encoded local pocket embedding vector.
9. The protein ligand affinity prediction method based on multi-source feature fusion according to claim 8, characterized in that, In step S5, the specific processing procedure for the interaction between ligand fusion features and local pocket features is as follows: S51: The ligand fusion features and local pocket embedding vectors are input into a Transformer neural network to achieve mutual interaction and information fusion between the ligand fusion features and local pocket features; S52: Output the interaction feature representation obtained across the attention fusion module for subsequent affinity prediction.
10. The protein ligand affinity prediction method based on multi-source feature fusion according to claim 9, characterized in that, The TTC network, one-dimensional convolutional neural network, cross-transformer neural network, and MLP are combined to form a protein-ligand binding affinity prediction model. During the model training phase, the Adam optimizer is used to iteratively update all parameters of the model, with an initial learning rate set to 0.0003, 20 training epochs, and mean squared error (MSE) as the loss function. During training, the model performance in each epoch is evaluated using a validation set, and the model with the smallest MSE on the validation set is selected as the final protein-ligand binding affinity prediction model. Finally, the test set is input into the final protein-ligand binding affinity prediction model to obtain the protein-ligand binding affinity prediction results.