Domain adaptation and graph network based prediction of ligand-specific binding residues

Abstract

A ligand specific binding residue prediction method based on domain adaptation and graph network, according to the physicochemical properties of the ligand, the ligands in the pre-training data set are clustered into multiple domains, and then a relationship perception classifier is trained for each domain using the data of all ligands contained in the domain, which is used to learn the shared binding mode of the ligands contained in the domain. In the fine-tuning stage, for each ligand, the pre-trained relationship perception classifier is fine-tuned through the ligand-specific data set, and the prediction module based on the domain adaptation neural network is used to infer the weights of the relationship perception classifiers corresponding to different domains, and the weighted sum of their results is obtained to obtain the final predicted binding probability.

Description

Technical Field

[0001] This invention relates to a technology in the field of bioengineering, specifically a method for predicting ligand-specific binding residues based on domain adaptation and graph networks. Background Technology

[0002] Existing computational methods for studying protein-ligand interactions based on protein structure include ligand-general methods and ligand-specific methods. The former is a method for predicting binding residues regardless of ligand type, collecting proteins that interact with multiple ligands to form a joint dataset for model design. Ligand-general methods are widely used when there is no specific target ligand type, or when the number of proteins binding to a particular ligand is too small to train a model. However, different ligands interact with proteins in diverse ways, and simply mixing all ligands may ignore specific binding patterns, leading to a decrease in the effectiveness of ligand-general models. For example, metal complexation is crucial for metal ion-protein interactions, while in protein-nucleic acid interactions, protein topological features and water solubility are more critical. The latter predicts the specific binding residues of a particular ligand. These ligands interact with a sufficient number of proteins to construct a ligand-specific dataset, which can be used to train a ligand-specific predictor. Previous ligand-specific prediction algorithms, limited by the availability of sufficient ligand-specific datasets, often only predict a few ligand types, thus restricting their applicability. Summary of the Invention

[0003] This invention addresses the shortcomings of existing technologies, which can only be applied to ligands with a relatively large number of known binding proteins, cannot be extended to small ligand-specific datasets, and lack universality. It proposes a ligand-specific binding residue prediction method based on domain adaptation and graph networks. Through the transfer learning model of domain adaptation and graph networks, it can predict the specific binding residues of more than 1,000 ligands based on protein structure.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a method for predicting ligand-specific binding residues based on domain adaptation and graph network transfer learning, comprising:

[0006] Step 1: Collect the ligand universality dataset and cluster it to obtain the number of ligand domains N. Extract the physicochemical features and graphical representations of protein residues from the ligand universality dataset.

[0007] The aforementioned ligand generality dataset contains data on the interaction between various ligands and proteins. Using the RDKit software package, based on the SMILES encoding of the ligands, each pair of interacting proteins and ligands in the dataset is clustered into N domains according to the calculated 200-dimensional physicochemical features.

[0008] The physicochemical characteristics include: molecular mass, charge, water solubility, and molecular surface descriptor.

[0009] The clustering described herein uses the k-means clustering algorithm to separate N domains from the ligand features.

[0010] The graph representation of protein residues is achieved by representing each residue of the protein in the ligand universality dataset with a 44-dimensional feature vector, and representing the local information of the residues as a graph structure. Specifically, for a protein of length L, HHblits are used to calculate L×30-dimensional evolutionary information, and the hydrogen bond estimation algorithm (DSSP) is used to calculate L×14-dimensional secondary structure information, resulting in a total of L×44-dimensional protein features. Using the centroid position of each residue in the protein as the spatial coordinate of the residue, for each residue s, a graph is constructed based on all residues within a radius of 15 angstroms. Where: residues are nodes The 44-dimensional feature vector of a residue is a node feature. An edge is defined between two residues when the Euclidean distance between them is less than 10 angstroms. The 2D edge features are the distance and cosine of the included angle between two adjacent residues.

[0011] Step 2: Construct a residue binding probability prediction model that includes a ligand coding layer, a residue coding layer, and a relation-aware classifier. Set the number of relation-aware classifiers according to the number of ligand domains N. Input the physicochemical features and the graph representation of protein residues into the ligand coding layer and the residue coding layer respectively, and concatenate the results into ligand residue pair features. Train each relation-aware classifier using the ligand residue pair features contained in each domain of the ligand universality dataset.

[0012] The ligand encoding layer is based on the input 200-dimensional residue features. Output ligand characterization ,in: , , , BatchNorm represents the model parameters to be optimized. This is a layer normalization layer.

[0013] The residue coding layer, based on the input graph After unifying the 2D edge features and 44D node features to the same feature dimension through batch normalization and linear transformation layers, the residue encoding is extracted and the residue representation is output through five stacked graph network layers. Specifically, the graph network at layer k extracts the features of edge (i, j), node i, and graph features of layer k, respectively: , , ,in: , , These are the features of the edge (i, j) at the k-th layer, the features of node i, and the graph features, respectively. [To splice on the last dimension] and Features Let i be the set of neighboring nodes of node i. Let be the number of neighboring nodes of node i. For the image The number of nodes in the middle, It is a nonlinear transformation layer. For recurrent neural network layers, the last graph network layer outputs graph features. That is, residue characterization.

[0014] The training of each relation-aware classifier refers to training a corresponding relation-aware classifier in each domain using the representation of ligand residue pairs as input. Specifically, for each of the N domains obtained after ligand clustering, the data of interacting ligands and proteins within that domain are used as samples for training. The input is the representation of the ligand residue pairs obtained by concatenating the encoded ligand representation with the residue representation, and the output is the binding probability of the two. This represents the characteristics of the spliced ​​ligand residue pairs.

[0015] Step 3: Using a ligand-specific dataset, repeat the extraction of physicochemical features and graph representations of protein residues in Step 1, and perform the same training on the residue binding probability prediction model obtained in Step 2 to achieve model fine-tuning and optimization.

[0016] Step 4: In the online phase, the residue binding probability prediction model after fine-tuning the input of the ligand residue pairs of the test data is used to obtain the classification output results corresponding to N domains. At the same time, the domain prediction layer is randomly initialized by inputting the input of the ligand residue pairs of the test data, and the corresponding domain membership probability is used as a weight to perform weighted summation on the classification output results, thereby obtaining the final predicted binding probability of residue pairs to specific ligands.

[0017] The domain dependency probability is obtained by: based on ligand residue pairs ,in: This represents the characterization of ligand and residue pairs obtained through the ligand coding layer, residue coding layer, and domain prediction layer for specific ligands t, ligand l, and residues s.

[0018] The weighted summation mentioned above is specifically as follows: ,in: Let be the domain membership probability of ligand l and residue s belonging to domain n. This represents the prediction result obtained by the relation-aware classifier corresponding to domain n for ligand l and residue s.

[0019] Technical effect

[0020] This invention uses a domain-adaptive transfer method to predict the specific binding residues of ligands in a small sample, thereby achieving the prediction of binding residues on a small sample ligand-specific dataset. Attached Figure Description

[0021] Figure 1 This is a flowchart of the present invention;

[0022] Figure 2 This is a schematic diagram of the ligand coding layer;

[0023] Figure 3 A schematic diagram of the residue coding layer;

[0024] Figure 4 A schematic diagram of a relation-aware classifier;

[0025] Figure 5 The image shown is a rendering of an example. Detailed Implementation

[0026] This embodiment relates to a ligand-specific binding residue prediction system based on domain adaptation and graph networks, comprising: a ligand feature extraction unit, a protein feature extraction unit, a ligand clustering unit, a ligand encoding unit, a graph representation unit of residues, a residue encoding unit, a relation-aware classifier unit, and a domain prediction layer unit. Specifically: the ligand feature extraction unit extracts ligand features based on the ligand's SMILES information; the protein feature extraction unit extracts protein features based on the protein's sequence and structural information; and the ligand clustering unit clusters the ligands in the pre-training dataset based on the ligand features. The ligands are clustered into N domains. The ligand encoding unit obtains the ligand characterization through a nonlinear transformation network based on the ligand characteristics. The residue graph representation unit generates a graph representation for each residue based on the protein characteristics and residue position information. The residue encoding unit learns the residue characterization through a graph network based on the residue graph representation. The relation-aware classifier unit performs nonlinear transformation processing based on the ligand characteristics and residue characterization to obtain the residue-ligand binding probability. The domain prediction layer unit performs nonlinear transformation processing based on the ligand characteristics and residue characterization to obtain the probability that the residue-ligand binding pattern belongs to each domain.

[0027] like Figure 1 As shown in the figure, this embodiment relates to a method for predicting ligand-specific binding residues based on domain adaptation and graph networks, which includes the following steps:

[0028] Step 1) Calculate the physicochemical features of 1301 ligands in the ligand universality dataset, and then divide the ligands into N domains using the k-means clustering algorithm. Specifically, using the RDKit software package, calculate 200-dimensional physicochemical features based on the SMILES encoding of the ligands, including molecular mass, charge, water solubility, molecular surface descriptor, etc.

[0029] Step 2) Extract protein features from the ligand universality dataset and represent each residue in the protein with a 44-dimensional feature vector. Specifically, for a protein of length L, use HHblits to calculate L×30-dimensional evolutionary information and use DSSP to calculate L×14-dimensional secondary structure information.

[0030] Step 3) Model the local information of the residues as a graph representation. Specifically, use the centroid position of the residue as atomic coordinates, and for each residue s, construct a graph based on all residues within its radius of 15 angstroms. Residues are represented as nodes The feature vectors of the residues obtained in step 2 are represented as node features. When the Euclidean distance between two residues is less than 10 angstroms, an edge is defined between them. And the distance and cosine of the included angle between residues, i.e., edge features;

[0031] Step 4) Input the physicochemical characteristics of the ligands obtained in Step 1 into the following: Figure 2 The ligand-coding layer shown is used to obtain ligand characterization. The graph representation of protein residues obtained in step 3 is then input into the image. Figure 3 The residue coding layer shown yields residue representations. Subsequently, the ligand representation and residue representations are concatenated as features of the ligand-residue pair. For each domain in step 1, a corresponding [database name] is trained. Figure 4 The relation-aware classifier shown predicts the binding probability of ligands in this domain.

[0032] The calculation of the ligand coding layer is as follows: ,in: , , , The parameters of the model to be optimized are... This represents the characteristic features of the ligand.

[0033] The residue coding layer consists of 5 stacked graph network layers. The calculation of the k-th graph network layer is as follows: , , ,in: , , These are the features of the edge (i, j) at the k-th layer, the features of node i, and the graph features, respectively. It is a nonlinear transformation layer. It is a recurrent neural network layer.

[0034] The calculation of the relation-aware classifier is as follows: ,in: This represents the characteristics of the spliced ​​ligand residue pairs.

[0035] Step 5) Fine-tune the parameters of the residue binding probability prediction model and perform a weighted summation of the results of all relation-aware classifiers in the residue binding probability prediction model. Specifically, the parameters of the ligand coding layer, residue coding layer, and all relation-aware classifiers obtained by training with the ligand generality dataset in Step 4 are used as the starting point for parameter optimization in the fine-tuning training stage. For each ligand-specific dataset, the parameters of the ligand coding layer, residue coding layer, and all relation-aware classifiers are further optimized using this dataset. A randomly initialized domain prediction layer is also optimized to perform a weighted summation of the prediction results obtained by all relation-aware classifiers, thus obtaining the final predicted binding probability of the residue to the specific ligand.

[0036] The calculation of the domain prediction layer is as follows: ,in: Let represent the probability that a pair of ligands l and residues s belong to each domain.

[0037] The calculation of the weighted summation of the prediction results obtained from all N relation-aware classifiers to predict the combination probability is as follows: .

[0038] Through specific experiments, pre-training datasets and fine-tuned ligand-specific datasets were collected from the BioLip database. Ligand-binding proteins from February 2021 onwards were collected. For each ligand, the sequence redundancy of the bound protein was first reduced to 30% using CD-HIT software. Based on the protein release date, the datasets were divided into training and testing sets, one before and one after January 2017, resulting in a dataset of 1159 ligand-specific proteins, comprising a total of 27738 proteins. All proteins from BioLip before January 2017 were collected, and redundancy was removed from the 1159 ligand-specific test sets to ensure redundancy was below 30%, resulting in a pre-training dataset containing 6093 proteins and 1301 ligands.

[0039] In the pre-training phase, the physicochemical features of the ligands were first extracted, and the 1301 ligands were clustered into 10 domains using the k-means algorithm. Each domain corresponded to a classification layer during pre-training. Then, evolutionary and secondary structure information of the proteins were extracted, and the residues were represented as a graph structure. Next, in the model's pre-training phase, ligand feature extraction layers and protein feature extraction layers were used to extract features of the ligands and proteins, respectively. The corresponding domain-related classifiers were then used to predict the binding probability of the ligand-residue pair. Finally, optimized ligand feature extraction layers, protein feature extraction layers, and the 10 domain-related classifiers were obtained.

[0040] During the fine-tuning phase, for each ligand-specific dataset, features of both the ligand and protein are first extracted, and then the protein residues are plotted as a graph structure. Pre-trained parameters are used to initialize the ligand feature extraction layer, the protein feature extraction layer, and 10 domain-related classifiers. For the training data in the fine-tuning phase, the ligand and protein feature extraction layers are used to extract features of the ligand and protein respectively, and then concatenated to obtain ligand residue pair features. Subsequently, a domain predictor is used to predict the probability that the ligand belongs to one of the 10 domains. The 10 domain-related classifiers then predict 10 sets of binding probabilities, and the 10 results are weighted and summed according to the domain belonging probability to obtain the final prediction result. The fine-tuned model is used for ligand-specific binding probability prediction.

[0041] In the experimental phase, this embodiment was compared with two other deep learning models, namely ligand-specific deep models: 1) TargetS: Yu, D.-J., Hu, J., Yang, J., Shen, H.-B., Tang, J. and Yang, J.-Y. (2013) Designing template-free predictor for targeting protein-ligand binding sites with classifier ensemble and spatial clustering. IEEE / ACM Transactions on Computational Biology and Bioinformatics, 10, 994-1008; 2) IonCom: Hu, X., Dong, Q., Yang, J. and Zhang, Y. (2016) Recognizing metal and acid radical ion-binding sites by integrating abinitio modeling with template-based transferals. Bioinformatics, 32, 3260-3269; 3) DELIA: Chun-Qiu Xia, Xiaoyo ngPan,Hong-BinShen,Protein-ligandbindingresiduepredictionenhancementthroughhybriddeepheterogeneouslearningofsequenceandstruc turedata, Bioinformatics, 2020, 36: 3018-3027. and ligand universality model: 1) COACH: Yang, J., Roy, A. and Zhang, Y. (2013) Protein–ligand binding site recognition nusingcomplementarybinding-specificsubstructurecomparisonandsequenceprofilealignment.Bioinformatics,29,2588-2595.2)P2Rank:Krivák,R.andHoksza,D.(2018)P2Rank:machinelearningbasedtoolforrapidandaccuratepredictionofligandbindingsitesfromproteinstructure.Journal of cheminformatics,10,1-12.3) DeepPocket:Aggarwal,R.,Gupta,A.,Chelur,V.,Jawahar,C.andPriyakumar,UD(2021)Deeppocket:ligandbindingsitedetectionandsegmentationusing3dconvolutionalneuralnetworks.JournalofChemicalInformationandModeling.

[0042] Statistical violin plots of MCC on all 1159 specific test sets, as shown Figure 5 As shown in Table 1.

[0043] Table 1

[0044] TargetS IonCom DELIA COACH P2Rank DeepPocket This method Nucleicacids - - - 0.214 0.040 0.097 0.442 <![CDATA[Ca 2+ ]]> 0.253 0.220 0.321 0.185 0.029 0.021 0.465 <![CDATA[Mg 2+ ]]> 0.268 0.246 0.308 0.195 0.082 0.082 0.354 <![CDATA[Mn 2+ ]]> 0.405 0.291 <![CDATA[ 0.561 ]]> 0.359 0.144 0.138 0.608 <![CDATA[Zn 2+ ]]> <![CDATA[ 0.546 ]]> 0.299 - 0.370 0.046 0.043 0.750 <![CDATA[Fe 3+ ]]> <![CDATA[ 0.519 ]]> 0.377 - 0.391 0.151 0.178 0.679 <![CDATA[Fe 2+ ]]> - <![CDATA[ 0.498 ]]> - 0.432 0.163 0.142 0.678 <![CDATA[Cu 2+ ]]> - <![CDATA[ 0.416 ]]> - 0.351 0.038 0.016 0.552 <![CDATA[Na + ]]> - 0.066 - <![CDATA[ 0.091 ]]> 0.027 0.034 0.166 <![CDATA[K + ]]> - 0.114 - <![CDATA[ 0.154 ]]> 0.110 0.058 0.232 <![CDATA[CO3 2- ]]> - 0.129 - <![CDATA[ 0.198 ]]> 0.086 0.082 0.213 <![CDATA[NO2 - ]]> - <![CDATA[ 0.196 ]]> - 0.179 0.145 0.092 0.369 <![CDATA[SO4 2- ]]> - 0.197 - <![CDATA[ 0.234 ]]> 0.198 0.202 0.364 <![CDATA[PO4 3- ]]> - 0.207 - <![CDATA[ 0.255 ]]> 0.190 0.172 0.422 ADP <![CDATA[ 0.559 ]]> - - 0.489 0.277 0.369 0.663 AMP <![CDATA[ 0.422 ]]> - - 0.393 0.278 0.352 0.497 ATP 0.479 - <![CDATA[ 0.535 ]]> 0.464 0.255 0.352 0.578 GDP <![CDATA[ 0.543 ]]> - - 0.470 0.266 0.342 0.628 GTP <![CDATA[ 0.459 ]]> - - 0.400 0.273 0.368 0.554 HEME 0.542 - 0.596 0.459 0.454 0.529 0.728

[0045] Compared with existing technologies, this invention, through a domain-adaptive transfer learning method, not only improves prediction accuracy on large datasets but also extends the prediction model to more specific ligands, especially small-sample ligand-specific datasets. Specifically, from Figure 5 As can be seen, for the three baseline ligand general methods COACH, P2Rank, and DeepPocket, LigBind achieves a higher Matthews correlation coefficient (MCC) for 745–804 ligands. Further comparison of the average MCC for a test set of 1159 ligands shows that the average MCC for the three general methods is 0.283, 0.284, and 0.286, while LigBind's average MCC is 0.371, an improvement of approximately 30%. This demonstrates the method's advantage on a wide range of ligand-specific datasets. Table 1 shows the Matthews correlation coefficients (MCCs) for different methods on common specific ligand datasets. It can be seen that this method achieves the highest MCC across all datasets, indicating that it performs well on common ligands including nucleic acids, ions, acid radicals, and nucleotides.

[0046] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A method for predicting ligand-specific binding residues based on domain adaptation and graph networks, characterized in that, include: Step 1: Collect the ligand universality dataset and cluster it to obtain the number of ligand domains N. Extract the physicochemical features and graph representations of protein residues from the ligand universality dataset. Step 2: Construct a residue binding probability prediction model that includes a ligand-coding layer, a residue-coding layer, and a relation-aware classifier. Set the number of relation-aware classifiers according to the number of ligand domains N. Input the physicochemical features and the graph representation of protein residues into the ligand-coding layer and the residue-coding layer respectively, and then concatenate the results into ligand-residue pair features. Each relation-aware classifier is trained using the ligand residue pairs features contained in each domain of the ligand generality dataset. Step 3: Using a ligand-specific dataset, repeat the extraction of physicochemical features and graph representations of protein residues in Step 1, and perform the same training on the residue binding probability prediction model obtained in Step 2 to achieve model fine-tuning and optimization. Step 4: In the online phase, the residue binding probability prediction model after fine-tuning the input of the ligand residue pairs of the test data is used to obtain the classification output results corresponding to N domains. At the same time, the domain prediction layer is randomly initialized by inputting the input of the ligand residue pairs of the test data, and the corresponding domain membership probability is used as a weight to perform weighted summation on the classification output results, thereby obtaining the final predicted binding probability of residue pairs to specific ligands.

2. The ligand-specific binding residue prediction method based on domain adaptation and graph networks according to claim 1, characterized in that, The aforementioned ligand generality dataset contains data on the interaction between various ligands and proteins. Using the RDKit software package, based on the SMILES encoding of the ligands, each pair of interacting proteins and ligands in the dataset is clustered into N domains according to the calculated 200-dimensional physicochemical features. The physicochemical characteristics include: molecular mass, charge, water solubility, and molecular surface descriptor.

3. The ligand-specific binding residue prediction method based on domain adaptation and graph networks according to claim 1, characterized in that, The graph representation of protein residues is achieved by representing each residue of the protein in the ligand universality dataset with a 44-dimensional feature vector, and representing the local information of the residues as a graph structure. Specifically, for a protein of length L, HHblits are used to calculate L×30-dimensional evolutionary information, and the hydrogen bond estimation algorithm (DSSP) is used to calculate L×14-dimensional secondary structure information, resulting in a total of L×44-dimensional protein features. Using the centroid position of each residue in the protein as the spatial coordinate of the residue, for each residue s, a graph is constructed based on all residues within a radius of 15 angstroms. Where: residues are nodes The 44-dimensional feature vector of a residue is a node feature. An edge is defined between two residues when the Euclidean distance between them is less than 10 angstroms. The 2D edge features are the distance and cosine of the included angle between two adjacent residues.

4. The ligand-specific binding residue prediction method based on domain adaptation and graph networks according to claim 1, characterized in that, The ligand encoding layer is based on the input 200-dimensional residue features. Output ligand characterization ,in: , , , BatchNorm represents the model parameters to be optimized. This is a layer normalization layer.

5. The ligand-specific binding residue prediction method based on domain adaptation and graph networks according to claim 3, characterized in that, The residue coding layer, based on the input graph After unifying the 2D edge features and 44D node features to the same feature dimension through batch normalization and linear transformation layers, the residue encoding is extracted and the residue representation is output through five stacked graph network layers. Specifically, the graph network at layer k extracts the features of edge (i, j), node i, and graph features of layer k, respectively: , , ,in: , , Let be the features of edge (i,j) at the k-th layer, the features of node i, and the graph features, respectively. [To splice on the last dimension] and Features Let i be the set of neighboring nodes of node i. Let be the number of neighboring nodes of node i. For the image The number of nodes in the middle, It is a nonlinear transformation layer. For recurrent neural network layers, the last graph network layer outputs graph features. That is, residue characterization.

6. The ligand-specific binding residue prediction method based on domain adaptation and graph networks according to claim 1, characterized in that, The training of each relation-aware classifier refers to training a corresponding relation-aware classifier in each domain using the representation of ligand residue pairs as input. Specifically, for each of the N domains obtained after ligand clustering, the data of interacting ligands and proteins within that domain are used as samples for training. The input is the representation of the ligand residue pairs obtained by concatenating the encoded ligand representation with the residue representation, and the output is the binding probability of the two. Features of the spliced ​​ligand residue pairs , , , These are the model parameters to be optimized.

7. The ligand-specific binding residue prediction method based on domain adaptation and graph networks according to claim 1, characterized in that, The domain dependency probability is obtained by: based on ligand residue pairs ,in: This represents the characterization of ligand-residue pairs obtained through the ligand coding layer, residue coding layer, and domain prediction layer for specific ligands t, ligand l, and residues s. , , , The model parameters to be optimized are... Residue characteristics, The input image.

8. The ligand-specific binding residue prediction method based on domain adaptation and graph networks according to claim 1, characterized in that, The weighted summation mentioned above is specifically as follows: ,in: Let be the domain membership probability of ligand l and residue s belonging to domain n. This represents the prediction result obtained by the relation-aware classifier corresponding to domain n for ligand l and residue s.

9. A system for implementing the ligand-specific binding residue prediction method based on domain adaptation and graph networks as described in any one of claims 1-8, characterized in that, include: The system comprises a ligand feature extraction unit, a protein feature extraction unit, a ligand clustering unit, a ligand encoding unit, a residue graph representation unit, a residue encoding unit, a relation-aware classifier unit, and a domain prediction layer unit. Specifically: the ligand feature extraction unit extracts ligand features based on ligand SMILES information; the protein feature extraction unit extracts protein features based on protein sequence and structural information; the ligand clustering unit clusters ligands in the pre-training dataset into N domains based on ligand features; the ligand encoding unit obtains ligand representations through a nonlinear transformation network based on ligand features; the residue graph representation unit generates a graph representation for each residue based on protein features and residue position information; the residue encoding unit learns residue representations through a graph network based on the graph representations; the relation-aware classifier unit performs nonlinear transformation processing based on ligand features and residue representations to obtain the binding probability of residues with ligands; and the domain prediction layer unit performs nonlinear transformation processing based on ligand features and residue representations to obtain the probability that the binding pattern of a residue with ligand belongs to each domain.

Images