A method and system for predicting the binding affinity of perfluoro and polyfluoroalkyl substances to nuclear receptors based on multi-modal deep learning

Abstract

The application discloses a perfluoro and polyfluoro alkyl substance and nuclear receptor binding affinity prediction method and system based on a multi-modal deep learning, and the method comprises the following steps: acquiring a SMILES string of the perfluoro and polyfluoro alkyl substance, a PDB structure file and sequence information of the nuclear receptor; converting the perfluoro and polyfluoro alkyl substance into a two-dimensional molecular graph and encoding to extract three-dimensional structure features; constructing a k-neighbor graph for a nuclear receptor binding pocket, extracting three-dimensional geometric structure features, and fusing with sequence features; generating a normalized interaction matrix between perfluoro and polyfluoro alkyl substance atoms and nuclear receptor amino acid residues through an interactive attention mechanism, and generating a global interaction embedding accordingly; inputting the global interaction embedding into a multi-layer perception machine to output a binding affinity prediction value. The application realizes high-precision regression prediction of the perfluoro and polyfluoro alkyl substance and nuclear receptor binding affinity, has atom-residue-level interpretability, and can be used for perfluoro and polyfluoro alkyl substance toxicity virtual screening.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics technology, and in particular to a method and system for predicting the binding affinity of perfluorinated and polyfluoroalkyl substances to nuclear receptors based on multimodal deep learning. Background Technology

[0002] Perfluoroalkyl and polyfluoroalkyl substances (PFAS) are a class of synthetic chemical substances composed of carbon-fluorine bonds, widely used in industrial and consumer products. Due to the extremely high thermal and chemical stability of carbon-fluorine bonds, PFAS are very difficult to degrade in the environment and are known as "permanent chemicals." A growing body of research indicates that PFAS exposure is associated with a variety of health hazards, including kidney and liver damage, reproductive and developmental toxicity, endocrine and immune system disruption, and metabolic and cardiovascular diseases.

[0003] At the molecular level, most toxic effects of PFAS stem from their preferential binding to the ligand-binding domain (LBD) of nuclear receptor (NR) proteins, thereby interfering with NR signaling pathways and triggering toxic responses. Therefore, predicting the binding affinity of PFAS to NR is a crucial method for assessing the toxic risks of PFAS. Regulatory agencies such as the U.S. Environmental Protection Agency (EPA) and the Organisation for Economic Co-operation and Development (OECD) have emphasized the importance of NR screening in their PFAS hazard assessments.

[0004] Currently, predictive models for PFAS-nuclear receptor binding affinity have the following shortcomings: First, there is a lack of universal regression prediction models. In existing technologies, most machine learning models for PFAS-NR binding affinity are classification models (such as NR-ToxPred), which can only predict whether a compound binds to the nuclear receptor and cannot output continuous affinity values. While a few multiple linear regression models can predict binding affinity, they use different equation forms and parameters for different nuclear receptors, resulting in inconsistent models and poor generalization ability. Second, their feature representation capabilities are limited. Most existing PFAS toxicity prediction models employ traditional shallow machine learning methods (such as random forests, support vector machines, and XGBoost), requiring manual feature engineering and heavily relying on domain expert knowledge. These models typically use only molecular fingerprints and SMILES-based molecular descriptors or two-dimensional molecular graphs as input features, neglecting the crucial three-dimensional structural information that is essential for prediction accuracy: including both the three-dimensional conformational information of the PFAS molecule and the three-dimensional spatial structural information of the nuclear receptor protein binding pocket. Third, the models have poor interpretability. While existing deep learning models have made some progress in the field of drug-target affinity prediction, most only output affinity values ​​and cannot provide an atomic-residue level interpretation of interactions. For PFAS toxicity assessment, understanding which atoms in the PFAS molecule interact with which key residues in NR is crucial for elucidating the toxicity mechanism and designing low-toxicity alternatives. Fourth, existing drug-target affinity prediction models are not applicable to PFAS. Although various deep learning models have been developed in the field of drug-target affinity prediction, these models are mainly designed for traditional drug molecules and have not been optimized for environmental pollutants such as PFAS with special fluorination structures. Their three-dimensional structure utilization efficiency is low and the interactive data is limited, making it difficult to directly transfer them to the PFAS toxicity prediction scenario.

[0005] In summary, there is a lack of a general prediction method in the prior art that can accurately predict the binding affinity of PFAS to NR, make full use of three-dimensional structural information, and has atomic-residue level interpretability. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, this invention provides a method and system for predicting the binding affinity of perfluorinated and polyfluoroalkyl substances to nuclear acceptors based on multimodal deep learning, aiming to solve the following technical problems: 1. Existing PFAS-nuclear receptor binding affinity prediction models suffer from a lack of regression prediction ability, inconsistent models, and poor generalization ability; 2. Existing models fail to fully utilize the three-dimensional structural information of PFAS molecules and NR proteins, resulting in limited prediction accuracy; 3. Existing models lack atomic-residue level interpretability, failing to provide fine-grained interaction information for elucidating toxicity mechanisms.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for predicting the binding affinity of perfluorinated and polyfluoroalkyl substances to nuclear acceptors based on multimodal deep learning, the method comprising: Step 1: Obtain the SMILES strings of the perfluorinated and polyfluoroalkyl substances to be predicted, as well as the PDB structure file and amino acid sequence information of the target nuclear acceptor binding pocket; Step 2: Convert the SMILES strings of the perfluorinated and polyfluoroalkyl substances into two-dimensional molecular graphs, and encode the two-dimensional molecular graphs using a pre-trained molecular graph neural network model to obtain three-dimensional structure embedding vectors. Then, extract the three-dimensional structure feature matrices of the perfluorinated and polyfluoroalkyl substances using a 3D-EC decoder. ; Step 3: Construct the three-dimensional structure of the nuclear receptor binding pocket as a k-nearest neighbor graph, and extract the three-dimensional geometric feature matrix of the nuclear receptor using a geometric bottleneck perception network. Simultaneously, the amino acid sequence of the nuclear receptor is used to extract sequence features using a protein language model and embedded. The three-dimensional geometric feature matrix is ​​then fused with the sequence feature embedding to obtain the final nuclear receptor feature matrix. ; Step 4: Obtain the three-dimensional structural feature matrix of the perfluorinated and polyfluoroalkyl substances. With the final feature matrix of the nuclear receptor The input to the PFAS-NR fusion module generates a normalized interaction matrix between perfluorinated and polyfluoroalkyl substances and nuclear acceptor amino acid residues through an interactive attention mechanism. Based on the normalized interaction matrix, a weighted aggregation of the joint embedding of perfluorinated and polyfluoroalkyl substances with nuclear acceptor residues is performed to generate a global interaction embedding. ; Step 5: Embed the global interaction Input a multilayer sensor and output predicted binding affinity values ​​for perfluorinated and polyfluoroalkyl substances to nuclear acceptors.

[0008] In one embodiment of the present invention, the 3D-EC decoder in step 2 includes; Embed the three-dimensional structure into a vector After linear transformation and projection onto a unified dimension d, we obtain ; Will The decoded features are obtained by sequentially passing through a first one-dimensional convolutional layer, an ELU activation function, a second one-dimensional convolutional layer, and a gated linear unit activation. ; By residual connection, the With the Add and scaled. By performing layer scaling, the three-dimensional structural feature matrix is ​​obtained. ; Wherein, the uniform dimension d=320, and the scaling factor s=0.5.

[0009] In one embodiment of the present invention, the geometric bottleneck sensing network in step 3 includes two encoder layers and one decoder layer connected in sequence; the geometric bottleneck sensing network processes the scalar features and vector features of nodes and edges in the k-nearest neighbor graph through the geometric bottleneck sensing module, and aggregates spatial information through the SO(3) equivariant message passing network to generate the three-dimensional geometric structure feature matrix. .

[0010] In one embodiment of the present invention, step 3, which involves fusing the three-dimensional geometric structure feature matrix with the sequence feature embedding, specifically includes: The three-dimensional geometric structure feature matrix and the sequence feature embedding Each is unified to dimension d through a linear transformation; After unifying dimensions and Concatenating along the feature dimensions yields a result with dimension . splicing characteristics; The spliced ​​features are passed through a gated linear unit activation layer to output the final feature matrix of the nuclear receptor. .

[0011] In one embodiment of the present invention, the normalized interaction matrix in step 4 is generated in the following manner: Three-dimensional structural feature matrix of the perfluorinated and polyfluoroalkyl substances and the final feature matrix of the nuclear receptor Perform layer normalization separately; Normalized Learnable weight matrix Transformed and normalized Multiplying by the transpose of the matrix and then performing double softmax activation yields the normalized interaction matrix. ; in, The weight matrix is ​​a learnable matrix; the dual softmax activation includes: first performing softmax normalization on the interaction matrix along the residue dimension, and then performing softmax normalization along the atom dimension.

[0012] In one embodiment of the present invention, the global interaction embedding in step 4 is generated in the following manner: Characterize the i-th atom of perfluorinated and polyfluoroalkyl substances Features of the j-th residue of the nuclear receptor After element-wise multiplication and tanh activation, atom-residue co-intercalation is obtained. ; The element in the i-th row and j-th column of the normalized interaction matrix As weights, a weighted sum is performed on the joint embeddings of all residues to obtain the global interactive embedding of the i-th atom. ; The global interaction embeddings of all atoms are combined into the global interaction embedding. .

[0013] In one embodiment of the present invention, the multilayer perceptron in step 5 includes a first fully connected layer, an ELU activation layer, a Dropout layer, a second fully connected layer, an ELU activation layer, and a third fully connected layer connected in sequence; the first fully connected layer has an input dimension of 320 and an output dimension of 160; the second fully connected layer has an input dimension of 160 and an output dimension of 64; and the third fully connected layer has an input dimension of 64 and an output dimension of 1.

[0014] In one embodiment of the present invention, the nuclear receptor is any one or more of the following: androgen receptor, estrogen receptor α, estrogen receptor β, glucocorticoid receptor, progesterone receptor, peroxisome proliferator-activated receptor α, peroxisome proliferator-activated receptor δ, peroxisome proliferator-activated receptor γ, mineralocorticoid receptor, and retinoic acid X receptor α.

[0015] Secondly, this invention provides a multimodal deep learning-based system for predicting the binding affinity of perfluorinated and polyfluoroalkyl substances to nuclear acceptors, used to implement the method described above, the system comprising: The data acquisition module is used to acquire the SMILES strings of the perfluorinated and polyfluoroalkyl substances to be predicted, as well as the PDB structure file and amino acid sequence information of the target nuclear acceptor binding pocket; The perfluorinated and polyfluoroalkyl substances feature extraction module is used to convert the SMILES strings of the perfluorinated and polyfluoroalkyl substances into two-dimensional molecular graphs, and extract the three-dimensional structural feature matrices of the perfluorinated and polyfluoroalkyl substances through a pre-trained molecular graph neural network model and a 3D-EC decoder. The nuclear receptor feature extraction module includes a geometric feature extraction submodule and a sequence feature extraction submodule. The geometric feature extraction submodule is used to construct a k-nearest neighbor graph of the three-dimensional structure of the nuclear receptor binding pocket and extract a three-dimensional geometric structure feature matrix through a geometric bottleneck perception network. The sequence feature extraction submodule is used to extract sequence feature embeddings from the amino acid sequence of the nuclear receptor using a protein language model. The nuclear receptor feature extraction module also includes a feature fusion submodule, which is used to fuse the three-dimensional geometric structure feature matrix with the sequence feature embeddings to obtain the final nuclear receptor feature matrix. The PFAS-NR fusion module is used to generate a normalized interaction matrix between perfluorinated and polyfluoroalkyl substances and nuclear acceptor amino acid residues through an interactive attention mechanism, and to generate a global interactive embedding based on the normalized interaction matrix. An affinity prediction module is used to embed the global interaction into the input multilayer perceptron and output predicted values ​​of the binding affinity between perfluorinated and polyfluoroalkyl substances and nuclear acceptors.

[0016] Thirdly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described thereon.

[0017] The beneficial effects of this invention are: 1. Significantly Improved Prediction Accuracy. This invention enhances feature extraction of the three-dimensional structure of PFAS molecules by introducing a 3D-EC decoder, performs isovariant representation learning of the three-dimensional geometry of the nuclear receptor binding pocket using GBPNet, and fully integrates the structural and sequence information of PFAS and the nuclear receptor through a multimodal fusion mechanism, significantly improving the prediction accuracy of binding affinity. Experimental results show that on a test set (13,635 test samples) containing 4545 PFAS molecules and 10 types of NR, this invention achieves an RMSE of 3.956 kcal / mol, a MAE of 2.588 kcal / mol, and an R² of 0.910, all significantly outperforming existing models such as DeepDTA, GraphDTA, NHGNNDTA, and PocketDTA.

[0018] 2. Atom-residue level interpretability. This invention utilizes a normalized interaction matrix generated through an interactive attention mechanism to quantify the interaction strength between each atom in the PFAS molecule and each amino acid residue in the nuclear acceptor, achieving fine-grained interpretability at the atom-residue level. This interaction matrix exhibits good consistency with molecular docking results and can accurately identify key non-covalent interaction sites such as hydrogen bonds, hydrophobic interactions, π-π stacking, and halogen bonds, providing direct guidance for elucidating the toxicity mechanism of PFAS and designing low-toxicity alternatives.

[0019] 3. Excellent generalization ability. This invention not only performs excellently on the PFAS-NR dataset but also demonstrates good generalization performance on the publicly available PDBbind-v2016 general protein-ligand binding affinity dataset, achieving an RMSE of 1.271 and an R² of 0.657. This generalization performance surpasses baseline models such as DeepDTA, GraphDTA, NHGNNDTA, and PocketDTA. This indicates that the method of this invention has cross-dataset versatility and can be extended to a wider range of environmental pollutant toxicity prediction scenarios. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 The overall model architecture diagram of the method for predicting the binding affinity of perfluorinated and polyfluoroalkyl substances to nuclear acceptors based on multimodal deep learning provided in the embodiments of the present invention; Figure 2 This is a performance comparison chart of the embodiments of the present invention and existing technical models on multiple evaluation indicators for different nuclear receptors; Figure 3 This is a visualization of the molecular docking verification results and interaction matrix of the interaction between DTXSID50462159 and estrogen receptor α (ERA) in an embodiment of the present invention. Figure 4 This is a visualization of the molecular docking verification results and interaction matrix of the interaction between DTXSID4047672 and peroxisome proliferator-activated receptor α (PPARA) in an embodiment of the present invention. Detailed Implementation The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the scope of protection of the present invention is not limited to the following embodiments. For those skilled in the art, equivalent substitutions, combinations or conventional modifications made without departing from the concept of the present invention should all fall within the scope of protection of the present invention.

[0022] Example 1 like Figure 1 As shown, this embodiment of the invention provides a method for predicting the binding affinity of perfluorinated and polyfluoroalkyl substances to nuclear acceptors based on multimodal deep learning, the method comprising: Step 1: Obtain the SMILES strings of the perfluorinated and polyfluoroalkyl substances to be predicted, as well as the PDB structure file and amino acid sequence information of the target nuclear acceptor binding pocket; The nuclear receptor binding pocket is determined by selecting all amino acid residues within a 10 Å range from any atom of the ligand in the PDB structure as the center to form the binding pocket.

[0023] Step 2: Convert the SMILES strings of the perfluorinated and polyfluoroalkyl substances into two-dimensional molecular graphs, and encode the two-dimensional molecular graphs using a pre-trained molecular graph neural network model to obtain three-dimensional structure embedding vectors. Then, extract the three-dimensional structure feature matrices of the perfluorinated and polyfluoroalkyl substances using a 3D-EC decoder. ; The 3D-EC decoder is a three-dimensional enhanced coordinate decoder (3D-EC).

[0024] Step 3: Construct the three-dimensional structure of the nuclear receptor binding pocket as a k-nearest neighbor graph, and extract the three-dimensional geometric feature matrix of the nuclear receptor using a geometric bottleneck perception network. Simultaneously, the amino acid sequence of the nuclear receptor is used to extract sequence features using a protein language model and embedded. The three-dimensional geometric feature matrix is ​​then fused with the sequence feature embedding to obtain the final nuclear receptor feature matrix. ; Step 4: Obtain the three-dimensional structural feature matrix of the perfluorinated and polyfluoroalkyl substances. With the final feature matrix of the nuclear receptor The input to the PFAS-NR fusion module generates a normalized interaction matrix between perfluorinated and polyfluoroalkyl substances and nuclear acceptor amino acid residues through an interactive attention mechanism. Based on the normalized interaction matrix, a weighted aggregation of the joint embedding of perfluorinated and polyfluoroalkyl substances with nuclear acceptor residues is performed to generate a global interaction embedding. ; Step 5: Embed the global interaction Input a multilayer sensor and output predicted binding affinity values ​​for perfluorinated and polyfluoroalkyl substances to nuclear acceptors.

[0025] Optionally, step 2 includes: The SMILES string of the PFAS is converted into a two-dimensional molecular graph using RDKit and the Open Graph Benchmark library. , where the matrix Representing atomic properties, matrix Indicates key attributes; This refers to the number of heavy atoms in perfluorinated and polyfluoroalkyl substances. For atomic attribute dimensions, For the number of keys, For key attribute dimensions.

[0026] The two-dimensional molecular graph is encoded into a three-dimensional structure embedding vector using a pre-trained GraphMVP framework. ( ), keeping the pre-trained parameters of GraphMVP unchanged.

[0027] The 3D structure embedding vector D is input into the 3D-EC decoder for feature extraction. The working process of the 3D-EC decoder is as follows: First of all After linear transformation and projection onto a unified dimension d, we obtain ( =320):

[0028] in This is the weight matrix. This is the bias vector.

[0029] Then The decoded features are obtained by sequentially passing the first one-dimensional convolutional layer (with a kernel size of 1), the ELU activation function, the second one-dimensional convolutional layer, and the gated linear unit (GLU) activation. :

[0030] The GLU operation is defined as follows: , ⊙ indicates element-wise multiplication.

[0031] Finally, residual join is used to... and Add and scaled. By performing layer scaling, the three-dimensional structural feature matrix of PFAS is obtained. :

[0032] Where s = 0.5.

[0033] Optionally, step 3 utilizes a feature extraction and fusion module to extract and fuse features of the NR protein. The module includes: three-dimensional geometric structure feature extraction, sequence feature extraction, and structure-sequence feature fusion. The three-dimensional geometric structure feature extraction represents the three-dimensional structure of the NR-combined pocket as a k-nearest neighbor graph. .in This represents a set of amino acid nodes, each node... Corresponding to one amino acid residue atom; Represents the set of edges, if residues It is a residue One of the k=30 nearest neighbors (based on (the Euclidean distance between atoms), then construct from arrive directed edges .

[0034] Node features include scalar components (11-dimensional) and vector components (3×3 dimensions); wherein, the node scalar features include 11 dimensions such as amino acid type, hydrophobicity, charge, and solvent-accessible surface area; Edge features contain scalar components (32-dimensional) and vector components (1×3-dimensional); wherein the edge scalar features include 32-dimensional features such as inter-atomic distance encoding and relative orientation encoding.

[0035] The k-nearest neighbor graph is input into the Geometric Bottleneck Perception Network (GBPNet) for processing. GBPNet consists of two encoder layers and one decoder layer. First, the geometric bottleneck perception module processes the scalar and vector features of nodes and edges respectively. Then, the spatial information is aggregated through an SO(3) equal-variable message passing network to generate the 3D geometric structure feature matrix of NR. ( ),in To incorporate the number of residues in the pocket, the GBPNet has a hidden dimension of 320 for both the encoder and decoder layers, 8 attention heads, a feedforward network dimension of 640, and employs layer normalization.

[0036] The sequence feature extraction module inputs the amino acid sequence of the NR binding pocket into the ESM-2 protein language model to extract sequence feature embeddings. ( ); The structure-sequence feature fusion module will and Each dimension was uniformly reduced to d=320 through linear transformation; the dimension-unified and Concatenating along the feature dimensions yields a result with dimension . The splicing features are then processed through a gated linear unit activation layer to output the final feature matrix of the nuclear receptor. .

[0037] Optionally, step 4 involves processing the three-dimensional structural feature matrix of the PFAS. and nuclear receptor final feature matrix Perform layer normalization:

[0038] The mean and variance are calculated along the feature dimension d. =10 -5 , and These are learnable parameters.

[0039] Normalized Learnable weight matrix Transformed and normalized Multiplying by the transposes of the matrix and then performing double softmax activation yields the normalized interaction matrix. :

[0040] The normalization direction of the dual softmax activation is as follows: first, the interaction matrix is ​​normalized along the residue dimension, and then the interaction matrix is ​​normalized along the atom dimension.

[0041] in The element in the i-th row and j-th column quantifies the interaction strength between the i-th atom of PFAS and the j-th residue of the nuclear acceptor.

[0042] The characteristics of the i-th atom of PFAS Features of the j-th residue of the nuclear receptor After element-wise multiplication and tanh activation, we obtain atom-residue co-intercalation:

[0043] Using the element in the i-th row and j-th column of the normalized interaction matrix as weights, the joint embeddings of all residues are weighted and summed to obtain the global interaction embedding of the i-th atom:

[0044] Combine the global interaction embeddings of all atoms into a global interaction embedding. .

[0045] Optionally, step 5 embeds the global interaction. The input is a multilayer perceptron, and the output is a predicted value of the binding affinity between PFAS and NR. The multilayer perceptron includes a first fully connected layer (320-dimensional input, 160-dimensional output), an ELU activation layer, a Dropout layer, a second fully connected layer (160-dimensional input, 64-dimensional output), an ELU activation layer, and a third fully connected layer (64-dimensional input, 1-dimensional output) connected in sequence.

[0046] Optionally, the nuclear receptor (NR) is any one or more of the following: androgen receptor (AR), estrogen receptor α (ERA), estrogen receptor β (ERB), glucocorticoid receptor (GR), progesterone receptor (PR), peroxisome proliferator-activated receptor α (PPARA), peroxisome proliferator-activated receptor δ (PPARD), peroxisome proliferator-activated receptor γ (PPARG), mineralocorticoid receptor (MR), and retinoic acid X receptor α (RXRA).

[0047] Example 2 This embodiment provides a method for predicting the binding affinity of perfluorinated and polyfluoroalkyl substances (PFAS) to nuclear receptors (NR) based on multimodal deep learning. The method includes: Step 1: Collect the dataset This embodiment uses the PFAS-NR binding affinity dataset by Singam et al. for model training and validation. This dataset contains binding affinity data (MM-GBSA free energy, in kcal / mol) between 4,545 PFAS molecules and 10 human nuclear receptors. The 10 nuclear receptors are: androgen receptor (AR), estrogen receptor α (ERA), estrogen receptor β (ERB), glucocorticoid receptor (GR), progesterone receptor (PR), peroxisome proliferator-activated receptor α (PPARA), peroxisome proliferator-activated receptor δ (PPARD), peroxisome proliferator-activated receptor γ (PPARG), mineralocorticoid receptor (MR), and retinoic acid X receptor α (RXRA).

[0048] The binding pocket PDB structures for each nuclear acceptor were obtained from the PDBbind-v2016 database. The selection criteria were: prioritizing structures with ligand binding pocket information, and choosing the structure with the highest resolution when multiple structures met the criteria. The PDB files were preprocessed using PyMOL to remove non-structural water molecules and irrelevant atoms. The binding pocket was determined by selecting all amino acid residues within a 10 Å radius of any ligand atom in the PDB structure, centered on the ligand binding site.

[0049] The dataset for each nuclear receptor was randomly divided into a training set and a test set in a 7:3 ratio, with one-third of the training set further used as a validation set. The total number of training samples was 31,815, and the number of test samples was 13,635.

[0050] Step 2: Input representation Input representation of perfluorinated and polyfluoroalkyl substances: Converting the SMILES string of PFAS into a two-dimensional molecular diagram using RDKit. The atomic properties include 9-dimensional features (including atom type, formal charge, chirality, degree, number of hydrogen atoms, hybridization type, aromaticity, number of heavy atoms, and atomic mass), and the bond properties include 3-dimensional features (including bond type, stereochemistry, and conjugation). The two-dimensional molecular graph is encoded into a 300-dimensional three-dimensional structure embedding vector through a pre-trained GraphMVP framework. Nuclear receptor input representation: The PDB structure of the nuclear receptor binding pocket is converted into a k-nearest neighbor graph (k=30) as a three-dimensional geometric input. The node features contain 11-dimensional scalars and 3×3-dimensional vectors, and the edge features contain 32-dimensional scalars and 1×3-dimensional vectors. On the other hand, the amino acid sequence of the binding pocket is input into the ESM-2 protein language model to extract 480-dimensional sequence features for embedding.

[0051] Step 3: Set the model hyperparameters In this embodiment, the uniform feature dimension d is set to 320. In the 3D-EC decoder, the first one-dimensional convolutional layer expands the input channels from 320 to 640, and the second one-dimensional convolutional layer compresses the channels back to 320 from 640. GBPNet contains two encoder layers and one decoder layer, with a hidden dimension of 320, eight attention heads, and a feedforward network dimension of 640. The input / output dimensions of the three fully connected layers of the multilayer perceptron are 320 / 160, 160 / 64, and 64 / 1, respectively.

[0052] The training used mean squared error (MSE) as the loss function, the Adam optimizer, a learning rate of 0.0001, weight decay of 0.00001, and 180 training epochs. The model was implemented using the PyTorch framework and trained on an NVIDIA GeForce RTX 3080Ti GPU (12 GB VRAM).

[0053] The following six indicators are used to comprehensively evaluate the model performance: root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), Pearson correlation coefficient (R), coefficient of determination (R²), and standard deviation (SD).

[0054] Step 4: Make predictions and evaluate the test set data. To verify the effectiveness of this invention, the model of this invention (named 3D-ECPA, i.e., 3D-EC-based Prediction of Affinity) was compared with representative models in the prior art, including DeepDTA, GraphDTA, NHGNNDTA, and PocketDTA. Each model was trained and evaluated on the same training and test sets, and each model was run five times independently (using different random seeds) to ensure statistical robustness, as shown in Table 1.

[0055] Table 1. Performance comparison of different models on the Singam dataset.

[0056] Compared to the second-best performing PocketDTA model, the model of this invention reduced RMSE by 16.38%, MAE by 17.37%, and R² by 4.45%. The model of this invention ranked first in performance across all 10 nuclear receptors. Figure 2 ).

[0057] Step 5: Ablation Experiment To verify the effectiveness of each module of the present invention, an ablation experiment was conducted, including: 1. Remove the GBPNet module; 2. Remove the 3D-EC decoder; 3. Only two-dimensional molecular graph features are used; 4. Add an additional one-dimensional molecular fingerprint feature.

[0058] Table 2 Ablation Experiment Results

[0059] The 2D-ECPA is a control model that uses only two-dimensional molecular graph features and not three-dimensional structural features.

[0060] As shown in Table 2, the results show that removing any key module or feature leads to a significant performance decrease, verifying the necessity and effectiveness of each module (3D-EC decoder, GBPNet, and three-dimensional structural features) in this invention.

[0061] Step 6: PDBbind-v2016 Generalization Verification To verify the cross-dataset generalization ability of the method of this invention, a generalization experiment was conducted on the publicly available PDBbind-v2016 general protein-ligand binding affinity dataset. The same network architecture, loss function, optimizer, and hyperparameter settings as the PFAS-NR dataset were used, with only the input data replaced. The PDBbind-v2016 dataset was randomly divided into training and test sets in a 7:3 ratio. The training run consisted of 180 epochs with a learning rate of 0.0001 and a weight decay of 0.00001, as shown in Table 3.

[0062] Table 3 Performance comparison of different models on the PDBbind-v2016 dataset

[0063] The results show that the RMSE of the present invention reaches 1.271 and the R reaches 0.819 on the test set, both of which are better than the existing baseline models, proving that the method of the present invention has cross-dataset universality and is applicable to the prediction of binding affinity of different protein receptor targets.

[0064] Step 7: Molecular docking verification This invention generates a normalized interaction matrix between perfluorinated and polyfluoroalkyl substances and nuclear acceptor residues through an interactive attention mechanism, enabling interpretable analysis at the atom-residue level. To verify the accuracy of the interaction matrix, two high-affinity PFAS-nuclear acceptor complexes were randomly selected for molecular docking verification.

[0065] Example 1: DTXSID50462159 and estrogen receptor α (ERA) Molecular docking was performed using QuickVina-W software with a grid size of 29 × 26 × 37 ų and a exhaustiveness of 40. The docking results show ( Figure 3The PFAS forms various non-covalent interactions with the ERA: hydrogen bonds (between the hydroxyl oxygen atom of PFAS and ARG394, HIS524), hydrophobic interactions (involving residues such as LEU387, LEU391, TRP383), π-π stacking (between the aromatic ring of PFAS and PHE404), and halogen bonds (between the fluorine atom of PFAS and LYS531, ASP351).

[0066] The high-weighted atom-residue pairs identified by the interaction matrix generated by this invention are highly consistent with the key interaction sites identified by the above molecular docking, verifying the accuracy of the interaction matrix in identifying key binding sites.

[0067] Example 2: DTXSID4047672 and peroxisome proliferator-activated receptor α (PPARA) Molecular docking was performed using QuickVina-W software with a grid size of 29×30×34 ų and a exhaustiveness of 40. The docking results show ( Figure 4 The main non-covalent interactions between the PFAS and PPARA include hydrophobic interactions (LEU321, LEU258, VAL255), hydrogen bonds (ALA333, TYR334, THR279), and halogen bonds (CYS275, ILE272). The amino group (-NH-) in the PFAS also participates in hydrogen bond formation with THR279.

[0068] The interaction matrix generated by this invention also identifies the aforementioned key interaction sites, further verifying the reliability of the interaction matrix.

[0069] The above results demonstrate that the present invention can not only accurately predict binding affinity values, but also provide atomic-residue level interaction information, providing direct guidance for the elucidation of PFAS toxicity mechanisms and the design of low-toxicity alternatives.

[0070] Example 3 This embodiment provides a multimodal deep learning-based system for predicting the binding affinity of perfluorinated and polyfluoroalkyl substances to nuclear acceptors, used to implement the aforementioned method for predicting the binding affinity of perfluorinated and polyfluoroalkyl substances to nuclear acceptors based on multimodal deep learning. The system includes: The data acquisition module is used to acquire the SMILES strings of the perfluorinated and polyfluoroalkyl substances to be predicted, as well as the PDB structure file and amino acid sequence information of the target nuclear acceptor binding pocket; The perfluorinated and polyfluoroalkyl substances feature extraction module is used to convert the SMILES strings of the perfluorinated and polyfluoroalkyl substances into two-dimensional molecular graphs, and extract the three-dimensional structural feature matrices of the perfluorinated and polyfluoroalkyl substances through a pre-trained molecular graph neural network model and a 3D-EC decoder. The nuclear receptor feature extraction module includes a geometric feature extraction submodule and a sequence feature extraction submodule. The geometric feature extraction submodule is used to construct a k-nearest neighbor graph of the three-dimensional structure of the nuclear receptor binding pocket and extract a three-dimensional geometric structure feature matrix through a geometric bottleneck perception network. The sequence feature extraction submodule is used to extract sequence feature embeddings from the amino acid sequence of the nuclear receptor using a protein language model. The nuclear receptor feature extraction module also includes a feature fusion submodule, which is used to fuse the three-dimensional geometric structure feature matrix with the sequence feature embeddings to obtain the final nuclear receptor feature matrix. The PFAS-NR fusion module is used to generate a normalized interaction matrix between perfluorinated and polyfluoroalkyl substances and nuclear acceptor amino acid residues through an interactive attention mechanism, and to generate a global interactive embedding based on the normalized interaction matrix. An affinity prediction module is used to embed the global interaction into the input multilayer perceptron and output predicted values ​​of the binding affinity between perfluorinated and polyfluoroalkyl substances and nuclear acceptors.

[0071] Example 4 This embodiment provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements a method for predicting the binding affinity of perfluorinated and polyfluoroalkyl substances to nuclear acceptors based on multimodal deep learning.

[0072] In summary, the present invention provides a method and system for predicting the binding affinity of perfluorinated and polyfluoroalkyl substances to nuclear acceptors based on multimodal deep learning. Its core innovations are reflected in the following aspects: First, based on obtaining two-dimensional molecular maps of perfluorinated and polyfluoroalkyl substances and encoding them using a pre-trained molecular map neural network to obtain three-dimensional structure embeddings, a 3D-EC decoder is introduced to enhance the extraction of the three-dimensional structural features of perfluorinated and polyfluoroalkyl substances, thereby overcoming the deficiency of existing models in fully utilizing the three-dimensional conformational information of perfluorinated and polyfluoroalkyl substances; Second, the three-dimensional structure of the nuclear acceptor binding pocket is constructed as a k-nearest neighbor graph, and through... The SO(3) Geometric Bottleneck Perception Network (GBPNet) performs isovariant geometric representation learning on the binding pocket and fuses it with sequence feature embeddings extracted from a protein language model to fully capture the structural and sequence information of the nuclear receptor binding pocket. Third, a normalized interaction matrix between perfluorinated and polyfluoroalkyl substances (PFAS) atoms and nuclear receptor amino acid residues is generated through an interactive attention mechanism. Global interaction embeddings are then generated based on weighted aggregation of this matrix, thereby achieving fine-grained interpretability at the atom-residue level and providing direct evidence for the analysis of toxicity mechanisms of PFAS and the design of low-toxicity alternatives. The synergistic effect of the above techniques significantly improves the accuracy of binding affinity prediction and cross-dataset generalization ability, enabling this invention to be widely applied to virtual toxicity screening and risk assessment of PFAS.

[0073] Although the invention has been described with reference to preferred embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, the technical features mentioned in the various embodiments can be combined in any manner as long as there is no structural conflict. The invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for predicting the binding affinity of perfluorinated and polyfluoroalkyl substances to nuclear acceptors based on multimodal deep learning, characterized in that, The method includes: Step 1: Obtain the SMILES strings of the perfluorinated and polyfluoroalkyl substances to be predicted, as well as the PDB structure file and amino acid sequence information of the target nuclear acceptor binding pocket; Step 2: Convert the SMILES strings of the perfluorinated and polyfluoroalkyl substances into two-dimensional molecular graphs, and encode the two-dimensional molecular graphs using a pre-trained molecular graph neural network model to obtain three-dimensional structure embedding vectors. Then, extract the three-dimensional structure feature matrices of the perfluorinated and polyfluoroalkyl substances using a 3D-EC decoder. ; Step 3: Construct the three-dimensional structure of the nuclear receptor binding pocket as a k-nearest neighbor graph, and extract the three-dimensional geometric feature matrix of the nuclear receptor using a geometric bottleneck perception network. Simultaneously, the amino acid sequence of the nuclear receptor is used to extract sequence features using a protein language model and embedded. The three-dimensional geometric feature matrix is ​​then fused with the sequence feature embedding to obtain the final nuclear receptor feature matrix. ; Step 4: Obtain the three-dimensional structural feature matrix of the perfluorinated and polyfluoroalkyl substances. With the final feature matrix of the nuclear receptor The input to the PFAS-NR fusion module generates a normalized interaction matrix between perfluorinated and polyfluoroalkyl substances and nuclear acceptor amino acid residues through an interactive attention mechanism. Based on the normalized interaction matrix, a weighted aggregation of the joint embedding of perfluorinated and polyfluoroalkyl substances with nuclear acceptor residues is performed to generate a global interaction embedding. ; Step 5: Embed the global interaction Input a multilayer sensor and output predicted binding affinity values ​​for perfluorinated and polyfluoroalkyl substances to nuclear acceptors.

2. The method according to claim 1, characterized in that, The 3D-EC decoder mentioned in step 2 includes: Embed the three-dimensional structure into the vector After linear transformation and projection onto a unified dimension d, we obtain ; Will The decoded features are obtained by sequentially passing through a first one-dimensional convolutional layer, an ELU activation function, a second one-dimensional convolutional layer, and a gated linear unit activation. ; By residual connection, the With the Add and scaled. By performing layer scaling, the three-dimensional structural feature matrix is ​​obtained. ; Wherein, the uniform dimension d=320, and the scaling factor s=0.

5.

3. The method according to claim 1, characterized in that, The geometric bottleneck perception network in step 3 includes two encoder layers and one decoder layer connected in sequence. The geometric bottleneck perception network processes the scalar and vector features of nodes and edges in the k-nearest neighbor graph through the geometric bottleneck perception module, and aggregates spatial information through the SO(3) equivariant message passing network to generate the three-dimensional geometric structure feature matrix. .

4. The method according to claim 1, characterized in that, Step 3, which involves fusing the 3D geometric structure feature matrix with the sequence feature embedding, specifically includes: The three-dimensional geometric structure feature matrix and the sequence feature embedding Each is unified to dimension d through a linear transformation; After unifying dimensions and Concatenating along the feature dimensions yields a result with dimension . splicing characteristics; The spliced ​​features are passed through a gated linear unit activation layer to output the final feature matrix of the nuclear receptor. .

5. The method according to claim 1, characterized in that, The normalized interaction matrix mentioned in step 4 is generated in the following way: Three-dimensional structural feature matrix of the perfluorinated and polyfluoroalkyl substances and the final feature matrix of the nuclear receptor Perform layer normalization separately; Normalized Learnable weight matrix Transformed and normalized Multiplying by the transpose of the matrix and then performing double softmax activation yields the normalized interaction matrix. ; in, The weight matrix is ​​a learnable matrix; the dual softmax activation includes: first performing softmax normalization on the interaction matrix along the residue dimension, and then performing softmax normalization along the atom dimension.

6. The method according to claim 1, characterized in that, The global interaction embedding mentioned in step 4 is generated in the following way: Characterize the i-th atom of perfluorinated and polyfluoroalkyl substances Features of the j-th residue of the nuclear receptor After element-wise multiplication and tanh activation, atom-residue co-intercalation is obtained. ; The element in the i-th row and j-th column of the normalized interaction matrix As weights, a weighted sum is performed on the joint embeddings of all residues to obtain the global interactive embedding of the i-th atom. ; The global interaction embeddings of all atoms are combined into the global interaction embedding. .

7. The method according to claim 1, characterized in that, The multilayer perceptron in step 5 includes a first fully connected layer, an ELU activation layer, a Dropout layer, a second fully connected layer, an ELU activation layer, and a third fully connected layer connected in sequence; the first fully connected layer has an input dimension of 320 and an output dimension of 160; the second fully connected layer has an input dimension of 160 and an output dimension of 64; and the third fully connected layer has an input dimension of 64 and an output dimension of 1.

8. The method according to claim 1, characterized in that, The nuclear receptor is any one or more of the following: androgen receptor, estrogen receptor α, estrogen receptor β, glucocorticoid receptor, progesterone receptor, peroxisome proliferator-activated receptor α, peroxisome proliferator-activated receptor δ, peroxisome proliferator-activated receptor γ, mineralocorticoid receptor, and retinoic acid X receptor α.

9. A system for predicting the binding affinity of perfluorinated and polyfluoroalkyl substances to nuclear acceptors based on multimodal deep learning, characterized in that, The system for implementing the method as described in any one of claims 1 to 8 includes: The data acquisition module is used to acquire the SMILES strings of the perfluorinated and polyfluoroalkyl substances to be predicted, as well as the PDB structure file and amino acid sequence information of the target nuclear acceptor binding pocket; The perfluorinated and polyfluoroalkyl substances feature extraction module is used to convert the SMILES strings of the perfluorinated and polyfluoroalkyl substances into two-dimensional molecular graphs, and extract the three-dimensional structural feature matrices of the perfluorinated and polyfluoroalkyl substances through a pre-trained molecular graph neural network model and a 3D-EC decoder. The nuclear receptor feature extraction module includes a geometric feature extraction submodule and a sequence feature extraction submodule. The geometric feature extraction submodule is used to construct a k-nearest neighbor graph of the three-dimensional structure of the nuclear receptor binding pocket and extract a three-dimensional geometric structure feature matrix through a geometric bottleneck perception network. The sequence feature extraction submodule is used to extract sequence feature embeddings from the amino acid sequence of the nuclear receptor using a protein language model. The nuclear receptor feature extraction module also includes a feature fusion submodule, which is used to fuse the three-dimensional geometric structure feature matrix with the sequence feature embeddings to obtain the final nuclear receptor feature matrix. The PFAS-NR fusion module is used to generate a normalized interaction matrix between perfluorinated and polyfluoroalkyl substances and nuclear acceptor amino acid residues through an interactive attention mechanism, and to generate a global interactive embedding based on the normalized interaction matrix. An affinity prediction module is used to embed the global interaction into the input multilayer perceptron and output predicted values ​​of the binding affinity between perfluorinated and polyfluoroalkyl substances and nuclear acceptors.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 8.

Images