Precision prediction method for pharmaceutical adjuvant-protein interaction based on cross attention mechanism and deep transfer learning strategy

Abstract

The invention discloses a method for accurately predicting the interaction between pharmaceutic adjuvants and proteins based on a cross attention mechanism and a deep transfer learning strategy. The method comprises the following steps: firstly, collecting large-scale drug-like molecule-protein activity data and pharmaceutic adjuvant-protein interaction data; secondly, constructing an interaction prediction model framework based on a cross attention mechanism; then, inputting the drug-like molecular structure and the protein sequence into the prediction model, and training to obtain a pre-trained model PreDIG; then, the pharmaceutic adjuvant structure and the protein sequence serve as input, a deep transfer learning strategy is used, a PreDIG model is subjected to fine tuning training, and a BioDIG model is obtained; and finally, inputting the to-be-detected pharmaceutic adjuvant structure and the protein sequence into BioDIG to obtain the to-be-detected pharmaceutic adjuvant-protein interaction prediction probability, and judging whether the pharmaceutic adjuvant is combined with the protein or not according to the pharmaceutic adjuvant-protein interaction prediction probability.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of adverse drug reactions and computational biology, and specifically relates to a precise prediction method for drug inactive ingredients-protein interactions based on a cross-attention mechanism and a deep transfer learning strategy. BACKGROUND

[0002] Drug inactive ingredients (DIGs, also known as pharmaceutical excipients) are generally considered to be biologically inert substances. However, recent studies have shown that some DIGs can have biological activity with human proteins. Since the content of DIGs in drug formulations is usually 10 to 100 times that of bioactive molecules (BAMs), if they have potential activity, they may affect the efficacy of the drug, and even cause toxic reactions. Therefore, in the early stages of drug development, it is of great significance to predict the interaction between pharmaceutical excipients and target proteins using computational methods to ensure drug safety and effectiveness.

[0003] The human proteome is large, containing about 22,000 proteins. Traditional experimental detection methods are not only time-consuming and labor-intensive, but also difficult to systematically and comprehensively analyze the activity spectrum of DIGs, so there is an urgent need for efficient computational prediction methods. Existing computational methods can be broadly divided into two categories: compound-based methods and protein-based methods. Compound-based methods do not require protein information, but are modeled for specific protein activity and inactive molecules. However, this method is highly dependent on a large amount of known activity data, and its application is limited due to the lack of relevant molecular information for many human proteins. Protein-based methods input both compound and protein data and train the model based on known compound-protein interactions, which can compensate for the above shortcomings to some extent. However, this type of method still faces two challenges: on the one hand, existing models generally rely on BAMs data for training, and there are significant differences in activity distribution between DIGs and BAMs, so directly using such models will lead to a decrease in prediction accuracy; on the other hand, the number of interactions in the publicly available DIG database is extremely limited, and if the CPI (Compound-Protein Interaction) strategy is used and only limited data is used for modeling, the resulting model usually lacks generalization ability. The above problems jointly restrict their application in large-scale pharmaceutical excipient activity research. SUMMARY

[0004] To address the aforementioned problems, this invention proposes a precise prediction method for pharmaceutical excipient-protein interactions based on cross-attention mechanisms and deep transfer learning strategies. This aims to overcome the limitations of existing methods and provide a more accurate predictive tool for adverse reactions of pharmaceutical excipients. This method effectively improves the model's generalization ability and prediction accuracy, thus providing a new solution for the systematic analysis and research of pharmaceutical excipient activity.

[0005] A method for accurately predicting pharmaceutical excipient-protein interactions based on cross-attention mechanism and deep transfer learning strategy includes the following steps:

[0006] Step S1: Collect drug molecule-protein activity data and pharmaceutical excipient-protein interaction data;

[0007] Step S2: Construct an interaction prediction model framework based on the cross-attention mechanism;

[0008] Step S3: Input the drug-like molecule data and protein data from the drug-like molecule-protein activity data in Step S1 into the interaction prediction model framework based on the cross-attention mechanism constructed in Step S2, and train to obtain the pre-trained model PreDIG;

[0009] Step S4: Using the pharmaceutical excipient data and protein data from the pharmaceutical excipient-protein interaction data in Step S1 as input, train the PreDIG model using a deep transfer learning strategy to obtain the BioDIG model.

[0010] Step S5: Input the structure and protein sequence of the pharmaceutical excipient to be tested into the BioDIG model to obtain the predicted probability of the interaction between the pharmaceutical excipient and the protein. Based on the predicted probability of the interaction between the pharmaceutical excipient and the protein, determine whether the pharmaceutical excipient binds to the protein.

[0011] The model proposed in this invention combines cross-attention mechanism and deep transfer learning strategy, which effectively overcomes the problem of inaccurate prediction caused by activity distribution bias and sparse activity data of pharmaceutical excipients, thereby significantly improving the accuracy and reliability of prediction of pharmaceutical excipient-target interaction activity.

[0012] In step S1, the drug-like molecule-protein activity data is the BindingBD dataset, and the pharmaceutical excipient-protein interaction data is the ACDINA dataset. The drug-like molecule-protein activity data includes: drug-like molecule data, protein data, and compound-protein activity data; the pharmaceutical excipient-protein interaction data includes: pharmaceutical excipient data, protein data, and pharmaceutical excipient-protein interaction data.

[0013] Furthermore, the drug-like molecule-protein interaction data is the BindingBD dataset. The pharmaceutical excipient-protein interaction data is the ACDINA dataset. The drug-like molecule-protein interaction dataset includes: drug-like molecule data, protein data, and drug-like molecule-protein interaction data. The pharmaceutical excipient-protein interaction dataset includes: pharmaceutical excipient data, protein data, and pharmaceutical excipient-protein interaction data.

[0014] Further, in step S2, constructing the interaction prediction model framework based on the cross-attention mechanism refers to building the model framework using a compound encoder, a protein encoder, an interaction decoder, and a binary classifier. The compound encoder encodes features of drug-like molecule data to obtain a compound encoding vector; the protein encoder encodes features of protein data to obtain a protein encoding vector; the interaction decoder based on the cross-attention mechanism takes the compound encoding vector and the protein encoding vector as input and the interaction matrix as output; the binary classifier takes the interaction matrix as input and is trained using the cross-entropy loss function.

[0015] Specifically, using a protein encoder to encode protein sequence features involves first encoding the protein data in the pharmaceutical excipient-protein interaction dataset using the natural language processing technique Word2Vec. The protein residue encoding matrix is ​​then transformed into an L×64 matrix using a fully connected layer (FC), and subsequently input into a gating convolutional neural network (GatedCNN). Each gating convolutional layer contains one convolutional unit and one gated linear unit. GatedCNN contains a total of l gating convolutional layers, and the calculation formula for the i-th layer is shown below:

[0016]

[0017] in It is the input of the i-th layer. and Here are the trainable parameters: l is the number of gated convolutional layers, n is the length of the protein sequence, m1 and m2 are the dimensions of the input and hidden features of each gated convolutional layer, k is the size of the one-dimensional convolutional layer (Conv1D), and σ refers to the Sigmoid function. This represents the element-wise product of matrices. In this patent, l is 3, m1 is 64, m2 is 64, and k is 7.

[0018] Specifically, in step S2, the compound encoder performs feature encoding of the compound structure. This involves using RDKit to encode the compounds in the pharmaceutical excipient-protein interaction dataset, obtaining a 34-dimensional atomic feature matrix. Subsequently, an adjacency matrix is ​​constructed based on the covalent bonds between atoms. A single-layer graph convolutional network (GCN) is used to aggregate adjacent features and update the atomic features, thereby obtaining the compound encoding vector.

[0019] Furthermore, in step S2, the interaction decoder consists of a self-attention layer, a cross-attention layer, and a feedforward layer.

[0020] The formula for calculating attention weights is as follows:

[0021]

[0022] Where d k This is a scaling factor for the hidden layer dimension and the number of heads. For the first decoder layer, a self-attention mechanism is first applied to extract updated compound features. Q, K, and V are all calculated from the compound encoding vector in this layer. Since the order of the atom vectors does not affect the modeling, positional embeddings are removed in PreDIG. Then, a cross-attention layer is used to capture the interaction between the compound and the protein, where Q comes from the self-attention layer output of the compound, and K and V come from the protein encoding vector, ensuring that the interaction decoder can access the entire sequence information. Subsequently, a feedforward layer is used to further refine the features, enhancing expressive power through nonlinear transformations.

[0023] Further, in step S2, the compound encoding vector and protein encoding vector are input into the interaction decoder to obtain the interaction matrix, which is then input into the classifier composed of fully connected layers to construct the initial model. After training is completed, the pre-trained model PreDIG is obtained.

[0024] Specifically, in step S2, the classifier consists of two fully connected layers. The output of the first fully connected layer undergoes a non-linear transformation using the ReLU activation function; the last layer normalizes the output to the probabilities of positive and negative classes using the Softmax function. Subsequently, the classification loss Loss is calculated using the binary cross-entropy loss function based on the difference between the predicted probability and the true label. The calculation formula is as follows:

[0025]

[0026] Where y i It is a true label for drug-protein interaction pairs. is the predicted probability of the interaction between the compound and the protein, and N is the number of all interaction pairs.

[0027] Furthermore, in step S3, the training data employs a stratified sampling strategy, dividing the drug-like molecule activity data in the BindingDB dataset into a 4:1 ratio for training and validation sets, respectively. These sets are used for pre-training and validation of the PreDIG model, enabling the protein encoder and interaction decoder to learn a large number of protein characterizations and compound-protein interaction patterns. When the model achieves optimal performance on the validation dataset, the PreDIG model parameters are saved.

[0028] To comprehensively evaluate model performance, this invention also employs the following two cross-validation strategies: Random 5-fold Cross-Validation: a) The entire dataset is randomly divided into five non-overlapping subsets. In each training iteration, four subsets are selected as the training set, and the remaining subset is used as the validation set. This process is repeated five times, ensuring that each subset is used as the validation set once. The final model performance metric is the average of the five validation results. This method effectively reduces the model's dependence on specific data partitions; b) Zero-shot Protein 5-fold Cross-Validation: First, all protein IDs are divided into five groups, and then five rounds of training are performed. In each round, the data corresponding to four protein IDs is selected as the training set, and the data corresponding to the remaining protein ID is used as the validation set. This strategy aims to evaluate the model's generalization ability to unknown proteins.

[0029] Furthermore, in step S3, the present invention comprehensively evaluates the model's performance by considering four evaluation metrics: Matthews Correlation Coefficient (MCC), Area Under the Receiver Operating Characteristic Curve (AUROC), Area Under the Precision-Recall Curve (AUPRC), and Accuracy (ACC).

[0030] MCC: A comprehensive model performance evaluation metric used to assess the overall predictive performance of a binary classification model. Values ​​range from -1 to 1, where 1 represents completely correct predictions, 0 represents random predictions, and -1 represents completely incorrect predictions. The calculation formula is:

[0031]

[0032] AUROC: Used to measure the ability of a classification model to distinguish between positive and negative classes. It represents the area under the ROC curve; the closer the value is to 1, the better the model performance. The horizontal axis of the ROC curve represents the false positive rate (FPR), and the vertical axis represents the true positive rate (TPR). The calculation formula is:

[0033] AUROC=∫0 1 TPR(FPR)d(FPR)

[0034]

[0035] AUPRC: Used to measure the balance between precision and recall of a classification model at different thresholds; the higher the value, the better the performance. The horizontal axis of the curve represents recall, and the vertical axis represents precision. The calculation formula is:

[0036] AUPRC=∫0 1 Precision(Recall)d(Recall)

[0037]

[0038] In the above formula, TP represents the number of true positive samples (the number of correctly predicted positive samples), FP represents the number of false positive samples (the number of negative samples that were incorrectly predicted as positive), TN represents the number of true negative samples (the number of correctly predicted negative samples), and FN represents the number of false negative samples (the number of positive samples that were incorrectly predicted as negative).

[0039] ACC: Used to measure the overall accuracy of a classification model's predictions. It represents the proportion of samples correctly predicted by the model out of the total number of samples. A higher accuracy indicates that the model predicts more accurately overall. The calculation formula is:

[0040]

[0041] In the formula, TP represents the number of true positive samples, which are correctly predicted as positive by the model; TN represents the number of true negative samples, which are correctly predicted as negative by the model; FP represents the number of false positive samples, which are incorrectly predicted as positive by the model; and FN represents the number of false negative samples, which are incorrectly predicted as negative by the model.

[0042] Further, in step S4, a deep transfer learning strategy is used to train the PreDIG model to obtain the BioDIG model. Specifically, this includes loading the pre-trained parameters of the PreDIG model, initializing the parameters, and then fine-tuning the training to obtain the BioDIG model. Fine-tuning includes training the parameters of the compound encoder and binary classifier of the PreDIG model using pharmaceutical excipient-protein interaction data. Using pharmaceutical excipient structures and protein sequences as input, a deep transfer learning strategy is used to fine-tune the PreDIG model to obtain the BioDIG model. Specifically, a model named BioDIG, consistent with PreDIG in framework, is constructed, and the BioDIG model is initialized by loading the pre-trained parameters of PreDIG. Subsequently, using pharmaceutical excipient structures and protein sequences as input, the compound encoder and classifier of the BioDIG model are fine-tuned using pharmaceutical excipient-protein interaction data, while employing the same data partitioning strategy and performance evaluation metrics as in step S3, thereby achieving model optimization and performance improvement.

[0043] Further, in step S5, the sequences of the pharmaceutical excipient and the protein to be tested are input into BioDIG to obtain the predicted probability of the interaction between the pharmaceutical excipient and the protein, and the binding of the pharmaceutical excipient and the protein is determined based on the predicted probability of the interaction.

[0044] Compared with the prior art, the present invention has the following advantages:

[0045] Due to the significant differences in activity distribution between BAMs and DIGs, traditional CPI prediction methods are insufficiently accurate in predicting DIG activity, and the scarcity of DIG-related activity data further limits the model's generalization ability. This invention effectively addresses the problems of insufficient model generalization and inaccurate predictions caused by activity distribution bias under sparse DIG data conditions by introducing a cross-attention Transformer architecture and a deep transfer learning strategy, thereby significantly improving the reliability of pharmaceutical excipient-protein interaction prediction. Simultaneously, this invention validates the actual biological activities of several DIGs through activity assays. For example, it discovers that β-carotene can inhibit organic cation transporter 1 (OCT1) and reveals for the first time the regulatory role of dodecyl gallate on epidermal growth factor receptor (EGFR), further demonstrating the accuracy of the model's prediction results. In summary, the BioDIG model developed in this invention can efficiently and accurately predict the biological activities of DIGs within the proteome, overcoming the limitations of existing methods and providing a novel computational tool for drug formulation design and clinical application. Attached Figure Description

[0046] Figure 1This is a flowchart illustrating the precise prediction method for pharmaceutical excipient-target interaction based on transfer learning strategy in this invention.

[0047] Figure 2 This is a schematic diagram of the model structure of the accurate prediction method for pharmaceutical excipient-target interaction based on transfer learning strategy in this invention;

[0048] Figure 3 This invention is compared with six other prediction methods.

[0049] Figure 4 This invention aims to verify the predicted interaction between propyl gallate-ALOX5 and methylene blue-CHRM2 in an independent test set and its molecular docking experiments.

[0050] Figure 5 This study aims to verify the predicted interactions of linoleic acid-SPHK1, β-carotene-OCT1, gallate dodecyl ester-EGFR, and chlorhexidine-EGFR in this invention and to conduct bioactivity testing experiments. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0052] A method for accurately predicting pharmaceutical excipient-protein interactions based on cross-attention mechanism and deep transfer learning strategy, characterized by the following steps:

[0053] Step S1: Collect large-scale drug-molecule-protein activity data and pharmaceutical excipient-protein interaction data;

[0054] Step S2: Construct an interaction prediction model framework based on the cross-attention mechanism;

[0055] Step S3: Input the drug-like molecule structure and protein sequence into the prediction model to train and obtain the pre-trained model PreDIG;

[0056] Step S4: Using the structure of pharmaceutical excipients and protein sequences as input, fine-tune the PreDIG model using a deep transfer learning strategy to obtain the BioDIG model;

[0057] Step S5: Input the structure and protein sequence of the pharmaceutical excipient to be tested into BioDIG to obtain the predicted probability of the interaction between the pharmaceutical excipient and the protein. Based on the predicted probability of the interaction between the pharmaceutical excipient and the protein, determine whether the pharmaceutical excipient binds to the protein.

[0058] In step S1, the large-scale drug-molecule-protein activity data is the BindingBD dataset. The pharmaceutical excipient-protein interaction data is the ACDINA dataset.

[0059] The aforementioned drug-like molecule-protein activity dataset includes: drug-like molecule data, protein data, and compound-protein activity data. The aforementioned pharmaceutical excipient-protein interaction dataset includes: pharmaceutical excipient data, protein data, and pharmaceutical excipient-protein interaction data.

[0060] In step S2, an interaction prediction model framework based on the cross-attention mechanism is constructed, which uses a compound encoder, a protein encoder, an interaction decoder, and a binary classifier to build the model framework. Specifically, this includes:

[0061] 2.1) The compound encoder and protein encoder were used to encode the features of the drug-like molecule structure and protein sequence, respectively, to obtain the compound encoding vector and the protein encoding vector, respectively;

[0062] 2.2) After inputting the compound encoding vector and protein encoding vector into the interaction decoder based on the cross-attention mechanism, the interaction matrix is ​​obtained;

[0063] 2.3) Input the interaction matrix into a binary classifier consisting of fully connected layers and train the model using the cross-entropy loss function.

[0064] In step 2.1), the compound encoder and protein encoder are used respectively to encode the features of the compound and protein sequences. Specifically, this includes:

[0065] 3.1) The protein data in the compound-protein interaction dataset is encoded using the natural language processing technology Word2Vec to obtain protein encoding vectors.

[0066] 3.2) The compounds in the compound-protein interaction dataset are encoded using RDKit to obtain a 34-dimensional atomic feature matrix. Then, an adjacency matrix is ​​constructed based on the covalent bonds between atoms. A Graph Convolutional Network (GCN) is used to aggregate the adjacency matrix and update the atomic features to obtain the compound encoding vector.

[0067] In step 2.2), the interaction decoder consists of a self-attention layer, a cross-attention layer, and a feedforward layer. The formula for calculating the attention weights is as follows:

[0068]

[0069] Where d k It is a scaling factor for the hidden layer dimension and the number of heads. In the self-attention layer, Q, K, and V are all calculated from the compound encoding vector; in the cross-attention layer, Q is calculated from the output of the self-attention layer, and K and V are derived from the protein encoding vector.

[0070] In step 2.3), the cross-entropy loss function calculates the classification loss Loss based on the difference between the probability value output by the binary classifier and the true label. The calculation formula is as follows:

[0071]

[0072] Among them, y i It is a true label for compound-protein interaction pairs. is the predicted probability of the interaction between the compound and the protein, and N is the number of all interaction pairs.

[0073] In step S3, the prediction model is the interaction prediction model framework in step S2. The drug-like molecule structure and protein sequence are input into the prediction model to train and obtain the pre-trained model PreDIG.

[0074] In step S4, the framework of the BioDIG model is consistent with that of PreDIG; the deep transfer learning strategy refers to loading the pre-trained model parameters of PreDIG to initialize the parameters of the BioDIG model; the fine-tuning training refers to using pharmaceutical excipient-protein interaction data to fine-tune the parameters of the compound encoder and binary classifier of BioDIG.

[0075] In this implementation case, such as Figure 1 As shown, this invention provides BioDIG, a method for accurately predicting pharmaceutical excipient-target interactions based on transfer learning strategies, comprising the following steps:

[0076] Step S1: Collect large-scale drug-molecule-protein activity data and pharmaceutical excipient-protein interaction data.

[0077] The bioactivity data of drug-like molecules were obtained from the BindingDB dataset. The main data processing steps included: a) removing data lacking a clear drug-like molecule name, SMILES sequence, or a clear protein name and amino acid sequence; b) removing pairing data containing multiple protein chains in the target; c) selecting data with binding activity indicators (IC50).50 EC 50 d) Remove data with no literature support; e) Label pairs with activity values ​​less than 1 μM as positive data and pairs with activity values ​​greater than 10 μM as negative data; f) Exclude compound-protein combinations with activity values ​​between 1 μM and 10 μM to avoid interference from ambiguous data in model training; g) Remove pairs with contradictory labels and remove duplicate pairs to ensure data consistency and accuracy.

[0078] Bioactivity data for pharmaceutical excipients were obtained from the ACDINA dataset. Data processing was performed using the same workflow described above. Furthermore, to ensure the comprehensiveness of DIG activity, a complete list of pharmaceutical excipients was first retrieved from ACDINA. Then, all activity data related to pharmaceutical excipients were extracted from the processed BindingDB dataset and integrated into the ACDINA dataset.

[0079] Ultimately, the BindingDB dataset used for pre-training the PreDIG model contains 996,265 positive samples and 398,059 negative samples, involving over 7,600 unique proteins and 840,000 drug-like molecules. Meanwhile, the ACDINA benchmark dataset used for fine-tuning the BioDIG model includes 182 positive and 813 negative samples, covering 499 proteins and 222 pharmaceutical excipients. Furthermore, there is no overlap in compounds between the BindingDB and ACDINA datasets to prevent information leakage during model evaluation.

[0080] Step S2: Construct an interaction prediction model framework based on the cross-attention mechanism.

[0081] like Figure 2 As shown, the interaction prediction model framework based on the cross-attention mechanism constructed in this invention includes a compound encoder, a protein encoder, an interaction decoder, and a binary classifier. The detailed components are as follows:

[0082] 1) Protein Encoder: The target data in the pharmaceutical excipient-target interaction dataset is encoded using the natural language processing technique Word2Vec to obtain target encoding vectors. The protein residue encoding matrix is ​​first transformed into an L×64 matrix using a fully connected layer (FC), and then input into a gated convolutional neural network (GatedCNN). Each gated convolutional layer contains one convolutional unit and one gated linear unit. GatedCNN contains a total of l gated convolutional layers, and the calculation formula for the i-th layer is shown below.

[0083]

[0084] in It is the input of the i-th layer. and Here are the trainable parameters: l is the number of gated convolutional layers, n is the length of the protein sequence, m1 and m2 are the dimensions of the input and hidden features of each gated convolutional layer, k is the size of the one-dimensional convolutional layer (Conv1D), and σ refers to the Sigmoid function. This represents the element-wise product of matrices. In this patent, l is 3, m1 is 64, m2 is 64, and k is 7.

[0085] 2) Compound Encoder: The RDKit is used to encode the atoms of drug-like molecules or pharmaceutical excipients in the pharmaceutical excipient-target interaction dataset, resulting in a 34-dimensional atomic feature matrix. Subsequently, an adjacency matrix is ​​constructed based on the covalent bonds between atoms. A single-layer graph convolutional network (GCN) is used to aggregate the adjacency matrix and update the atomic features, thereby obtaining the encoded vector of the drug-like molecule or pharmaceutical excipient.

[0086] 3) Interaction Decoder: The interaction decoder in this invention is an improved Transformer decoder, consisting of a self-attention layer, a cross-attention layer, and a feedforward layer.

[0087] The formula for calculating attention weights is as follows:

[0088]

[0089] Where d k This is a scaling factor for the hidden layer dimension and the number of heads. For the first decoder layer, a self-attention mechanism is first applied to extract updated compound features. Q, K, and V are all calculated from the compound encoding vector in this layer. Since the order of the atom vectors does not affect the modeling, positional embeddings are removed in PreDIG. Then, a cross-attention layer is used to capture the interaction between the compound and the protein, where Q comes from the self-attention layer output of the compound, and K and V come from the protein encoding vector, ensuring that the interaction decoder can access the entire sequence information. Subsequently, a feedforward layer is used to further refine the features, enhancing expressive power through nonlinear transformations.

[0090] The drug-like molecule encoding vector and protein encoding vector are input into the interaction decoder to obtain the interaction matrix, which is then input into a classifier composed of fully connected layers to build the initial model. After training, the pre-trained model PreDIG is obtained.

[0091] 4) Binary Classifier: The binary classifier in this invention consists of two fully connected layers. The output of the first fully connected layer is nonlinearly transformed using the ReLU activation function, and the last layer normalizes the output to the probability that the interaction pair belongs to one of the two classes (positive and negative) using the softmax function. The calculation formula is as follows:

[0092]

[0093] Where y i It is a true label for drug-like molecule-target interaction pairs. is the predicted probability of the interaction between the compound and the target, and N is the number of all interaction pairs.

[0094] Step S3: Input the drug-like molecule structure and protein sequence into the prediction model to train and obtain the pre-trained model PreDIG.

[0095] like Figure 2 As shown, the training method for the pre-trained model PreDIG is as follows: The training data employs a stratified sampling strategy, dividing the drug-like molecule activity data in the BindingDB dataset into a 4:1 ratio for training and validation sets. These sets are used for pre-training and validation of the PreDIG model, respectively, allowing the protein encoder and interaction decoder to learn a large number of protein characterizations and compound-protein interaction patterns. When the model achieves optimal performance on the validation dataset, the PreDIG model parameters are saved.

[0096] To comprehensively evaluate model performance, this invention also employs the following two cross-validation strategies: Random 5-fold Cross-Validation: a) The entire dataset is randomly divided into five non-overlapping subsets. In each training iteration, four subsets are selected as the training set, and the remaining subset is used as the validation set. This process is repeated five times, ensuring that each subset is used as the validation set once. The final model performance metric is the average of the five validation results. This method effectively reduces the model's dependence on specific data partitions; b) Zero-shot Protein 5-fold Cross-Validation: First, all protein IDs are divided into five groups, and then five rounds of training are performed. In each round, the data corresponding to four protein IDs is selected as the training set, and the data corresponding to the remaining protein ID is used as the validation set. This strategy aims to evaluate the model's generalization ability to unknown proteins.

[0097] This invention comprehensively considers four evaluation metrics: Matthews Correlation Coefficient (MCC), Area Under the Receiver Operating Characteristic Curve (AUROC), Area Under the Precision-Recall Curve (AUPRC), and Accuracy (ACC). Higher values ​​for these metrics indicate better model predictive performance.

[0098] MCC: A comprehensive model performance evaluation metric used to assess the overall predictive performance of a binary classification model. Values ​​range from -1 to 1, where 1 represents completely correct predictions, 0 represents random predictions, and -1 represents completely incorrect predictions. The calculation formula is:

[0099]

[0100] AUROC: Used to measure the ability of a classification model to distinguish between positive and negative classes. It represents the area under the ROC curve; the closer the value is to 1, the better the model performance. The horizontal axis of the ROC curve represents the false positive rate (FPR), and the vertical axis represents the true positive rate (TPR). The calculation formula is:

[0101] AUROC=∫0 1 TPR(FPR)d(FPR)

[0102]

[0103] AUPRC: Used to measure the balance between precision and recall of a classification model at different thresholds; the higher the value, the better the performance. The horizontal axis of the curve represents recall, and the vertical axis represents precision. The calculation formula is:

[0104] AUPRC=∫0 1 Precision(Recall)d(Recall)

[0105]

[0106] In the above formula, TP represents the number of true positive samples (the number of correctly predicted positive samples), FP represents the number of false positive samples (the number of negative samples that were incorrectly predicted as positive), TN represents the number of true negative samples (the number of correctly predicted negative samples), and FN represents the number of false negative samples (the number of positive samples that were incorrectly predicted as negative).

[0107] ACC: Used to measure the overall accuracy of a classification model's predictions. It represents the proportion of samples correctly predicted by the model out of the total number of samples. A higher accuracy indicates that the model predicts more accurately overall. The calculation formula is:

[0108]

[0109] In the formula, TP represents the number of true positive samples, which are correctly predicted as positive by the model; TN represents the number of true negative samples, which are correctly predicted as negative by the model; FP represents the number of false positive samples, which are incorrectly predicted as positive by the model; and FN represents the number of false negative samples, which are incorrectly predicted as negative by the model.

[0110] Step S4: Using the pharmaceutical excipient structure and protein sequence as input, fine-tune the PreDIG model using a deep transfer learning strategy to obtain the BioDIG model.

[0111] like Figure 2 As shown, this invention uses a transfer learning strategy to construct a BioDIG model. The pharmaceutical excipient sequence and target sequence are input into the BioDIG model, and the model is fine-tuned during training to obtain the BioDIG model. The specific steps are as follows:

[0112] 1) Use transfer learning strategy: Build a model BioDIG that is structurally consistent with PreDIG, and initialize the BioDIG model by loading the pre-trained parameters of PreDIG.

[0113] 2) Model Fine-tuning Training Strategy: The parameters of the BioDIG compound encoder and binary classifier were fine-tuned using DIG activity data. Specifically, a five-fold cross-validation task was used to train and evaluate the performance of BioDIG. In this task, all DIG activity data in the ACDINA dataset were randomly divided into five folds, with four folds used for training BioDIG and the remaining fold used as the validation dataset for evaluation. This process was repeated five times, and the average of the five results was calculated to obtain the overall evaluation result. Based on these results, the optimal hyperparameters were selected, and the best average result was reported. After determining the optimal hyperparameters, the final BioDIG model was retrained using the complete ACDINA dataset. During training, an early stopping strategy based on the AUROC metric (patience of 10 epochs) was introduced to reduce the risk of overfitting and decrease training costs. Simultaneously, regularization methods such as dropout and weight decay were employed to accelerate model convergence and improve generalization ability. The final optimal combination of hyperparameters was determined to be: learning rate 0.0008, dropout rate 0.1, and batch size 32.

[0114] Step S5: Input the structure and protein sequence of the pharmaceutical excipient to be tested into BioDIG to obtain the predicted probability of the interaction between the pharmaceutical excipient and the protein. Based on the predicted probability of the interaction between the pharmaceutical excipient and the protein, determine whether the pharmaceutical excipient binds to the protein.

[0115] For any pharmaceutical excipient-target pair, the sequences of the pharmaceutical excipient and the target are input into the final BioDIG model to obtain the predicted probability of the interaction between the pharmaceutical excipient and the target. Based on the predicted probability of the interaction between the pharmaceutical excipient and the target, it is determined whether the pharmaceutical excipient and the target have bound together.

[0116] In this embodiment, as shown in Table 1 and Figure 3 As shown in Figure a, the developed BioDIG model outperforms six existing state-of-the-art models across all evaluation metrics in a randomly partitioned five-fold cross-validation task, fully validating the effectiveness of the model proposed in this invention. Furthermore, as shown in Table 2, Figure 3 b and Figure 3 As shown in c, in the five-fold cross-validation task for zero-sample proteins and the independent sample testing task, the BioDIG model also comprehensively outperformed the other five state-of-the-art models in all indicators, demonstrating that the model also possesses extremely high accuracy and reliability in predicting pharmaceutical excipient-target interactions with unseen targets. Furthermore, as... Figure 3 As shown in d and Table 1, the final model BioDIG of this invention significantly outperforms the pre-trained model PreDIG in overall performance, fully demonstrating the key role of transfer learning strategy in this invention.

[0117] Furthermore, in independent sample testing tasks, such as Figure 4 As shown in Table 2, the predictions of the BioDIG model proposed in this invention regarding the interaction between propyl gallate-ALOX5 and methylene blue-CHRM2 were validated in molecular docking experiments. The results show that the PHE556 interaction between propyl gallate and ALOX5... ASN555 and TYR559 The residues form hydrogen bonds; methylene blue, on the other hand, forms hydrogen bonds in CHRM2. The protein interacts with ILE72, SER76, PHE181, and PHE195 residues within the target range. This demonstrates that the BioDIG model can effectively capture key interactions between pharmaceutical excipient atoms and protein residues, and these visualization results prove the superiority of this invention in predicting pharmaceutical excipient-target interactions.

[0118] In this embodiment, to evaluate the ability of the BioDIG model to predict the interaction between pharmaceutical excipients and targets in the real world, we selected 222 DIGs from the ACDINA dataset and 5 clinically significant human protein targets, and used BioDIG to predict the binding probability of each pharmaceutical excipient to its target. These 5 protein targets include arachidonic acid 5-lipoxygenase (ALOX5), muscarinic acetylcholine receptor M2 subtype (CHRM2), sphingosine kinase 1 (SPHK1), organic cation transporter 1 (OCT1), and epidermal growth factor receptor (EGFR). Experimental results are as follows: Figure 5 As shown: Linoleic acid, as a surfactant, had a prediction score of 0.989 for the SPHK1 target, corresponding to an experimental IC50 value of 0.989. 50 The concentration was 80.7 μM; β-carotene as a colorant had a prediction score of 0.995 for the OCT1 target, and the IC50 was 80.7 μM. 50 The concentration was 138.4 μM; dodecyl gallate as an antioxidant had a prediction score of 0.993 for the EGFR target, and an IC50 value of 138.4 μM. 50 The concentration was 27.0 μM; chlorhexidine, as an antibacterial preservative, had a prediction score of 0.991 for the EGFR target, and an IC50 value of 27.0 μM. 50 The concentration was 59.2 μM. The prediction results of the BioDIG model for pharmaceutical excipients-targets were highly consistent with the experimentally measured bioactivity, indicating that the model has a strong predictive ability in identifying the potential bioactivity of pharmaceutical excipients.

[0119] In summary, the interactions between pharmaceutical excipients and targets predicted by BioDIG have been validated by molecular docking and bioactivity testing experiments. This result further demonstrates the superior performance and reliability of the BioDIG model in accurately predicting pharmaceutical excipient-target interactions in this invention.

[0120] Table 1. Comparison of experimental data between the BioDIG model and other models in a randomly partitioned five-fold cross-validation task.

[0121]

[0122] Table 2. Comparison of experimental data between the BioDIG model and other models in the five-fold cross-validation task for zero-sample proteins.

[0123]

[0124] Table 3. BioDIG model prediction scores and actual biological activities for the interaction of a given target in two pharmaceutical excipients.

[0125]

[0126] Finally, it should be noted that although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention.

[0127] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

[0128] The application of this invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for accurately predicting pharmaceutical excipient-protein interactions based on cross-attention mechanism and deep transfer learning strategy, characterized in that, Includes the following steps: Step S1: Collect drug molecule-protein activity data and pharmaceutical excipient-protein interaction data; Step S2: Construct an interaction prediction model framework based on the cross-attention mechanism; Step S3: Input the drug-like molecule data and protein data from the drug-like molecule-protein activity data in Step S1 into the interaction prediction model framework based on the cross-attention mechanism constructed in Step S2, and train to obtain the pre-trained model PreDIG; Step S4: Using the pharmaceutical excipient data and protein data from the pharmaceutical excipient-protein interaction data in Step S1 as input, train the PreDIG model using a deep transfer learning strategy to obtain the BioDIG model. Step S5: Input the structure and protein sequence of the pharmaceutical excipient to be tested into the BioDIG model to obtain the predicted probability of the interaction between the pharmaceutical excipient and the protein. Based on the predicted probability of the interaction between the pharmaceutical excipient and the protein, determine whether the pharmaceutical excipient binds to the protein.

2. The method for accurate prediction of pharmaceutical excipient-protein interactions based on cross-attention mechanism and deep transfer learning strategy according to claim 1, characterized in that, In step S1, the drug-protein activity data is the BindingBD dataset, and the pharmaceutical excipient-protein interaction data is the ACDINA dataset.

3. The method for accurate prediction of pharmaceutical excipient-protein interactions based on cross-attention mechanism and deep transfer learning strategy according to claim 1, characterized in that, In step S1, the drug-like molecule-protein activity data includes: drug-like molecule data, protein data, and compound-protein activity data; The pharmaceutical excipient-protein interaction data includes: pharmaceutical excipient data, protein data, and pharmaceutical excipient-protein interaction data.

4. The method for accurate prediction of pharmaceutical excipient-protein interactions based on cross-attention mechanism and deep transfer learning strategy according to claim 1, characterized in that, In step S2, the interaction prediction model framework based on the cross-attention mechanism includes: a compound encoder, a protein encoder, an interaction decoder based on the cross-attention mechanism, and a binary classifier. The compound encoder performs feature encoding on drug-like molecule data to obtain a compound encoding vector; The protein encoder performs feature encoding on the protein data to obtain a protein encoding vector; The interaction decoder based on the cross-attention mechanism takes compound coding vectors and protein coding vectors as inputs and the interaction matrix as output. The binary classifier is trained using the interaction matrix as input and the cross-entropy loss function.

5. The method for accurate prediction of pharmaceutical excipient-protein interactions based on cross-attention mechanism and deep transfer learning strategy according to claim 4, characterized in that, The interaction decoder based on the cross-attention mechanism includes a self-attention layer, a cross-attention layer, and a feedforward layer connected in sequence.

6. The method for accurate prediction of pharmaceutical excipient-protein interactions based on cross-attention mechanism and deep transfer learning strategy according to claim 4, characterized in that, Training is performed using the cross-entropy loss function, specifically including: The classification loss (Loss) is calculated based on the difference between the probability values ​​output by the binary classifier and the true labels. The formula is as follows: Among them, y i It is a true label for compound-protein interaction pairs. is the predicted probability of the interaction between the compound and the protein, and N is the number of all interaction pairs.

7. The method for accurate prediction of pharmaceutical excipient-protein interactions based on cross-attention mechanism and deep transfer learning strategy according to claim 1, characterized in that, In step S4, a deep transfer learning strategy is used to train the PreDIG model to obtain the BioDIG model, specifically including: Load the pre-trained model parameters of the PreDIG model, initialize the parameters, and then fine-tune the training to obtain the BioDIG model.

8. The method for accurate prediction of pharmaceutical excipient-protein interactions based on cross-attention mechanism and deep transfer learning strategy according to claim 1, characterized in that, In step S4, fine-tuning the training includes: The parameters of the compound encoder and binary classifier of the PreDIG model were trained using pharmaceutical excipient-protein interaction data.

Images