Drug-target interaction prediction method and system based on dual-channel feature consistency and course pseudo-label

Abstract

The present application relates to a kind of based on double channel feature consistency and course pseudo label cross-domain drug-target interaction prediction method and system.The method comprises: constructing double channel coding network with different inductive bias, by sequence feature channel, drug sequence, protein sequence are encoded, obtain sequence feature representation;By semantic feature channel, drug sequence, protein sequence are extracted to high-level semantic feature, obtain semantic feature representation;By multilayer perception, obtain drug-target interaction prediction score;For unlabelled sample, generate pseudo label, according to prediction confidence from high to low, unlabelled sample is sorted, and using course learning strategy, teacher-student model architecture is trained based on prediction confidence, obtain the prediction model after training, output drug-target interaction prediction result.Can improve the accuracy and stability of drug-target interaction prediction.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to a method and system for predicting cross-domain drug-target interactions based on dual-channel feature consistency and course pseudo-labels. Background Technology

[0002] Drug-target interactions refer to the binding or regulatory relationships between drug molecules and specific protein targets in organisms, and are a key fundamental issue in drug discovery and drug retargeting research. Accurately identifying potential drug-target interactions helps narrow down experimental screening, reduce drug development costs, and accelerate the new drug development process. With the continuous growth of biological data, computational methods have become an important tool for studying drug-target interactions. By modeling drug molecule structural information and protein sequence information, predicting the existence of interactions between them can provide effective candidate results before experimental validation. Traditional computational methods mainly include: sequence modeling-based methods, which encode drug molecule sequences or fingerprint features and protein amino acid sequences, and use models such as convolutional neural networks and recurrent neural networks to learn sequence features, and infer whether there is an interaction between the drug and the target based on the model output; graph structure modeling-based methods, which represent drug molecules as molecular graph structures, use graph neural networks to extract structural features between atoms and chemical bonds, and combine them with protein sequence features to predict drug-target interactions; and pre-trained model and semi-supervised learning-based methods, which mainly introduce models pre-trained on large-scale molecular or protein data to obtain high-level semantic features, or use unlabeled data in the target domain to generate pseudo-labels, thereby improving the model's predictive ability in scenarios with insufficient data or changing data distribution.

[0003] However, in practical applications, new drug molecules or targets often differ significantly from existing training data in terms of structural or sequence features, leading to changes in data distribution. This distributional difference results in insufficient generalization ability of prediction models trained on existing data in new scenarios, leading to a significant decline in prediction performance and becoming a major factor limiting the practical application of drug-target interaction prediction methods. Specifically, prediction methods based on sequence modeling and graph structure modeling are prone to inaccurate predictions or spurious interactions when facing new drugs or targets; methods based on pre-trained models and semi-supervised learning often suffer from uncertain target domain data distribution in cross-domain scenarios, resulting in model-generated pseudo-labels containing significant noise.

[0004] Therefore, traditional drug-target interaction prediction methods usually assume that the training data and the test data follow the same distribution and do not fully consider the consistency constraints between different feature perspectives, which often results in low prediction accuracy and stability. Summary of the Invention

[0005] Based on this, in order to solve the above-mentioned technical problems, a method and system for predicting cross-domain drug-target interactions based on dual-channel feature consistency and course pseudo-labels is provided, which can improve the accuracy and stability of drug-target interaction prediction.

[0006] A method for predicting cross-domain drug-target interactions based on dual-channel feature consistency and course pseudo-labels, the method comprising:

[0007] Obtain source domain data containing known drug-target interaction tags and unlabeled target domain data, and standardize the drug molecules and target proteins in the source domain data and target domain data respectively to obtain a unified input representation;

[0008] A dual-channel encoding network with different inductive biases is constructed. The drug sequence and protein sequence in the input representation are encoded through the sequence feature channel of the dual-channel encoding network to obtain the sequence feature representation. Then, the high-level semantic features of the drug sequence and protein sequence are extracted through the semantic feature channel of the dual-channel encoding network to obtain the semantic feature representation.

[0009] The sequence feature representation and semantic feature representation are concatenated and fused after global average pooling. The fused features are then nonlinearly mapped by a multilayer perceptron to obtain the drug-target interaction prediction score.

[0010] A teacher-student model architecture is constructed. For unlabeled samples in the target domain data, pseudo-labels are generated based on the drug-target interaction prediction scores. The unlabeled samples are sorted from high to low prediction confidence. A course learning strategy is adopted to train the teacher-student model architecture across domains based on the prediction confidence to obtain the trained prediction model.

[0011] The drug-target sample to be predicted is input into the trained prediction model, and the drug-target interaction prediction result is output.

[0012] In one embodiment, the drug molecules and target proteins in the source domain data and target domain data are standardized, including:

[0013] The drug molecule is structurally normalized and cleaned to convert it into a SMILES sequence. The SMILES sequence is then length-adjusted and character-encoded to obtain the drug sequence.

[0014] The target protein is cleaned to remove non-standard amino acid characters, and the standard amino acids are standardized in length and encoded to obtain the protein sequence.

[0015] The drug sequence and protein sequence are respectively subjected to numerical normalization and embedding mapping to obtain a unified input representation.

[0016] In one embodiment, the drug sequence and protein sequence in the input representation are encoded through a sequence feature channel to obtain a sequence feature representation, including:

[0017] After the drug sequence and protein sequence in the input representation are vectorized, they are sequentially input into a bidirectional long short-term memory network and a Mamba module to capture the local structure and contextual dependency information of the drug sequence and protein sequence, respectively, to obtain drug sequence features and protein sequence features.

[0018] The drug sequence features and protein sequence features are interactively modeled using a bidirectional cross-attention mechanism to generate a sequence feature representation that integrates drug-target sequence association information.

[0019] In one embodiment, high-level semantic features are extracted from the drug sequence and protein sequence respectively through a semantic feature channel to obtain semantic feature representations, including:

[0020] The drug sequence is input into a pre-trained molecular language model, and a drug semantic feature matrix is ​​obtained through semantic encoding; the protein sequence is input into a pre-trained protein language model, and a protein semantic feature matrix is ​​obtained through semantic encoding.

[0021] Global average pooling is performed along the sequence dimension on the drug semantic feature matrix and the protein semantic feature matrix respectively to extract the high-level semantic representations of the drug and protein;

[0022] After inputting the high-level semantic representations of drugs and proteins into a lightweight multilayer perceptron for linear mapping, the semantic correlation between drugs and proteins is obtained through a bidirectional cross-attention mechanism, generating a semantic feature representation that integrates the semantic association information between drugs and targets.

[0023] In one embodiment, the sequence feature representation and semantic feature representation are globally averaged and then concatenated and fused. A multilayer perceptron is used to perform a nonlinear mapping on the fused features to obtain a drug-target interaction prediction score, including:

[0024] Global average pooling is performed on the sequence feature representation and semantic feature representation respectively to obtain the sequence feature vector and the semantic feature vector.

[0025] The sequence feature vector and the semantic feature vector are concatenated and fused along the feature dimension to obtain a cross-view fused feature vector;

[0026] After normalizing the cross-view fusion feature vector, it is input into a multilayer perceptron. After nonlinear transformation by the multilayer perceptron, the drug-target interaction prediction score is output after passing through the Sigmoid activation function.

[0027] In one embodiment, constructing a teacher-student model architecture includes:

[0028] Construct a teacher model and a student model with consistent architecture; wherein the student model is a model composed of the dual-channel coding network and feature fusion prediction; the initial parameters of the teacher model are copied from the parameters of the student model, and the parameters of the teacher model are dynamically updated by the parameters of the student model through exponential moving average.

[0029] In one embodiment, for unlabeled samples in the target domain data, pseudo-labels are generated based on the drug-target interaction prediction scores. The unlabeled samples are then sorted from high to low prediction confidence. A course-learning strategy is employed to perform cross-domain training on the teacher-student model architecture based on the prediction confidence, resulting in a trained prediction model, including:

[0030] The unlabeled samples in the target domain data are input into the teacher model, and the dual-channel coding network is used to fuse and predict with features to obtain the sequence perspective prediction score and the semantic perspective prediction score.

[0031] The sequence perspective prediction score and the semantic perspective prediction score are evaluated for consistency. If the consistency evaluation result is reliable, the prediction score is fused as a pseudo-label, and the confidence level is calculated.

[0032] The unlabeled samples with pseudo-labels are sorted from high to low confidence. Using a course learning strategy, high-confidence pseudo-labeled samples are first mixed with labeled source domain data and input into the student model. The supervised loss is calculated to guide the basic learning of the student model. Then, low-confidence pseudo-labeled samples are gradually included, and the pseudo-label loss is calculated to complete cross-domain self-training, resulting in the trained prediction model.

[0033] In one embodiment, the method further includes:

[0034] The model performance evaluation index was determined, and the drug-target interaction prediction results were binarized to obtain the processed results.

[0035] Based on the processing results, the trained prediction model is evaluated using the model performance evaluation metrics to obtain the evaluation results.

[0036] A cross-domain drug-target interaction prediction system based on dual-channel feature consistency and course pseudo-labels, the system comprising:

[0037] The data processing module is used to acquire source domain data containing known drug-target interaction tags and unlabeled target domain data, and to standardize the drug molecules and target proteins in the source domain data and target domain data respectively to obtain a unified input representation.

[0038] The feature extraction module is used to construct a dual-channel encoding network with different inductive biases. The sequence feature channel in the dual-channel encoding network encodes the drug sequence and protein sequence in the input representation to obtain a sequence feature representation. The semantic feature channel in the dual-channel encoding network is used to extract high-level semantic features from the drug sequence and protein sequence to obtain a semantic feature representation.

[0039] The feature fusion and mapping module is used to perform global average pooling on the sequence feature representation and semantic feature representation and then concatenate and fuse them. The fused features are then nonlinearly mapped by a multilayer perceptron to obtain the drug-target interaction prediction score.

[0040] The model training module is used to construct a teacher-student model architecture. For unlabeled samples in the target domain data, pseudo-labels are generated based on the drug-target interaction prediction scores. The unlabeled samples are sorted from high to low prediction confidence. A course learning strategy is adopted to train the teacher-student model architecture across domains based on the prediction confidence to obtain the trained prediction model.

[0041] The prediction module is used to input the drug-target sample to be predicted into the trained prediction model and output the drug-target interaction prediction result.

[0042] The aforementioned cross-domain drug-target interaction prediction method and system based on dual-channel feature consistency and curriculum pseudo-labels enhances the completeness of feature representation by constructing dual-channel encoding networks with different inductive biases to jointly model drugs and targets from both sequence and semantic perspectives. Furthermore, by screening the reliability of pseudo-labels generated from unlabeled samples in the target domain and employing a curriculum learning strategy to gradually introduce high-confidence samples for training, pseudo-label noise and confirmation bias issues are effectively suppressed. This significantly improves the model's prediction accuracy, generalization ability, and training stability for unknown drugs or new targets in cross-domain scenarios. Attached Figure Description

[0043] Figure 1 This is an application environment diagram of a cross-domain drug-target interaction prediction method based on dual-channel feature consistency and course pseudo-labels in one embodiment.

[0044] Figure 2This is a flowchart illustrating a cross-domain drug-target interaction prediction method based on dual-channel feature consistency and course pseudo-labels in one embodiment.

[0045] Figure 3 This is a schematic diagram of the model training strategy architecture in one embodiment;

[0046] Figure 4 This is a flowchart illustrating a cross-domain drug-target interaction prediction method based on dual-channel feature consistency and course pseudo-labels in another embodiment.

[0047] Figure 5 This is a block diagram of a cross-domain drug-target interaction prediction system based on dual-channel feature consistency and course pseudo-labels in one embodiment.

[0048] Figure 6 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0050] The cross-domain drug-target interaction prediction method based on dual-channel feature consistency and course pseudo-labeling provided in this application can be applied to, for example... Figure 1 The application environment shown. For example... Figure 1As shown, the application environment includes computer device 110. Computer device 110 can acquire source domain data containing known drug-target interaction tags and unlabeled target domain data. It can standardize the drug molecules and target proteins in the source and target domain data respectively to obtain a unified input representation. Computer device 110 can construct a dual-channel encoding network with different inductive biases. The sequence feature channel in the dual-channel encoding network encodes the drug sequence and protein sequence in the input representation to obtain a sequence feature representation. Furthermore, the semantic feature channel in the dual-channel encoding network extracts high-level semantic features from the drug sequence and protein sequence respectively to obtain a semantic feature representation. Computer device 110 can process the sequence feature representation... The semantic feature representations are globally averaged and then concatenated and fused. A multilayer perceptron is used to perform a nonlinear mapping on the fused features to obtain the drug-target interaction prediction score. Computer device 110 can construct a teacher-student model architecture. For unlabeled samples in the target domain data, pseudo-labels are generated based on the drug-target interaction prediction score. The unlabeled samples are sorted from high to low prediction confidence. A course learning strategy is adopted to train the teacher-student model architecture across domains based on the prediction confidence, resulting in a trained prediction model. Computer device 110 can input the drug-target sample to be predicted into the trained prediction model and output the drug-target interaction prediction result. The computer device 110 can be, but is not limited to, various personal computers, laptops, smartphones, robots, unmanned aerial vehicles, tablets, etc.

[0051] In one embodiment, such as Figure 2 As shown, a method for predicting cross-domain drug-target interactions based on dual-channel feature consistency and course pseudo-labels is provided, including the following steps:

[0052] Step 202: Obtain source domain data containing known drug-target interaction tags and unlabeled target domain data. Standardize the drug molecules and target proteins in the source domain data and target domain data respectively to obtain a unified input representation.

[0053] Computer devices can collect drug-target interaction datasets containing drug molecules and target proteins, where the source domain data contains known interaction labels and the target domain data is unlabeled data.

[0054] In one embodiment, a cross-domain drug-target interaction prediction method based on dual-channel feature consistency and course pseudo-labels may further include a data processing procedure, specifically including: cleaning the drug molecule according to its structure and converting it into a SMILES sequence; adjusting the length and encoding the SMILES sequence to obtain the drug sequence; cleaning the target protein, removing non-standard amino acid characters from the target protein, and performing length unification and character encoding on the standard amino acids to obtain the protein sequence; and performing numerical normalization and embedding mapping on the drug sequence and protein sequence respectively to obtain a unified input representation.

[0055] Computer equipment can simultaneously standardize source domain data containing known drug-target interaction tags and unlabeled target domain data, achieving format unification and feature normalization for both drug molecule and target protein data, ultimately resulting in a standardized representation that can be directly input into a model for feature encoding. The computer equipment can collect source domain data and target domain data separately; both types of data contain pairing information between drug molecules and target proteins. The source domain data includes explicit drug-target interaction tags (positive: interaction exists; negative: no interaction), while the target domain data consists of drug-target pairing samples without interaction tags.

[0056] Next, the computer equipment can standardize the drug molecules, converting them into SMILES sequences or molecular diagram structures. Specifically, the computer equipment can perform standardized cleaning of the drug molecules to eliminate differences in drug structure representations from different data sources, and then convert them into SMILES sequences. SMILES sequences are character-based one-dimensional sequences that can accurately characterize the atomic composition, chemical bonding patterns, and spatial structural features of drug molecules. Simultaneously, the computer equipment can also perform length normalization on the generated SMILES sequences and uniformly encode sequences that meet the requirements to form drug sequences.

[0057] Computer equipment can perform standardized encoding and length processing on target proteins, converting them into standardized amino acid sequences. Specifically, the raw data of the target protein can be cleaned, and non-standard amino acid residues can be uniformly converted into standard amino acid characters to obtain a standardized amino acid sequence of the target protein. Then, the length distribution of all standardized amino acid sequences is statistically analyzed, a reasonable sequence length threshold is set, and the standardized amino acid sequences with uniform lengths are then encoded, mapping each standard amino acid character to a unique numerical identifier to form the target protein sequence.

[0058] Then, the drug sequence and protein sequence are mapped to the same common feature space through a linear projection layer to obtain a standardized input representation with a unified dimension that can be used for subsequent model training.

[0059] Step 204: Construct a dual-channel coding network with different inductive biases. Encode the drug sequence and protein sequence in the input representation through the sequence feature channel of the dual-channel coding network to obtain the sequence feature representation. Then, extract high-level semantic features from the drug sequence and protein sequence through the semantic feature channel of the dual-channel coding network to obtain the semantic feature representation.

[0060] Computer devices can be equipped with dual-channel coding networks with different inductive biases, including sequence feature channels and semantic feature channels, for feature extraction.

[0061] In one embodiment, a cross-domain drug-target interaction prediction method based on dual-channel feature consistency and course pseudo-labels can also be provided through the process of extracting feature representations via sequence feature channels. The specific process includes: vectorizing the drug sequence and protein sequence in the input representation, and then inputting them sequentially into a bidirectional long short-term memory network and a Mamba module to capture the local structure and contextual dependency information of the drug sequence and protein sequence, respectively, to obtain drug sequence features and protein sequence features; and using a bidirectional cross-attention mechanism to interactively model the drug sequence features and protein sequence features to generate a sequence feature representation that integrates drug-target sequence association information.

[0062] The sequence feature channel employs a two-layer heterogeneous network structure, consisting of a bidirectional long short-term memory network (BiLSTM) and a Mamba module, balancing the ability to model contextual dependencies and capture long-range information. The sequence feature channel encodes drug and protein sequences using a BiLSTM and Mamba modules, capturing local patterns and long-range dependencies to obtain task-relevant sequence feature representations. Specifically, drug and protein sequences are vectorized through an embedding layer and then sequentially input into the BiLSTM and Mamba modules to capture local structure and contextual dependency information. A bidirectional cross-attention mechanism is used to interactively model the drug and protein sequence features, generating a joint feature representation at the sequence level.

[0063] The computer device can input the vectorized drug sequence and protein sequence into independent BiLSTM networks. Forward LSTM and backward LSTM are used to model the forward and reverse context of the sequence, respectively, to mine local association patterns between adjacent characters or residues in the sequence, and obtain drug local feature vectors and protein local feature vectors containing bidirectional contextual information. Then, the drug and protein local feature vectors output by BiLSTM can be input into independent Mamba modules. Mamba, as a linear time series model, models the long-range features of the sequence through a selective state-space mechanism, effectively capturing long-range structural associations across functional groups in drug sequences and long-range spatial conformation features across peptide chains in protein sequences. This makes up for the shortcomings of BiLSTM in long sequence dependency modeling and outputs drug sequence features and protein sequence features containing local patterns and long-range dependencies.

[0064] Next, the computer device can input the obtained drug sequence features and protein sequence features into the bidirectional cross-attention interaction module to realize bidirectional association modeling between drug and target sequence features, explore the potential interaction features between the two, and finally generate a sequence feature representation that integrates drug-target sequence association information.

[0065] In one embodiment, a cross-domain drug-target interaction prediction method based on dual-channel feature consistency and course pseudo-labels may further include a process of extracting feature representations through semantic feature channels. Specifically, this process includes: inputting a drug sequence into a pre-trained molecular language model and obtaining a drug semantic feature matrix through semantic encoding; inputting a protein sequence into a pre-trained protein language model and obtaining a protein semantic feature matrix through semantic encoding; performing global average pooling along the sequence dimension on both the drug and protein semantic feature matrices to extract high-level semantic representations of the drug and protein; inputting the high-level semantic representations of the drug and protein into a lightweight multilayer perceptron for linear mapping, and then obtaining the semantic correlation between the drug and protein through a bidirectional cross-attention mechanism to generate a semantic feature representation that integrates drug-target semantic association information.

[0066] Computer equipment can input drug molecules into a pre-trained molecular language model (MolFormer) and protein sequences into a pre-trained protein language model (ESM-2). Utilizing semantic models pre-trained on large-scale molecular and protein sequence data, high-level semantic features are extracted from the drugs and targets to obtain representations with universal biochemical semantic information. These two types of features form drug-target representations from different perspectives, providing a foundation for subsequent consistency assessment. Next, the semantic representations can be input into a lightweight MLP for mapping, and a bidirectional cross-attention mechanism can be used to model drug-protein semantic correlations, resulting in a fused semantic feature representation.

[0067] As another core branch of the dual-channel coding network, the semantic feature channel relies on a large-scale pre-trained biomolecular language model. It takes the standardized drug SMILES sequence and protein amino acid sequence as input representations and performs three core operations in sequence: feature extraction from the pre-trained model, lightweight MLP linear mapping, and bidirectional cross-attention feature interaction. This process is used to mine the biochemical semantic information of the drug and the target and the semantic relationship between them, and finally outputs a semantic feature representation that integrates the semantic relationship information between the drug and the target.

[0068] Specifically, computer equipment can input standardized drug SMILES sequences into the pre-trained molecular language model MolFormer. The MolFormer model is trained on a large-scale molecular dataset and has the ability to semantically represent drug molecular structure, functional group combination, and biochemical properties. The model captures global correlation information in the drug sequence through a self-attention mechanism, automatically extracts high-level molecular semantic features of drug molecules, and outputs a drug semantic feature matrix with fixed dimensions.

[0069] Computer equipment can input standardized protein sequences into the pre-trained protein language model ESM-2. The ESM-2 model is trained based on massive protein sequence data and can accurately represent information such as the spatial conformation, domain features, and biological functions of amino acid sequences. The model mines long-range semantic associations of protein sequences through a deep Transformer architecture, extracts high-level protein semantic features, and outputs a protein semantic feature matrix with fixed dimensions.

[0070] Next, the computer device can perform global average pooling on the drug semantic feature matrix and the protein semantic feature matrix along the sequence dimension to extract the high-level semantic representations of the drug and protein. After inputting them into a lightweight multilayer perceptron (MLP) for linear mapping and feature optimization, a bidirectional cross-attention mechanism is used to realize the bidirectional association model of the semantic features of the drug and the target, and finally generate the semantic feature representation of the drug-target semantic association.

[0071] Step 206: After global average pooling, the sequence feature representation and semantic feature representation are concatenated and fused. The fused features are then nonlinearly mapped using a multilayer perceptron to obtain the drug-target interaction prediction score.

[0072] The computer device incorporates a dual-consistency feature interaction and evaluation mechanism. Drug and target representations from both sequence and semantic feature channels are input into the interaction module, where a cross-perspective attention mechanism models the potential interactions between the drug and target. Based on this, the features from the semantic and sequence branches are globally averaged and then concatenated and fused before being input into the prediction module. The prediction module uses a multilayer perceptron to perform a nonlinear mapping on the fused features, ultimately outputting a drug-target interaction prediction score. The consistency of prediction results obtained from the same sample under different perspectives is evaluated. When the prediction results from the two perspectives are consistent in confidence and class judgment, the prediction result is considered to have high reliability.

[0073] In one embodiment, a cross-domain drug-target interaction prediction method based on dual-channel feature consistency and course pseudo-labels may further include a prediction score generation process. This process includes: performing global average pooling on the sequence feature representation and semantic feature representation respectively to obtain a sequence feature vector and a semantic feature vector; concatenating and fusing the sequence feature vector and semantic feature vector along the feature dimension to obtain a cross-perspective fused feature vector; normalizing the cross-perspective fused feature vector and inputting it into a multilayer perceptron; performing a nonlinear transformation through the multilayer perceptron; and outputting the drug-target interaction prediction score after passing through a Sigmoid activation function.

[0074] The computer device can perform global average pooling along the sequence dimension on the drug sequence features and protein sequence features output from the sequence feature channel to obtain drug sequence feature vectors and protein sequence feature vectors; and perform global average pooling along the sequence dimension on the drug semantic features and protein semantic features output from the semantic feature channel to obtain drug semantic feature vectors and protein semantic feature vectors; then, the drug sequence feature vector, protein sequence feature vector, drug semantic feature vector, and protein semantic feature vector are mapped to the same fusion feature space through a linear projection layer to obtain four feature vectors with unified dimensions; the four feature vectors with unified dimensions are concatenated and fused to form a fusion feature vector; the fusion feature vector is input into a multilayer perceptron, and nonlinear transformation is performed through at least two fully connected layers; finally, the drug-target interaction prediction score located in the (0,1) interval is output through the Sigmoid activation function.

[0075] Step 208: Construct a teacher-student model architecture. For unlabeled samples in the target domain data, generate pseudo-labels based on the drug-target interaction prediction scores. Sort the unlabeled samples from high to low prediction confidence and use a course learning strategy to train the teacher-student model architecture across domains based on the prediction confidence to obtain the trained prediction model.

[0076] In one embodiment, a computer device can construct a teacher model and a student model with consistent architecture; wherein, the student model is a model composed of a dual-channel coding network and feature fusion prediction; the initial parameters of the teacher model are copied from the parameters of the student model, and the teacher model parameters are dynamically updated by the student model parameters through exponential moving average. That is, the teacher model is obtained by updating the student model parameters through EMA, does not participate in gradient backpropagation, and its parameters remain frozen.

[0077] The computer device can generate pseudo-labels for unlabeled samples in the target domain based on the double consistency evaluation mechanism, and sort the samples according to the prediction confidence from high to low. It adopts a course learning strategy, which prioritizes the introduction of high-confidence samples for training, and gradually incorporates more difficult and lower-confidence samples as the training process progresses, thereby effectively reducing the negative impact of noisy pseudo-labels on model training and avoiding the accumulation of confirmation bias.

[0078] In one embodiment, a cross-domain drug-target interaction prediction method based on dual-channel feature consistency and course pseudo-labels may further include a model training process. The specific process includes: inputting unlabeled samples from the target domain data into the teacher model, and fusing them with features through a dual-channel encoding network to obtain sequence-perspective prediction scores and semantic-perspective prediction scores; performing a consistency evaluation on the sequence-perspective prediction scores and semantic-perspective prediction scores; if the consistency evaluation result is reliable, fusing the prediction scores as pseudo-labels and calculating the confidence level; sorting the unlabeled samples with pseudo-labels from high to low confidence levels, and using a course learning strategy, first mixing high-confidence pseudo-label samples with labeled source domain data and inputting them into the student model, calculating the supervised loss to guide the student model's basic learning, and then gradually incorporating low-confidence pseudo-label samples, calculating the pseudo-label loss to complete cross-domain self-training, and obtaining the trained prediction model.

[0079] like Figure 3 As shown, labeled source domain data is input into the student model to calculate supervised loss and guide basic learning; unlabeled target domain data is input into the teacher model, and prediction results are generated from the semantic branch and sequence branch obtained earlier; pseudo-labels are filtered using uncertainty ranking and consistency check, retaining only pseudo-labels with high confidence and consistent bi-branch; pseudo-label loss is gradually introduced according to the course learning strategy, and the student model is trained from easy to difficult, while the teacher model is updated synchronously through EMA to achieve cross-domain self-training.

[0080] During model training, such as Figure 3As shown, unlabeled samples from the target domain can be input into the teacher model. A dual-channel encoding network is used to obtain prediction scores from both the sequence and semantic perspectives. A dual consistency evaluation is performed on these scores, including category consistency and confidence consistency checks. When the predicted categories from both perspectives are the same and the difference in prediction scores is less than a preset threshold, the sample is deemed reliable, and its fused perspective prediction score is used as a pseudo-label. Simultaneously, a confidence score is calculated. Next, all target domain samples that pass the dual consistency evaluation are sorted from highest to lowest confidence score to construct a target domain sample queue. A confidence threshold for the current stage is determined from this queue, and samples with confidence scores higher than the threshold are selected for training in the current round, achieving a learning progression from easy to difficult.

[0081] Next, as Figure 3 As shown, the selected target domain samples and their pseudo-labels can be input into the student model along with the labeled source domain samples. The supervised loss of the source domain and the unsupervised loss of the target domain can be calculated. The parameters of the student model are updated through backpropagation using the joint loss function, and the parameters of the teacher model are updated simultaneously through EMA to complete the model training.

[0082] Step 210: Input the drug-target sample to be predicted into the trained prediction model and output the drug-target interaction prediction result.

[0083] In one embodiment, a cross-domain drug-target interaction prediction method based on dual-channel feature consistency and course pseudo-labels may further include a model evaluation process, which specifically includes: determining model performance evaluation indicators and binarizing the drug-target interaction prediction results to obtain the processed results; and evaluating the trained prediction model based on the processed results using the model performance evaluation indicators to obtain the evaluation results.

[0084] Computer equipment can binarize the drug-target interaction prediction results, with the presence of an interaction being positive and the absence of an interaction being negative, thus constructing a binary classification task.

[0085] After model training, various evaluation metrics were used to assess the performance of the drug-target interaction prediction model. These metrics included Area Under the Curve (AUC), Area Under the Precision-Recall Curve (AUPR), and F1 score. AUC measures the model's overall ability to distinguish between positive and negative samples at different discrimination thresholds, comprehensively reflecting the model's ability to discriminate between drug-target interacting and non-interacting samples. AUPR focuses on the prediction quality of positive samples, particularly suitable for scenarios where the distribution of positive and negative samples is unbalanced in drug-target interaction prediction. The F1 score considers both precision and recall, evaluating the model's ability to balance prediction accuracy and coverage. The calculation formulas are as follows: ; where TP represents true positive, TN represents true negative, FP represents false positive, and FN represents false negative.

[0086] Finally, the prediction results or pseudo-labels can be visualized to analyze the reconstruction effect of the drug-target interaction network.

[0087] In one embodiment, a cross-domain drug-target interaction prediction method based on dual-channel feature consistency and course pseudo-labels is provided, such as... Figure 4 As shown, the process can be divided into three stages: semantic branching encoding, sequence branching encoding, and feature fusion and prediction. The semantic branching encoding stage includes: inputting drug molecules into a pre-trained molecular language model MolFormer for semantic encoding to obtain high-level semantic features of the drug molecules; inputting protein sequences into a pre-trained protein language model ESM-2 for semantic encoding to obtain high-level semantic features of the proteins; the high-level semantic features of the drug and protein are respectively linearly mapped, and then the semantic correlation between the two is modeled through a cross-attention mechanism to generate semantic feature representations at the semantic level. The sequence branching encoding stage includes: inputting drug sequence embeddings and protein sequence embeddings into a bidirectional long short-term memory network (BiLSTM) to capture local contextual information of the sequences; the output of BiLSTM is then input into a Mamba module to strengthen the modeling of long-range dependencies; the sequence features of the drug and protein are modeled through a cross-attention mechanism to generate sequence feature representations at the sequence level. The feature fusion and prediction stage includes: the interaction features of the semantic branch and the sequence branch are respectively subjected to global average pooling to obtain fixed-dimensional feature vectors; the pooled features of the two branches are added together and fused into a unified multi-view feature representation; the fused features are input into a fully connected layer, and finally output as a drug-target interaction prediction score. At the same time, the semantic branch and the sequence branch can also independently output their respective prediction scores.

[0088] This application constructs a dual-channel encoding structure combining semantic feature branches and sequence feature branches to achieve multi-perspective joint modeling of drug and protein information. This enables the model to simultaneously possess high-level semantic prior expression capabilities and task-related sequence pattern modeling capabilities, thereby improving the completeness and discriminativeness of feature expression. Simultaneously, this application proposes a dual-consistency course pseudo-label learning strategy. This strategy uses a multi-perspective consistency verification mechanism to screen unlabeled samples and gradually introduces target domain data into training according to a learning approach from easy to difficult. This effectively reduces interference from noisy pseudo-labels, suppresses confirmation bias problems during self-training, and improves the stability and reliability of the model training process.

[0089] To verify the impact of each module on the overall performance of this application, ablation experiments were conducted on the model and tested on a dataset. AUC, AUPR, and F1 were used as evaluation metrics. The experimental results are shown in the table below:

[0090]

[0091] Experimental results show that the complete model outperforms the comparative model by removing any module across all evaluation metrics, demonstrating that the proposed dual-channel coding structure and dual-consistency course pseudo-label strategy can synergistically improve the model's prediction accuracy, generalization ability, and robustness. Therefore, this application can achieve stable, efficient, and reliable prediction results in cross-domain drug-target interaction prediction tasks, providing an effective technical means for drug discovery and biomolecular interaction research, and possesses promising application prospects and widespread value.

[0092] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0093] In one embodiment, such as Figure 5 As shown, a cross-domain drug-target interaction prediction system based on dual-channel feature consistency and course pseudo-labels is provided, including: a data processing module 510, a feature extraction module 520, a feature fusion and mapping module 530, a model training module 540, and a prediction module 550, wherein:

[0094] The data processing module 510 is used to acquire source domain data containing known drug-target interaction tags and unlabeled target domain data, and to standardize the drug molecules and target proteins in the source domain data and target domain data respectively to obtain a unified input representation.

[0095] The feature extraction module 520 is used to construct a dual-channel encoding network with different inductive biases. The sequence feature channel in the dual-channel encoding network encodes the drug sequence and protein sequence in the input representation to obtain the sequence feature representation. The semantic feature channel in the dual-channel encoding network is used to extract high-level semantic features from the drug sequence and protein sequence to obtain the semantic feature representation.

[0096] The feature fusion and mapping module 530 is used to perform global average pooling on sequence feature representation and semantic feature representation and then splice and fuse them. The fused features are then nonlinearly mapped by a multilayer perceptron to obtain the drug-target interaction prediction score.

[0097] The model training module 540 is used to build a teacher-student model architecture. For unlabeled samples in the target domain data, pseudo-labels are generated based on the drug-target interaction prediction scores. The unlabeled samples are sorted from high to low prediction confidence. A course learning strategy is adopted to train the teacher-student model architecture across domains based on the prediction confidence to obtain the trained prediction model.

[0098] The prediction module 550 is used to input the drug-target sample to be predicted into the trained prediction model and output the drug-target interaction prediction result.

[0099] In one embodiment, the data processing module 510 is further configured to perform structural standardization cleaning on the drug molecule and convert it into a SMILES sequence, perform length adjustment and character encoding on the SMILES sequence to obtain the drug sequence; perform cleaning on the target protein, remove non-standard amino acid characters from the target protein, and perform length unification and character encoding on the standard amino acids to obtain the protein sequence; and perform numerical normalization and embedding mapping on the drug sequence and the protein sequence respectively to obtain a unified input representation.

[0100] In one embodiment, the feature extraction module 520 is further configured to vectorize the drug sequence and protein sequence in the input representation, and then input them sequentially into a bidirectional long short-term memory network and a Mamba module to capture the local structure and contextual dependency information of the drug sequence and protein sequence, respectively, to obtain drug sequence features and protein sequence features; and to perform interactive modeling of the drug sequence features and protein sequence features through a bidirectional cross-attention mechanism to generate a sequence feature representation that integrates drug-target sequence association information.

[0101] In one embodiment, the feature extraction module 520 is further configured to input the drug sequence into a pre-trained molecular language model and obtain a drug semantic feature matrix through semantic encoding; input the protein sequence into a pre-trained protein language model and obtain a protein semantic feature matrix through semantic encoding; perform global average pooling on the drug semantic feature matrix and the protein semantic feature matrix along the sequence dimension to extract high-level semantic representations of the drug and protein; input the high-level semantic representations of the drug and protein into a lightweight multilayer perceptron for linear mapping, and obtain the semantic correlation between the drug and protein through a bidirectional cross-attention mechanism to generate a semantic feature representation that integrates the semantic association information between the drug and the target.

[0102] In one embodiment, the feature fusion and mapping module 530 is further configured to perform global average pooling operations on the sequence feature representation and the semantic feature representation respectively to obtain the sequence feature vector and the semantic feature vector; to concatenate and fuse the sequence feature vector and the semantic feature vector in the feature dimension to obtain the cross-view fused feature vector; to normalize the cross-view fused feature vector and input it into the multilayer perceptron; to perform nonlinear transformation through the multilayer perceptron and output the drug-target interaction prediction score through the Sigmoid activation function.

[0103] In one embodiment, the model training module 540 is further used to construct a teacher model and a student model with consistent architecture; wherein the student model is a model composed of a dual-channel coding network and feature fusion prediction; the initial parameters of the teacher model are copied from the parameters of the student model, and the parameters of the teacher model are dynamically updated by the parameters of the student model through an exponential moving average.

[0104] In one embodiment, the model training module 540 is further configured to input unlabeled samples from the target domain data into the teacher model, and fused them with features through a dual-channel encoding network to obtain sequence-perspective prediction scores and semantic-perspective prediction scores; perform consistency evaluation on the sequence-perspective prediction scores and semantic-perspective prediction scores; if the consistency evaluation result is reliable, fuse the prediction scores as pseudo-labels and calculate the confidence level; sort the unlabeled samples with pseudo-labels from high to low confidence level, and adopt a course learning strategy to first mix high-confidence pseudo-label samples with labeled source domain data and input them into the student model, calculate the supervised loss to guide the student model's basic learning, and then gradually incorporate low-confidence pseudo-label samples, calculate the pseudo-label loss to complete cross-domain self-training, and obtain the trained prediction model.

[0105] In one embodiment, the system further includes a model evaluation module for determining model performance evaluation metrics and binarizing the drug-target interaction prediction results to obtain processing results; based on the processing results, the trained prediction model is evaluated using the model performance evaluation metrics to obtain evaluation results.

[0106] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 6 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a cross-domain drug-target interaction prediction method based on dual-channel feature consistency and curriculum pseudo-labels. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device casing, or an external keyboard, touchpad, or mouse.

[0107] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0108] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of a cross-domain drug-target interaction prediction method based on dual-channel feature consistency and course pseudo-labels.

[0109] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program being executed by a processor to implement the steps of a cross-domain drug-target interaction prediction method based on dual-channel feature consistency and course pseudo-labels.

[0110] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0111] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0112] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for predicting cross-domain drug-target interactions based on dual-channel feature consistency and course pseudo-labels, characterized in that, The method includes: Obtain source domain data containing known drug-target interaction tags and unlabeled target domain data, and standardize the drug molecules and target proteins in the source domain data and target domain data respectively to obtain a unified input representation; A dual-channel encoding network with different inductive biases is constructed. The drug sequence and protein sequence in the input representation are encoded through the sequence feature channel of the dual-channel encoding network to obtain the sequence feature representation. Then, the high-level semantic features of the drug sequence and protein sequence are extracted through the semantic feature channel of the dual-channel encoding network to obtain the semantic feature representation. The sequence feature representation and semantic feature representation are concatenated and fused after global average pooling. The fused features are then nonlinearly mapped by a multilayer perceptron to obtain the drug-target interaction prediction score. A teacher-student model architecture is constructed. For unlabeled samples in the target domain data, pseudo-labels are generated based on the drug-target interaction prediction scores. The unlabeled samples are sorted from high to low prediction confidence. A course learning strategy is adopted to train the teacher-student model architecture across domains based on the prediction confidence to obtain the trained prediction model. The drug-target sample to be predicted is input into the trained prediction model, and the drug-target interaction prediction result is output.

2. The method for predicting cross-domain drug-target interactions based on dual-channel feature consistency and course pseudo-labels according to claim 1, characterized in that, The drug molecules and target proteins in the source domain data and target domain data are standardized, including: The drug molecule is structurally normalized and cleaned to convert it into a SMILES sequence. The SMILES sequence is then length-adjusted and character-encoded to obtain the drug sequence. The target protein is cleaned to remove non-standard amino acid characters, and the standard amino acids are standardized in length and encoded to obtain the protein sequence. The drug sequence and protein sequence are respectively subjected to numerical normalization and embedding mapping to obtain a unified input representation.

3. The method for predicting cross-domain drug-target interactions based on dual-channel feature consistency and course pseudo-labels according to claim 1, characterized in that, The drug sequence and protein sequence in the input representation are encoded through sequence feature channels to obtain a sequence feature representation, including: After the drug sequence and protein sequence in the input representation are vectorized, they are sequentially input into a bidirectional long short-term memory network and a Mamba module to capture the local structure and contextual dependency information of the drug sequence and protein sequence, respectively, to obtain drug sequence features and protein sequence features. The drug sequence features and protein sequence features are interactively modeled using a bidirectional cross-attention mechanism to generate a sequence feature representation that integrates drug-target sequence association information.

4. The method for predicting cross-domain drug-target interactions based on dual-channel feature consistency and course pseudo-labels according to claim 1, characterized in that, High-level semantic features are extracted from the drug sequence and protein sequence through semantic feature channels to obtain semantic feature representations, including: The drug sequence is input into a pre-trained molecular language model, and a drug semantic feature matrix is ​​obtained through semantic encoding; the protein sequence is input into a pre-trained protein language model, and a protein semantic feature matrix is ​​obtained through semantic encoding. Global average pooling is performed along the sequence dimension on the drug semantic feature matrix and the protein semantic feature matrix respectively to extract the high-level semantic representations of the drug and protein; After inputting the high-level semantic representations of drugs and proteins into a lightweight multilayer perceptron for linear mapping, the semantic correlation between drugs and proteins is obtained through a bidirectional cross-attention mechanism, generating a semantic feature representation that integrates the semantic association information between drugs and targets.

5. The method for predicting cross-domain drug-target interactions based on dual-channel feature consistency and course pseudo-labels according to claim 1, characterized in that, The sequence feature representation and semantic feature representation are concatenated and fused after global average pooling. A multilayer perceptron is then used to perform a nonlinear mapping on the fused features to obtain a drug-target interaction prediction score, including: Global average pooling is performed on the sequence feature representation and semantic feature representation respectively to obtain the sequence feature vector and the semantic feature vector. The sequence feature vector and the semantic feature vector are concatenated and fused along the feature dimension to obtain a cross-view fused feature vector; After normalizing the cross-view fusion feature vector, it is input into a multilayer perceptron. After nonlinear transformation by the multilayer perceptron, the drug-target interaction prediction score is output after passing through the Sigmoid activation function.

6. The method for predicting cross-domain drug-target interactions based on dual-channel feature consistency and course pseudo-labels according to claim 1, characterized in that, Constructing a teacher-student model architecture includes: Construct a teacher model and a student model with consistent architecture; wherein the student model is a model composed of the dual-channel coding network and feature fusion prediction; the initial parameters of the teacher model are copied from the parameters of the student model, and the parameters of the teacher model are dynamically updated by the parameters of the student model through exponential moving average.

7. The method for predicting cross-domain drug-target interactions based on dual-channel feature consistency and course pseudo-labels according to claim 6, characterized in that, For unlabeled samples in the target domain data, pseudo-labels are generated based on the drug-target interaction prediction scores. The unlabeled samples are then sorted from high to low prediction confidence. A course-learning strategy is employed to train the teacher-student model architecture across domains based on the prediction confidence, resulting in a trained prediction model, including: The unlabeled samples in the target domain data are input into the teacher model, and the dual-channel coding network is used to fuse and predict with features to obtain the sequence perspective prediction score and the semantic perspective prediction score. The sequence perspective prediction score and the semantic perspective prediction score are evaluated for consistency. If the consistency evaluation result is reliable, the prediction score is fused as a pseudo-label, and the confidence level is calculated. The unlabeled samples with pseudo-labels are sorted from high to low confidence. Using a course learning strategy, high-confidence pseudo-labeled samples are first mixed with labeled source domain data and input into the student model. The supervised loss is calculated to guide the basic learning of the student model. Then, low-confidence pseudo-labeled samples are gradually included, and the pseudo-label loss is calculated to complete cross-domain self-training, resulting in the trained prediction model.

8. The method for predicting cross-domain drug-target interactions based on dual-channel feature consistency and course pseudo-labels according to claim 1, characterized in that, The method further includes: The model performance evaluation index was determined, and the drug-target interaction prediction results were binarized to obtain the processed results. Based on the processing results, the trained prediction model is evaluated using the model performance evaluation metrics to obtain the evaluation results.

9. A cross-domain drug-target interaction prediction system based on dual-channel feature consistency and course pseudo-labeling, characterized in that, The system includes: The data processing module is used to acquire source domain data containing known drug-target interaction tags and unlabeled target domain data, and to standardize the drug molecules and target proteins in the source domain data and target domain data respectively to obtain a unified input representation. The feature extraction module is used to construct a dual-channel encoding network with different inductive biases. The sequence feature channel in the dual-channel encoding network encodes the drug sequence and protein sequence in the input representation to obtain a sequence feature representation. The semantic feature channel in the dual-channel encoding network is used to extract high-level semantic features from the drug sequence and protein sequence to obtain a semantic feature representation. The feature fusion and mapping module is used to perform global average pooling on the sequence feature representation and semantic feature representation and then concatenate and fuse them. The fused features are then nonlinearly mapped by a multilayer perceptron to obtain the drug-target interaction prediction score. The model training module is used to construct a teacher-student model architecture. For unlabeled samples in the target domain data, pseudo-labels are generated based on the drug-target interaction prediction scores. The unlabeled samples are sorted from high to low prediction confidence. A course learning strategy is adopted to train the teacher-student model architecture across domains based on the prediction confidence to obtain the trained prediction model. The prediction module is used to input the drug-target sample to be predicted into the trained prediction model and output the drug-target interaction prediction result.

Images