DTI Prediction Methods Based on LLM and RBMO

Abstract

This invention belongs to the field of artificial intelligence algorithm application—drug target interaction prediction, and involves a DTI prediction method based on LLM and RBMO. First, through data collection, integration, and redundancy removal, a batch of drug target sequences with significant positive-negative sample imbalance was obtained as input data. Next, the K-BERT model was used to extract deep features representing drug molecule SMILES, and the ProstT5 model was used to analyze the high-dimensional information of the protein sequences. Then, a feature selection method combining RBMO and KNN was introduced to intelligently screen protein features, ultimately selecting 513 key features with the strongest biological discriminative power. Finally, a multi-core feature fusion mechanism was combined with the emerging Kolmogorov–Arnold network (KAN) to enhance the model's ability to analyze and represent complex biomolecular features.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence algorithm application - drug target interaction prediction, and involves a DTI prediction method based on LLM and RBMO. Background Technology

[0002] Drug-target interaction (DTI) describes the physical binding process between drug molecules and target macromolecules through non-covalent bonds (such as hydrogen bonds, van der Waals forces, and electrostatic interactions) or covalent bonds. This interaction is the molecular basis for the action of all drugs. A successful DTI can activate or inhibit the biological function of the target, thereby correcting pathological states and producing therapeutic effects. Therefore, identifying and validating new DTIs is the starting point and core link in innovative drug discovery. It not only reveals the mechanism of drug action but also provides a fundamental basis for understanding the efficacy, selectivity, and even potential side effects (off-target effects) of drugs. Traditionally, the discovery of DTIs has mainly relied on experimental methods, such as high-throughput screening (HTS). Although these wet experimental techniques are the gold standard for DTI validation, their inherent high cost, long cycle, and low success rate have become major bottlenecks in new drug development. Statistics show that it takes an average of more than ten years and billions of dollars to bring a new drug from concept to market. In addition, the chemical and target spaces that experimental methods can explore are relatively limited. These challenges have greatly driven computational methods.

[0003] In recent years, significant progress has been made in predicting drug-target interactions. Although numerous computational tools have been developed for this purpose, providing powerful tools for research on drug-target interaction prediction, most models still face the following insurmountable challenges: Incomplete feature extraction: Most current model feature extraction methods fail to capture global information and ignore the 3D spatial structure that determines molecular interactions. This results in incomplete feature information, preventing the model from learning the true interaction relationships.

[0004] Limited Feature Selection Methods: Current feature selection methods are mostly basic methods such as PCA and Lasso. These methods have poor generalization and cannot effectively identify noise and important data in drug and target features.

[0005] Imbalanced data: In drug target interaction prediction methods, the distribution of positive and negative samples is extremely unbalanced. This imbalance can cause bias during model training, leading to distorted evaluation metrics, shifted decision boundaries, and ultimately reduced model stability and prediction accuracy.

[0006] Model interpretability: In the field of biological function prediction, while model accuracy is crucial, model interpretability is equally important. An easily interpretable model clearly demonstrates its predictive logic, which not only improves the model's transparency but also enhances its credibility. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a DTI prediction method based on LLM and RBMO.

[0008] The technical solution of the present invention is as follows: The dataset used in this study comes from BindingDB, an authoritative database resource dedicated to collecting, organizing, and publishing experimental data on the binding affinity of small molecules to protein targets. Given the severe imbalance in the ratio of positive to negative samples in the original dataset (exceeding 1:5 in the independent test set) and the existence of data redundancy, this invention systematically optimized and reconstructed the dataset. Using this dataset, this invention proposes an interpretable drug-target interaction prediction method based on a fusion of K-BERT and ProstT5 features, using RBMO combined with KNN for target feature selection, and combining multi-kernel feature fusion with KAN. This method extracts sequence features of drugs and targets based on K-BERT and ProstT5 models respectively, utilizes a feature selection method based on RBMO, employs a feature fusion method based on a multi-kernel feature fusion module, uses a KAN-based classification model, and incorporates SHAP for interpretable analysis.

[0009] The DTI prediction method based on LLM and RBMO has the following steps: Step 1: Data Preprocessing

[0010] This phase includes dividing the training and testing data, and using the large language model K-BERT and ProstT5 to extract features from the drug and target sequences, respectively.

[0011] 1.1 Dataset Processing Methods For the dataset built on the BindingDB database, there are a total of 9166 positive samples and 23427 negative samples. After removing duplicate sequences, there are 6588 positive samples and 19759 negative samples. 80% of the dataset is used as training samples and 20% is used as test samples to train the prediction model. 1.2 Feature extraction of drug sequences using the K-BERT large language model Each drug sequence in SMILES format is sequentially fed into the K-BERT large model. Features are extracted from all sequences, and each sequence is transformed into a 768-dimensional vector as a feature. 1.3 Feature extraction of target sequences using the ProstT5 large language model Each target sequence in FASTA format is sequentially fed into the ProstT5 large model. Features are extracted from all sequences, and each sequence is transformed into a 1024-dimensional vector as a feature. Step 2: Feature selection of target features

[0012] The feature vectors generated by the ProtT5 target feature extraction method have a dimension as high as 1024. Such a high-dimensional feature space may cause the "curse of dimensionality" problem. This invention uses the Red-billed BlueMagpie Optimizer (RBMO) method to reduce the dimensionality and select target features.

[0013] 2.1 The "Curse of Dimensions" Problem The "curse of dimensionality" is a series of problems that arise when models process high-dimensional data. These problems negatively impact model learning, specifically as follows: 1. Decreased generalization ability: As the dimensionality increases, the volume of data grows exponentially, resulting in extremely sparse data in the limited training samples.

[0014] 2. Distance metric failure: In high-dimensional space, the distance between any two points (such as Euclidean distance) will become very similar and lose its distinguishability.

[0015] 3. Low computational complexity and efficiency: The computational complexity of the algorithm increases exponentially with the feature dimension.

[0016] 4. Increased redundancy and noise: Many high-dimensional features are often unrelated, or even redundant or noisy.

[0017] Therefore, feature selection of high-dimensional data is necessary before model construction. This invention utilizes the RBMO method, based on the K-nearest neighbor metric, to perform feature selection on 1024-dimensional target data.

[0018] 2.2 Red-billed Blue Magpie Optimizer The Red-billed Magpie Optimization (RBMO) is a novel swarm intelligence optimization algorithm based on the collaborative foraging and vigilance behaviors of red-billed blue magpies. Its core idea is to achieve a dynamic balance between global exploration and local exploitation during the search process by simulating the leader-follower mechanism and reconnaissance behavior among individuals.

[0019] In RBMO, the position of the i-th candidate solution (the red-billed blue magpie individual) can be represented as: (1) in For the number of iterations, For feature dimensions.

[0020] The first phase will be the reconnaissance (global exploration) phase. The red-billed blue magpie will expand its search area through reconnaissance activities, and the individual locations will be updated as follows: (2) in For random coefficients, As the current globally optimal individual, The individuals are randomly selected. This mechanism allows the algorithm to maintain strong global search capabilities.

[0021] After finding food, the red-billed blue magpie enters a cooperative foraging (local development) phase. As it approaches the optimal solution area, the red-billed blue magpie uses cooperative foraging to complete a more refined search. (3) in It serves as a local perturbation factor, used to enhance local exploration.

[0022] 2.3 Decision-making methods for feature selection The optimization objective of RBMO is to minimize a carefully designed objective function that evaluates the "quality" of each candidate feature subset. An ideal feature subset should enable the prediction model to achieve the highest possible performance while containing the fewest possible features. To this end, the objective function combines two mutually constraining metrics: classification error rate and feature subset size. The objective function F(X) is defined as shown in Equation (4): (4) in: It is composed of individual position vectors The feature subset obtained from decoding. It uses feature subsets The average classification error rate is obtained by training a K-nearest neighbor classifier and performing k-fold cross-validation on the validation set. It is a subset The number of features. It is the total number of original features. It is a weighting coefficient used to balance the goals of classification performance and model simplicity. A higher weighting coefficient results in a lower weighting. A higher value indicates a greater focus on improving classification accuracy, while a lower value indicates a greater focus on improving classification accuracy. Values ​​tend to select fewer features.

[0023] Step 3: Multi-core feature fusion To effectively extract and integrate information from the high-dimensional features resulting from the merging of drug and target data, this invention employs a Multi-Kernel Feature Fusion (MKFF) module. MKFF significantly enhances the quality of feature representation through parallel multi-scale feature extraction and a global-local dynamic fusion mechanism. The specific process is as follows: 3.1 Multi-scale feature extraction After the feature data is input into this module, the Multi-Kernel Feature Extraction Function (MDF) captures comprehensive contextual information from different scales and regions. This module has four parallel paths to process the input embedding vector. The representation of each path is shown in equations (5)-(8): (5) (6) (7) (8) in It is a 1x1 pointwise convolution kernel used for feature transformation and information integration; It consists of a 1x1 convolution to reduce the number of channels and reduce model parameters and complexity, and a 3x3 convolution kernel to extract local region features; and Similarly, but using a 5x5 convolutional kernel to capture a larger range of receptive field information; First, a 3x3 max pooling layer is used to capture the most salient local features, followed by a 1x1 convolution for channel transformation. These four paths output feature maps at four different scales (k1, k2, k3, k4), which together constitute a multi-view, multi-granular observation of the input features.

[0024] 3.2 Global-Local Dynamic Fusion The data obtained afterward, containing multi-view features, enters the Global-Local Dynamic Fusion (GLDF) module. Simply concatenating multi-kernel features fails to consider the differences and importance between different feature maps. The GLDF module, however, intelligently fuses these multi-kernel features in a dynamic, attention-based manner, preserving broad contextual information and fine local details. MKFF comprises four parts: multi-scale convolutional extraction, global channel dynamic fusion, local spatial dynamic fusion, and residual enhancement. The process is as follows: (1) Multi-scale convolution extraction First, four sets of multi-scale representations are extracted from the input sequence, including convolution and pooling operations at different receptive fields. These multi-scale features are then concatenated and compressed using a 1×1 convolution to obtain the basic fused features. (9) Where k1, k2, k3, and k4 are features after convolution or pooling at different scales.

[0025] (2) Dynamic fusion of global channels To extract the overall global dependencies, we first need to... Perform average pooling: (10) Then based on the number of channels Calculate the adaptive convolution kernel size: (11) in This indicates taking the closest odd number to ensure the center alignment of the convolution kernel.

[0026] The convolutional kernel is then used to perform channel transformation on the global pooling result, and channel attention weights are obtained through the Sigmoid function, ultimately yielding the globally enhanced features. The formula is shown in equation (12).

[0027] (12) (3) Local spatial dynamic fusion To further enhance the expressive power of local structures, we first... An adaptive convolution is applied to obtain spatially compressed features, followed by another convolution to obtain local spatial attention weights, which are then normalized using Softmax in the spatial dimension. The local enhanced features are represented as follows: (13) in This is the activation layer, used to generate attention weights. This is a compression layer used to compress data.

[0028] (4) Residual enhancement After concatenating the global attention enhancement features and the local attention enhancement features, a 1×1 convolution is used for further fusion. To maintain the continuity of information in the input features and enhance the stability of the model during deep propagation, MKFF employs residual connections, adding the initial input features to the fused features. Finally, average pooling is applied to the residual results to obtain the final output. : (14) (15) Step 4: Building a KAN-based classification model In this phase, this research aims to construct a robust and effective classification model that captures sequence dependency structures using the KAN method. This phase mainly includes key steps such as loss function definition, model architecture design, and model training and validation. The specific steps are as follows: 4.1 Loss Function Selection Loss functions play a crucial role in machine learning and deep learning. In this study, the loss function quantifies the error magnitude between the model's predicted sample labels and the actual labels. This invention employs a loss function, BCEWithLogitsLoss, which combines the sigmoid function and binary cross-entropy loss, to calculate the model's loss for each sample label prediction. The specific calculation formula is as follows: (16) in Indicates the number of samples. Indicates the first The true label of each sample The model represents the first The predicted probability of a sample, i.e., the value output by the model. This is the Sigmoid function. Compared to loss functions such as mean squared error, the binary cross-entropy loss penalizes misclassification more severely and is better able to guide the model to learn the correct classification decision.

[0029] 4.2 Model Architecture Design To ensure the model fully considers the dependency structure of sequences, this invention employs a Multilayer Perceptron (MLP) model composed of KANLinear-like fully connected layers. Its architecture is designed as follows: 1. Function: The KAN classification model is responsible for integrating the features extracted by the previous modules and making the final classification decision.

[0030] 2. Design: Fully connected layers flatten the features and apply linear transformations, then introduce non-linearity through activation functions to perform complex classification tasks.

[0031] 3. Parameters: First layer: KANLinear(mkff_channels * pooled_length, 128, wavelet_type='dog') Second layer: KANLinear(128, 64, wavelet_type='dog') Third layer: KANLinear(64, 1, wavelet_type='dog') 4. Output: The final fully connected layer outputs a one-dimensional vector, corresponding to the probability prediction. The Sigmoid activation function is used to ensure that the output value is between 0 and 1.

[0032] 4.3 Model Training and Validation Testing A sample outputs a (1, 1) matrix through the model, with values ​​between 0 and 1, representing the probability that the sample contains the corresponding label. For example, (0.9) represents the predicted probability of drug target interaction. The loss function used to train the classifier is set to cross-entropy loss, suitable for classification problems, to measure the difference between the target and the predicted output. This model can fully consider the dependency structure of sequences. In the model training experiment of this invention, the Adam optimizer is used with a learning rate lr of 0.0025. Based on five-fold cross-validation and seven evaluation metrics. Select the optimal model. Test the model's performance on an independent test set and obtain evaluation metrics for the independent test results.

[0033] Step 5: Interpretable Analysis To gain a deeper understanding of how the model works during the prediction process and how predicted values ​​are generated, the SHAP method was used to explain the model's prediction process. The following is an introduction to this method: SHAP is a game theory-based model interpretation technique. Its core objective is to evaluate the influence of each feature on the prediction result and assign a corresponding interpretability score to each feature. The calculation of SHAP values ​​is based on Shapley value theory, a fair allocation scheme that reasonably attributes the model's final output (e.g., predicted value) to the contribution of each feature. By considering all possible feature combinations, Shapley values ​​can accurately assess the specific impact of each individual feature on the model's prediction. In interpreting deep learning models, Shapley values ​​quantify the contribution of each input feature to the model's predictive performance. A higher absolute value of the Shapley value indicates that the corresponding feature has a greater impact on the prediction result. By applying SHAP analysis, we can gain insights into the underlying causes of the prediction results. In this invention, an in-depth interpretive analysis was performed on the extracted 20-dimensional features, providing a clear explanation of the model's prediction results.

[0034] The beneficial effects of this invention are: 1. Unlike many traditional methods, this invention realizes an interpretable, efficient prediction framework that integrates a deep language model with optimized feature selection, which can capture deeper interaction patterns between drugs and targets.

[0035] 2. This invention employs the K-BERT model to extract deep features from the SMILES characterization of drug molecules and utilizes the ProstT5 model to analyze the high-dimensional information of protein sequences. By fully combining knowledge-enhanced molecular semantics with deep context-driven protein characterization, it achieves superior discriminative ability and generalization performance in drug-target prediction tasks.

[0036] 3. This invention innovatively combines RBMO with K-Nearest Neighbor, and through an intelligent optimization-driven feature selection mechanism, it not only overcomes the shortcomings of traditional methods that are prone to getting trapped in local optima, but also significantly enhances the model's discrimination ability and generalization performance in high-dimensional and complex drug-target data, providing a solid foundation for the performance improvement of subsequent prediction modules.

[0037] 4. This invention achieves adaptive integration of cross-scale information and effective suppression of feature redundancy by fusing the MKFF multi-scale feature fusion model and the novel KAN framework, enabling the model to learn the nonlinear relationship between drugs and targets under multi-level representation.

[0038] 5. By comparing prediction results on both five-fold cross-validation and independent test sets, this invention demonstrates significantly superior performance compared to other methods.

[0039] 6. The interpretability analysis of the model based on SHAP values ​​shows that the prediction model constructed in this invention is effective, enhancing the transparency and credibility of the model. Attached Figure Description

[0040] Figure 1 This is the overall framework diagram of the present invention.

[0041] Figure 2 This is the SHAP feature interpretation diagram of the present invention, which is a feature importance ranking diagram for the prediction result of a certain sample.

[0042] Figure 3 This is a comparison of the metrics of the model proposed in this invention with previous models on an independent test set. Detailed Implementation

[0043] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0044] This invention implements a DTI prediction method based on LLM and RBMO, the architecture of which is as follows: Figure 1As shown. Step 1: Data preprocessing. This stage includes dividing the training and testing data and extracting features from drug and target sequences using the large language models K-BERT and ProstT5. Step 2: Feature selection using a combination of RBMO and KNN to remove noise and select important information, making it easier for the model to acquire correct feature information. Step 3: Multi-kernel feature extraction to extract features from the processed feature data at multiple scales and perform global-local dynamic fusion, obtaining rich and complementary feature information extracted from multiple different receptive fields, improving data quality and the model's generalization ability. Step 4: Learning and testing the results using a multilayer perceptron composed of KAN linear layers. Then, using the divided test set, the accuracy, precision, and recall are compared. The model with the highest specificity, AUC (area under the receiver operating characteristic curve), and PR (area under the precision-recall curve) is pre-selected, and then validated using a pre-defined independent test set. Step 5: Interpretability Analysis. To improve the interpretability of the model, this invention uses the SHAP method to analyze the model's prediction process, revealing the contribution of different features to the model's prediction results. Finally, to facilitate users' use of the drug target interaction prediction model constructed in this invention, a publicly available prediction tool has been developed.

[0045] Please see the following implementation example for details: Example 1: Data Preprocessing and Feature Extraction We first performed thorough preprocessing on the raw data, including redundancy removal and data standardization. This ensures the quality and consistency of the input data, providing a reliable foundation for subsequent model training. Next, we fed the processed drug and target sequences one by one into the pre-trained K-BERT and ProstT5 models. Each drug sequence received a corresponding 768-dimensional vector, and each target sequence received a 1024-dimensional vector. A total of 6588 positive samples and 19759 negative samples were used for downstream tasks.

[0046] Example 2: RBMO Feature Selection This invention combines RBMO and K-Nearest Neighbor as fitness evaluation functions to perform feature selection on high-dimensional target feature vectors, ultimately selecting 513 features with the highest biological discriminative power. To verify the effectiveness of the RBMO method, a comprehensive performance evaluation was conducted on RBMO with several classic feature selection techniques, including Principal Component Analysis (PCA), Lasso Regression, SHAP, Chi-squared test, and F-test.

[0047] The comparison results are shown in Table 1. Compared with traditional feature selection methods, RBMO achieved the highest Accuracy (0.8491), AUC (0.9132), and PR (0.7903), demonstrating its significant advantages in preserving key discriminative features and suppressing noise redundancy. In particular, it achieved a better balance between Recall (0.8258) and Specificity (0.8568), indicating that this method can effectively capture positive class sample features (potentially interacting drug-target pairs) while avoiding negative class misclassification, ensuring the robustness and reliability of the model prediction. In contrast, traditional statistical or linear methods, such as PCA, Lasso, F-test, and Chi-squared test, while having certain advantages in local accuracy or interpretability, generally lack global search capabilities and are difficult to adapt to the complex nonlinear relationships between drug-protein features. RBMO, by simulating the social cooperation and dynamic search behavior of the red-billed blue magpie during foraging, achieves an adaptive balance between global and local search in the feature space. By combining the classification feedback mechanism of KNN, RBMO can continuously adjust the candidate feature set and improve the fitness function value, thereby selecting a more representative optimal feature subset.

[0048] Table 1. Results of five-fold cross-validation under different feature selection methods

[0049] Example 3: Sampling Strategy Selection Because the number of positive and negative samples in the re-partitioned data is highly imbalanced, the strategy for balancing the impact of positive and negative samples on the model is crucial. This invention selects a strategy to balance the data according to sample weights. To verify the effectiveness of this strategy, a comprehensive evaluation was conducted on this sampling strategy and several classic undersampling methods, including ClusterCentroids, AllKNN, Instance Hardness Threshold, Edited Nearest Neighbours, Neighbourhood Cleaning Rule, Repeated Edited Nearest Neighbours, and the NearMiss method.

[0050] The comparison results are shown in Table 2. Overall, the different sampling strategies show significant differences in model performance. The Sample Weight strategy performs best across all metrics, demonstrating strong overall discriminative ability and stability with good Accuracy (0.848), AUC (0.9198), and F1-Score (0.7355). Its Recall (0.8448) and Specificity (0.8491) maintain a good balance, indicating that the model can identify minority class samples without sacrificing accuracy for the majority class, effectively avoiding the bias and overfitting problems common in traditional oversampling methods. Precision (0.6517) and PR (0.7913) further confirm the balanced performance of this strategy between positive and negative classes, demonstrating high classification accuracy and generalization ability. In contrast, while Cluster Centroids performed exceptionally well in Precision (0.8392) and Recall (0.8553), its Accuracy (0.8191) and AUC (0.9005) were slightly lower due to the forced equalization of sample ratios, indicating that oversampling may have resulted in the loss of some feature information. The Neighbourhood Cleaning Rule followed closely in Accuracy (0.7110) and AUC (0.7719), but its low F1-Score (0.7015) reflected the continued presence of noise interference in its boundary sample cleaning process. Other methods, such as ALLKNN and Edited Nearest Neighbours, failed to effectively improve model performance, with AUCs all below 0.60, indicating limited generalization ability under extremely imbalanced data.

[0051] In summary, Sample Weight, by dynamically adjusting sample weights, enables the model to automatically strengthen the learning of the minority class during training while maintaining the integrity of the data structure, effectively improving the model's robustness and discriminative power in imbalanced scenarios. This method demonstrates significant advantages in complex biological data tasks and provides an efficient and scalable sampling optimization approach for future research.

[0052] Table 2. Results of five-fold cross-validation under different sampling methods

[0053] Example 4: Model Performance Evaluation and Comparison To enhance the professionalism of this invention, standardized and professional evaluation metrics were used to assess the model's performance. These metrics are: Accuracy (ACC), Precision, and Recall. Fraction Specificity, Area Under the Receiving Operator Feature (AUC), and Area Under the Precision-Recall Curve (PR) are all measured. The proposed model KPRMKDTI was compared with three other existing prediction models (BarlowDTI, MolTrans, and Kang et al.) using a five-fold cross-validation experiment. The results are as follows: Figure 3As shown in the comparative experiments, KPRMKDTI demonstrated the best overall performance across all seven core metrics, showcasing the model's strong discriminative power and stability in drug-target interaction prediction tasks. Compared to other methods, KPRMKDTI achieved the highest values ​​in both Accuracy (0.8624) and AUC (0.9303), indicating that the model outperforms mainstream methods in both overall prediction accuracy and global classification ability. Particularly noteworthy is its high levels of both Recall (0.8657) and Specificity (0.8613), demonstrating a good balance in identifying positive and negative samples. It efficiently captures potential interactions while effectively suppressing false positives, exhibiting excellent robustness and generalization ability. In comparison, BarlowDTI had a slightly higher Precision (0.6851), but its Recall was only 0.8156, indicating that it still had omissions in boundary sample identification; MolTrans had the weakest overall performance (AUC = 0.8482, PR = 0.6469), indicating that its feature representation was limited in capturing the complex relationship between drugs and targets; Kang et al.

[32] model performed moderately in Accuracy (0.8226) and Recall (0.8035), but its Precision (0.6104) was low, the prediction results were not stable enough, and it was easy to have class shift.

[0054] Overall, KPRMKDTI, by integrating the MKFF multi-scale feature fusion module and the RBMO feature optimization mechanism, achieves adaptive integration of cross-scale information and effective suppression of feature redundancy, enabling the model to learn the nonlinear relationship between drugs and targets under multi-level representations. It comprehensively surpasses existing methods in all metrics, fully demonstrating the structural advantages and application potential of this model in complex biological data scenarios, and providing an important reference direction for the structural optimization of future DTI prediction models.

[0055] Example 5: Model Interpretation Figure 2(A) presents a heatmap of the top 20 key features × the top 20 representative samples. The horizontal axis represents the feature number, and the vertical axis represents the sample. The color reflects the magnitude and direction of the SHAP value: red indicates that the feature makes a positive contribution to the prediction (i.e., promotes the discrimination of drug-target interaction), while blue indicates that it pushes the prediction result towards the negative class (no effect). The results show heterogeneity in the contribution patterns of different samples on the same features, indicating that the model not only relies on a few global features for prediction but also captures the differences in the importance of local features in different drug-protein combinations. The heatmap not only reveals the key features of the model's discrimination but also demonstrates the diversity of feature utilization among different samples. This shows that the proposed model can identify universal patterns of drug-target interaction globally while capturing the differential contributions of specific pairings locally.

[0056] Figure 2 (B) shows the SHAP summary plot of the top 20 important features. As can be seen from the figure, several features make significant contributions to the model's discrimination, with Features 1149, 934, 790, and 1150 exhibiting strong explanatory power across different samples. The SHAP values ​​on the horizontal axis reflect the positive or negative impact of features on the predicted output, while the color distribution reveals the coupling relationship between feature values ​​and the direction of contribution. For example, some features with higher values ​​(red dots) correspond to positive SHAP values, indicating that they enhance the likelihood of interaction between the drug and the target; while with lower values ​​(blue dots), they tend to push the prediction results towards the negative class. This pattern aligns with the known structure-function coupling relationship in biology, where an increased frequency of specific chemical fragments or amino acid sequence patterns is usually associated with increased interaction strength. Notably, the distribution of feature importance is highly consistent with the model's structural design. The RBMO module effectively filters out redundant information through feature selection, making key discriminative features stand out; the MKFF embedding layer further ensures the fusion and alignment of drug and protein features in the shared representation space; the final KAN classifier learns the nonlinear discriminative boundary on this basis, thus achieving better interpretability while maintaining high prediction performance.

[0057] Overall, the SHAP analysis results indicate that the model in this study can capture the core discriminant factors in drug-target interactions, and the obtained key features show a certain degree of consistency with the underlying biological mechanisms. This not only improves the reliability of the prediction results but also provides a reference for subsequent drug design, site of action inference, and experimental validation.

Claims

1. A DTI prediction method based on LLM and RBMO, characterized in that, The steps are as follows: Step 1: Data Preprocessing This includes dividing the training data and test data, and using the large language model K-BERT and ProstT5 to extract features from drug and target sequences, respectively. Step 2: Feature selection of target points: The red-billed blue magpie optimizer method is used to reduce the dimensionality and select target features; 2.1 Solving the "Curse of Dimensions" problem: Using the RBMO method and based on the K-nearest neighbor metric, feature selection is performed on the target data. 2.2 Red-billed Blue Magpie Optimizer: In RBMO, the position of the i-th candidate solution is represented as: (1) Where t is the number of iterations and d is the feature dimension; The first phase involves reconnaissance. The red-billed blue magpie expands its search area through reconnaissance activities, and the individual locations are updated as follows: (2) in For random coefficients, As the current globally optimal individual, The individuals are randomly selected; this mechanism enables the algorithm to maintain a strong global search capability. When approaching the optimal solution area, the red-billed blue magpie completes a more refined search by coordinating foraging: (3) in It serves as a local perturbation factor, used to enhance local exploration; 2.3 Feature selection decision-making methods: The objective function combines two mutually constraining metrics: classification error rate and feature subset size; the objective function F(X) is defined as shown in formula (4): (4) in: It is a subset of features obtained by decoding the individual position vector X; It uses feature subsets The average classification error rate is obtained by training a K-nearest neighbor classifier and then performing k-fold cross-validation on the validation set. It is a subset The number of features; It is the total number of original features; It is a weighting coefficient used to balance the two objectives of classification performance and model simplicity; Step 3: Multi-core feature fusion: 3.1 Multi-scale feature extraction: After the feature data is input into the multi-kernel feature extraction module, the module captures comprehensive contextual information from different scales and regions. This module has four parallel paths to process the input embedding vector; the representation of each path is shown in formulas (5)-(8): (5) (6) (7) (8) Where k1 is a 1x1 convolutional kernel used for feature transformation and information integration; k2 consists of a 1x1 convolution to reduce the number of channels and a 3x3 convolutional kernel; k3 uses a 5x5 convolutional kernel to capture a larger receptive field; k4 first uses a 3x3 max pooling layer to capture the most salient local features, and then uses a 1x1 convolution to perform channel transformation; the four paths output feature maps k1, k2, k3, and k4 at four different scales, which together constitute a multi-view, multi-granular observation of the input features; 3.2 Global-Local Dynamic Fusion: MKFF comprises four parts: multi-scale convolutional extraction, global channel dynamic fusion, local spatial dynamic fusion, and residual enhancement; the process is as follows: 3.2.1 Multi-scale convolution extraction: First, four sets of multi-scale representations are extracted from the input sequence, including convolution and pooling operations at different receptive fields. These multi-scale features are then concatenated and compressed using a 1×1 convolution to obtain the basic fused features. (9) 3.2.2 Global Channel Dynamic Fusion: First of all Perform average pooling: (10) Next, the adaptive convolutional kernel size is calculated based on the number of channels C: (11) in This indicates taking the closest odd number to ensure the center alignment of the adaptive convolution kernel; Then, an adaptive convolutional kernel is used to perform channel transformation on the global pooling result, and channel attention weights are obtained through a sigmoid function, ultimately yielding the globally enhanced features. The formula is shown in equation (12); (12) 3.2.3 Local Spatial Dynamic Fusion: First of all An adaptive convolution is applied to obtain spatially compressed features, followed by another convolution to obtain local spatial attention weights, which are then normalized using Softmax in the spatial dimension; the local enhanced features are represented as: (13) in This is the activation layer, used to generate attention weights; This is a compression layer used to compress data; 3.2.4 Residual Enhancement: After concatenating the global attention enhancement features and the local attention enhancement features, a 1×1 convolution is used for further fusion. To maintain the continuity of information in the input features and enhance the stability of the model during deep propagation, MKFF employs residual connections, adding the initial input features to the fused features. Finally, average pooling is performed on the residual results to obtain the final output. : (14) (15) Step 4: Building a KAN-based classification model: It includes loss function definition, model architecture design, and model training and validation; Step 5: Interpretable Analysis: The SHAP method is used to explain the model's prediction process.

2. The DTI prediction method based on LLM and RBMO as described in claim 1, characterized in that the first step specifically operates as follows: 1.1 Dataset processing methods: For the dataset built on the BindingDB database, there are a total of 9166 positive samples and 23427 negative samples. After removing duplicate sequences, there are 6588 positive samples and 19759 negative samples. 80% of the dataset is used as training samples and 20% is used as test samples to train the prediction model. 1.2 Feature extraction of drug sequences using the K-BERT large language model: Each drug sequence in SMILES format is sequentially fed into the K-BERT large model, and features of all sequences are extracted. Each sequence is transformed into a 768-dimensional vector as a feature. 1.3 Feature extraction of target sequences using the ProstT5 large language model: Each target sequence in FASTA format is sequentially fed into the ProstT5 large model. Features are extracted from all sequences, and each sequence is transformed into a 1024-dimensional vector as a feature.

3. The DTI prediction method based on LLM and RBMO as described in claim 1, characterized in that, The fourth step, as described above, is performed as follows: 4.1 Loss Function Selection: A loss function, BCEWithLogitsLoss, which combines the Sigmoid function and the binary cross-entropy loss, is used to calculate the model's loss for each sample label prediction; the specific calculation formula is as follows: (16) in Indicates the number of samples. Indicates the first The true label of each sample The model represents the first The predicted probability of a sample, i.e., the value output by the model. For the Sigmoid function; 4.2 Model Architecture Design: A multilayer perceptron model consisting of quasi-fully connected layers is adopted; its architecture design is as follows: 4.2.

1. Function: The KAN classification model is responsible for integrating the features extracted by the previous modules and making the final classification decision; 4.2.

2. Design: Fully connected layers flatten the features and apply linear transformations, then introduce non-linearity through activation functions to perform complex classification tasks; 4.2.

3. Parameters: First layer: KANLinear(mkff_channels * pooled_length, 128, wavelet_type='dog') Second layer: KANLinear(128, 64, wavelet_type='dog') Third layer: KANLinear(64, 1, wavelet_type='dog') 4.2.

4. Output: The final fully connected layer outputs a one-dimensional vector, corresponding to the probability prediction. The Sigmoid activation function is used to ensure that the output value is between 0 and 1. 4.3 Model Training and Validation Testing: The sample outputs a (1, 1) matrix through the model, with values ​​between 0 and 1, representing the probability that the sample contains the corresponding label.

Images