ProtFPreDTI: Drug-Target Interaction Prediction Study and LIME Interpretability Analysis Based on the ProtBERT Deep...

The interaction mechanism between drugs and targets is a core topic of modern pharmaceutical research, and its precise analysis can not only significantly improve the efficiency of innovative drug development but also reveal the mechanism of interaction between active compounds and biomolecules. Currently, this field is facing a double challenge: on the one hand, the traditional experimental methods have a technical bottleneck of high time consumption and high cost; on the other hand, the existing computational models still have obvious limitations in terms of the comprehensiveness of feature characterization and the handling of data imbalance. In this context, new prediction methods based on machine learning are becoming an important technical path to break through the existing research dilemmas due to their superior computational efficiency and economic advantages. In this study, the Mol2Vec algorithm was used to extract features from SMILES representations of drug molecules, while the ProtBERT deep language model was applied to obtain high-dimensional features of protein sequences. To address the problem of high dimensionality and redundancy of protein sequence features, SHAP value analysis was innovatively introduced to achieve quantitative screening of feature importance, and the most judgmental 300-dimensional features were finally retained for modeling. To solve the challenge of sample category imbalance, the study designed an intelligent undersampling strategy based on fuzzy logic-fuzzy undersampling, which effectively constructs a balanced training data set. In terms of prediction model architecture, this study creatively complements the advantages of two integrated learning methods, XGBoost and random forest, and significantly improves the generalization ability of the model through an adaptive weighted fusion mechanism. In order to enhance the transparency of the model, the LIME interpretable analysis framework is specially integrated to provide an intuitive prediction basis from the perspective of local feature contribution, so that the decision-making process of the model is both scientific and traceable. The whole set of methodology systems ensures the prediction accuracy and realizes the optimization of the whole process from feature engineering to result interpretation. The integrated prediction model ProtFPreDTI constructed in this study demonstrated excellent performance in independent validation, with an area under the subject operating characteristic curve (AUC) of 0.92, which significantly improved the core metrics of prediction accuracy, sensitivity, and specificity compared with the traditional method. Through the innovative feature screening algorithm and multimodel fusion technology, the system effectively enhances the robustness of prediction and the ability of cross-data set generalization, providing a breakthrough solution for drug-target interaction research. The code and data sets related to this study are available at the following URL: https://github.com/flying-peanut/ProtFPreDTI.