FKSUDDAPre: A drug-disease association prediction framework based on F-TEST feature selection and AMDKSU resampling...

Area of Science:

• Computational biology and bioinformatics
• Pharmacology and drug discovery
• Machine learning in healthcare

Background:

• Predicting drug-disease associations (DDAs) is crucial for drug discovery, offering insights into molecular mechanisms, drug repositioning, and personalized medicine.
• Traditional DDA prediction methods are time-consuming and resource-intensive.
• Existing machine learning approaches face challenges with feature complexity, data sparsity, and sample imbalance, limiting their practical application.

Purpose of the Study:

• To develop an efficient and accurate framework, FKSUDDAPre, for predicting drug-disease associations (DDAs).
• To overcome limitations of existing machine learning methods in feature construction, data sparsity, and sample imbalance.
• To improve the accuracy and generalization performance of DDA prediction models.

Main Methods:

• Employed a multi-modal feature fusion strategy combining Mol2vec and K-BERT for drug molecular fingerprints and Medical Subject Headings (MeSH) with DeepWalk for disease features.
• Developed the AMDKSU optimization algorithm to address class imbalance through clustering and an improved distance metric strategy.
• Utilized F-test for feature importance ranking, an ensemble of XGBoost, Decision Tree, Random Forest, and HyperFast with dynamic weight allocation for prediction, and LIME for interpretability.

Main Results:

• Achieved an average AUC of 0.9725, outperforming baseline models by approximately 3.88%.
• Demonstrated practical applicability with 80% and 60% accuracy in identifying top candidate drugs for Alzheimer's and Parkinson's diseases, respectively, confirmed by literature.
• The developed framework showed strong interpretability via LIME analysis and included a user-friendly visualization tool.

Conclusions:

• FKSUDDAPre offers an efficient and accurate solution for drug-disease association prediction, addressing key challenges in current machine learning models.
• The framework's multi-modal feature fusion, imbalance handling, and ensemble architecture significantly enhance predictive performance and generalization.
• The model's practical utility and interpretability suggest its potential to accelerate drug discovery and therapeutic research.