REPRODUCIBLE DRUG REPURPOSING: WHEN SIMILARITY DOES NOT SUFFICE

Area of Science:

• Computational pharmacology
• Drug discovery and development
• Machine learning in medicine

Background:

• Drug repurposing accelerates the identification of new therapeutic uses for existing medications.
• Supervised computational models often use drug chemical, target, and side effect similarity for predictions.
• Current methods face limitations when applied to conditions with no known similar drugs, hindering practical use.

Purpose of the Study:

• To evaluate the prediction accuracy of supervised drug repurposing models with disjoint training and testing datasets.
• To investigate the impact of independent similarity information usage within training and test sets.
• To explore the necessity of unsupervised approaches for novel drug repurposing.

Main Methods:

• Developed a Python platform for reproducible, similarity-based drug repurposing models.
• Employed machine learning methods utilizing chemical, target, and side effect similarity.
• Implemented a disjoint cross-validation strategy to ensure non-overlapping training and test sets.

Main Results:

• A standard similarity-based machine learning approach achieved good performance on benchmark datasets.
• Prediction accuracy significantly dropped when cross-validation folds were disjoint.
• Independent use of similarity information between training and test sets led to performance degradation.

Conclusions:

• Revisiting validation scenarios for similarity-based drug repurposing methods is crucial.
• Supervised methods may overestimate performance due to overlapping data.
• Unsupervised approaches are needed to explore novel drug uses in uncharted pharmacological spaces.