Reciprocal perspective as a super learner improves drug-target interaction prediction (MUSDTI)

Area of Science:

• Computational chemistry
• Machine learning
• Pharmacology

Background:

• Accurate drug-target interaction (DTI) prediction is crucial for drug discovery and repurposing, especially for emergent diseases.
• Traditional binary classification for DTI prediction is less informative than predicting physiochemical binding affinity.
• Deep learning models show promise for DTI prediction with increasing experimental data availability.

Purpose of the Study:

• To develop a novel meta-model for DTI prediction by ensembling student-generated deep learning models.
• To evaluate the effectiveness of the Reciprocal Perspective (RP) multi-view learning framework in DTI prediction.
• To establish a new benchmark for DTI prediction, particularly for emergent challenges.

Main Methods:

• Formulated a DTI prediction competition for an undergraduate machine learning course.
• Students developed 28 component deep learning DTI models.
• Combined component models using the Reciprocal Perspective (RP) multi-view learning framework to create the MUSDTI meta-model.
• Employed a double-cold experimental design for evaluating model performance.

Main Results:

• The MUSDTI meta-model, utilizing RP, achieved state-of-the-art (SOTA) performance in DTI prediction.
• RP significantly improved upon existing SOTA DTI prediction models.
• The double-cold experimental design proved suitable for emergent DTI prediction scenarios.
• RP demonstrated effectiveness as an ensembling method and for low-computation transfer learning.

Conclusions:

• The MUSDTI meta-model represents a significant advancement in DTI prediction.
• The Reciprocal Perspective (RP) framework offers a powerful approach for ensembling and improving DTI prediction models.
• This work provides valuable insights for sequence-based, pairwise prediction tasks in computational biology and drug discovery.