CMFMTL

CMFMTL predicts drug-disease associations using collective matrix factorization and multi-task learning to jointly model therapeutic and non-therapeutic (marker/mechanism) relationships and prioritize candidate indications and potential adverse associations.

Key Features:

• Multi-Task Learning Framework: Employs multi-task learning with two tasks targeting therapeutic effects and non-therapeutic effects (marker/mechanism), incorporating prior knowledge about correlations between association types.
• Collective Matrix Factorization: Represents drug-disease associations as a bipartite network with two distinct link types and approximates association matrices via matrix tri-factorization.
• Shared Latent Representations: Shares low-dimensional latent representations for drugs and diseases across tasks to enable collective learning between related association types.
• Unified Framework and Optimization Algorithm: Integrates both tasks into a unified computational framework and solves the resulting optimization problem with an efficient algorithm.

Scientific Applications:

• Improved prediction accuracy: Demonstrates superior performance compared to several state-of-the-art methods in predicting drug-disease associations.
• Novel association discovery: Identifies novel drug-disease associations not documented in databases such as the Comparative Toxicogenomics Database (CTD).
• Drug repositioning: Prioritizes candidate indications to support drug repositioning efforts.
• Safety profiling: Anticipates potential side effects and non-therapeutic associations to inform drug safety profiles.

Methodology:

Uses multi-task learning with two tasks (therapeutic and non-therapeutic), collective matrix factorization implemented via matrix tri-factorization on a bipartite network with two link types, shared low-dimensional drug and disease latent representations, and a unified optimization algorithm.