This paper introduces DualView-DTI, a novel framework that extends latent dual-view representations to few-shot drug-target interaction prediction. It integrates graph neural networks with sequence-based encoders to capture complementary information from drugs and targets, employs meta-learning for rapid adaptation, and uses contrastive learning to align representations for robust interaction modeling.
Key findings
DualView-DTI integrates graph and sequence-based encoders to capture structural and sequential information.
Contrastive learning aligns dual-view representations in a shared latent space for interaction modeling.
Meta-learning strategy enables rapid adaptation to new drug-target pairs with limited labeled examples.
Comprehensive experimental validation on standard benchmarks including BindingDB, Davis, and KIBA datasets.
Limitations & open questions
The framework's performance in scenarios with extremely limited data is yet to be fully explored.
Potential failure modes and risk mitigation strategies require further investigation.