MLDTA an Ensemble-Driven Multimodal Model with Dynamic Fusion for Drug-Target Affinity Prediction

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning

Background:

• Existing drug-target binding affinity (DTA) models struggle with fixed fusion strategies and single architectures, limiting adaptive relationship capture.
• Over-reliance on single learning algorithms reduces model robustness and generalization, leading to significant prediction errors.

Purpose of the Study:

• To introduce MLDTA, a multimodal framework for DTA prediction.
• To address limitations in current DTA models by integrating dynamic feature fusion and ensemble-inspired principles.

Main Methods:

• MLDTA utilizes structural information, Geary autocorrelation descriptors, and tripeptide composition for drug and target representations.
• It incorporates five representative DTA models as auxiliary predictive modules (APMs).
• A dynamic fusion mechanism with attention modules adaptively integrates APMs and learned representations.

Main Results:

• MLDTA surpasses existing methods in drug-target binding affinity prediction on standard datasets.
• The dynamic fusion mechanism adaptively learns feature importance and enhances cross-modal interaction.
• Case studies demonstrate MLDTA's effectiveness in drug screening.

Conclusions:

• MLDTA offers a robust and adaptive framework for DTA prediction.
• The integration of dynamic fusion and ensemble principles improves model generalization and reduces prediction errors.
• MLDTA shows significant potential for accelerating drug discovery and screening processes.