Multimodal Fusion-Based Lightweight Model for Enhanced Generalization in Drug-Target Interaction Prediction

Area of Science:

• Computational Biology
• Drug Discovery
• Bioinformatics

Background:

• Precise prediction of drug-target interactions (DTIs) is vital for efficient drug discovery.
• Current DTI models demand substantial computational resources due to complex protein sequences.
• There is a need for computationally efficient and generalizable DTI prediction methods.

Purpose of the Study:

• To introduce MMF-DTI, a lightweight model for enhanced DTI prediction.
• To improve the generalizability of DTI predictions without compromising computational efficiency.
• To explore the utility of multimodal data fusion, including natural language descriptions, for DTI prediction.

Main Methods:

• Developed MMF-DTI, a multimodal fusion model for DTI prediction.
• Integrated four data modalities: molecular sequence, molecular properties, target sequence, and target function descriptions.
• Utilized natural language processing for target function descriptions – a novel approach in DTI prediction.

Main Results:

• MMF-DTI demonstrated comparable performance to state-of-the-art models in generalizability, particularly with limited drug or target information.
• The model achieved this using significantly fewer parameters (50% reduction) and VRAM (17% usage) compared to existing methods.
• MMF-DTI proved effective in computationally constrained environments.

Conclusions:

• Multimodal data fusion offers a promising avenue for improving DTI prediction model efficiency and generalizability.
• MMF-DTI presents a computationally efficient solution for DTI prediction, facilitating broader applicability in drug discovery.
• The model's performance highlights the potential for leveraging diverse data types, including textual information, in computational drug discovery.