MIFS: An adaptive multipath information fused self-supervised framework for drug discovery

Area of Science:

• Artificial Intelligence
• Computational Chemistry
• Drug Discovery

Background:

• AI-driven drug discovery faces challenges in generating expressive molecular representations with limited labeled data.
• Current methods often neglect diverse information propagation within molecules and lack chemical constraints during pre-training.

Purpose of the Study:

• To develop an adaptive multipath information fused self-supervised framework (MIFS) for improved molecular representation learning.
• To address limitations in current pre-training strategies for molecular encoders in drug discovery.

Main Methods:

• Proposed an adaptive multipath information fused self-supervised framework (MIFS) utilizing a novel molecular graph encoder (Mol-EN).
• Mol-EN employs three information propagation pathways (atom-to-atom, bond-to-atom, group-to-atom) for comprehensive semantic understanding.
• Implemented an adaptive pre-training strategy on 11 million unlabeled molecules using a topological contrastive loss and incorporated an elemental knowledge graph (ElementKG) during fine-tuning.

Main Results:

• MIFS achieved competitive performance across 14 drug discovery benchmark datasets, including property prediction and interaction prediction tasks.
• The framework demonstrated the ability to provide chemically plausible explanations for its predictions.
• Pre-training on large unlabeled datasets with scaffold-based strategies and ElementKG integration enhanced molecular representation quality.

Conclusions:

• The proposed MIFS framework effectively generates expressive molecular representations for AI-driven drug discovery.
• MIFS offers a promising approach to overcome data scarcity and incorporate chemical knowledge into molecular modeling.
• The method provides a foundation for more interpretable and accurate AI models in pharmaceutical research.