Qimai: a multi-agent framework for zero-shot DNA-protein interaction prediction.

Area of Science:

• Genomics
• Bioinformatics
• Artificial Intelligence

Background:

• Predicting DNA-protein interactions (DPI) is crucial in genomics but current models struggle with novel proteins.
• Existing deep learning models lack generalization capabilities for unseen protein targets.

Purpose of the Study:

• To develop a novel AI framework, Qimai, for accurate and interpretable prediction of DNA-protein interactions.
• To enhance model generalization to previously unseen proteins by integrating diverse biological evidence.

Main Methods:

• Qimai utilizes a modular AI agent framework with a Large Language Model (LLM) as a reasoning engine.
• It integrates direct motif evidence, indirect motif evidence from protein interactors, and predictions from a transformer-based DPI model.
• The framework provides explainable predictions with confidence scores.

Main Results:

• Qimai significantly outperformed standalone deep learning models on a benchmark of 78 unseen proteins.
• Key performance metrics including AUC-PR, AUC-ROC, and MCC showed substantial improvements (17.6%, 15.6%, and 244% respectively).
• Ablation studies highlighted the LLM's role in dynamically weighing evidence, with indirect co-factor motif data being critical for novel proteins.

Conclusions:

• Qimai establishes a generalizable and interpretable paradigm for integrating heterogeneous data in predictive genomics.
• The framework demonstrates superior performance and robustness for predicting DNA-protein interactions, especially for novel proteins.
• Qimai offers a valuable tool for advancing genomics research and is accessible via a web portal.