Uncertainty-Aware Adaptation of Large Language Models for Protein-Protein Interaction Analysis.

Area of Science:

• Computational Biology
• Bioinformatics
• Artificial Intelligence in Medicine

Background:

• Protein-protein interactions (PPIs) are crucial for understanding cellular mechanisms in diseases like cancer and neurodegenerative disorders.
• Large Language Models (LLMs) show promise in predicting PPIs by analyzing biomedical literature, but their uncertainty poses a challenge for reproducibility.
• Reliable PPI identification is essential for advancing precision medicine and biomedical research.

Purpose of the Study:

• To develop an uncertainty-aware adaptation of LLMs for more reliable protein-protein interaction (PPI) analysis.
• To enhance the trustworthiness and reproducibility of computational biology findings derived from LLMs.
• To leverage advanced AI techniques for improved understanding of protein behavior in disease contexts.

Main Methods:

• Fine-tuning LLaMA-3 and BioMedGPT models for PPI prediction.
• Integrating LoRA (Low-Rank Adaptation) ensembles for enhanced prediction reliability.
• Employing Bayesian LoRA models for robust uncertainty quantification (UQ).

Main Results:

• Achieved competitive performance in identifying protein-protein interactions across various disease contexts.
• Successfully addressed the challenge of model uncertainty in LLM-based PPI analysis.
• Provided confidence-calibrated insights into protein behavior, improving trustworthiness.

Conclusions:

• Uncertainty-aware LLM adaptation offers a promising approach for reliable PPI identification.
• The developed methods enhance reproducibility in computational biology, crucial for biomedical applications.
• This work contributes to advancing precision medicine through more trustworthy AI-driven biological insights.