This research introduces a novel framework for aligning Large Language Models with Geometric Deep Models to create more effective protein representations.

Key Contributions:

• Proposes systematic evaluation metrics for protein representation alignment quality
• Demonstrates that mapping geometric models to LLM space outperforms the reverse approach
• Introduces a new contrastive learning strategy that significantly improves alignment
• Achieves state-of-the-art performance on protein structure understanding tasks

These advances enable more powerful multimodal protein analysis tools that combine sequence and structural information, potentially accelerating drug discovery and protein engineering applications in biology and medicine.

Aligning Large Language Models and Geometric Deep Models for Protein Representation