Aligning large language models and geometric deep models for protein representation.

Area of Science:

• Computational biology and bioinformatics
• Artificial intelligence in drug discovery and protein science
• Machine learning for structural biology

Background:

• Large language models (LLMs) show promise in biological applications.
• Geometric deep models (GDMs) excel at capturing structural information in proteins.
• Integrating these models requires understanding their representational alignment.

Purpose of the Study:

• To investigate the alignment between multimodal representations of LLMs and GDMs in the protein domain.
• To identify factors influencing this alignment from both model and protein perspectives.
• To propose strategies for enhancing alignment and its impact on downstream tasks.

Main Methods:

• Comprehensive evaluation of three LLMs against four protein-specialized GDMs.
• Analysis of alignment based on model architectures, protein characteristics, and embedding strategies.
• Experimental validation of proposed alignment enhancement techniques.

Main Results:

• GDMs utilizing both graph and 3D structural data show superior alignment with LLMs.
• Larger LLMs exhibit enhanced alignment capabilities.
• Protein rarity significantly affects alignment; increasing GDM embedding dimensions and using two-layer projection heads improve alignment quality.

Conclusions:

• Alignment between LLMs and GDMs is achievable and can be optimized through specific architectural and data strategies.
• Fine-tuning LLMs on protein-specific data and enhancing GDM embeddings are effective methods.
• Improved alignment directly correlates with enhanced downstream task performance and reduced hallucination in protein-focused multimodal LLMs.