Efficiently Bridging Protein Language Model and Large Language Model with a Cross-modal Lightweight Adapter

ACL ARR 2025 February Submission3665 Authors

Abstract: In natural language processing and biology, large language models (LLMs) and protein language models (PLMs) have advanced significantly. Despite similarities in their organizational form, protein sequences and natural language lack direct semantic association due to domain differences. Thus, efficiently connecting LLMs and PLMs to leverage cross-field benefits and promote large model toolization remains a challenge. To bridge this gap, we propose a lightweight cross-modal adapter that aligns protein sequences with natural language representations through contrastive learning, effectively reducing modality difference, thereby bridging PLMs and LLMs and enhancing the performance of both. In the experiments, we first evaluated the performance of the PLM integrated with the adapter across multiple tasks. The experimental results show that the adapter achieved better results in many cases compared to using the PLM alone. Additionally, given the significant progress in protein-related LLMs, we further explored how the adapter can enhance this paradigm. In this experiment, we not only demonstrated that the adapter can enhance the LLM’s ability to analyze protein sequences, outperforming other baseline models, but also proved the adapter’s applicability in different base models.

Paper Type: Long

Research Area: Multimodality and Language Grounding to Vision, Robotics and Beyond

Research Area Keywords: cross-modal application; cross-modal pretraining; cross-modal content generation; multimodality;

Contribution Types: Model analysis & interpretability, Approaches low compute settings-efficiency

Languages Studied: English

Submission Number: 3665