Large language models for biomedical relation extraction and explanation

Due to the rapid growth of biomedical literature, the extraction of useful information from biomedical texts using natural language processing (NLP) and text mining techniques is in high demand. One of the most important text mining tasks is relation extraction (RE), which involves identifying relationships between entities. Inspired by the success of Large Language Models (LLMs) in various NLP tasks, we propose an LLM-based approach to extract and explain protein-protein interactions (PPIs) from biomedical texts. For relation extraction, we fine-tuned the masked language models BioBERT, SciBERT and PubMedBERT on five PPI benchmark datasets and showed that they can effectively identify PPIs mentioned in a given text. Next, we developed a novel approach to further analyze the extracted PPIs using causal LLMs. We applied different learning strategies, namely in-context learning and parameter-efficient instruction fine-tuning for the Llama-2 chat models, to identify keywords in the text that reveal an interaction between a protein pair. Our results show that parameter-efficient fine-tuning leads to a performance gain even when the domain is new. The smaller fine-tuned models outperformed the zero-shot performance of much larger models. Our study suggests that a pipeline consisting of a masked LLM for relation extraction and a causal LLM for relation explanation can be an effective strategy for building a PPI analyser in real-world scenarios.