Reproduce “LLM4GRN: Discovering Causal Gene Regulatory Networks with LLMs -- Evaluation through Synthetic Data Generation” on OpenResearch

About this paper

Tejumade Afonja, Ivaxi Sheth, Ruta Binkyte, Waqar Hanif, Thomas Ulas, Matthias Becker, Mario Fritz

Researchers introduce LLM4GRN, a framework leveraging large language models (LLMs) to infer gene regulatory networks (GRNs) from single-cell RNA sequencing data. The work evaluates the derived networks through causal synthetic data generation and demonstrates superior performance when LLMs provide biological priors to statistical inference algorithms, particularly with open-source models like Llama-3.1-70B.

Key insights

• LLM-derived GRNs demonstrate higher self-overlap (robustness) across runs and generate more biologically meaningful connections than random graphs.
• Combining LLM-derived prior knowledge (specifically, TF lists) with traditional statistical causal inference algorithms (hybrid approach) consistently yields superior GRN inference performance compared to either method alone.
• Open-source LLMs, such as Llama-3.1-70B, can provide highly effective biological priors, with its hybrid integration achieving the best overall statistical performance for the PBMC-ALL dataset, surpassing proprietary models.

Summary by alphaXiv.