Gero-LLM: A Multimodal Large Language Model for Geroprotector Discovery via Cross-Modal Differentiated Mutual Learning.

Abstract

Geroprotectors underpin therapeutic strategies to intervene in aging pathologies and extend lifespans. Unfortunately, geroprotector discovery remains a significant challenge due to data quality and pathway redundancy. Existing methods often rely on single data modalities, which fail in capturing the intricate structure-activity relationships in geroprotector molecules. Therefore, we present Gero-LLM, a multimodal framework that synergizes the reasoning capabilities of pre-trained large language models (LLMs) with the topological modeling of Graph Isomorphism Network with Edge features (GINE) for geroprotector discovery. By fusing textual representations with structural embeddings, Gero-LLM leverages multimodal chemical information to enhance predictive ability. To overcome the limitations of standard fine-tuning, we utilize a cross-modal differentiated deep mutual learning (CM-Diff-DML) strategy. This training paradigm enforces the diversity between modalities, preventing mode collapse and improving model prediction ability. Gero-LLM achieves state-of-the-art performance, demonstrating promising robustness on highly imbalanced external datasets, resembling the real-world geroprotector screening scenarios. Furthermore, in silico mutagenesis confirms that Gero-LLM captures fundamental chemical pharmacophores beyond summary statistics. This work attempts to bridge the gaps between LLMs and multimodal molecule information, providing a robust platform to accelerate the discovery of therapeutic interventions on aging.