LLM-guided prediction of riboswitch-ligand binding affinity under data-limited conditions.
TL;DR
A neuro-symbolic framework in which LLM-derived embeddings encode riboswitch sequences, secondary structures, and ligand representations into a unified representation that captures riboswitch-ligand interaction context, while domain-informed rules encode explicit biochemical priors is presented.
Abstract
Riboswitches are RNA regulatory elements that sense small molecules and regulate gene expression through ligand-induced conformational changes in an aptamer domain. They enable synthetic biology applications such as biosensing and gene circuit design. Accurate prediction of riboswitch-ligand binding affinity, quantified by the dissociation constant (Kd), is essential for rational riboswitch engineering. We present a neuro-symbolic framework in which LLM-derived embeddings encode riboswitch sequences, secondary structures, and ligand representations into a unified representation that captures riboswitch-ligand interaction context, while domain-informed rules encode explicit biochemical priors. Despite limited labeled data, this framework achieves strong predictive performance and improved data efficiency. It predicts min-max log-scaled pKd values and supports binary classification of binding strength, particularly by mitigating systematic overestimation of affinity. More broadly, the results suggest that LLM-based neuro-symbolic representations provide an effective route for modeling riboswitch-ligand interactions under data-limited conditions.