ThermoRL: Structure-Aware RL for Protein Mutation Design to Enhance Thermostability

ICLR 2026 Conference Submission13257 Authors

Keywords: reinforcement learning, protein design

Abstract:

Designing mutations to optimize protein thermostability remains challenging due to the complex relationship between sequence variations, structural dynamics, and thermostability, often assessed by $\Delta\Delta G$ (the change in free energy of unfolding).
Existing methods rely on experimental random mutagenesis or one-shot predictions over fixed mutation libraries, which limits design space exploration and lacks iterative refinement capabilities.
We present \textbf{ThermoRL}, a reinforcement learning (RL) framework that integrates graph neural networks (GNN) with hierarchical Q-learning to sequentially design thermostabilizing mutations.
ThermoRL combines a pre-trained GNN-based encoder with a hierarchical Q-learning network and employs a surrogate model for reward feedback, to guide the agent in selecting both mutation positions and amino acid substitutions.
Experimental results show that ThermoRL achieves higher or comparable rewards than baselines while maintaining computational efficiency. It effectively avoids destabilizing mutations, recovers experimentally validated stabilizing variants, and generalizes to unseen proteins by identifying context-dependent mutation sites.
These results highlight ThermoRL as a scalable, structure-informed framework for adaptive and transferable protein design.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 13257