RNA-EFM: energy-based flow matching for protein-conditioned RNA sequence-structure co-design.

Motivation: Designing RNA molecules that can specifically bind to target proteins is fundamental to numerous biological and therapeutic applications. However, existing approaches to protein-conditioned RNA design primarily focus on structural alignment or sequence recovery, often ignoring essential biophysical factors such as molecular stability and thermodynamic feasibility. Results: To address this gap, we propose RNA-EFM, a novel deep learning framework that integrates energy-based refinement with flow matching for protein-conditioned RNA sequence and structure co-design. RNA-EFM consists of two complementary components: a flow matching objective that supervises geometric alignment between predicted and native RNA backbone structures, and an energy-based idempotent refinement that iteratively improves RNA structure predictions by minimizing both structural error and physical energy. The energy refinement is guided by biophysical priors including the Lennard-Jones potential and sequence-derived free energy, ensuring that the generated RNAs are not only geometrically plausible but also thermodynamically stable. We demonstrate the effectiveness of RNA-EFM through extensive experiments. RNA-EFM significantly outperforms state-of-the-art baselines in terms of RMSD, lDDT, sequence recovery, and binding energy improvement. These results highlight the importance of incorporating biophysical constraints into RNA design and establish RNA-EFM as a promising framework. Availability and implementation: The source code for RNA-EFM is available at: https://github.com/abrarrahmanabir/RNA-EFM.