Towards Absolute Stability-Conditioned Protein Sequence Design

Current protein sequence design models do not explicitly control thermodynamic properties. We introduce DECODER, a work-in-progress model which designs proteins conditioned on a target backbone and unfolding free energy, $\Delta G_{\mathrm{unfold}}$ (kcal/mol, $\Delta G$ hereafter). We designed sequences for RFdiffusion-generated backbones at target $\Delta G$ values from $-10$ to $+10$, with higher values indicating greater stability. Across the evaluated conditions, an orthogonal $\Delta G$ predictor using ESM3 embedding (ESM3$\Delta$G) separated negative and positive designs, while the $+5$ kcal/mol condition yielded the strongest folding metrics from ESMFold2, providing preliminary evidence that $\Delta G$ conditioning affects sequence-design outcomes. DECODER does not yet outperform ProteinMPNN, motivating broader labeled data and experimental validation.