EnerBridge-DPO

Inverse folding asks for sequences that will fold into a given backbone, and nearly every deep learning model for the task is trained to maximize one number: how often it recovers the residue that nature actually put at each position. Native sequence recovery is easy to measure and correlates loosely with design success, but it is not what a protein engineer wants. A designed sequence has to be thermodynamically stable and, for a complex, has to hold the interface together. Optimizing recovery says nothing directly about either.

EnerBridge-DPO, from a group at Shanghai Jiao Tong University, retargets inverse folding at energy. It builds on the Markov bridge formulation, in which generation starts from an informative structure-derived prior sequence rather than from noise and is iteratively refined toward a plausible design. That starting point matters here for a specific reason: it means every intermediate the model produces is already structurally reasonable, so the candidate pool used for preference learning consists of real design alternatives rather than random sequences.

Onto that base the authors graft direct preference optimization, the alignment technique from language modeling, with preferences derived from measured or computed energies instead of human judgments. An explicit energy-constraint loss is added alongside the DPO objective, which has a useful side effect: the model learns to output calibrated energy values, so the same network that designs sequences also predicts ΔΔG between variants. Relative to catalog inverse-folding entries such as ProteinMPNN and PRISM, the distinguishing move is optimizing the objective designers actually care about rather than a proxy.

#Key Features

• Energy-guided preference tuning: DPO is adapted to the Markov bridge process, with winner/loser pairs drawn from mutational data ranked by binding or stability score, so fine-tuning pushes the generator toward lower-energy sequences.
• Structure-conditioned prior: PiFold supplies the prior sequence that initializes the bridge, giving the optimization a pool of structurally plausible candidates instead of an uninformative starting distribution.
• Parameter-efficient backbone: A pretrained protein language model is kept frozen and conditioned on timestep and structure through AdaLN-Bias modulation and structural cross-attention adapters.
• Dual output: An explicit energy-constraint loss makes the network predict quantitative sequence energies, turning one model into both a sequence designer and a ΔΔG predictor.
• Short generation trajectory: The discrete bridge runs over 25 timesteps, so designs are produced in a small number of refinement passes.

#Technical Details

Pretraining uses the ProteinMPNN dataset clustered at 30% sequence identity into 25,361 clusters, split into 23,358 training and 1,464 validation clusters with the remainder held out, ensuring no target chain or its biological assembly partners cross the split. Preference fine-tuning draws on BindingGym, whose 10M mutational data points yield pairs from the top and bottom 10% of mutants per protein, and on SKEMPI v2, contributing roughly 3,744 pairs from the top and bottom 30% by binding energy. Pretraining runs 50 epochs at a 40,000-residue batch size; DPO fine-tuning runs 150 epochs at a learning rate of 1e-5 with the energy loss weighted at 0.5. On the held-out test set the model reaches 60.91% overall sequence recovery and 3.87 perplexity, against 60.56% and 3.89 for Bridge-IF and 57.28% and 4.25 for ProteinMPNN. The energy gap is larger than the recovery gap: across 26 protein complexes, mean FoldX energy for EnerBridge-DPO designs is 130.03 versus 163.02 for Bridge-IF and 163.58 for ProteinMPNN. On SKEMPI v2 with three-fold cross-validation, ΔΔG predictions reach Pearson 0.7487 and Spearman 0.6447, ahead of the BA-DDG and ProMIM baselines.

#Applications

The primary use is fixed-backbone sequence design where stability is the binding constraint: redesigning protein complexes and interfaces, engineering thermostable enzymes, and generating candidate sequences that survive expression and purification. Because the same model scores ΔΔG, it can also be used to triage mutations for affinity maturation or stability engineering without a separate predictor, keeping design and scoring in one consistent energy representation.

#Impact

The work makes a clean argument that preference optimization transfers from language model alignment to structural biology, provided the preference signal comes from a physically meaningful quantity. The evidence is entirely in silico — energies come from FoldX, Rosetta, and a binding-affinity cycle rather than from wet-lab measurement, and no designed sequence was expressed or characterized experimentally, so the stability claims remain computational. Neither source code nor pretrained weights have been released, and the preprint has not been peer reviewed, which together limit independent evaluation and reuse.