ProteinZero

Inverse folding models — those that read a protein backbone and write a sequence likely to fold into it — are trained by maximum likelihood on curated sequence–structure pairs. That objective rewards reproducing the native sequence, which is not the same thing as producing a sequence that actually folds, stays folded, and differs usefully from what nature already made. ProteinZero, from the University of Illinois Urbana-Champaign, closes that gap by wrapping an existing inverse folding model in an online reinforcement learning loop whose rewards are the design goals themselves.

The central engineering problem in such a loop is the cost of feedback. Evaluating whether a generated sequence folds correctly normally means running a structure predictor, and evaluating stability normally means a physics-based free-energy calculation, either of which makes online RL prohibitively slow. ProteinZero uses ESMFold for the structural reward and replaces physics-based stability scoring with a self-derived ddG estimator built from the policy's own backbone-conditioned and unconditional sequence likelihoods, roughly 236× faster than FoldX. A third term, an embedding-level diversity regularizer, penalizes cosine similarity among generated sequences to prevent the mode collapse that reward optimization otherwise induces.

The preprint was posted in June 2025 by Ziwen Wang, Jiajun Fan, Ruihan Guo, Thao Nguyen, Heng Ji, and Ge Liu. The authors state that source code and model checkpoints will be released upon publication; as of now neither is public.

#Key Features

• Rewards that match design goals: Optimization targets designability, predicted stability, sequence recovery, and diversity directly, rather than the likelihood surrogate used in supervised inverse folding training.
• Cheap stability feedback: The self-derived ddG predictor is computed from the model's own likelihoods and takes about two seconds of GPU time per evaluation, making stability a usable online reward signal.
• Diversity regularization: An explicit embedding-space penalty keeps the policy from collapsing onto a narrow band of high-reward sequences, preserving functionally meaningful variation.
• No labeled data required: The loop improves the model from its own generations plus computed rewards, so it needs no additional curated sequence–structure pairs.
• Modest compute: A complete RL run over CATH-4.3, reward computation and data generation included, finishes in under three days on a single eight-GPU node.

#Technical Details

ProteinZero fine-tunes InstructPLM, a structure-conditioned protein language model, as the policy. Two online RL algorithms are implemented — RAFT (reward-ranked fine-tuning) and GRPO (group relative policy optimization) — with GRPO producing the stronger results. The general objective balances the multi-reward return against a KL penalty toward the reference model and the diversity term. Evaluation uses the CATH-4.3 benchmark split into two length regimes, 0–150 and 150–300 residues, with designs assessed by sequence recovery, self-consistency RMSD and TM-score, pLDDT, FoldX ddG, and pairwise diversity; success is scored against FoldX-measured ddG rather than the model's own predictor. Against ProteinMPNN, ESM-IF, and InstructPLM, ProteinZero cuts design failure rates by 36–48% and holds success rates above 90%, with TM-score gains of 0.05–0.06 alongside higher sequence recovery and higher diversity at the same time — a trade-off inverse folding methods normally have to make. The margin widens on the longer 150–300 residue proteins. On timing, the predicted-ddG reward takes about two seconds on GPU where FoldX takes 472 seconds for a 0–150 residue design, and the ESMFold structural reward takes 18.7 seconds where AlphaFold 3 takes 705 seconds.

#Applications

The practical target is any structure-based design campaign where the backbone is fixed and the bottleneck is the fraction of proposed sequences that survive downstream filtering: scaffolding a functional motif, redesigning a de novo backbone from a hallucination or diffusion pipeline, or resurfacing a natural fold for stability. A lower failure rate translates directly into fewer genes ordered and fewer expression tests run per successful design. More broadly, the recipe is model-agnostic — any inverse folding policy that exposes sequence likelihoods can supply the ddG signal — so the same loop could be applied to other structure-conditioned generators.

#Impact

ProteinZero demonstrates that the RL-from-feedback paradigm that reshaped language model alignment carries over to protein generation, provided the reward pipeline is engineered for speed. Its most transferable idea is the self-derived stability estimator, which turns an expensive physics calculation into a forward pass through the model already in the loop. The evaluation is computational throughout — self-consistency RMSD and FoldX ddG are proxies for folding and stability, not measurements, and no experimental characterization is reported. Because neither code nor checkpoints have been released, the work currently functions as a method others must reimplement rather than a model that can be picked up and run.