Loading…
Protein inverse folding by categorical diffusion, tuned with reinforcement learning on structural consistency. Samples diverse refoldable sequences.
No providers recorded yet. Browse all providers
Inverse folding is a one-to-many problem: many sequences fold into the same backbone, and a design campaign wants several of them, because the criteria that decide which candidate is worth synthesising — expression, stability, aggregation, immunogenicity — are invisible to the coordinates. Yet these models are ranked on sequence recovery, the fraction of positions matching the residue nature happened to use, which rewards concentrating on one native-like answer; the usual workaround of raising sampling temperature buys variety by degrading the structures the sequences fold into.
RL-DIF optimises structural consistency directly, under a metric that refuses to credit diversity bought at that structure's expense. Foldable diversity is the average pairwise sequence dissimilarity among designs, counted only over pairs whose predicted structures both stay above a TM-score threshold — 0.7 by default — of the fold predicted from the native sequence. Sequence diversity can be inflated arbitrarily; foldable diversity cannot.
The model reaches it in two phases. A structure-conditioned categorical diffusion model is first pretrained on sequence recovery, then fine-tuned with denoising diffusion policy optimisation (DDPO), which treats the reverse denoising chain as a Markov decision process and rewards each finished sample by how well its predicted structure matches the target. Yasha Ektefaie, Olivia Viessmann and colleagues at Flagship Pioneering and Harvard University released it in 2024, benchmarked against ProteinMPNN, PiFold, KWDesign and ESM-IF.
The denoiser is a discrete denoising diffusion probabilistic model over a 20-letter amino acid vocabulary with uniform transition matrices and 150 timesteps, trained under the full D3PM hybrid loss. Its network is a modified PiFold architecture: backbone coordinates become a k-nearest-neighbour graph with k = 30, the PiFold featuriser supplies distance, dihedral and direction features, ten PiGNN layers process them, and added MLPs inject the noised sequence and the timestep. Pretraining used CATH 4.2 under the Ingraham splits — 18,025 training, 1,637 validation and 1,911 test chains — with Adam at a learning rate of 1e-3 and batch size 64 across four A10 GPUs for 200 epochs. The RL phase ran 1,000 DDPO steps at a learning rate of 1e-5 with batches of 32 backbones, four samples each; ESMFold served the reward, chosen over AlphaFold 2 for throughput.
On the full CATH 4.2 test split, sampling four sequences per backbone, RL-DIF reaches 29% foldable diversity against 23% for the best prior model trained on the same data, at a self-consistency TM (sc-TM) score of 0.78 and 44% sequence recovery. Recovery and consistency are competitive rather than leading, and the RL phase itself trades diversity for consistency: the pre-RL ablation reaches 32% foldable diversity but only 0.72 sc-TM. Pretraining instead on 100,000 predicted structures curated with Foldseek raises foldable diversity to 34%, approaching ESM-IF's 37% with roughly 70 times less data. TS50, TS500 and CASP15 results carry 42–84% cross-split overlap with the training set, so the paper focuses there.
The model fits the sequence design step of a structure-based pipeline: given a crystal structure, a predicted model or a de novo scaffold, it returns a panel of candidates to filter on the campaign's own criteria, rather than one confident suggestion. Positions can be left free while the rest are held fixed, for partial redesign. The released checkpoints also serve as a general-protein starting point for domain-specific work — AntiDIF fine-tunes them for antibodies.
RL-DIF's durable contribution is the argument that sequence recovery is the wrong scoreboard for a one-to-many task, together with a metric that makes the alternative measurable and hard to game. It also shows that reinforcement learning against a folding oracle transfers to conditional generation, where the goal is generalising to structures never seen in training, rather than the unconditional design settings where diffusion RL had mostly been tried. The work remains a preprint, evaluation is entirely computational with no expressed or assayed designs, and the single folding oracle bounds the reward's fidelity. Weights are archived on Zenodo rather than a model hub, and the repository is small and lightly maintained.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.