Inverse FoldDir

A protein designer rarely starts from a blank sequence. Some positions are non-negotiable — a catalytic histidine, a disulfide cysteine, an interface hot spot — while others carry only a preference: this solvent-exposed face should lean polar, this candidate metal site should favor histidine or aspartate. Autoregressive inverse-folding decoders such as ESM-IF1 and ProteinMPNN honor the first kind of constraint by clamping residues, but they draw positions in sequence and never revisit one once sampled, so a soft preference at one position cannot propagate back into choices already committed.

Inverse FoldDir keeps the whole sequence in a continuous, mutable state until generation ends. Instead of emitting residue identities, it carries a probability distribution over amino acids at every position and denoises all of them together along a learned Dirichlet flow on the probability simplex. Fixed residues are clamped to exact identities; soft preferences enter as a custom initial Dirichlet distribution at selected positions, biasing where the trajectory starts without dictating where it lands. Because every update conditions on the evolving whole-sequence state, positions co-adapt: across the test set the flow direction reverses at an average of 27% of positions, and about 4% of predicted identities change in the second half of denoising.

It was developed at Stanford University with Microsoft's Office of the Chief Scientific Officer, and posted as a preprint in September 2026 alongside an MIT-licensed implementation and released weights.

#Key Features

• Whole-sequence denoising: All positions update jointly at every step, so residue identities stay revisable until the trajectory ends rather than fixed by an irreversible decoding order.
• Fixed-residue inpainting: User-specified positions are held to exact identities throughout generation, preserving catalytic residues, disulfides or a validated motif while the rest is redesigned.
• Soft residue priors: Selected positions can start from a custom amino-acid distribution — favoring polar or metal-coordinating residues, say — expressing a preference without fixing the position to one identity.
• SE(3)-equivariant structure encoder: A Geometric Vector Perceptron graph network reads the backbone into a residue graph, so conditioning does not depend on how the input structure is oriented.
• Runs without a GPU: The released checkpoint is 38 MB and designs can be generated on CPU.

#Technical Details

Inverse FoldDir pairs the Dirichlet flow-matching formulation of Stärk et al. with a GVP-based SE(3)-equivariant residue-graph encoder, and contains 9.7 million trainable parameters. Interleaved GVP graph-attention layers propagate structural information across the backbone while the model predicts denoised amino-acid probabilities at each residue; auxiliary heads supervising residue charge and steric class serve the training objective only. Training combined the CATH 4.2 training split with 2,824,736 AlphaFold DB structures gated at pLDDT 70, with clusters within a Foldseek E-value of 0.1 of any held-out chain removed in whole. On the 1,120-chain held-out CATH 4.2 test set, generated sequences refolded with ESMFold reached a mean TM-score of 84.5 and a mean Cα RMSD of 1.76 Å, against 83.3 / 1.86 Å for ESM-IF1, 82.7 / 1.89 Å for ProteinMPNN, 81.8 / 1.95 Å for FAMPNN, 79.9 / 2.06 Å for PiFold and 73.0 / 2.52 Å for MapDiff. Ablations trace most of the gain to two ingredients: removing the AlphaFold-predicted training structures drops TM-score to 81.0, and removing iterative denoising drops it to 83.6.

#Applications

The model targets settings where a backbone exists but a sequence does not: sequences for computationally generated backbones, enzyme and interface redesign, fold stabilization, and nanobody engineering. The developers tested it on an X-ray anti-GFP nanobody scaffold (PDB 3OGO) held out of training, designing from the nanobody backbone alone without the antigen present. Thirty-five designs, a wild-type control and four randomly substituted baselines were expressed cell-free and assayed by AlphaLISA against sfGFP. Two designs reproduced binding above background across all measurements at roughly 2.9-fold signal while retaining 57% identity to the native nanobody, and three more cleared the threshold in three of four; the wild type gave a substantially stronger signal than any redesign. Position-wise inpainting also yields mutant-versus-wild-type log-likelihood ratios, competitive in zero-shot positive predictive value across 35 ProteinGym deep mutational scanning assays and agreeing best with stability phenotypes.

#Impact

Inverse FoldDir is a preprint awaiting peer review, and its wet-lab validation is narrow: a five-in-35 result on one nanobody scaffold shows that heavily rewritten designs can retain function, not a success rate for the method. The contribution is structural: a continuous simplex representation buys two things autoregressive decoding cannot offer together, whole-sequence revision and a slot for soft position-level preferences, with differentiable property guidance the next step. The study also reports a useful negative control: Boltz-2 cofolding confidence showed no correlation with measured AlphaLISA signal, with or without an MSA, and no other structure-confidence metric tested ranked the confirmed binders either. Code, documentation, notebooks and pretrained weights are released under the MIT license.