ADFLIP

ADFLIP (All-atom Discrete FLow matching Inverse Protein folding) is a generative model for inverse protein folding — the task of finding an amino acid sequence that adopts a given three-dimensional structure. It was developed by Kai Yi, Kiarash Jamali, and Sjors Scheres at the MRC Laboratory of Molecular Biology and presented at ICML 2025.

Most inverse folding methods reason only about the protein backbone, which leaves out exactly the chemistry that makes many proteins interesting: bound small molecules, nucleic acids, metal ions, and the side chains that coordinate them. LigandMPNN closed part of that gap by conditioning on non-protein atoms, and AlphaFold 3 made all-atom complex structures routinely available as design targets. ADFLIP goes further in two directions. It treats sequence generation as a discrete flow-matching process, so that as residues are decoded their predicted side chains are packed and fed back in as additional structural context. And because that process is iterative rather than single-pass, it can average the denoiser's predictions over several conformational states, letting a single designed sequence be conditioned on an entire structural ensemble rather than one static snapshot.

The model also supports training-free classifier guidance: any pretrained regressor that scores complete sequences can steer sampling toward a desired property, with no retraining of either model.

#Key Features

• All-atom structural conditioning: Sequences are designed against protein backbones together with small-molecule ligands, nucleotides, and metal ions represented at the atomic level, rather than against backbone geometry alone.
• Progressive side-chain context: After each denoising step, the side chains of already-decoded residues are packed with PIPPack and added to the input, so later positions see the physical environment created by earlier choices.
• Ensemble sampling over conformations: The denoiser's predictions are averaged across multiple structural states of the same complex, which is what allows design against NMR ensembles of dynamic proteins.
• Training-free property guidance: External predictors — the paper uses the DSMBind binding-affinity model — reweight the sequence distribution during sampling, raising the fraction of designs that beat wild-type predicted affinity from 41.9% to 58.1% while keeping foldability above 90%.
• Adaptive sampling schedule: A purity score (the maximum probability of the per-position categorical distribution) decides which positions to commit first, spending more sampling steps on ambiguous residues and fewer on constrained buried ones.

#Technical Details

The denoiser is a multi-scale graph neural network with both residue nodes and atom nodes, drawing architectural elements from AlphaFold 3 and LigandMPNN: Fourier and word embeddings in the atom encoder, a local atom attention module with frame averaging, message-passing GNN blocks with diffusion-time modulation, and a three-layer Transformer decoder over residue nodes that emits amino acid logits. Training uses cross-entropy against the clean sequence under a masked discrete flow, following the discrete flow matching formulation of Campbell et al.

Training data follows the LigandMPNN curation protocol: PDB assemblies as of 16 December 2022 solved by X-ray crystallography or cryo-EM at better than 3.5 Å with fewer than 6,000 residues, clustered at 30% sequence identity with MMseqs2, giving 27,818 training clusters. Held-out test sets comprise 317 protein–small molecule, 74 protein–nucleotide, and 83 protein–metal ion complexes. On interaction residues (within 5.0 Å of a non-protein atom), ADFLIP recovers 62.2% of native residues at ligand sites, 50.2% at nucleotide sites, and 75.7% at metal sites, against 59.2%, 46.1%, and 69.3% for a retrained LigandMPNN and 54.5%, 40.3%, and 54.1% for ProteinMPNN. Refolding designs with Chai-1 gives 100% foldability and 1.15 Å RMSD for small-molecule complexes. On a purpose-built benchmark of 219 NMR ensembles averaging 18 states each, conditioning on the full ensemble instead of a single state raises ligand-site recovery from 41.5% to 50.1%.

#Applications

ADFLIP targets design problems where the chemical environment is the point: engineering metalloenzyme active sites, redesigning small-molecule binding pockets, designing DNA- and RNA-binding proteins, and stabilizing or reshaping proteins that cycle through multiple conformational states. Because guidance is training-free, groups with an in-house property predictor — affinity, stability, or expression — can steer sampling with it directly. The code and an ADFLIP_v1.pt checkpoint are available under an MIT license, with a sampling API and a benchmark script for reproducing the reported recovery rates.

#Impact

ADFLIP extends all-atom inverse folding toward structural dynamics rather than a single deposited model, and shows that an iterative generative decoder can exploit side-chain feedback that single-pass designers discard. The gains are incremental at ligand and nucleotide sites, and nucleotide-binding recovery remains the weakest case, which the authors attribute to conformational heterogeneity a single state does not capture. The reported results are entirely computational: no ADFLIP-designed sequence has been experimentally characterized, and foldability is assessed with a structure predictor rather than in the lab. Practical use also depends on obtaining a meaningful conformational ensemble, which for most design targets is not straightforward.