AntiDIF

An antibody backbone does not pin down a single sequence. Many residue assignments fold into the same CDR loops, and which one a designer wants depends on properties the coordinates never encode: expression titre, thermal stability, immunogenicity, developability liabilities. Yet inverse folding models are graded on sequence recovery — how often they reproduce the native residue — and that metric rewards a model for collapsing onto one confident answer. A designer handed four near-identical predictions has, in practice, been handed one.

AntiDIF treats inverse folding as generative sampling rather than per-position classification. It is a discrete diffusion model: a native antibody sequence is corrupted step by step through amino-acid transition matrices, and a denoising network conditioned on the backbone coordinates learns to run that chain in reverse. Because every generation starts from a fresh noise draw and follows a stochastic trajectory, sampling the same backbone repeatedly yields sequences that differ from one another, rather than temperature-jittered copies of a single mode.

Nikhil Branson and Charlotte Deane built the model at the University of Oxford and presented it at the ICML 2025 Workshop on Generative AI for Biology. It is an antibody-specialised fine-tune of RL-DIF, a general-protein discrete diffusion inverse folding method, and is benchmarked against the antibody-specific AntiFold and AbMPNN as well as the general-purpose ProteinMPNN lineage and ESM-IF.

#Key Features

• Diversity treated as a reported metric: alongside sequence recovery and structural agreement, the work measures the fraction of positions that differ between sampled sequences for the same backbone — a quantity earlier antibody inverse folding papers did not report.
• Discrete diffusion over residue identities: a D3PM-style forward process with uniform transition matrices over a 20-letter amino acid vocabulary, run for 150 timesteps, its reverse posterior parameterised by a learned prediction of the clean sequence.
• Geometric graph denoiser: PiGNN layers operate on a k-nearest-neighbour graph of residues with k = 30, using node and edge features that describe inter-residue distances, angles and directions.
• Antibody-specific fine-tuning: training starts from released general-protein RL-DIF weights and continues on solved antibody structures plus a far larger pool of predicted repertoire structures.
• Refolding as a validity check: sequences are folded back with ABodyBuilder2 and compared to the experimental backbones, so diversity is only credited when the sequences still describe the target structure.

#Technical Details

The denoiser is a 10-layer PiGNN encoder with 128-dimensional node, edge and hidden representations, mapping backbone coordinates, a noised sequence and a timestep to the denoised sequence. Training used the dataset curated for AbMPNN: 3,500 experimentally solved structures from SAbDab and 147,919 predicted structures of OAS repertoire sequences, clustered by CDR similarity and split 80/10/10 with cluster-level separation. Each epoch drew all SAbDab training examples plus a random half that number of OAS examples, using AdamW at a peak learning rate of 1e-3 under a degree-2 polynomial schedule with 1,000 warm-up steps.

Sampling four sequences per test backbone, AntiDIF reaches CDRH3 sequence recovery of 0.598 against 0.570 for AntiFold at its default sampling temperature, while raising CDRH3 diversity from 0.092 to 0.236; on CDRH1 the diversity gap is 0.030 to 0.135. Recovery exceeds AbMPNN, ESM-IF and the RL-DIF starting point across all six CDRs. Raising AntiFold's temperature buys diversity at a cost in recovery, and AntiDIF sits above that trade-off curve. On the 56 test structures resolved better than 2.5 Å, refolding with ABodyBuilder2 gives CDRH3 backbone RMSD of 1.498 Å for AntiDIF versus 1.670 Å for AntiFold, against 0.846 Å for the native sequences.

#Applications

The intended use is the sequence design step of a computational antibody pipeline: given a backbone — a crystal structure, a predicted model, or a de novo designed scaffold — produce a panel of candidates rather than a single suggestion, then filter it on whatever secondary criteria the campaign cares about. Because framework regions are largely germline-encoded and the CDR loops carry binding specificity, a method that varies loop residues while holding the fold constant maps onto affinity maturation and liability-removal workflows. Inference runs from the shipped checkpoint against user-supplied PDB files listing heavy and light chains.

#Impact

AntiDIF makes a narrow but useful argument: sequence recovery alone is the wrong scoreboard for a one-to-many problem, and a generative formulation can widen the candidate set without giving up structural plausibility. That reframing matters more than the margins, since a design team's throughput is bounded by how many distinct sequences it can test. The work remains a preprint that has not been peer reviewed, evaluation is entirely computational — no expressed or assayed antibodies — and diversity is measured as sequence dissimilarity, a proxy for rather than a guarantee of useful spread in biophysical properties. Distribution is the GitHub repository alone, with the checkpoint committed there and no hosted inference.