AffinityFlow

AffinityFlow is a guided-generation framework for antibody affinity maturation — the iterative mutation of a binder to tighten its grip on a target antigen. Developed at Mila and Amazon and presented at ICML 2025, it operates in the setting most groups actually face: antibody and antigen sequences are known, but no co-crystal structure of the complex is available.

The method builds on AlphaFlow, which wraps AlphaFold inside flow matching to turn a single-structure predictor into a sequence-conditioned generator of structural ensembles. AffinityFlow leaves that model entirely frozen and instead intervenes at sampling time. A structure-based affinity predictor supplies a gradient that steers coordinate denoising toward conformations it scores as high-affinity; the resulting structure is then passed to inverse folding, which proposes mutations in the complementarity determining regions, and a sequence-based predictor selects which candidates survive to the next round. Structure generation and sequence mutation alternate for nine rounds, so mutations accumulate a few at a time rather than all at once.

The hard part is training the two predictors, since labeled antibody-antigen affinity data is scarce. AffinityFlow's answer is a co-teaching scheme that mines supervision from computed biophysical energies, which are plentiful but noisy.

#Key Features

• Sequence-only inputs: Neither an experimental complex structure nor a docking pose is required; the antibody and antigen sequences are sufficient to begin.
• Guidance restricted to the CDRs: The affinity gradient is applied only to complementarity determining region coordinates, so the framework scaffold is not distorted during sampling.
• Reciprocal co-teaching: The sequence predictor and the structure predictor each select the samples on which their own prediction agrees with the computed energy label, and pass those to the other, filtering noise that would otherwise be learned as signal.
• Pairwise energy labels: Rather than regressing absolute binding energies, training uses the sign of the energy difference within antibody pairs against a shared antigen — a more robust target than the raw values.

#Technical Details

The sequence-based predictor is built on ESM-2 embeddings of the antibody and antigen; the structure-based predictor combines those embeddings with a geometric vector perceptron over interface residues. Mutations are proposed by inverse folding with ProteinMPNN, introducing one, two, or three point mutations per round. Experiments use single-domain antibodies from SAbDab: 120 labeled sdAb-antigen pairs initialize the predictors, 77 sdAbs against 54 antigens yield 4,158 pairs scored with Rosetta for co-teaching, and 60 sdAb-antigen complexes form the maturation benchmark. Designs are folded with IgFold, docked with GeoDock, and scored with Rosetta; the headline metric is Improvement Percentage (IMP), the fraction of designs with lower binding energy than the parent. Across CDR-H1, CDR-H2, CDR-H3, and the full CDR region, AffinityFlow scores between 88.9% and 93.3% IMP, higher in every setting than protein language model baselines (AbLang, nanoBERT), the discrete walk-jump sampler, and diffusion-based designers (DiffAb, AbDPO, GearBind), while producing more distinct sequences across antigens. Ablations on CDR-H3 are decisive: removing AlphaFlow-guided structure generation drops IMP from 93.3% to 63.3%, and removing the energy-based co-teaching drops it to 66.7%.

#Applications

The target user is a therapeutic antibody program at the lead optimization stage, where a binder exists and the task is to propose a small, testable set of affinity-improving variants. Because AffinityFlow needs no structure of the complex and no retraining of its generative backbone, it can be pointed at a new antibody-antigen pair directly, and the restriction of mutations to the CDRs keeps proposals within the space a developability pipeline will accept. A case study applies it to a single-domain antibody against the SARS-CoV-2 receptor binding domain.

#Impact

AffinityFlow is a clear demonstration that a frozen structure generator plus inference-time guidance can substitute for training a bespoke design model, and its co-teaching module offers a reusable pattern for tasks where computed energies are abundant and measured affinities are not. The caveat is important: evaluation is entirely in silico, with binding energies estimated by Rosetta on IgFold-predicted, GeoDock-docked structures, so reported improvement percentages reflect those oracles rather than measured binding. The paper states an intent to open-source the code after acceptance; no release has followed.