FlowDesign

An antibody's specificity is carried almost entirely by six complementarity-determining region loops, and CDRH3 in particular. Diffusion models made it possible to design those loops as sequence and structure jointly, but they inherit three awkward properties from the formulation. Generation must start from a standard Gaussian, so whatever is known about plausible loop conformations is thrown away; amino acid identity is discrete and has to be forced through a continuous noising process that makes consecutive steps jump; and every sample needs a long denoising chain, prohibitive when a campaign requires tens of thousands of candidates.

FlowDesign replaces diffusion with flow matching, which learns a direct transport between an arbitrary source distribution and the data distribution. Because the source is arbitrary, it can be an informative guess rather than noise — the authors tried structures from the end-to-end antibody model dyMEAN, conformations from kinematic closure, sequence distributions from ESM-2, and random initialisation, finding the data-driven structural prior clearly best. Flow matching also matches discrete distributions directly and reaches a usable sample in far fewer network evaluations.

The work came out of the Institute for AI Industry Research at Tsinghua University with the University of Science and Technology of China and Helixon — the group behind PepMimic — and was published in Cell Systems in 2025.

#Key Features

• Flexible priors: Generation can start from any distribution, so structural predictions, kinematic-closure ensembles, or language model sequence distributions can be injected as prior knowledge.
• Discrete matching: Amino acid types are matched as discrete distributions rather than relaxed into a continuous space, removing the abrupt type changes diffusion shows between consecutive steps.
• Large-scale sampling: Cheaper generation makes it practical to produce tens of thousands of variants and filter them by energy, feeding a yeast display library.
• Sequence-only design: With no crystal structure, the framework can be predicted first — with dyMEAN or AlphaFold 3 — and the CDRs designed on it.
• Released checkpoints: The repository ships a general CDRH3 model and the HIV antibody sampling model used in the wet-lab campaign.

#Technical Details

Training minimises the difference between an antibody-antigen complex from SAbDab and an initial complex drawn from the prior, learning the flow that carries one to the other; side chains are rebuilt with Rosetta packing, and designs are relaxed and scored with PyRosetta. Comparing the four priors on CDRH3, the dyMEAN-derived structural prior gave the highest sequence recovery — close to 15 percentage points above random initialisation — and the lowest average RMSD at 2.291 Å, while the ESM-2 sequence prior underperformed for want of structural information.

Against HERN, RefineGNN, RosettaAb, the diffusion model DiffAb, and dyMEAN, with all CDR data removed and all six loops designed simultaneously, FlowDesign led on amino acid recovery across the CDRs and by the largest margin on CDRH3, where the top 1% of samples by Rosetta energy exceeded 60% recovery. Widening the selection added under 5 percentage points. On CDRH3 backbone accuracy it cut average RMSD by 0.693 Å relative to DiffAb, and the share of designs with lower energy than the native loop rose 13.11% on CDRH1.

The model was then applied to the HIV-1 receptor CD4. After removing antibodies sharing more than 50% CDRH3 identity with ibalizumab and retraining, FlowDesign generated about 50,000 CDRH3 variants on the ibalizumab framework; 12,000 energy-ranked sequences entered a yeast display library, where CD4-binding clones rose from 0.19% to 25.9% to 61.7% across three sorting rounds. Three candidates were characterised by biolayer interferometry, with dissociation constants of 0.19 nM, 0.56 nM, and 0.90 nM against ibalizumab's 0.26 nM, so one designed variant bound more tightly than the parent. Across 11 pseudoviruses spanning five subtypes and tiers 1 through 3, neutralisation IC50 values were comparable in order of magnitude to ibalizumab, with imab-mut-2 beating the wild type on six of the eleven — 0.006 µg/ml against CNE4, versus 0.024 µg/ml.

#Applications

FlowDesign fits antibody engineering workflows that end in a display library: it produces a large, diverse, energy-ranked candidate pool from a fixed framework, which is what high-throughput screening consumes. Affinity maturation of an existing therapeutic is the demonstrated case, and the sequence-only mode extends it to antibodies known only from repertoire sequencing. The same machinery applies to protein loop design generally.

#Impact

The paper's transferable result is that the prior matters as much as the generative formulation: swapping Gaussian noise for a predicted structure improved every metric tested, which reframes structure predictors as initialisation for design rather than only as evaluation tools. The CD4 campaign runs from generation through yeast display to neutralisation assays, at a scale that only became affordable because sampling is cheap. Code is Apache-2.0 licensed with both checkpoints in the repository, though the SAbDab structures and the PyRosetta dependency carry their own terms, and the wet-lab validation covers a single antigen.