GeoFlow-V2

Structure prediction and generative protein design have advanced on separate tracks. Co-folding models like AlphaFold 3 and its open reimplementations predict a structure from a complete sequence, while diffusion designers like RFdiffusion generate backbones and hand the sequence problem to an inverse-folding model. Repurposing a predictor for design works, but only against the grain of its architecture.

GeoFlow-V2, released as a technical report in May 2025 by BioGeometry — a protein design company founded by Jian Tang, a core academic member of Mila — collapses that distinction. The insight is that both tasks are conditional generation over the same joint sequence-structure object: prediction inpaints coordinates given a full sequence, and design inpaints both coordinates and residue identities given a masked one. GeoFlow-V2 implements this with a single atomic diffusion process that spans proteins, nucleic acids, and small molecules, so the same checkpoint folds an antibody-antigen complex or generates a nanobody against a specified epitope depending only on how much of the input is masked.

Its second distinguishing feature is controllability. The model accepts explicit epitope constraints, pairwise contact constraints, and apo or holo target structures as conditioning, which lets experimental knowledge — epitope mapping, alanine scanning, a solved target structure — steer both folding and design.

#Key Features

• Masking unifies prediction and design: Masked residues become "UNK" pseudo-residues carrying only the four backbone atoms, so a complete sequence yields structure prediction and a partially or fully masked one yields de novo design with no architectural switch.
• Bidirectional sequence-structure flow: A sequence co-design module predicts amino acid identities for masked regions at each diffusion step, and a structure-conditioning module encodes noise-perturbed target distance maps, letting the model take either modality as input and emit both.
• Constraint-guided docking: Supplying four epitope constraints lifts the top-1 acceptable antibody-antigen success rate from 45% to 75%, and holo-antigen conditioning is the strongest setting for high-accuracy (DockQ > 0.80) predictions.
• A distilled antibody-specialist variant: GeoFlow-V2-ab trims the input embedding stack from 48 layers to 8 and runs 150-250x faster than AlphaFold Multimer v2.3 while matching or beating ABodyBuilder3 and NanoBodyBuilder2 on CDR accuracy.
• Design recycling for in-silico screening: Because designed sequences and structures can be fed back as inputs, a second folding pass produces an independent confidence score for ranking candidate binders.

#Technical Details

The architecture and training recipe follow AlphaFold 3, with two configurations: a benchmarking model with a 2021-09-30 training cutoff, and a production model trained through 2024-06-30. Epitope and contact constraints are encoded by Gaussian smearing over six distance bins (6-20 Å and 6-30 Å respectively); structure conditioning bins a noise-perturbed distance map into 0-4, 4-8, 8-16, and >16 Å. Each constraint type is active with 20% probability during training to prevent over-reliance.

On 104 low-homology antibody-antigen complexes released after the training cutoff, GeoFlow-V2 reaches a 45.19% top-1 DockQ success rate, ahead of Protenix, Chai-1, Boltz-1, AlphaFold-Multimer v2.3, and GeoFlow-V1. On the PoseBusters ligand set it achieves 77% success at the 2 Å pocket-aligned RMSD threshold, benchmarked against Chai-1, AlphaFold 3, and RoseTTAFold All-Atom. The antibody variant was pretrained on curated SAbDab structures, then retrained from scratch on a 1:1 mixture of SAbDab and 1.86 million distilled OAS paired-antibody predictions filtered at a confidence threshold of 0.70, and evaluated on 285 held-out antibodies and nanobodies.

#Applications

The intended workflow is therapeutic antibody and nanobody discovery: specify a target structure and hotspot residues, generate a library of CDR-grafted binders on a standard humanized framework such as Trastuzumab or h-NbBcII10FGLA, then rank them by folding confidence before committing to synthesis. The report also demonstrates binder design conditioned on other modalities — flavin-binding proteins, an OKT3-masking peptide, and a competitive binder against NhaR-mediated transcriptional activation. The ultra-fast antibody variant suits high-throughput screening and structure-based developability assessment where per-structure runtime dominates.

#Impact

GeoFlow-V2 is a practical demonstration that a co-folding architecture can be trained natively as a generative designer rather than coaxed into one, and its constraint interface is a useful template for injecting wet-lab knowledge into diffusion models. Several caveats bound the claims. The report is a preprint that has not undergone peer review, and the de novo design results are in-silico only: libraries were synthesized as oligo pools, but experimental validation is deferred to a future update. The model cannot yet generate nucleic acids or small molecules, only condition on them, and as a pure all-atom model without frame-based inductive biases it occasionally produces ligands with incorrect chirality. Neither code nor weights have been released; access is through the prot.design web server for non-commercial research use only.