DiffAbOpt

Therapeutic antibodies are almost never engineered in one shot. A lead is carried through successive rounds of design and in vitro screening, each round mutating the best molecules from the last. Three-dimensional structure is a powerful inductive bias for that mutation step, but it is rarely used in practice: an experimentally determined antibody-antigen complex usually exists only for the starting molecule, not for the evolving leads that follow.

DiffAbOpt, developed at BigHat Biosciences, is a strategy for putting a sequence-structure generative model inside that loop anyway. The team retrained DiffAb, a diffusion model that jointly generates the sequence and backbone geometry of antibody complementarity-determining regions (CDRs) conditioned on an antibody-antigen complex. At each design round, the current top performers are folded with an antibody structure predictor, rigidly aligned back onto the original crystal complex to preserve the binding pose, partially noised across the CDRs, and denoised into new candidates. DiffAbOpt also folds campaign assay measurements directly into sampling through a product-of-experts guidance term.

This sets it apart from single-shot structure-based designers such as RFdiffusion and ProteinMPNN, none of which are built to absorb experimental data accumulated over multiple rounds. The work was published at the GEM workshop at ICLR 2025 and posted to arXiv that September.

#Key Features

• Predicted complexes as the design substrate: Evolving leads are folded and aligned to the starting crystal pose, so structure-conditioned generation continues after the last experimental structure runs out.
• Round-over-round noise-and-denoise: Applying 8 noising steps across all six CDRs at once, then denoising for 8, gave the best yield of variants within a 4-edit budget.
• Oracle-guided sampling: A product-of-experts distribution reweights the residue-type denoising posterior by an oracle fit to the campaign's binding data, factorized per-position so it can be sampled exactly.
• Multi-property steering: Guiding with a polyreactivity oracle lowers predicted polyreactivity at comparable affinity gains, trading off two objectives otherwise in tension.
• Wet-lab validation in an active campaign: Designs were synthesized as scFvs by cell-free protein synthesis and affinities measured by biolayer interferometry at two stages of a live optimization program.

#Technical Details

The generator follows the DiffAb formulation: a denoising network predicts residue types (multinomial), Cα positions (Gaussian), and residue orientations (isotropic Gaussian over SO(3)) for the masked CDR region, conditioned on the unmasked framework and antigen context. BigHat retrained it on a SAbDab snapshot carrying roughly 30% more data than the original release, filtered to protein antigens and structures better than 4 Å and clustered at 50% HCDR3 identity; substituted an expanded CDR definition taking the union of Chothia, IMGT, and Kabat indices; and corrected a variance term in the residue-type denoising posterior, reweighting the loss to λ_type = 10. Training ran 200,000 Adam steps at a learning rate of 1e-4 on 128-residue structural patches, and ablations attribute a lift in top-ranked design quality to each change.

Two oracle families supply the lab signal. Ranking uses an ensemble of ten ByteNet CNNs in the CARP family, pretrained on about 75,000 in-house NGS sequences and fine-tuned on the 4,000-8,000 affinity measurements available at a given round; guidance uses a lightweight ridge regression for speed, with reweighting factor γ = 2. A random forest trained on 250 flow-cytometry measurements serves as the polyreactivity oracle (γ = 10). In vitro, mid-campaign designs from ranked DiffAbOpt (22 of 22 synthesized) outperformed a ranked sequence-only masked-language-model baseline (66 of 72 synthesized); late-campaign, ranked designs clearly beat unranked ones. Guidance gave only a marginal, statistically non-significant affinity improvement in vitro. Predicted antibody structures proved nearly as good an input as the crystal structure, while perturbing the binding pose or dropping the antigen degraded designs substantially.

#Applications

DiffAbOpt targets antibody engineering programs that begin with a solved antibody-antigen complex and then run several rounds of affinity maturation and developability optimization. It fits teams already operating a design-build-test loop who want their structural prior to keep contributing past round one, and it offers a mechanism for pushing a specific liability — polyreactivity, in the reported experiments — down without giving up binding. The recipe extends to any property with enough campaign assay data to fit a sequence-level regressor.

#Impact

The contribution is less a new architecture than a demonstration that structure-conditioned generative models can be operated inside an active therapeutic campaign, with wet-lab readouts at multiple stages rather than a single retrospective benchmark. Its limits are worth stating plainly: the method assumes the binding pose is preserved as the lead evolves, the guidance layer is fit to one campaign's assay data and does not transfer, the in vitro cohorts number in the tens of molecules, and the benefit attributable to guidance specifically was not statistically significant. Neither code nor trained weights have been released, so the transferable artifact is the design pattern rather than a downloadable model.