ConformAb

ConformAb is a guided discrete diffusion model for antibody sequence lead optimization that explicitly conditions sequence generation on the canonical backbone conformation of the complementarity-determining regions (CDRs). It was developed by Prescient Design at Genentech and released as a bioRxiv preprint in November 2025. The model targets a persistent failure mode in antibody engineering: as candidate sequences are mutated to improve affinity, developability, or other properties, those mutations can reshape the CDR loops — particularly the hypervariable CDR-H3 — and silently abolish binding to the intended epitope.

Rather than predicting structure explicitly or relying on target-specific experimental data, ConformAb steers a sequence-only generative process so that designed variants are predicted to retain the seed binder's CDR canonical loop classes. This makes it well suited to the common industrial scenario in which no repertoire or screening data exists for a given target and only a single lead binder is available as a starting point. The authors position ConformAb as a one-shot optimizer: a single pretrained checkpoint is used at inference to propose variants of a new seed without any per-target fine-tuning.

ConformAb extends the guided discrete diffusion lineage developed at Prescient Design (the NOS / LaMBO family) into the antibody domain, adding conformational awareness as the conditioning signal that keeps generated sequences on the binding-competent region of sequence space.

#Key Features

• Conformation-conditioned generation: Design is guided toward sequences whose predicted CDR canonical class probabilities match those of the seed, preserving the binding-competent loop backbone through mutation.
• One-shot, no per-target fine-tuning: A single pretrained checkpoint is applied at inference to any new seed binder, with no target-specific training data required.
• Wet-lab validated: Designs achieved binding rates of 15-60% across targets using fewer than 100 designs per target, and 3-5x improvements in binding affinity over the seed for two of three targets.
• Structure-validated conformation retention: Experimental structure prediction for binders of two targets confirmed that designs maintain the seed's CDR canonical backbone even with biophysically non-trivial mutations.
• Six per-CDR classifiers: The diffusion model is jointly trained with six CDR-specific canonical-class classifiers that supply the guidance signal during sampling.

#Technical Details

ConformAb is a discrete diffusion model operating directly on antibody sequence, using a masking (absorbing) noise process in which tokens are progressively masked and the network learns to denoise them. The denoiser is trained jointly with six CDR-specific classifiers that predict canonical conformational classes for each CDR loop. At generation time, classifier guidance biases the reverse diffusion trajectory toward sequences whose per-CDR canonical class probabilities match those of the seed binder, so designs remain consistent with the lead's loop geometry without ever explicitly modeling 3D coordinates. Critically, no target-specific or repertoire data is used in training, isolating the optimization signal to general antibody sequence statistics plus the conformational constraint. Validation combined retrospective benchmarks with prospective wet-lab campaigns across three targets, reporting per-target binding rates of 15-60% from fewer than 100 designs and affinity gains of 3-5x for two of the three.

#Applications

ConformAb is aimed at antibody lead optimization in therapeutic discovery, especially the data-poor regime where only a single seed binder is available and no large screening or repertoire dataset exists to train a target-specific model. Discovery teams can use it to propose a small batch of variants predicted to improve affinity while preserving the original epitope engagement, then validate that batch experimentally — turning expensive, large-scale affinity-maturation campaigns into compact, high-hit-rate design rounds. Because it requires no fine-tuning, it can be applied across diverse targets from a single checkpoint.

#Impact

ConformAb addresses a core obstacle in machine-learning-driven antibody engineering: keeping generative sequence design tethered to binding-competent loop geometry without target-specific data. By demonstrating wet-lab binding and affinity improvements from one-shot, conformation-guided designs, it offers evidence that conformational conditioning can substantially raise hit rates in realistic optimization settings. As a 2025 preprint from Prescient Design, its long-term influence is still emerging, and at the time of writing no public code or model weights have been released, which limits independent reproduction and adoption outside the originating group.