GenCDR

An antibody's six hypervariable loops do not sit together in sequence: CDR1, CDR2 and CDR3 are interleaved between four conserved framework regions, so a model that writes residues strictly left to right must commit to CDR1 before it has seen FR2, FR3 or FR4 — the scaffolding that constrains what CDR1 can be. The ordering also blurs scoring: a sequence log-likelihood mixes largely germline-determined framework tokens with the CDR tokens that are the object of design.

GenCDR reorders the input instead of changing the architecture. All four framework regions are presented first as a conditioning prompt, a generation-boundary token marks the hand-off, and CDR1, CDR2 and CDR3 follow as a single variable-length response. Three things fall out of that change. Every CDR is generated with the complete framework context rather than a truncated prefix. Loop lengths are proposed by the model, so a campaign need not specify them in advance. And framework likelihood separates cleanly from CDR likelihood, turning the boundary into a prompt–response interface: a reward computed on a design is attributable to the generated loops alone, which makes preference optimisation on CDRs well posed.

GenCDR was developed by AstraZeneca with the Oxford Protein Informatics Group and released as a preprint in September 2026. It is a family of three checkpoints sharing one architecture and differing in training corpus: IgGenCDR for unpaired chains, p-IgGenCDR for paired heavy–light chains, and NanoGenCDR for camelid VHH nanobodies. It is a distinct model from the same groups' earlier p-IgGen, whose training data it reuses.

#Key Features

• Framework-first prompting: All four frameworks are read before any CDR token is emitted, giving every loop the context that left-to-right decoding withholds.
• Joint variable-length CDR generation: All three loops are produced in one forward pass at lengths the model proposes, rather than filling a pre-specified window.
• Separable CDR likelihoods: Framework and CDR log-probabilities factorise, so likelihoods can be scored over the designed loops alone — a cleaner zero-shot fitness signal than full-sequence scoring.
• Scaffold-grouped weighted DPO: Generation is steered with scalar rewards over CDR response tokens only, comparing samples within a framework group against a frozen reference policy.
• Three antibody formats: One architecture covers unpaired chains, natively paired VH–VL, and single-domain nanobodies.

#Technical Details

All variants are decoder-only transformers following the LLaMA design — rotary positional embeddings, RMSNorm, SwiGLU activations and grouped-query attention — with residue-level tokenisation, 4 layers and roughly 25M parameters. IgGenCDR was pretrained for 10 epochs on about 254M quality-filtered unpaired human sequences from the Observed Antibody Space (≈118M heavy, ≈136M light), with 20% of batches rendered as randomly paired inputs and the numbering scheme sampled uniformly per example across IMGT, Kabat and Chothia. p-IgGenCDR was fine-tuned from that checkpoint on ≈1.8M natively paired VH–VL sequences, and NanoGenCDR on a 1:1 mixture of ≈6.5M non-redundant INDI camelid VHH sequences and human heavy-chain OAS, the human half retained to preserve humanness. On a held-out set built to be disjoint at the CDR3 level, IgGenCDR recovers 91.5% of heavy CDR1, 89.8% of CDR2 and 49.8% of CDR3 residues, exceeding p-IgGen, ProGen2-OAS and IgLM on every loop and beating an identically sized left-to-right ablation on five of six. Likelihoods track experiment: Spearman ρ of 0.227 on FLAb's CDR-driven humanisation assay, and 0.331 and 0.325 on the two NbBench nanobody thermostability sets.

#Applications

The immediate use is CDR grafting — holding a validated or humanised framework fixed and designing compatible loops into it. Because CDR likelihood is separable, the same checkpoints serve as zero-shot ranking functions for expression, immunogenicity and thermostability triage before synthesis. The alignment interface is generic: a team supplies a CSV of sequences with scalar rewards from whatever objective it cares about, and the released code performs the preference optimisation, so in-house assay readouts can steer a public checkpoint without the assay leaving the lab.

#Impact

In the paper's case study NanoGenCDR was aligned over four rounds against a PD-L1 target using a composite reward built from Boltz-2 co-folding confidence, CDR contact fraction, loop geometry and discrete liability rules. Interface confidence rose from 0.404 to 0.530, past the 0.474 of the structure-conditioned BoltzGen pipeline run on the same scaffolds and co-folding budget, with a far higher fraction of the paratope on the CDRs (0.943 versus 0.791) and roughly 36,000 generations needed against BoltzGen's 63,000. The weight of the result is that a sequence-only model with no antigen encoder gets there by learning the target through reward alone. The caveats are real: every design is evaluated in silico with no expression or binding data, the alignment study targets a single antigen with a single offline preference method, and the work is a preprint. The inference code is released under Apache-2.0, but the trained weights had not been posted at the time of writing.