DoggifAI

Dogs get many of the same cancers people do, and the human therapeutic antibodies already developed against targets like PD-1, PD-L1, EGFR, and CTLA-4 are in principle reusable in veterinary oncology. The obstacle is immunogenicity: a mouse- or human-derived antibody given to a dog provokes an anti-drug response unless its framework regions are converted to canine ones. Conventional species conversion does this by grafting the CDRs onto the closest matching canine germline framework, a bioinformatics heuristic that treats framework and CDRs as independent parts. They are not — framework residues shape the loops they scaffold, so a mismatched graft can distort the paratope and cost binding.

DoggifAI, developed at the University of Edinburgh with collaborators at the University of Gdańsk and the University of Calgary, reframes caninisation as a conditional generation problem. A T5-style encoder-decoder receives the CDRs of the source antibody as input and generates the four framework regions around them, so the framework is written in the context of the loops it must support rather than selected from a germline catalogue. Because canine antibody sequence data is far scarcer than human, the model is first pretrained by span-denoising on a large multispecies antibody corpus, then fine-tuned on canine sequences alone.

The work's second contribution is that canine corpus. Alongside the model, the authors release over 430,000 unique canine antibody chain sequences, derived from a phage-display scFv library built from pooled dog spleen tissue and sequenced on PacBio long-read instruments — a substantial expansion of a public repertoire previously too thin to support species-specific modelling.

#Key Features

• CDR-conditioned framework generation: The model's input is the extracted CDR set and its output is the framework regions, making framework-CDR compatibility a property the model learns rather than an assumption the grafting heuristic makes.
• Cross-species transfer: T5 span-corruption pretraining on 10 million sequences from the Observed Antibody Space — 92% human, 7% murine, the rest rabbit, rat, and rhesus — supplies general antibody grammar that the small canine dataset cannot.
• Released canine repertoire: 272,535 heavy, 79,545 light kappa, and 81,881 light lambda unique sequences, filtered for productive rearrangements with IMGT High-V-Quest and deposited under CC-BY on Zenodo along with the trained model weights.
• Structure-level validation: Generated and natural sequence pairs were folded with AlphaFold 3, relaxed in GROMACS, and docked to their antigens with Rosetta SnugDock rather than being assessed on sequence recovery alone.

#Technical Details

Four encoder-decoder sizes were compared, from Tiny (2 layers, 4 heads, model dimension 128) to Large (8 layers, 8 heads, model dimension 512, feed-forward dimension 2048, roughly 25M parameters). Pretraining corrupts 15% of tokens in randomly chosen spans; regions are split by IMGT numbering using Abnum. Three regimes were compared — canine-only fine-tuning, pretraining and fine-tuning both on canine data, and OAS pretraining followed by canine fine-tuning — with the third clearly best and re-using the fine-tuning data for pretraining offering little. Greedy decoding outperformed beam search, since minimising cross-entropy does not track the sequence alignment objective the task cares about.

Across 1,000 generated test sequences, the best model reproduced CDRs intact in all but nine cases and failed region extraction in eight, all of them light lambda chains, which are underrepresented at 19% of the training data. Canine-likeness was scored with a reimplemented T20 metric against framework-only canine databases: generated sequences exceeded the 0.80 threshold and generally scored above the natural sequences themselves, except in light lambda. Folding 100 generated sequences and their natural counterparts gave a mean TM-score of 0.91 and mean RMSD of 0.66 Å. Caninised versions of five clinical antibodies all scored above 95 on canine T20; docking outcomes varied — the SARS-CoV-2 spike binder scored clearly above the negative controls though below the original antibody, while the EGFR binder did not separate from the negatives.

#Applications

The direct use is veterinary oncology, where converting an existing human or murine therapeutic to a canine-compatible format is cheaper than discovering a canine antibody de novo, and where the resulting treatment also serves comparative oncology studies that use naturally occurring canine tumours as models of human disease. The released sequence dataset is independently useful for anyone training canine-specific antibody models.

#Impact

DoggifAI shows that species conversion can be posed as CDR-conditioned generation, and that pretraining on abundant human and murine repertoires transfers well enough to make a data-poor species tractable — a template that extends to feline, equine, and other veterinary targets. The released canine dataset is likely the more durable contribution, since it removes the bottleneck that made canine-specific modelling impractical. The limits are honest ones: validation is entirely computational, with docking results the authors themselves flag as inconclusive for several targets and no expressed antibodies or measured affinities, light lambda chains remain the model's weak point, and the code repository carries no license.