DoggifAI: a transformer based approach for antibody caninisation

Dominik Grabarczyk, Mikołaj Kocikowski, Maciej Parys, Douglas R. Houston, Ted R. Hupp, Javier A. Alfaro, Shay B. Cohen

Abstract

Antibody translation across species offers a compelling strategy to extend the vast and expensive investments in human therapeutic antibodies to veterinary oncology, with applications in both veterinary medicine and comparative oncology. While precise, low-immunogenic treatments are essential for canine cancer care, traditional species conversion methods rely on ad hoc bioinformatics modifications. These methods often implicitly decouple the framework (FR) and complementarity-determining regions (CDRs), ignoring how structural changes in FRs can affect the conformation and function of CDRs. This can compromise binding specificity and require costly high-throughput in vitro screening. To address this, we present DoggifAI, a transformer model that translates non-canine antibody sequences into canine ones by generating species-appropriate framework regions (FRs) based on desired CDRs. This allows the model to better preserve structural compatibility between FRs and CDRs. The model is pretrained in a T5-style text-to-text denoising task on a large multispecies antibody dataset, which allows further finetuning on a much smaller species-specific dataset. DoggifAI generates highly canine-like antibodies and shows promising results in preserving binding specificity. To support further progress in this field, we also release a curated dataset of over 430,000 unique canine antibody chain sequences, significantly expanding the public sequence repertoire.

Introduction

The "gold rush" of human immunotherapy has produced transformative treatments for cancer and autoimmune diseases. Extending these benefits to veterinary medicine, particularly for canine patients, requires a process known as caninisation. Simply injecting human antibodies into dogs triggers an immune response against the "foreign" protein, rendering the treatment ineffective or dangerous.

Traditional caninisation involves "CDR grafting"—swapping human structural framework regions (FRs) for canine ones while attempting to preserve the complementarity-determining regions (CDRs) that bind to the target. However, this method often fails to account for the structural interplay between the framework and the binding loops. If the new framework does not perfectly support the CDRs, the antibody may lose its binding affinity. DoggifAI addresses this by treating antibody design as a contextual generation task, ensuring structural compatibility between the generated framework and the specific binding loops.

Methods and Model Architecture

DoggifAI utilizes a Transformer architecture, specifically a T5-style text-to-text model, to learn the underlying "grammar" of antibody sequences. The pipeline consists of three primary stages: pretraining, finetuning, and application.

The DoggifAI pipeline, encompassing pretraining on multi-species data, finetuning on canine sequences, and therapeutic application.

Training Objectives

The model was trained using a two-step approach:

  1. Pretraining (Denoising): The model was first trained on a subset of 10 million sequences from the Observed Antibody Space (OAS) database. It performed a denoising task, reconstructing sequences where 15% of the amino acids were stochastically masked. This taught the model universal structural rules across species.
  2. Fine-tuning (Construction): The model was specialized using a new, curated dataset of 425,284 unique canine antibody sequences. In this stage, the task shifted to generating appropriate canine FRs given only the CDR sequences as input.
Sequence-level view of the pretraining (PT) denoising task and the finetuning (FT) framework construction task.

Key Insights and Scaling

A critical finding of this research is the importance of contextual generation. Unlike rule-based swapping, DoggifAI "sees" the CDRs and generates a framework tailored to them. Furthermore, the researchers observed that scaling laws apply to antibody models; the "Large" variant (approx. 25M parameters) consistently outperformed smaller versions (Tiny, Small, Medium) in sequence recovery and species-likeness.

Results

  • Species Likeness: The generated sequences achieved cT20 scores >95, indicating they are indistinguishable from (or even more "canine-like" than) natural canine antibodies.
  • Structural Fidelity: In silico validation using AlphaFold3 demonstrated high structural preservation, with a TM-score of 0.91 and a low RMSD of 0.66 Å compared to natural counterparts.
  • Binding Affinity: Simulations using Rosetta Snugdock on five major therapeutic targets (including PD-1 and EGFR) showed that DoggifAI-generated antibodies maintained promising binding affinities, suggesting successful functional translation.

Limitations and Future Work

While these in silico results are promising, the primary limitation is the lack of "wet-lab" experimental validation. Future work will focus on biological testing to confirm the safety and efficacy of these candidates in vivo. Additionally, the model showed slightly lower performance on light lambda chains due to data imbalance in the pretraining set, an area for future optimization through rebalanced training data.

Summary updated

Abstract

Antibody translation across species offers a compelling strategy to extend the vast and expensive investments in human therapeutic antibodies to veterinary oncology, with applications in both veterinary medicine and comparative oncology. While precise, low-immunogenic treatments are essential for canine cancer care, traditional species conversion methods rely on ad hoc bioinformatics modifications. These methods often implicitly decouple the framework (FR) and complementarity-determining regions (CDRs), ignoring how structural changes in FRs can affect the conformation and function of CDRs. This can compromise binding specificity and require costly high-throughput in vitro screening. To address this, we present DoggifAI, a transformer model that translates non-canine antibody sequences into canine ones by generating species-appropriate framework regions (FRs) based on desired CDRs. This allows the model to better preserve structural compatibility between FRs and CDRs. The model is pretrained in a T5-style text-to-text denoising task on a large multispecies antibody dataset, which allows further finetuning on a much smaller species-specific dataset. DoggifAI generates highly canine-like antibodies and shows promising results in preserving binding specificity. To support further progress in this field, we also release a curated dataset of over 430,000 unique canine antibody chain sequences, significantly expanding the public sequence repertoire.
BibTeX
@inproceedings{grabarczyk2025doggifai,
  author = {Dominik Grabarczyk and Mikołaj Kocikowski and Maciej Parys and Douglas R. Houston and Ted R. Hupp and Javier A. Alfaro and Shay B. Cohen},
  title = {{DoggifAI: a transformer based approach for antibody caninisation}},
  booktitle = {ImmunoInformatics},
  year = {2025},
  doi = {10.1016/j.immuno.2025.100064},
}
Figure