p-IgGen

p-IgGen is a generative language model for designing paired antibody sequences, producing the heavy (VH) and light (VL) variable domains together rather than in isolation. Most antibody language models operate on single chains, but antigen binding and developability are properties of the assembled Fv, where the two chains pack against one another. By generating both chains jointly, p-IgGen captures the pairing statistics of natural antibodies and produces complete, self-consistent variable-region pairs suitable as starting points for discovery campaigns.

The model targets a central bottleneck in antibody drug development: candidate sequences must not only bind their target but also avoid developability liabilities such as aggregation, poly-specificity, poor expression, and low solubility. p-IgGen addresses this by pairing an autoregressive generator with an explicit developability bias. A fine-tuned "developable" variant is trained on sequences whose predicted three-dimensional biophysical properties fall within the distributions observed for clinical-stage therapeutic antibodies, steering generation toward more manufacturable candidates while retaining sequence diversity.

p-IgGen was developed by the Oxford Protein Informatics Group (OPIG) in the Department of Statistics at the University of Oxford together with AstraZeneca, by Oliver M. Turnbull, Dino Oglic, Rebecca Croasdale-Wood, and Charlotte M. Deane, and published in Bioinformatics in November 2024.

#Key Features

• Paired-chain generation: p-IgGen generates VH and VL as a single joined sequence, reproducing the heavy-light pairing patterns of natural antibodies rather than sampling chains independently.
• Flexible conditional sampling: The model can generate full paired antibodies de novo, complete a partial sequence, or generate one chain conditioned on the other (heavy from light, or light from heavy).
• Developability-conditioned variant: A fine-tuned "developable p-IgGen" is trained on sequences flagged as favorable by the Therapeutic Antibody Profiler (TAP), biasing output toward clinical-stage biophysical property ranges while preserving diversity.
• Zero-shot property scoring: Sequence log-likelihoods from the model serve as zero-shot predictors of properties such as immunogenicity and expression, without task-specific training.
• Open tooling: The inference code is released under a BSD-3-Clause license with a command-line interface for generation and likelihood scoring, and the cleaned OAS training data is deposited on Zenodo.

#Technical Details

p-IgGen is a decoder-only autoregressive transformer based on the GPT-2 architecture with rotary positional embeddings, totaling roughly 17.3 million parameters across 3 attention layers, 12 attention heads per layer, an embedding dimension of 768, and a feed-forward dimension of 2048. Sequences are tokenized at the residue level with dedicated start and end tokens. The model was pretrained on unpaired sequences from the Observed Antibody Space (OAS) — approximately 130.2 million heavy-chain and 117.4 million light-chain variable regions — then fine-tuned on about 1.8 million paired VH/VL sequences to learn chain pairing. The developable variant was further fine-tuned on 909,790 TAP-favorable sequences, with developability assessed via metrics including PSH, SFvCSP, PPC, and PNC. On zero-shot benchmarks, p-IgGen reaches a Pearson correlation of 0.53 for immunogenicity prediction and outperforms antibody language-model baselines including AntiBERTy, IgLM, and ProGen-OAS on expression prediction.

#Applications

p-IgGen is aimed at antibody discovery and engineering teams that need diverse, natural-like paired sequences as scaffolds or library members. Because it generates complete Fv pairs, it can seed de novo library design, propose paired partners for an existing single chain, or supply candidate sequences that are subsequently filtered by binding and developability assays. The developable variant is particularly useful for front-loading manufacturability considerations, enriching generated libraries for candidates likely to express well and resist aggregation before any wet-lab screening begins.

#Impact

p-IgGen extends antibody language modeling from single chains to full paired variable regions and demonstrates that a compact model can bias generation toward therapeutically relevant biophysical profiles while maintaining diversity. Its release with open inference code, HuggingFace-hosted weights, a public SAbPred web server, and a Zenodo dataset makes it broadly reusable within the growing ecosystem of antibody-focused generative models such as IgLM and AbGPT. As a sequence-only model, p-IgGen does not predict structure or binding affinity directly, and its developability bias reflects in-silico TAP metrics rather than experimental measurements, so generated candidates still require downstream structural and wet-lab validation.