peleke-1

peleke-1 is a suite of protein language models built for targeted antibody design: given an antigen sequence with its epitope highlighted, the models generate candidate antibody variable-fragment (Fv) sequences aimed at that antigen. It was developed by Nicholas Santolla, Trey Pridgen, Prbhuv Nigam, and Colby T. Ford at Silico Biosciences and Tuple, and released as a preprint in October 2025.

Rather than training a bespoke protein model from scratch, peleke-1 adapts general-purpose large language models to the antibody-design task. Each member of the suite is a parameter-efficient fine-tune of a widely used instruction-tuned LLM, teaching the base model to map from an antigen-plus-epitope prompt to plausible antibody Fv sequences. This positions peleke-1 within a growing effort to repurpose frontier language models for biological sequence generation, using curated structural data to specialize them for immunology.

The project is released as an open-source framework spanning training code, the curated antibody-antigen dataset, and multiple ready-to-use model checkpoints, lowering the barrier for others to build on or extend antibody-generation models.

#Key Features

• Antigen-conditioned generation: Produces antibody Fv sequences targeted to a specified antigen, with epitope residues marked in the input using <epi> tags or bracket notation so the model can focus on the intended binding region.
• Suite of base models: Provides three variants fine-tuned from Microsoft's Phi-4, Meta's Llama 3.1 8B Instruct, and Mistral 7B Instruct v0.2, letting users trade off model size and inference cost.
• Parameter-efficient fine-tuning: Uses LoRA via the PEFT library, adapting large base models with a small number of trainable parameters and enabling practical fine-tuning and deployment.
• Open framework and data: Ships merged and quantized (GGUF) checkpoints for local inference alongside the curated training dataset and code, supporting reproducible antibody-design experiments.

#Technical Details

Each peleke-1 model is a LoRA fine-tune of an instruction-tuned language model (Phi-4, Llama 3.1 8B Instruct, or Mistral 7B Instruct v0.2) applied with the PEFT framework. The training data is derived from paired antigen-antibody complexes in SAbDab, curated with the PandaProt tool to identify epitope residues; the resulting dataset pairs antigen sequences (with highlighted epitope residues) to antibody chain sequences. At inference, prompts encode the antigen with epitope residues delimited by <epi> tags, and the model autoregressively generates antibody Fv sequences. The repositories distribute both merged full-precision checkpoints (base model plus LoRA weights and a custom tokenizer) and GGUF builds for local inference. Code is released under GPL-3.0 and the preprint under a CC BY-NC license.

#Applications

peleke-1 is aimed at antibody discovery teams and computational immunologists who want to generate candidate binders against a target antigen in silico, before committing to laboratory synthesis and screening. By conditioning on a specific epitope, it can propose Fv sequences directed at a chosen region of an antigen, providing starting points for downstream filtering, structural modeling, and experimental validation. The accompanying open dataset and training scripts also make it a template for groups that want to fine-tune their own antibody-generation models.

#Impact

peleke-1 illustrates how instruction-tuned general LLMs can be repurposed for a specialized biological design task through parameter-efficient fine-tuning on curated structural data. Its main contribution is practical and open: a reusable framework, dataset, and set of checkpoints for antigen-conditioned antibody generation. As a preprint released under a non-commercial license, its generated antibodies await experimental validation, and its role in real discovery pipelines will become clearer as the community tests the models against measured binders.