peleke-1

peleke-1 is a suite of fine-tuned protein language models developed by Silico Biosciences, purpose-built for targeted antibody sequence generation. By accepting an antigen sequence as input, peleke-1 models generate heavy and light chain Fv sequences designed to bind a specified epitope — making them a practical tool for researchers and computational biologists working on antibody discovery and design. All models, code, and data are fully open-source and flexibly licensed, available via Hugging Face and GitHub.

The peleke-1 series transforms state-of-the-art large language models into Antibody Language Models (ALMs) through a rigorous pipeline of data curation, model fine-tuning, and in silico validation. The result is a collection of accessible, instruction-tuned models that integrate seamlessly with standard LLM frameworks such as Hugging Face Transformers.

Available Models

• peleke-phi-4: Based on Microsoft's 14B parameter Phi-4 model, this variant is optimised to handle large antigen inputs while maintaining a relatively compact model size.
• peleke-llama-3.1-8b-instruct: Built on Meta's Llama 3.1 with 8B parameters, this multilingual instruction-tuned generative model produces antibody sequences through an instruction-style modality.
• peleke-mistral-7b-instruct-v0.2: Derived from Mistral's 7B Instruct v0.2, this model leverages 32k token context windows, which are particularly well-suited for antibody sequence generation tasks.

How peleke-1 Was Built

• Data Curation: Training data was sourced from the Structural Antibody Database (SAbDab), resulting in a curated dataset of over 9,000 antibody-antigen complex sequences with annotated epitope residues.
• Model Tuning: A series of leading LLMs were selected and fine-tuned into ALMs capable of generating Fv sequences given an epitope-annotated antigen sequence as input.
• Output Evaluation: The training approach incorporates in silico validation of generated amino acid sequences through protein folding and docking simulations against the target antigen, ensuring biological plausibility of outputs.

How to Use peleke-1

• peleke-1 models are designed for compatibility with standard LLM tools and frameworks, including Hugging Face Transformers.
• Users provide an antigen sequence with the desired epitope residues annotated as input.
• The model then generates corresponding heavy and light chain Fv sequences targeting the specified binding site.
• Getting started resources are available on GitHub, and all models can be accessed through the Silico Biosciences Hugging Face collection.

The peleke-1 project was developed in collaboration with contributors from UNC Charlotte, NC School of Science and Mathematics, and Tuple. Personnel funding was provided by a NCBiotech Industrial Internship Program Grant, and cloud GPU resources were supported by the Microsoft Most Valuable Professionals program.