LinkLlama: chemically reasonable linker design with large language models

---

Overview

LinkLlama fine-tunes a Llama-class model to propose linkers between fragments using prompts with geometry and property constraints.

Training data are built from ChEMBL. For the data pipeline and LoRA setup, see the Retraining guide.

---

Usage

Prerequisites

Conda is recommended.

Installation

git clone https://github.com/THGLab/LinkLlama.git
cd linkllama
conda env create -f environment.yml
conda activate linkllama
pip install -e .

Optional geometry benchmarks: pip install -e ".[benchmark]" (installs spyrmsd).

Inference

• Checkpoint: THGLab/Llama-3.2-1B-Instruct-LinkLlama-Cap50 on the Hugging Face Hub.
• Default config: data/inference_config.yaml — sampling.model_path is set to that Hub id (same string as AutoModelForCausalLM.from_pretrained(...)).
• Auth: use huggingface-cli login or HF_TOKEN if the Hub requires it.
• Docs: Inference guide.
• Tiny example: data/zinc_minimal.csv plus the same YAML; see data/README.md.

Retraining

Build JSONL with the linkllama/llm pipeline, then run Axolotl LoRA from YAML under linkllama/training/.

Retraining guide — full steps and options.

Pretrained weights (Hub)

Resource	Link
Model weights	THGLab/Llama-3.2-1B-Instruct-LinkLlama-Cap50
Training JSONL	THGLab/LinkLlama-cap50-train

Benchmark data and processed ChEMBL (Figshare)

Benchmark splits and processed ChEMBL files for this project live on Figshare.

DOI: 10.6084/m9.figshare.32049072

Extract: tar -xzf <archive>.tar.gz → top-level folder linkllama_data/ with:

• 1k splits: 1k_hiqbind/, 1k_hiqbind_hard/, 1k_zinc/, 1k_zinc_hard/ (CSVs plus SDFs where used).
• ChEMBL-derived (repo root of linkllama_data/): chembl36_balanced_cap50.csv, chembl36_balanced_distinct_linkers.pkl, chembl36_cleaned.smi.

---

License

See LICENSE (UC Regents). Released model checkpoints follow Meta Llama and Hub terms where applicable.

---

Citation

If you use this work, please cite the bioRxiv preprint.

@article{sun_linkllama_2026,
	title = {{LinkLlama}: {Enabling} {Large} {Language} {Model} for {Chemically} {Reasonable} {Linker} {Design}},
	author = {Sun, Kunyang and Wang, Yingze Eric and Purnomo, Justin Clement and Cavanagh, Joseph M. and Alteri, Giovanni Battista and Head-Gordon, Teresa},
	year = {2026},
	doi = {10.64898/2026.04.15.718690},
	url = {https://www.biorxiv.org/content/10.64898/2026.04.15.718690v1},
	journal = {bioRxiv},
}