Chemical Language Model Linker

Running the code

First, install the dependencies from requirements.txt, preferably in a virtual environment. For instance, with conda run

conda create -n chemlml python=3.9
conda activate chemlml
pip install -r requirements.txt

Then, run python train_ChEBI-20.py to train on the ChEBI-20 dataset. Parameters include:

• accumulation_steps: Accumulation steps for gradient accumulation.
• warm_up_steps and lr_factor: Parameters for Noam Optimizer.
• text_encoder: Text encoder for ChemLML, "ChemT5" for encoder from Text + ChemT5, "SciBERT" for SciBERT model and "galactica-" + ["125m", "1.3b", "6.7b"] for different scales of Galactica model.
• molecule_decoder: Molecule decoder for ChemLML, "MolGen" and "MolGen-7B" for default MolGen and MolGen7B model.
• freeze_encoder: To freeze the text encoder or not.
• use_wandb: To use wandb or not.
• wandb_key: Your wandb key.
• skip_valid To skip the validation set or not.
• eval_epoch: Evaluate on validation set every eval_epoch epochs.

Run python test.py to test on different datasets. The only difference is the dataset_name parameters. Select from "ChEBI-20", "PubChem_filtered", or "PubChem_unfiltered". PubChem_unfiltered must be downloaded from Zenodo into the PubChem subdirectory first.

To evaluate the result, run python evaluation/fingerprint_metrics.py and specify the test file.

The pretrained ChemLML models are available on Zenodo.

Citation

Chemical Language Model Linker: blending text and molecules with modular adapters
Yifan Deng, Spencer S. Ericksen, Anthony Gitter
Journal of Chemical Information and Modeling (2025)

Datasets

See the ChEBI-20 and PubChem dataset subdirectories for details and licenses.

• The ChEBI-20 dataset is from MolT5
• PubChem dataset is from Mol-Instructions and PubChem. The unfiltered version is hosted on Zenodo.

Third-party code and models

The result evaluation code fingerprint_metrics.py is from the MolT5 repository, available under the BSD 3-Clause License Copyright (c) 2023, blender-nlp.

ChemLML uses the following models from Hugging Face:

• MolT5: https://huggingface.co/laituan245/molt5-base-smiles2caption
• Text+Chem T5: https://huggingface.co/GT4SD/multitask-text-and-chemistry-t5-base-standard
• MolGen: https://huggingface.co/zjunlp/MolGen-large
• MolGen-7B: https://huggingface.co/zjunlp/MolGen-7b
• Fine-tuned LLaMA2-7B: https://huggingface.co/zjunlp/llama2-molinst-molecule-7b In order to use LLaMA, please refer to the access request at https://huggingface.co/meta-llama/Llama-2-7b-hf. Then copy and paste the huggingface token to line 37 in model.py
• SCIBERT: https://huggingface.co/allenai/scibert_scivocab_uncased
• Galactica series: https://huggingface.co/facebook/galactica-125m
• MolXPT: https://huggingface.co/zequnl/molxpt

See the Hugging Face model cards for licenses, limitations, and citations.