Data and Codes for: Domain adaptation of a SMILES chemical transformer to SELFIES with limited computational resources
Description
This repository supports the manuscript “Domain adaptation of a SMILES chemical transformer to SELFIES with limited computational resources”, providing all necessary data, models, and code for reproducing the reported experiments and figures. It includes two core datasets (SMILES_to_SELFIES.csv and Filtered_QM9.csv) for SELFIES-based domain adaptation and QM9 regression, along with a zip archive (selfies_finetuned_model.zip) containing the final domain-adapted model. Five Jupyter notebooks are provided: Domain Adaptation and Figures.ipynb, QM9 regression: SELFIES DA model.ipynb, QM9 regression: ChemBERTa-77M-MLM model.ipynb, QM9 Regression: ChemBERTa-zinc-base-v1 model.ipynb, and Finetuning on Benchmark Datasets.ipynb. The last of these implements additional finetuning and performance evaluation of the SELFIES-repurposed model on three standard benchmark datasets: ESOL, FreeSolv, and Lipophilicity. Each notebook illustrates the full methodology, from data preparation through model training and evaluation.
Files
Steps to reproduce
1. Domain Adaptation and Figures.ipynb Shows how the original ChemBERTa-zinc-base-v1 model is domain-adapted to SELFIES, as well as the manuscript figures. Link: https://colab.research.google.com/drive/19OKVBugflvfIg_PFfeKclteiU7GypC4r?usp=sharing 2. QM9 regression: SELFIES DA model.ipynb Shows how the SELFIES-Domain-Adapted model was applied to predict 12 QM9 properties. Link: https://colab.research.google.com/drive/1ECwKUutl-eS3jAhRnAeJC1sJepY2X1qF?usp=sharing 3. QM9 regression: ChemBERTa-77M-MLM model.ipynb Outlines the same QM9 regression tasks using the larger ChemBERTa-77M-MLM model. Link: https://colab.research.google.com/drive/1LJIa1LPSpnt8xmI1FAE_6D1NH8b1hHO6?usp=sharing 4. QM9 Regression: ChemBERTa-zinc-base-v1 model.ipynb Provides a baseline comparison by performing QM9 regression with the original ChemBERTa-zinc-base-v1 model. Link: https://colab.research.google.com/drive/1O3nLWSnPaooTW1V7280tBLuwK7SEVVpe?usp=sharing 5. Finetuning on benchmark datasets.ipynb Showcases how the domain-adapted model was further finetuned on ESOL, FreeSolv, and Lipophilicity benchmarks. The fine-tuning protocol follows the same procedure used in the SELFormer paper. Link: https://colab.research.google.com/drive/1fzjWy01l75l9npjmdXV-6siFzxpATUAi?usp=sharing
Institutions
- Khalifa University of Science and Technology