This repository is the implementation for the paper RoFt-Mol: Benchmarking Robust Fine-Tuning with Molecular Graph Foundation Models by Shikun Liu, Deyu Zou, Nima Shoghi, Victor Fung, Kai Liu, Pan Li.

We introduce RoFt-Mol, which benchmarks robust fine-tuning methods on molecular property prediction tasks given pre-trained molecular graph foundation models to better understand how to fine-tune the pre-trained models to the downstream tasks robustly under limited samples and potential distribution shifts. We classify eight fine-tuning methods into three mechanisms: weight-based, representation-based, and partial fine-tuning and benchmark these methods on downstream regression and classification tasks across supervised and self-supervised pre-trained models in diverse labeling settings. This extensive evaluation provides valuable insights and informs the design of a refined robust fine-tuning method, ROFT-MOL. This approach combines the strengths of simple post-hoc weight interpolation with more complex weight ensemble finetuning methods, delivering improved performance across both task types while maintaining the ease of use inherent in post-hoc weight interpolation.

Figure 1. The overall framework of fine-tuning strategies evaluated in our benchmark, ROFT-MOL, and the proposed method, DWiSE-FT.

We include total 6 pretrained models and benchmark the fine-tuning methods on top of them. You can refer to the README file inside each pretrained model to see how to run the fine-tuning method in detail for that specific model. Here, the graphium folder refers to both the graphium-toy and graphium-large in our paper.

We utilize the standard MoleculeNet datasets as our downstream task evaluations, we include 8 classification tasks and 4 regression tasks.

The molecule_net datasets can be downloaded from chem data (2.5GB), unzip it, and put it under dataset/ for each pre-trained model. The two other regression malaria and cep datasets can be downloaded following the instruction here.

This repository inherits the code from Mole-BERT, Graphium, GraphGPS, GraphMAE and MoleculeSTM.

@article{liu2025roft,
  title={RoFt-Mol: Benchmarking Robust Fine-Tuning with Molecular Graph Foundation Models},
  author={Liu, Shikun and Zou, Deyu and Shoghi, Nima and Fung, Victor and Liu, Kai and Li, Pan},
  journal={arXiv preprint arXiv:2509.00614},
  year={2025}
}