UAM

Official implementation for paper "Uncertainty-aware Yield Prediction with Multi-modal Molecular Features".

Requirements

Required packages and recommended version:

python (>=3.10)
pytorch (>= 1.12.1_cuda11.3)
transformers (>=4.37.2)
rdkit (>=2023.9.5)
torch-geometric
torch-scatter
mixture_of_experts
wandb
...

Datasets

We use three datasets, two of which are High-throughput experiment (HTE) datasets and the third one is constructed from patent literature by expert chemists. For the HTE dataset, we provide the raw data under ./data. For the ACR dataset, pls check here

Pre-training

We use InfoNCE loss to pre-train the SMILES encoder and the Molecular Graph encoder. Pls run the following command:

python pretrain.py

Training & Evaluation

To train the model in the paper, run this command:

python train.py

Note that the original model is implemented using DGL and customized Transformer. While in this repo, we use PyG and Pytorch.transformer_encoder. So the best hyper-parmeters will be slightly different.

You can use following command to find the best configurations.

python sweep.py

References

If you find this repository useful in your research, please cite the following paper:

@inproceedings{chen2024uncertainty,
  title={Uncertainty-Aware Yield Prediction with Multimodal Molecular Features},
  author={Chen, Jiayuan and Guo, Kehan and Liu, Zhen and Isayev, Olexandr and Zhang, Xiangliang},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2024}
}