Pytorch implementation of “MF-SuP-pKa: multi-fidelity modeling with subgraph pooling mechanism for pKa prediction”
- MF-SuP-pKa is a novel pka prediction model that utilizes subgraph pooling, multi-fidelity learning and data augmentation. Compared with Attentive FP, MF-SuP-pKa achieves 23.83% and 20.12% improvement in terms of mean absolute error (MAE) on the DataWarrior acidic and basic sets, respectively.
AttentiveFP/: the implementation of Attentive FP for pka prediction.Graph_pka/: the implementation of Graph-pKa.MF_SuP_pka/: the source codes of MF-SuP-pKa.data/: the pre-training data set, fine-tuning data set, and external test data sets used in MF-SuP-pKa.model/: the pre-trained model weights of MF-SuP-pKa.prediction/: the prediction results on SAMPL6, SAMPL7, and Novartis data sets by MF-SuP-pKa and other counterparts.
- python 3.6
- rdkit 2021.09.4
- sklearn 0.24.2
- torch 1.10.1
- dgl 0.6.0
- Build graph data set
python build_pka_graph_dataset.py --dataset pka_acidic_2750 --type acid - Train the MF-SuP-pKa model
- Train from scratch
python MF_SuP_pka_model.py --task_name pka_acidic_2750 --type acid --k_hop 2 --stage before_transfer - Fine-tuning the pre-trained model
python MF_SuP_pka_model.py --task_name pka_acidic_2750 --type acid --k_hop 2 --stage transfer --pretrain_aug
- Train from scratch
If you use this code, please cite the following paper:
Wu, J., Wan, Y., Wu, Z., Zhang, S., Cao, D., Hsieh, C. Y., & Hou, T. (2022). MF-SuP-pKa: multi-fidelity modeling with subgraph pooling mechanism for pKa prediction. Acta Pharmaceutica Sinica B.
Links: https://www.sciencedirect.com/science/article/pii/S2211383522004622