scientific-deep-chemistry

Predict molecular properties from molecular structures using deep learning by running modular pipelines that train Graph Convolutional Networks, MPNN, and AttentiveFP models on MoleculeNet datasets. The workflow handles everything from data loading to prediction end-to-end.

ChemBERTa and GROVER pretrained representations are supported for transfer learning on downstream chemistry tasks. These embeddings enable you to leverage pretrained molecular knowledge for predicting specific molecular properties.

Running DeepChem graph convolution models requires Python libraries including DeepChem, PyTorch, and Transformers. These dependencies provide the foundational framework for training models like GCN, MPNN, and AttentiveFP on molecular graphs.

MoleculeNet benchmark support and evaluation is integrated to compare model performance across different molecular property prediction tasks. This allows standardized assessment of Graph Convolutional Networks, MPNN, and AttentiveFP architectures.

The best way to train an AttentiveFP model on a MoleculeNet dataset is through the exposed modular workflow, which provides an end-to-end pipeline from data loading to prediction on standard cheminformatics datasets.

Message passing neural networks like MPNN and AttentiveFP offer alternative molecular graph processing mechanisms compared to standard graph convolution. Use them when evaluating different model architectures against MoleculeNet benchmarks to find optimal property prediction performance.