Fragment-level graphs match atom-level molecular accuracy on smaller inputs
Original title: Fragment-Level Graph Encoding for Interpretable Molecular Property Prediction
Integrated Gradients recovered known pharmacophores at 1.5 times chance on BACE and 3.2 times on PTR1.
Researchers encode molecules as BRICS fragment graphs so message-passing GNNs attribute predictions to functional groups rather than atoms. Each node’s features are the element-wise sum of its atoms’ features, so existing GNNs and post-hoc explainers can be reused. Across four architectures, four tasks, and two explainers under scaffold splits, test accuracy matched atom-level models on roughly five-fold smaller graphs. Code is published on GitHub.
- Fragment-level training produces one attribution convention without post-hoc aggregation.
- On BACE, atom-level aggregation both beat and missed the fragment hit rate depending on two choices.
- No new model architecture is introduced; only the input representation changes.
From the original article
1 min read at arxiv.orgGraph neural networks (GNNs) are among the standard tools for molecular property prediction, but the atom-level explanations of their predictions are hard for medicinal chemists to act on.
By Rafael Lopes Almeida · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.