Property-driven localization and characterization in deep molecular representations
摘要
Representation learning via pre-trained deep learning models is emerging as an integral method for studying the molecular structure–property relationship, which is then leveraged to predict molecular properties or design new molecules with desired attributes. We propose an unsupervised method to localize and characterize representations of pre-trained models through the lens of non-parametric property-driven subset scanning (PDSS), to improve the interpretability of deep molecular representations. We assess its detection capabilities on diverse molecular benchmarks (ZINC-250K, MOSES, MoleculeNet, FlavorDB, M2OR) across predictive chemical language models (MoLFormer, ChemBERTa) and molecular graph generative models (GraphAF, GCPN). We further study how representations evolve due to domain adaptation, and we evaluate the usefulness of the extracted property-driven elements in the embeddings as lower-dimension representations for downstream tasks. Experiments reveal notable information condensation in the pre-trained embeddings upon task-specific fine-tuning. For example, among the property-driven elements found in the embedding (out of