Computational & AI Journal of Chemical Information and Modeling (ACS)
Fingerprint-similarity prior makes molecule-language embeddings more chemically faithful
- SAGA is a training-time regularizer that aligns a model's molecular embeddings with fingerprint-based similarity using kernel alignment, adding no parameters or fingerprint computation at inference.
- It was added to two existing molecule-text models and tested retrospectively on retrieval, analogue search, scaffold-split property prediction, captioning and generation.
- Retrieval, analogue quality and average property prediction improved without clear loss elsewhere; control experiments tied the gains to the chemical content of the prior.
Because it costs nothing at inference, SAGA is easy to try, and the control experiments are the most convincing part. Still, Tanimoto-aligned embeddings will naturally score well on analogue search and scaffold-split averages, which reward similarity. Activity-cliff pairs, where fingerprint similarity is exactly the wrong prior, are the test that matters for SAR use.
Paper: Journal of Chemical Information and Modeling (ACS) ↗ Classical Structural Priors Improve Neural Molecular Representations in Molecule Language Models