State-of-the-art molecular embeddings don't reliably capture important chemical properties better than classical fingerprints, and representation collisions affect billions of molecules in screening databases—highlighting a gap between benchmark performance and real-world utility.
This paper systematically evaluates molecular representations—fingerprints, embeddings, and descriptors—used in drug discovery and chemistry AI. The researchers found that modern learned embeddings don't outperform simpler methods at capturing molecular structure, 3D shape, or drug properties.