Fragment-level graph encoding makes molecular property prediction interpretable at functional-group resolution

A study proposes representing molecules as fragment-level graphs built by BRICS decomposition, so that message-passing graph neural networks and post-hoc attribution methods can explain property predictions at the resolution of functional groups rather than individual atoms. Under scaffold splits, the representation keeps test accuracy comparable to atom-level graphs while shrinking graph size to roughly one-fifth. Attribution quality also improves: Integrated-Gradients hit rates for known pharmacophores of BACE and PTR1 reached 1.5 and 3.2 times the random level, respectively. The work suggests a practical route to models that are both accurate and interpretable for drug-discovery applications, though broader validation on other targets and tasks is still open.

Written by a model from the reports · updated 4 hours ago