Answers are grounded in this paper and cite their sources.
Predicting interactions between odorant ligands and odorant receptor (OR) proteins is hindered by the scarcity of experimental structures and the wide sequence divergence within the OR family. We introduce ScentOR (Semantic Chemical Encoder via Neural Transfer for Odorant Receptors), a heterogeneous knowledge distillation framework that transfers the predictive distributions of a gradient boosting (GB) teacher into a neural network (NN) student using only one-dimensional (1D) semantic representations, achieving an area under the receiver operating characteristic curve (AUROC) of 0.954 and a Matthews correlation coefficient (MCC) of 0.602 on the curated M2OR data set of 39,723 odorant ligand-OR protein pairs, matching the current best-performing model while reducing the run-to-run standard deviation approximately 3-fold. A systematic comparison of eight scenarios shows that statically predicted three-dimensional (3D) topologies, encoded as geometry alone and at the accuracy currently attainable for ORs, add no complementary value: student models assign them near-zero importance, and the small gain they produce without distillation does not survive it. We further characterize what post hoc attribution reports for a model of this design, showing that integrated gradients (IG) and in silico alanine mutagenesis concentrate on the family-conserved DRY motif arginine yet carry no pair-specific information, even though pair-specific interaction dominates the predictions.
No takes yet. Share an insight, caveat, or question.
Kim et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: