Contrastive Learning Improves Generalization on M2OR Olfactory Binding Dataset

Contrastive learning on VOC-OR binding data using M2OR and iOBPdb achieves 0.88 AUCROC on every data split, despite motif-holdout tests that reduce performance for a two-layer neural network.

• Why it matters: Engineering vapor-phase biodetection depends on predicting VOC-OR binding, but existing models generalize poorly and evaluations can be inflated by identity leakage from overlapping molecular constituents.
• What they did: The authors evaluated data partitioning and neural network versus contrastive learning frameworks while varying training, architecture, and input distribution, using a contrastive objective based on shared molecular constituents with a stochastic sampler.
• The result: Holding out certain chemical or proteomic motifs diminished two-layer neural network performance, while the contrastive objective maintained state-of-the-art results across splits, suggesting a strong inductive bias and more robust evaluation.

AI summary of the abstract · may contain errors · not medical advice · read the paper.