Contrastive Learning Improves Generalization on M2OR Olfactory Binding Dataset

Olfaction-based biosensors have the potential to greatly impact areas like medical diagnostics, national security, and drug discovery. However, predicting how volatile organic compounds interact with specific olfactory receptors poses a challenge due to limited experimental data. Traditional machine learning techniques struggle to generalize well enough to overcome this hurdle.

In this study, researchers examined the Molecular to Olfactory Dataset (M2OR) and Insect Olfactory Binding Protein Database (iOBPdb) datasets, focusing on neural network models and contrastive learning approaches for VOC-OR binding. They explored different model training methods, architectures, and input data structures.

Surprisingly, merely excluding certain chemical or proteomic patterns from the training set negatively impacted the two-layer neural network's performance. But by employing a contrastive learning objective that emphasized shared molecular components and a random sampler, the model achieved state-of-the-art results (0.88 AUCROC) across all data subsets. This indicates that contrastive learning provides a valuable inductive bias for neural networks to accurately predict novel VOC-OR pairs.

The findings reveal that generalization issues in biological interaction datasets often stem from identity leakage caused by overlapping constituent data. Consequently, robust evaluation methods are crucial. Additionally, the study demonstrates that contrastive regularization helps overcome dataset degeneracy by promoting compatibility in the model's latent geometry, potentially benefiting evaluations on other binder target datasets where identity leakage skews results.