Physics-Inspired Latent Dynamics for Predicting Olfactory Mixture Similarity

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning for olfaction

Background:

• Predicting olfactory similarity in molecular mixtures is challenging due to the lack of universal structure-percept mapping.
• Existing methods require extensive labeled datasets, which are scarce for mixture similarity prediction.

Purpose of the Study:

• To introduce PhysSim, a physics-inspired latent dynamics network for predicting molecular mixture similarity.
• To evaluate PhysSim's performance against established baselines and its zero-shot transfer capabilities.

Main Methods:

• PhysSim models molecular embeddings evolving in a descriptor-initialized latent field.
• It utilizes three distance-dependent functional forms with end-to-end learned scaling constants.
• The network is implemented in PyTorch with 162,059 trainable parameters.

Main Results:

• The core PhysSim model achieved a Spearman correlation of 0.610 on molecule-level cross-validation.
• PhysSim demonstrated strong zero-shot transfer to new datasets (Ravia 2020, Bushdid 2014).
• Performance was specific to mixture similarity, with near-zero correlation for single-molecule odor character prediction.

Conclusions:

• PhysSim offers a promising, data-efficient approach for predicting molecular mixture similarity.
• The physics-inspired inductive bias is effective for mixture-level olfactory perception.
• Further research can explore component-level attributions and refine the model for broader applications.