UMN: Universe Multi Neural Network for Olfactory Mixture Similarity Prediction via Dierentiable N-body Simulation

Read full text Summarise with VIRAT
Authors: Junseong Kim
Cited by: 0
References: 10

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Predicting the perceptual similarity of olfactory mixtures from their molecular composition remains an open problem in computational olfaction. Existing approaches either rely on hand-crafted physicochemical descriptors with manual feature selection or employ graph neural networks that operate on individual molecules without modeling inter-molecular interactions at the mixture level. In this work, we propose Universe Multi Neural network (UMN), a physics-simulation-based architecture that maps molecular fingerprints to celestial bodies with learnable mass, position, and velocity, performs differentiable N-body gravitational simulation, and extracts orbital stability features as mixture-level representations. We evaluate UMN on the Snitz et al. (2013) mixture similarity dataset (360 pairs) using 5-seed × 5-fold cross-validation. UMN achieves a Pearson correlation of r = 0.780 (std = 0.039) without any manual feature engineering, compared to r >= 0.49 for a non-optimized descriptor-based model. Through systematic experiments across 9 versions (v17-v25), we demonstrate that (1) the physics engine architecture is near-optimal in its original form, (2) multi-restart training is the most effective optimization strategy, and (3) contrastive learning, physics-based loss functions, and architectural modifications all degrade performance on this small-data regime.

Fields of Study & Subjects

Computational Drug Discovery Methods Olfactory and Sensory Function Studies Advanced Chemical Sensor Technologies

Publication details

Journal / source
ChemRxiv
Publication date
26 March 2026
Language
EN
Document type
Preprint