Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AROMMA: Unifying Olfactory Embeddings for Single Molecules and Mixtures

Dayoung Kang, JongWon Kim, Jiho Park, Keonseock Lee, Ji-Woong Choi, Jinhyun So

TL;DR

AROMMA tackles the challenge of small, fragmented olfactory datasets by unifying representations for single molecules and two-molecule mixtures. It employs a chemical foundation model (SPMM) as the molecular embedder and an attention-based Aggregation Module with a learnable global query to capture asymmetric interactions while ensuring permutation invariance. Knowledge transfer via distillation from a reference odor map and class-aware pseudo-labeling expands the descriptor space, enabling descriptor prediction across both domains. The approach yields state-of-the-art AUROC on GS-LF and BP, demonstrates robust cross-domain generalization, and lays a scalable path toward modeling more complex olfactory mixtures and future 3D-aware representations.

Abstract

Public olfaction datasets are small and fragmented across single molecules and mixtures, limiting learning of generalizable odor representations. Recent works either learn single-molecule embeddings or address mixtures via similarity or pairwise label prediction, leaving representations separate and unaligned. In this work, we propose AROMMA, a framework that learns a unified embedding space for single molecules and two-molecule mixtures. Each molecule is encoded by a chemical foundation model and the mixtures are composed by an attention-based aggregator, ensuring both permutation invariance and asymmetric molecular interactions. We further align odor descriptor sets using knowledge distillation and class-aware pseudo-labeling to enrich missing mixture annotations. AROMMA achieves state-of-the-art performance in both single-molecule and molecule-pair datasets, with up to 19.1% AUROC improvement, demonstrating a robust generalization in two domains.

AROMMA: Unifying Olfactory Embeddings for Single Molecules and Mixtures

TL;DR

AROMMA tackles the challenge of small, fragmented olfactory datasets by unifying representations for single molecules and two-molecule mixtures. It employs a chemical foundation model (SPMM) as the molecular embedder and an attention-based Aggregation Module with a learnable global query to capture asymmetric interactions while ensuring permutation invariance. Knowledge transfer via distillation from a reference odor map and class-aware pseudo-labeling expands the descriptor space, enabling descriptor prediction across both domains. The approach yields state-of-the-art AUROC on GS-LF and BP, demonstrates robust cross-domain generalization, and lays a scalable path toward modeling more complex olfactory mixtures and future 3D-aware representations.

Abstract

Public olfaction datasets are small and fragmented across single molecules and mixtures, limiting learning of generalizable odor representations. Recent works either learn single-molecule embeddings or address mixtures via similarity or pairwise label prediction, leaving representations separate and unaligned. In this work, we propose AROMMA, a framework that learns a unified embedding space for single molecules and two-molecule mixtures. Each molecule is encoded by a chemical foundation model and the mixtures are composed by an attention-based aggregator, ensuring both permutation invariance and asymmetric molecular interactions. We further align odor descriptor sets using knowledge distillation and class-aware pseudo-labeling to enrich missing mixture annotations. AROMMA achieves state-of-the-art performance in both single-molecule and molecule-pair datasets, with up to 19.1% AUROC improvement, demonstrating a robust generalization in two domains.
Paper Structure (10 sections, 5 equations, 3 figures, 2 tables)

This paper contains 10 sections, 5 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Dataset
  4. Overall Architecture
  5. Training Phase
  6. Loss Function and Initial Training
  7. Augmenting Latent Labels and Re-Training
  8. Results
  9. Conclusion
  10. Acknowledgment

Figures (3)

  • Figure 1: Unified embedding space of single molecules and molecule pairs. Each point corresponds to a sample (single molecule or molecule pair), represented by its global embedding from our proposed framework, AROMMA.
  • Figure 2: Architecture of AROMMA. Each molecule is embedded with a chemical foundation model (SPMM) chang2024bidirectional and composed through an aggregation module with a learnable query. Knowledge distillation aligns single-molecule predictions with POM lee2023principal, while the aggregator captures molecular interactions in mixtures.
  • Figure 3: Training strategy of AROMMA. Missing and sparse labels in the BP dataset are augmented using class-distribution-aware pseudo labeling. The enriched datasets (Pseudo-78/152) are then used to re-train AROMMA, improving coverage of odor descriptors and performance.