AROMMA: Unifying Olfactory Embeddings for Single Molecules and Mixtures

Introduction

This is the pytorch implementation of our ICASSP 2026 paper "AROMMA: Unifying Olfactory Embeddings for Single Molecules and Mixtures". AROMMA (Aggregated Representations of Olfaction via Molecule and Mixture Alignment) is a novel framework that learns a unified embedding space for both single molecules and two-molecule mixtures by leveraging chemical foundation model (SPMM).

To address the label sparsity in the mixture dataset (BP), AROMMA employs a training strategy that combines:

• Knowledge distillation from a molecule-level teacher model (POM), and
• Class-distribution-aware pseudo-labeling.

Setup

Create and activate the conda environment:

conda env create -f environment.yml
conda activate aromma_env

Download trained checkpoints

The trained model checkpoints are available on Hugging Face:

Data	Model	Checkpoints
data/mixture	AROMMA	aromma_best_fold.pt
data/mixture_p78	AROMMA-P78	aromma_p78_best_fold.pt
data/mixture_p152	AROMMA-P152	aromma_p152_best_fold.pt

Training

Before training, download the pre-trained checkpoints following models/pom/README.md and models/spmm/README.md The directory structure should be:

models
├── pom
│   ├── gnn_embedder.pt
│   └── nn_predictor.pt
└── spmm
    ├── checkpoint_SPMM.ckpt
    ├── config_bert.json
    └── vocab_bpe_300.txt

Training Procedure:

1. python train.py --phase aromma
2. pseudo_labeling.ipynb
3. python train.py --phase aromma_p78 or python train.py --phase aromma_p152

Citation