batterdaysahead/molecular-odor-prediction

Odor Prediction Model

Predict odor descriptors and perceptual ratings from a molecule's SMILES string.

What it does:

• Predicts 112 odor descriptors (fruity, floral, woody, sweet, etc.)
• Predicts 3 perceptual ratings (Pleasantness, Intensity, Familiarity)

Results

Task	Metric	Score
Odor Classification	Top-20 Macro F1	0.45
Odor Classification	Mean AUC-ROC	0.84
Pleasantness	Pearson r	0.55
Intensity	Pearson r	0.34
Familiarity	Pearson r	0.44

Quick Start

pip install torch xgboost rdkit numpy

from src.predict import predict

predict("c1ccc(cc1)C=O")  # Benzaldehyde - smells like almonds

Output:

Odor descriptors:
  sweet      : 0.98
  spicy      : 0.98
  almond     : 0.98
  floral     : 0.96
  green      : 0.93

Perceptual ratings (0-1 scale):
  Pleasantness: 0.99
  Intensity   : 0.89
  Familiarity : 0.92

Data Sources

All training data comes from the Pyrfume project:

Dataset	What it contains
GoodScents	~4,500 molecules with expert odor annotations
Leffingwell	~3,200 molecules with odor descriptors
Keller et al. 2016/2012	Perceptual ratings (pleasantness, intensity, familiarity)
MA et al. 2021	Additional perceptual ratings
Arshamian et al. 2022	Pleasantness rankings

Model Details

Classifier (odor descriptors):

• XGBoost with 300 trees
• Morgan fingerprints (2048-bit) from RDKit
• Per-label threshold tuning

Regressor (perceptual ratings):

• PyTorch MLP (1024-512-256 hidden layers)
• Separate models for each rating

Limitations

• Only works on single molecules, not mixtures
• Intensity predictions are weaker (r=0.34)
• Perceptual ratings trained on ~500 molecules only

Citation

@misc{molecular-odor-prediction,
  title={Molecular Odor Prediction},
  author={batterdaysahead},
  year={2026},
  url={https://huggingface.co/batterdaysahead/molecular-odor-prediction},
}

License

MIT