Build an ML Model That Can Smell

How it works

1. 01

   Molecule → graph

   A molecule is treated as a graph, with atoms as nodes and bonds as edges.

2. 02

   Graph → fingerprint

   A message passing neural network (MPNN) learns a fingerprint of the molecule.

3. 03

   Fingerprint → smell

   It predicts odor labels like fruity, woody or minty, with probabilities.

Repo, paper & data

Setup guide

Requires Python 3.10. The first two commands use conda to create an environment with it. Run each command on its own, in order.

1. Create the environment

• $ conda create -n open_pom python=3.10Copy
• $ conda activate open_pomCopy

2. Install the libraries

• $ pip install --pre deepchemCopy
• $ pip install torch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 --index-url https://download.pytorch.org/whl/cu124Copy
• $ pip install dgl -f https://data.dgl.ai/wheels/torch-2.4/cu124/repo.htmlCopy
• $ pip install openpomCopy

No GPU? Edit the device_name setting in predict_odors.py (it's 'cuda' by default) to 'cpu' to load the models on your CPU.

Try it

• $ git clone https://github.com/BioMachineLearning/openpom.git && cd openpomCopy
• $ python predict_odors.pyCopy

Enter a SMILES string like CCO (ethanol) to see its predicted smells as a bar chart.