Exploring Machine Olfaction

Dataset

This study investigate machine learning for modeling continuous descriptor applicability from molecular representations. Two prediction tasks based on psychochemical olfactory dataset are of interests, which mainly differ in human participants' domain knowledge level in olfaction field:

• Dravnieks dataset is gathered on panelists with special olfactory training and selection.
• Keller dataset collect individuals odor descriptor ratings for each chemical from naive healthy human subjects.
• Further information please check data directory.

Experiment Code

Modeling Descriptor Applicability

Machine Learning Models gridsearch_ml_mordred.py fits descriptor-specific models (Support Vector Machine, K Nearest Neighbors, Random Forest, Gradient Boosting) in both prediction tasks.

Graph Neural Networks train.py trains proposed Graph Neural Networks on prediction task of interest.

Embeddings as feature input

gridsearch_ml_embeddings.py maintains a similar structure and functionality as gridsearch_ml_mordred.py with feature inputs being GNN-learned embeddings.

Transfer Learning

• transfer_ml.py applies descriptor-specific best-performing models fitted on Dravnieks dataset to predict descriptor applicability of matched odor labels in the Keller prediction task.
• transfer_dl.py fine-tuned pre-trained GNN model from one prediction task to learn and predict on another prediction task, where matching predicting targets are not required.

Results and Visualization

All results are stored in results directory. Visualization for the thesis is generated in results/performance.ipynb notebook and output figures are in results/figs for reference.