Experiments on Open Principal Odor Map - With intensity mappings

Please find the original readme from the main openpom repo.

All my work is present in openpom_playground.ipynb. Please bear with me, as it's not too organized at the moment. But you can find a high-level view of what's been done below.

The purpose of this Readme is to track my progress with different experiments to understand the inner-workings of the POM model hopefully towards narrowing down on my Main Goals. I'm also using this project as a way to improve and sharpen my skills in applied Statistics, so any feedback would be greatly appreciated!

Main goals:

Creating a systematic way to create an ‘engine’ capable of digitally conceptualizing odours (with the eventual goal of synthesis)

• Find a reliable minimum value for ‘n’ such that ‘n’ separate odour dimensions can represent the entire odour space.
• Find a chemical compound/mixture for each of these n dimensions.

QSOR relationship between compounds and odours:

We rely on the POM GNN model that seems to have shown promising abilities in mapping a QSOR between chemical structure of compound to some corresponding perception labels. Also, the dataset used to train this model appears to cover a large spectrum of odour labels in the human scent experience. Hence, we are assuming that this model has created a comprehensive understanding of odours. I’m working on work done from the openpom port/replication of the POM paper.

Testing out the setup and visualizing model learnings

• Visualizations of embedding space of different predetermined ‘primary odour’ categories to verify the learnings of the openpom port of the POM research paper. Here’s the Crocker Henderson categorization:

Odour Intensity

• In order to derive a ‘hex’ of ‘n’ minimum odours, odour ‘intensity’ data seems to be a valuable missing component. The idea is that, only with intensities can you hope to represent any odour as a function of varying intensities of these ‘n’ base odours. For this,I have explored a small ‘odour intensity’ dataset that captures the intensity of different odours from 24 separate classes against the name of the compound. Here’s a heatmap of the intensity of Phenolic odour from the Intensity dataset mapped against the 2-dim PCA coordinates of the POM model embedding:
• Attempts to model this regression dataset (from the data above) using different techniques ranging from gradient boosting to neural networks:

Some next steps (Activation maps):

It appears as though I have missed a critial step - Which is to analyze the Activation maps, rather than just the embedding activations from the forward pass. My next goal is to try to look at patterns (like number of neurons activated, for different intensities of certain odours.) from the uncompressed embedding layer through activations checked label-wise. This is a work in progress..

Contributors of the original openpom repo:

Aryan Amit Barsainyan, National Institute of Technology Karnataka, India: code, data cleaning, model development
Ritesh Kumar, CSIR-CSIO, Chandigarh, India: data cleaning, hyperparameter optimisation
Pinaki Saha, University of Hertfordshire, UK: discussions and feedback
Michael Schmuker, University of Hertfordshire, UK: conceptualisation, project lead

References:

[1] A Principal Odor Map Unifies Diverse Tasks in Human Olfactory Perception.

Brian K. Lee, Emily J. Mayhew, Benjamin Sanchez-Lengeling, Jennifer N. Wei, Wesley W. Qian, Kelsie A. Little, Matthew Andres, Britney B. Nguyen, Theresa Moloy, Jacob Yasonik, Jane K. Parker, Richard C. Gerkin, Joel D. Mainland, Alexander B. Wiltschko

Science381,999-1006(2023).DOI: 10.1126/science.ade4401
bioRxiv 2022.09.01.504602; doi: https://doi.org/10.1101/2022.09.01.504602