How this AI can Characterize Smell

Introduction

This paper will discuss a Graph Neural Network that can detect and classify smell. It will cover its significance, the parameters, the architecture, the results, the ability compared to other methods, and the conclusion.

The Significance of this Innovation

Have you ever wanted to analyze something you experienced? People have wanted to analyze the things around them for years. You analyze light by measuring the intensity of light or some other variables. You analyze sound by understanding frequency and wavelength. Have you ever wanted to analyze the smell around you? This article goes through the paper-A Principal Odor Map Unifies Diverse Tasks in Human Olfactory Perception. The reason this creation is so amazing is because the smell we experience is never just one particular smell. It’s a combination! It is also affected by the subjectivity involved in smell, which affects experiments and the difficulty in expressing smell in language. Descriptions are often inconsistent.

The Parameters

The characteristics under consideration include the following: valence, degree, hydrogen count, hybridization, formal charge, atomic charge, bond aromaticity, and the presence of a ring-like bond structure.

You may be wondering what each of these properties is. Here’s a quick run-through:

• Valence is the number of bonds an atom could potentially make.
• Degree is the number of bonds an atom has formed.
• Hydrogen count is the number of hydrogen atoms attached to an atom in a molecule.
• Hybridization is the process of mixing atomic orbitals to form new hybrid orbitals, with each hybrid orbital having specific properties.
  Formal charge is the electric charge of an atom in a molecule.
• Atomic charge is an electric charge that occurs when the number of protons in the nucleus differs from the number of electrons surrounding the nucleus.
• Aromaticity is a property among specific compounds that enhances the stability of a molecule because of delocalization of electrons present in specific orbitals.
• Ring-like bond structure refers to a cyclical compound where one or more groups of atoms are connected by a ring; these compounds have different specific characteristics.

The rationale behind this choice is simply ease. Every chemical compound has these properties, so it allows for a possible characterization system.

The Neural Network behind it

The dataset comprised 5000 molecules, complete with descriptive labels, sourced from the GoodScents and Leffingwell flavor and fragrance databases. The paper delved into a question that has long intrigued the scientific community: Is it possible to translate the physical attributes of smell into data?

The researchers developed a Graph Neural Network (GNN), specifically a Message Passing Neural Network. If you’re wondering what a GNN is, it’s a type of neural network designed to process data structured as a graph. A GNN consists of two fundamental components: nodes and edges. Nodes represent the object of interest and its properties, while edges symbolize the connections each node has with others, signifying relationships.

In the network architecture, the penultimate layer is engineered to establish associations among various odors, thereby creating clusters. This results in the formation of a Principal Odor Map (POM). What exactly is a POM? Think of it as a large, multicolored map where each point signifies a distinct smell. The distances between these points indicate the similarity between two smells.

Finally, the last layer of the network undertakes the task of predicting odors based on these established associations.

The Results

POM Vs Morgan Fingerprints(FP)

Morgan Fingerprints (FP) constitute a map constructed using standard chemoinformatic features. This method allows for the representation and analysis of molecular structures. Morgan Fingerprints encode groups of atoms in a chemical compound into a binary vector, with length and radius serving as parameters.

The researchers demonstrated that the Principal Odor Map (POM) is more representative of the data, exhibiting a higher correlation for distances on the perceptual map. The correlation coefficient for POM was R=0.73, while for FP it was R=-0.12. POM also displayed a higher cluster density (CD=0.51±0.19), compared to the FP map (CD=0.68±0.23).

The main issue with FP lies in the discontinuities present in the structure-odor map.

Potential Setbacks in the Research

One potential limitation of the research was the high complexity of odors, particularly when a smell is attributable to more than one specific chemical. Another significant limitation was the scarcity of training examples. The dataset, comprising 5000 molecules, may be limited and thus not as predictive as it could potentially be. Other limitations that emerged included impure chemical materials and odorous contaminations.

Conclusions

Thanks to that effort, we have AI to analyze images and detect them based on their features. This has made many industries more efficient and safer. A prime example would be the self-driving car industry (LiDAR). Imagine the potential applications of this new predictive modeling approach. One big potential advantage could be for an AI to detect whether toxic chemicals that are not visible to the human eye are in the air, for quicker evacuations in chemical labs. The applications of this technology are and will be plentiful and that’s why this paper is truly amazing.

References

Lee, B. K., Mayhew, E. J., Sanchez-Lengeling, B., Wei, J. N., Qian, W. W., Little, K., Andres, M., Nguyen, B. B., Moloy, T., Parker, J. K., Gerkin, R. C., Mainland, J. D., & Wiltschko, A. B. (2022). A Principal Odor Map Unifies Diverse Tasks in Human Olfactory Perception. https://doi.org/10.1101/2022.09.01.504602