Predicting odor from vibrational spectra: a data-driven approach.

Area of Science:

• Chemistry
• Computational Chemistry
• Sensory Science

Background:

• Olfaction, the sense of smell, is a complex modality with two main chemical theories: vibrational and docking.
• The vibrational theory of olfaction requires further validation through extensive research.
• Existing research often relies on traditional machine learning methods with limited feature sets.

Purpose of the Study:

• To systematically analyze vibrational spectra (VS) of 3018 molecules using data-driven techniques.
• To validate the vibrational theory of olfaction through advanced computational methods.
• To improve odor classification accuracy by integrating diverse molecular features.

Main Methods:

• Utilized atomistic simulations to generate VS data for 3018 molecules.
• Applied Gramian Angular Fields and Markov Transition Fields for VS image representation.
• Employed deep learning, classification, clustering (AHC, k-means), and Explainable AI techniques.
• Fused PCA-reduced fingerprint features with image-based features for enhanced analysis.
• Performed dimensionality reduction using UMAP, MDS, t-SNE, and PCA.

Main Results:

• Deep learning models significantly outperformed previous traditional machine learning approaches on the same dataset.
• Image representations of VS, combined with fingerprint features, improved odor classification accuracy.
• Clustering analyses provided insights into the relationship between molecular structure, VS, and perceived odor.
• The study confirmed the potential of deep learning in advancing olfactory research.

Conclusions:

• The findings provide strong support for the vibrational theory of olfaction.
• Deep learning and advanced computational techniques offer powerful tools for understanding smell.
• This research enhances the capability to predict and classify odors based on molecular properties.