Machine-Learning-Based Olfactometry: An Auxiliary System for Human Assessors in Olfactory Measurement

Liang Shang, Chuanjun Liu, Fengzhen Tang, Bin Chen, Lianqing Liu, Kenshi HayashibioRxiv 2022 · DOI

TL;DR

This paper introduces a novel machine learning approach to olfactometry, aiming to replace or augment human smell assessors. By clustering odor descriptors (ODs) into meaningful categories using Bayesian embedding and then predicting these categories from molecular structure, they achieve high accuracy, paving the way for more objective and scalable odor analysis.

What Problem Does This Paper Address?

Traditional gas chromatography-olfactometry (GC/O) relies on human panelists to identify and describe odors. This method, while valuable, suffers from significant drawbacks: variability between individuals, subjectivity in descriptions, and high costs associated with training and employing these human "detectors." Furthermore, the sheer number of potential odor descriptors makes it impractical to build individual predictive models for each one. This research tackles the challenge of creating a more objective, cost-effective, and scalable olfactometry system by leveraging machine learning to predict odor categories based on molecular structure.

Methods

The core of this work lies in a two-pronged approach:

  1. Odor Descriptor (OD) Clustering:
    • Bayesian Co-occurrence Embedding (BCE): This technique is used to embed odorants and their associated ODs into a vector space, capturing co-occurrence relationships from multiple odor databases. This helps to normalize and calibrate OD labels.
  2. Hierarchical Clustering Analysis (HCA): Based on the embedded OD vectors, HCA is applied to group similar ODs into categories. The authors settled on 20 distinct odor categories.
  3. Word2Vec Embedding: A pre-trained Word2Vec model is used to analyze the semantic relationships between ODs, providing an independent validation of the clustering results.
  4. Structure-Odor Relationship (SOR) Model Calibration:
    • Molecular Feature Extraction: Various molecular features are extracted from chemical structures, including molecular parameters, molecular fingerprints, and 2D graphic features derived from Convolutional Neural Networks (CNNs) (specifically VGG-16, Restnet, Densnet, and Alexnet).
  5. Machine Learning Models: Gradient Boosting Decision Trees (GBDT) and Generalized Learning Vector Quantization (GLVQ) are employed to predict the previously defined odor categories using the extracted molecular features.

Datasets

The study utilized three publicly available odor databases:

  • Flavors and Fragrances database (Sigma-Aldrich): Contained 1026 odorants.
  • Good Scents database: Contained 3673 odorants.
  • Odor Map database: Contained 321 odorants.

After data merging and cleaning, the final dataset comprised 2849 molecules with 510 ODs. The paper does not introduce a new benchmark dataset but rather consolidates and processes existing ones.

Key Insights and Ideas

  • Clustering ODs is Key: The paper demonstrates that grouping numerous odor descriptors into a manageable set of categories (20 in this case) is a crucial step towards building effective machine learning models for olfactometry. This overcomes the impracticality of modeling each individual descriptor.
  • Co-occurrence is Informative: The Bayesian co-occurrence embedding effectively captures relationships between odorants and their descriptors, providing a robust foundation for clustering.
  • Molecular Structure Predicts Odor Category: The study strongly suggests that molecular structure, particularly its spatial representation captured by CNNs, is a powerful predictor of odor categories, achieving high accuracy.
  • Flavor vs. Unpleasant Odors: The research hints that flavor-like impressions might be more complex to categorize than unpleasant ones, though both can be effectively modeled.
  • Semantic Validation: Using Word2Vec to analyze semantic relationships between ODs provides an independent and complementary way to validate the proposed odor categories.

Key Results

  • 20 Odor Categories: The hierarchical clustering successfully grouped 256 representative odor descriptors into 20 distinct categories.
  • High Prediction Accuracy: The best performing models (GBDT with Densenet-extracted features) achieved an Area Under the ROC curve (AUC) of 0.800±0.004, with precision of 0.595±0.004, recall of 0.721±0.003, and F1-score of 0.570±0.007.
  • CNNs Outperform: Molecular graphic features extracted by CNNs, especially Densenet and Restnet, significantly outperformed traditional molecular parameters and fingerprints in predicting odor categories.
  • GBDT Superiority: GBDT models generally showed better performance compared to GLVQ models.

Limitations and Future Work

  • Ambiguous ODs: The study acknowledges that some abundant ODs (e.g., "sweet," "fruity") were excluded due to ambiguity. Future work should focus on defining metrics to remove such ambiguous descriptors and identify more characteristic ones.
  • Mixtures of Odorants: The current SOR models are designed for single odorants. Future research should extend these models to handle mixtures of odorants, as molecular structures cannot be simply linearly combined. Novel approaches like 3D interaction embedding and topology graph representations are suggested.
  • Understanding Olfaction Biology: While the study provides a computational framework, further integration with biological understanding of olfaction is encouraged.
  • Auxiliary System: The authors envision their ML-GCO models serving as either an auxiliary system to human assessors or a complete replacement, highlighting the potential for practical application.

Summary updated

Abstract

Abstract Although gas chromatography/olfactometry (GC/O) has been employed as a powerful analytical tool in odor measurement, its application is limited by the variability, subjectivity, and high cost of the trained panelists who are used as detectors in the system. The advancements in data-driven science have made it possible to predict structure-odor-relationship (SOR) and thus to develop machine-learning-based olfactometry (ML-GCO) in which the human panelists may be replaced by machine learning models to obtain the sensory information of GC-separated chemical compounds. However, one challenge remained in ML-GCO is that there are too many odor descriptors (ODs) being used to describe the sensory characteristics of odorants. It is impractical to build a corresponding model for each OD. To solve this issue, we propose a SOR prediction approach based on odor descriptor clustering. 256 representative ODs are firstly classified into 20 categories using a co-occurrence Bayesian embedding model. The categorization effect is explained according to the semantic relationships using a pre-trained Word2Vec model. Various molecular structure features including molecularly parameters, molecular fingerprints, and molecular 2D graphic features extracted by convolutional neural networks, are employed to predict the aforementioned odor categories. High prediction accuracies (Area under ROC curve was 0.800±0.004) demonstrate the rationality of the proposed clustering scenario and molecular feature extraction. This study makes the ML-GCO models much closer to the practical application since they can be expected as either an auxiliary system or complete replacement of human panelists to perform the olfactory evaluation.
BibTeX
@misc{shang2022machine,
  author = {Liang Shang and Chuanjun Liu and Fengzhen Tang and Bin Chen and Lianqing Liu and Kenshi Hayashi},
  title = {{Machine-Learning-Based Olfactometry: An Auxiliary System for Human Assessors in Olfactory Measurement}},
  year = {2022},
  doi = {10.1101/2022.04.20.488973},
  howpublished = {bioRxiv},
}
Figure