• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Monday, November 16, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Agriculture

Machine Learning Accurately Predicts Odor Detection Thresholds for Food Flavor Compounds

by
November 16, 2026
in Agriculture
Reading Time: 7 mins read
0
Machine Learning Accurately Predicts Odor Detection Thresholds for Food Flavor Compounds

Machine Learning Accurately Predicts Odor Detection Thresholds for Food Flavor Compounds

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

Understanding the exact concentration at which a volatile molecule triggers human olfactory perception represents a foundational challenge in food flavor science, where key aroma contributors frequently exert powerful sensory impacts at trace levels. In a study published in npj Science of Food, researchers developed an interpretable machine learning framework to systematically predict aqueous odor detection thresholds directly from molecular structures across a curated dataset of 1003 discrete chemical compounds. The optimized gradient boosting model achieved a test-set coefficient of determination of 0.8268, providing a computational strategy to bypass laborious geometric dilution tests and predict odor potency with quantifiable accuracy. Odor detection threshold, defined strictly as the minimum physical concentration required to elicit an olfactory sensation, quantifies the perceptual potency of volatile molecules and serves as an indispensable parameter for determining odor activity values and identifying principal aroma drivers in complex food matrices. Despite decades of analytical investigation, reliable threshold values have remained sparse, heterogeneous, and experimentally unavailable for the vast majority of volatile metabolites routinely detected in non-targeted food metabolomics profiles.

The scarcity and variability of empirical olfactory thresholds stem directly from the methodological complexity of human sensory evaluation and historical inconsistencies in published scientific literature. Conventional determinations rely upon human sensory panels performing iterative geometric dilution series, an approach that is inherently resource-intensive, slow, and exceptionally vulnerable to variations in panelist sensitivity, environmental conditions, and experimental design. Consequently, reported detection thresholds for identical chemical entities frequently fluctuate across independent scientific investigations by more than three orders of magnitude. This extensive experimental discrepancy is further exacerbated by conflating distinct sensory endpoints and measurement media throughout historical compendia. A rigorous functional distinction exists between an odor detection threshold, which marks the lowest concentration at which an odor presence is discerned, and an odor recognition threshold, which requires the panelist to identify the specific odor quality. Recognition thresholds are systematically higher, sometimes exceeding detection thresholds by up to two orders of magnitude, while measurement matrices such as water, vegetable oil, or air radically alter compound volatility, gas-matrix partitioning, and molecular binding interactions with lipids, carbohydrates, and proteins.

To overcome the confounding noise introduced by conflicting matrices and definitions, the research team established an extensively standardized modeling dataset restricted exclusively to aqueous odor detection thresholds. Beginning with the comprehensive Odor Thresholds compilation edited by Van Gemert and augmenting it with subsequent peer-reviewed literature published after 2011, the investigators excluded recognition thresholds, measurements derived from non-aqueous matrices, ambiguous entries, chemical mixtures, organometallic salts, and undefined molecular entities. Canonical Simplified Molecular Input Line Entry System representations were retrieved from the PubChem database and curated using RDKit to resolve duplicate structures. For odorants documented by multiple independent investigations, values converted to milligrams per liter were evaluated in logarithmic order-of-magnitude histograms; only the modal frequency bin and its adjacent intervals were retained to eliminate irreconcilable outlier reports. The final consensus aqueous threshold was computed as the geometric mean of retained values, which mathematically corresponds to averaging records in logarithmic space. To compress the immense numerical dynamic range into a continuous, well-behaved target variable, thresholds were converted into negative logarithmic values denoted as pT, where higher values correspond directly to lower detection limits and substantially greater odor potency.

Structural profiling of the curated 1003-compound dataset revealed that aqueous odor thresholds follow a broadly unimodal statistical distribution centered near moderate potencies while spanning a diverse chemical space. Most examined odorants clustered between pT values of 1 and 5, corresponding to aqueous detection thresholds ranging approximately from 0.01 to 100 milligrams per liter, with peak sample density occurring around pT 3, or approximately 1 milligram per liter. To visualize the global structural diversity and chemical distribution of the cataloged compounds, the researchers converted molecular representations into high-dimensional Morgan fingerprints and projected them into two dimensions using unsupervised Uniform Manifold Approximation and Projection with cosine distance metrics. Rather than collapsing into an isolated structural cluster, the curated odorants populated broad, continuous sectors of chemical space. Overlaying experimental pT gradients onto this manifold demonstrated that molecules exhibiting high, intermediate, and low odor potencies are distributed widely across diverse functional classes, proving that aqueous odor potency cannot be explained by an isolated chemical scaffold, while exhibiting localized structural clustering that strongly indicated non-random, learnable relationships between molecular features and human sensory detection.

Systematic classification of the curated molecules into aliphatic and cyclic categories subdivided by functional groups established clear class-dependent potency trends across the chemical spectrum. Among aliphatic structures, carboxylic acids and nitrogen-containing compounds were heavily concentrated in the low-pT region with narrow dispersion, demonstrating that these highly polar functional groups consistently confer elevated detection thresholds and weak potency in water. Conversely, aliphatic hydrocarbons, aldehydes, ketones, and esters displayed broad threshold distributions spanning multiple logarithmic decades, demonstrating that while many aliphatic volatiles exhibit moderate sensory potency, specific structural configurations achieve exceptionally low thresholds. Aliphatic sulfur-containing compounds and alcohols also displayed wide potency ranges, with certain organosulfur volatiles extending prominently into the ultra-potent, high-pT region. Among cyclic architectures, cyclic alcohols occupied predominantly low-pT intervals, whereas cyclic ketones and heterocycles like furans and thiazoles exhibited broader dispersion. Notably, cyclic esters exhibited a striking architectural dichotomy: lactones clustered in the low-pT region, whereas compounds containing exocyclic ester groups were enriched at high pT values, and prominent high-potency outliers repeatedly emerged among methoxy-substituted pyrazines and poly-substituted benzene derivatives bearing both methoxy and halogen groups.

To identify the most effective predictive architecture, the researchers benchmarked nine distinct machine learning algorithms using an 80-20 train-test partition and 121 curated two-dimensional molecular descriptors computed via RDKit. These descriptors captured essential molecular attributes, including molecular weight, polar surface area, connectivity indices, electronic features, and hydrogen-bonding motifs, following the removal of zero-variance and excessively missing variables. Tree-based ensemble architectures, including Gradient Boosting Decision Trees, Random Forest, LightGBM, and extreme gradient boosting (XGBoost), systematically outperformed support vector machines, multi-layer perceptron neural networks, and classical statistical algorithms. The optimized XGBoost model delivered the highest predictive accuracy on the unseen test set, achieving a coefficient of determination of 0.8268, a mean squared error of 0.6939, a root mean squared error of 0.8323, and a mean absolute error of 0.6394. Five-fold cross-validation confirmed stable generalization with a mean cross-validated coefficient of determination of 0.7980, while k-nearest neighbors exhibited severe overfitting by recording an artificial training set R-squared of 0.9987 that collapsed to 0.7677 on test evaluations. Crucially, one hundred iterations of Y-randomization shuffling destroyed model performance, proving that the XGBoost predictions reflected genuine structure-potency associations rather than chance mathematical artifacts.

Elucidation of the underlying structure-activity relationships using SHapley Additive exPlanations (SHAP) identified the key physical and chemical determinants driving the model’s threshold estimations. Descriptors governing electrostatic charge distribution, hydrogen bonding, polar surface area, polarizability, and topological size emerged as dominant contributors to predicted pT values. The partial equalization of orbital electronegativity surface area descriptor PEOE_VSA1 exerted the strongest individual influence on model output, followed closely by the count of NH and OH groups, Burden eigenvalue indices weighted by molar refractivity and atomic mass, molecular weight, valence connectivity, and topological polar surface area. Elevated values for polar surface area, hydrogen-bond donors, and partial-charge surface metrics contributed negative SHAP values, driving down predicted pT and signaling higher aqueous detection thresholds. This model-derived relationship aligns closely with physical headspace partitioning mechanics: molecules possessing strong hydrogen-bonding capabilities and high polarities interact favorably with water molecules, impeding their release from the aqueous matrix into surrounding air. Conversely, descriptors reflecting molecular refractivity, topological connectivity, and molecular size contributed positive SHAP values, reflecting how molecular bulk and polarizability enhance perceived potency, with several electronic and topological metrics unexpectedly outranking the traditional lipophilicity parameter MolLogP.

To rigorously test real-world predictive validity beyond the historical modeling dataset, the researchers assembled an independent external validation set comprising 177 commercially available aroma chemicals whose aqueous detection thresholds were experimentally determined in their sensory laboratory. Sensory testing was executed in an ISO-compliant sensory evaluation facility maintained at 25 degrees Celsius, employing a panel of thirty trained assessors evaluated for normal olfactory function using n-butanol benchmarks. In accordance with ISO 13301 standards, panels conducted three-alternative forced-choice presentations across geometric dilution series, determining individual thresholds via the best-estimate threshold calculation and deriving group values through geometric averaging. Structural similarity analysis using Morgan fingerprints revealed an average maximum Tanimoto similarity coefficient of 0.5771 between the 177 validation odorants and the training set, classifying 58 compounds as in-domain, 85 as borderline, and 34 as out-of-domain. When evaluated against these independent sensory measurements, the XGBoost model achieved satisfactory predictions within a ten-fold error margin for 87.0 percent of the compounds, while 10.7 percent fell within acceptable hundred-fold deviations, and only 2.3 percent exhibited large deviations, demonstrating substantial predictive utility under rigorous experimental conditions.

Building upon the retrospective accuracy demonstrated across internal and external cohorts, the researchers implemented a prospective screening protocol designed to prioritize previously uncharacterized high-potency food odorants from expansive chemical libraries. Starting from the Flavor and Extract Manufacturers Association ingredient repository, the team curated 1679 compounds possessing retrievable molecular structures and excluded all chemicals already included in the training or external validation cohorts, yielding 1041 prospective flavor candidates. Applying the calibrated XGBoost model to this prospective screening library yielded 102 volatile compounds predicted to possess extreme olfactory potency, defined by predicted pT values of 5 or greater, representing aqueous detection thresholds at or below 0.01 milligrams per liter. To empirically verify the reliability of this prioritization pipeline, twelve commercially available candidates from the prioritized cohort were procured and evaluated through standardized three-alternative forced-choice sensory trials. Experimental validation confirmed eight compounds within satisfactory ten-fold deviations and four within acceptable hundred-fold boundaries, successfully corroborating trace-level detection thresholds among prioritized candidate esters, methoxypyrazines, aldehydes, ketones, and aliphatic organosulfur molecules.

While this machine learning framework provides a practical and interpretable mechanism for resolving long-standing threshold data gaps, the investigators noted several technical boundaries inherent to the computational architecture and physical methodology. Because the model relies exclusively upon two-dimensional molecular descriptors, it does not explicitly represent three-dimensional conformational dynamics, spatial stereochemistry, or direct physical docking interactions with specific human olfactory receptor proteins. Furthermore, predictions for compounds situated far beyond the established applicability domain or exhibiting extreme chemical instability, such as reactive polysulfides or easily hydrolyzable terpene esters, carry heightened uncertainty due to the physical difficulty of maintaining stable aqueous concentrations during sensory verification. Additionally, because the training set is calibrated exclusively upon aqueous solutions, the resulting potency projections cannot be transferred directly to complex lipid-dense or ethanol-rich food matrices where phase partitioning and vapor-liquid equilibria behave fundamentally differently. Nevertheless, by integrating rigorous endpoint curation, robust gradient boosting architectures, and prospective sensory confirmation, this computational framework establishes a reliable tool for characterizing volatile metabolomes and accelerating flavor discovery.

Subject of Research: Machine learning prediction of aqueous odor detection thresholds in food aroma chemistry

Article Title: Machine learning reveals structural determinants of odor detection thresholds and identifies high-potency food odorants

Article References: Hou, X., Zhou, Z., Shi, P., Wang, Y., Liu, S., Ji, Z., Zhang, D., & Mao, J. (2026). Machine learning reveals structural determinants of odor detection thresholds and identifies high-potency food odorants. npj Science of Food, 10(1), Article 276. https://doi.org/10.1038/s41538-026-01125-7

Image Credits: AI Generated

DOI: 10.1038/s41538-026-01125-7

Keywords: odor detection threshold, machine learning, flavor chemistry, XGBoost, food odorants, quantitative structure-activity relationship, sensory evaluation, chemical space, SHAP analysis, olfactory perception, aqueous thresholds, molecular descriptors, methoxypyrazines, organosulfur compounds, FEMA flavor ingredients, aroma metabolomics, three-alternative forced choice

News Source: Teresa Odom. (November 16, 2026). Machine Learning Accurately Predicts Odor Detection Thresholds for Food Flavor Compounds. Scienmag.

Tags: aqueous thresholdsaroma metabolomicschemical spaceFEMA flavor ingredientsflavor chemistryfood odorantsMachine Learningmethoxypyrazinesmolecular descriptorsodor detection thresholdolfactory perceptionorganosulfur compoundsQuantitative Structure-Activity RelationshipSensory evaluationSHAP analysisthree-alternative forced choiceXGBoost
Share12Tweet7Share2ShareShareShare1

Related Posts

Agricultural Information Systems Associate With Major Productivity and Efficiency Gains in Jordan

Agricultural Information Systems Associate With Major Productivity and Efficiency Gains in Jordan

November 16, 2026
National Team Launches 53-Million-Dollar Project to Develop Next-Generation Herbicides

National Team Launches 53-Million-Dollar Project to Develop Next-Generation Herbicides

November 16, 2026

Stomatal Traits Provide Rapid Ploidy Screening for Doubled Haploid Onion Breeding

November 16, 2026

Global Review Defines Safe Biochar Application Thresholds Across Six Major Ecosystems

November 16, 2026

POPULAR NEWS

  • Longitudinal Molecular Profiling in ALK-Rearranged Lung Cancer: A Case Report

    Longitudinal Molecular Profiling in ALK-Rearranged Lung Cancer: A Case Report

    29 shares
    Share 12 Tweet 7
  • Periodontal Ligament Simulation Variations Do Not Alter Root Fracture Resistance In Vitro

    29 shares
    Share 12 Tweet 7
  • Museum Records Reveal 33,000 Non-Asilid Insect Specimens Collected by Jason Londt

    29 shares
    Share 12 Tweet 7
  • State Reconstruction Model Slashes Epidemic Infection Costs by 88 Percent

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Advanced Lipid Bioinformatics Tools Connect Complex Lipidomics to Brain and Skin Health

ChatGPT Matched German Guidelines in Seventy Percent of Gastrointestinal Cancer Cases

China’s Carbon Trading Pilots Aggravated Energy Misallocation Under Distorted Allowance Prices

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 87 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.