Inconsistent odor-threshold measurements limit prediction from molecular structure
Abstract
Existing work treats odor detection threshold prediction as a supervised regression problem, assuming published thresholds are clean targets. In current public datasets, they are not: values come from different measurement sources and often disagree for the same molecule. Across the three largest public air-phase compilations, thresholds for identical molecules differ systematically by median factors of 3 to 11. Individual source studies vary around a compound’s underlying threshold by about 1.4 log 10 units, so under an independent-source approximation, a new study is expected to differ from the compiled value by about 1.6 log 10 units. This exceeds the best structure-based model’s error against that value (root-mean-square error 1.34), meaning the model matches the published compilation more closely than a new experiment would be expected to. Although averaging multiple studies makes the compiled value more precise, its remaining uncertainty still accounts for at least one-third of the model’s prediction-error variance, and a single source-study threshold has estimated reliability near 0.5. Changing the molecular representation, by comparison, changes root-mean-square error by only about 0.2 log 10 units. We therefore introduce a joint calibration model that places the compilations on a common scale, estimates source-specific offsets and noise, and improves cross-compilation prediction over a simple fixed-offset correction while providing calibrated uncertainty. These results show that progress in odor detection threshold prediction depends not only on better molecular models, but also on better measurements and methods that account for disagreement across data sources.
Related
- DEEPNET-LLM: Highly accurate structure prediction of unknown metabolites from mass spectra via their molecular neighborhoods
- A decision-based criterion and a resolution threshold for ranking genomic prediction models within trials
- Tri-Modality Representation Learning for Molecular Property Prediction
- Evaluating Autonomous LLM Agents Across Molecular Prediction and Optimization Benchmarks
- NuFold RNA Structure Database: Large-Scale RNA Tertiary Structure Prediction with Confidence Metrics
BibTeX
@misc{mysorekar2026inconsistent,
title = {{Inconsistent odor-threshold measurements limit prediction from molecular structure}},
author = {Anand R. Mysorekar and Lekha Challappa},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010039/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010039/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Inconsistent odor-threshold measurements limit prediction from molecular structure
AU - Anand R. Mysorekar
AU - Lekha Challappa
PY - 2026
DA - 2026/10/07
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010039/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010039/v1
AB - Existing work treats odor detection threshold prediction as a supervised regression problem, assuming published thresholds are clean targets. In current public datasets, they are not: values come from different measurement sources and often disagree for the same molecule. Across the three largest public air-phase compilations, thresholds for identical molecules differ systematically by median factors of 3 to 11. Individual source studies vary around a compound’s underlying threshold by about 1.4 log 10 units, so under an independent-source approximation, a new study is expected to differ from the compiled value by about 1.6 log 10 units. This exceeds the best structure-based model’s error against that value (root-mean-square error 1.34), meaning the model matches the published compilation more closely than a new experiment would be expected to. Although averaging multiple studies makes the compiled value more precise, its remaining uncertainty still accounts for at least one-third of the model’s prediction-error variance, and a single source-study threshold has estimated reliability near 0.5. Changing the molecular representation, by comparison, changes root-mean-square error by only about 0.2 log 10 units. We therefore introduce a joint calibration model that places the compilations on a common scale, estimates source-specific offsets and noise, and improves cross-compilation prediction over a simple fixed-offset correction while providing calibrated uncertainty. These results show that progress in odor detection threshold prediction depends not only on better molecular models, but also on better measurements and methods that account for disagreement across data sources.
ER -