Linking molecules to their odors
with supervised descriptive rules discovery


CNRS INSA de Lyon Université Claude Bernard Lyon 1

Understanding the physicochemical determinants of the olfactory percept

While it is generally agreed that the physicochemical characteristics of odorants affect the olfactory percept, no precise rules governing this "chemical structure / odor" relation have yet been determined. One obstacle to this is the large number of dimensions – several thousand – needed to describe the molecules and, above all, a lack of data-processing methods suitable for this level of complexity. Our first objective will therefore make use of the 3 data-bases established as part of the Prevalolf project, correlating the multiple molecular characteristics (olfactory chemical space) to olfactory qualities (fruity, floral, woody, etc.: odor perception space). Methods taken from data mining producing descriptive rules are used to test the hypothesis that certain olfactory qualities, on which there is strong inter-subject agreement, can be predicted from the physicochemical properties of the odorant molecule.

Explore Descriptive Rules

Implemented contributions

h(odor) WebbApp

Allows the olfaction experts to interact with the discovery algorithm to guide the search in the huge description space thanks to an ergonomic user interface.


ELMMut algorithm

Finds molecular description
a subset of olfactory qualities
with the Fβ Score


State of the art methods

Subgroup Discovery

Finds molecular description
targeting a unique quality


Redescription Mining

Finds couples (description, qualities)
covering almost the same molecules


Exceptional Model Mining

Finds molecular descriptions
targeting several qualities


The idea

Described by physicochemical properties,
some molecules have a perceived odor
Three databases are considered, entailing hundreds of molecules,
with odors associated by experts from a well defined vocabulary
Subgroup discovery consists in finding descriptions of molecules
that are strongly characteristic of a (several) percept(s)
Redescription mining consists in finding couples (description, qualities)
such as both queries support almost the same set of molecules.
Learn more about the methodology!

The projects

This research is partially supported by:

  • the CNRS (Préfute PEPS FASCIDO)
  • the Institut rhônalpin des systèmes complexes (IXXI)