International Flavours & Fragrances (IFF) used Cerella to retrospectively fill their sparse sensory property data set. By harnessing their existing minimal experimental data, Cerella could dramatically improve the prediction accuracy of complex in vivo responses such as odor detection threshold, even compared to other sophisticated multi-target deep learning methods.
The Challenge
Assessing sensory properties is extremely difficult, with huge time and expense put into recruiting panels of trained experts. The subjectivity of sensory properties and physiological differences between panelists leads to extremely variable data.
IFF wanted to see if Cerella could ‘fill in the gaps’ in their sparse sensory property data set collated over many years, and more importantly, achieve accurate predictions for these notoriously complex properties.
The Solution
Using Cerella, models were built for two physicochemical properties and five sensory properties. The performance of the models were compared with state-of-the-art machine learning approaches.
Cerella’s models performed extremely well compared to all other methods, particularly for complex properties like odor detection threshold (ODT).
By informing these models with physicochemical property and odor intensity assessment data, even though the data was incomplete, it dramatically improved the accuracy of prediction.
This is a clear demonstration of Cerella’s ability to use data from less expensive, early experiments to make a better selection of compounds for more costly late-stage studies (in this instance ODT), saving significant money and time.