Imputation of sensory properties using deep learning.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Predicting sensory properties of chemical compounds is challenging due to subjective human panel measurements.
• Current methods are expensive and time-consuming, limiting compound selection for further studies.

Purpose of the Study:

• To apply a deep learning method (Alchemite™) for imputing sparse physicochemical and sensory data.
• To compare the imputation model's performance against traditional quantitative structure-activity relationship (QSAR) methods and graph convolutional neural networks (GCNNs).
• To evaluate the model's ability to identify accurate predictions and activity cliffs.

Main Methods:

• Utilized Alchemite™, a deep learning approach, for data imputation.
• Compared imputation model with conventional QSAR and multi-target GCNN methods.
• Assessed prediction accuracy using R-squared values and analyzed uncertainty estimates.

Main Results:

• The Alchemite™ imputation model demonstrated substantially higher prediction accuracy, with R² improvements of 0.26–0.45 over the next best method.
• Robust uncertainty estimates from the model allowed for identification of the most accurate predictions.
• The model accurately predicted activity cliffs, where minor structural changes cause significant sensory property alterations.

Conclusions:

• Deep learning-based data imputation offers a more accurate and efficient approach to predicting compound sensory properties.
• This method enhances the selection of compounds for costly experimental studies, optimizing resource allocation and accelerating research.
• Alchemite™ provides valuable insights into structure-property relationships, including critical activity cliffs.