DeepFragrance: Artificial Intelligence Methods for designing fragrant molecules

The sense of smell plays an enormous role in our everyday life. We spend huge amounts of money on scent and aroma products and in essential areas like food, hygiene and lifestyle. But what makes a molecule smell like apple, or orange? How are these molecules different from others that smell like musk, wood, or leather? Predicting the smell properties of a molecule is currently an unsolved problem.

Decades ago, drug discovery faced a similar problem - how to predict whether a given molecule could be useful to treat a certain disease? This question has led to Quantitative Structure-Activity Relationships (QSAR), where computational and statistical models are used to relate molecular structure to biological activity. QSAR is used with great success in drug discovery, as it has enabled chemists to search vast parts of chemical space for drug candidates based on similarity to known compounds, using only a (powerful) computer and a database.

There have been efforts to use QSAR methodology to discover specific molecular properties accounting for specific perceptual descriptors (e.g. "musk"), with limited success. Likewise, it has proven to be challenging to predict the general perceptual properties of a molecule. The lack of a method to predict the smell of a novel molecule impedes progress in designing new molecules with a specific fragrance.

Deep Learning (DL) has revolutionised machine learning, and is now often used synonymously to "Artificial Intelligence" due to its huge success in image and speech recognition, and in data mining in general. Naturally, DL has also been one of the most important developments in the design of molecules for drug discovery and material design.

In this project, we will apply deep learning methods to scent prediction and the design of fragrant molecules. Our aim is to leverage the capability of DL models to learn latent structure in the vast input space, produce a sparse encoding of the original high-dimensional representation. Sparse feature sets facilitate learning and inference of scent and scent-likeness. Our ultimate goal is to enable the prediction of the smell of a novel molecule, thus assisting scent and aroma design.