Why does that molecule smell?

Abstract

Learning structure-scent relationships is a complex challenge due to both the large chemical space of odorous molecules and the molecular biology of a smell. We empirically fit structure-scent relationships by training an accurate graph neural network and then explaining its predictions. We use counterfactuals and descriptor attribution to generate explanations for the 112 scents in the Leffingwell Odor Dataset (Sanchez-Lengeling et al., 2019). Then we use natural language processing to summarize the quantitative explanations into text. The complete process goes from data to a natural language explanation with the aim of determining structure-scent relationships.

A Perspective on Explanations of Molecular Prediction Models

“…The resulting natural language explanation is: "The molecular property "fatty scent" can be explained by the presence of a heptanyl fragment, two CH2 groups separated by four bonds, and a C=O double bond, as well as the lack of more than one or two O atoms." [116] The importance of a heptanyl fragment aligns with that reported in the literature, as 'fatty' molecules often have a long carbon chain. [142] Furthermore, the importance of a C=O double bond is supported by the findings reported by Licon et al [143] , where in addition to a "larger carbon-chain skeleton", they found that 'fatty' molecules also had "aldehyde or acid functions".…”

“…[10] Then we demonstrate how counterfactuals and descriptor explanations can propose structure-property relationships in the domain of molecular scent. [116] 3.1 Blood-brain barrier permeation prediction…”

“…There are some exceptions though, like tert-amyl acetate which has a 'camphoraceous' rather than 'fruity' scent. [138,139] In Seshadri et al [116] , we trained a GNN model to predict the scent of molecules and utilized counterfactuals and descriptor explanations to quantify scent-structure relationships. The MMACE method was modified to account for the multi-label aspect of scent prediction.…”

“…[140,141] The resulting counterfactual, which has a shorter carbon chain and no carbonyl group, highlights the influence of these structural features on the 'fatty' scent of 2,4 decadienal. To generalize to other molecules, Seshadri et al [116] applied the descriptor attribution method to obtain global explanations for the scents. The global explanation for the 'fatty' scent was generated by gathering chemical spaces around many 'fatty' scented molecules.…”

“…[143] For the 'pineapple' scent, the following natural language explanation was obtained: "The molecular property "pineapple scent" can be explained by the presence of ester, ethyl/ether O group, alkene/ether O group, and C=O double bond, as well as the absence of an Aromatic atom." [116] Esters, such as ethyl 2-methylbutyrate, are present in many pineapple volatile compounds. [144,145] The combination of a C=O double bond with an ether could also correspond to an ester group.…”