Deciphering Smells from SMILES Notation of the Chemical Compounds: A Deep Learning Approach

Abstract

Discovering the alliance between the molecular arrangement and its corresponding smell has proved to be a laborious task in the field of neuroscience, olfactory research, perfumery, psychology, and chemistry. One of the constraints to demonstrate structureodor alliance is the fuzzy and incon-clusive molecular descriptors with distinct sources of smell-based molecular compounds. The graph models tend to suffer from the problem of over-smoothing whenever the number of layers is increased, node representations become fuzzy, performance degrades, and graph representations cannot be distinguished. To overcome this problem, we have proposed the Graphormer model- Bidirectional Encoder Representations from Transformers(BERT) with graph methods(GraphSAGE, Graph Isomorphism Network(GIN), Graph Attention Networks(GAT), and Graph Convolution Network(GCN)). It uses Atomic Co-ordinates(AC), Substring(SS), Structural Images(IMG) representations for the prediction of odor classes and predicts sharp smells using SMILES input. Multi-label prediction models used 5300 chemical compounds with 110 smell insights. Graphormer(GraphSAGE)+ HYB model got the highest accuracy of 98.59% and 98% AUC while with AC+SS it attained 97.8% accuracy and 98% AUC. Hence, the proposed Graphormer paradigms with physico-chemical properties motivate us to pin-point the broad correlation between the structure and smell percepts by outperforming the existing baseline models. © 2024 IEEE.