Molecular Similarity Used for Evaluating the Accuracy of Retention Index Predictions in Gas Chromatography Using Deep Learning

Abstract

When predicting retention indices using deep learning, there is typically no way to assess the reliability of predictions for specific molecules. The present study demonstrates, using stationary phases based on polyethylene glycol and NIST 17 database, that predictions are generally more accurate when the training dataset includes molecules structurally similar to the compound for which prediction is made. The Tanimoto similarity of “molecular fingerprints” ECFP is the most suitable algorithm for this task among the four algorithms considered. For several transformation products of unsymmetrical dimethylhydrazine whose structures were established using such predictions, the predictions were shown to be unreliable.