Applying molecular similarity used for evaluating the accuracy of retention index predictions in gas chromatography using deep learning

When predicting retention indices using deep learning, there is usually no way to assess the reliability of the prediction for a particular molecule. In this work, using stationary phases based on polyethylene glycol and the NIST 17 database as an example, it is shown that, on average, the closer the molecule in the training data set is to the compound being predicted, the more accurate the prediction. Tanimoto similarity of “molecular fingerprints” ECFP is the most appropriate molecular similarity calculation algorithm for this problem among the four considered. It is shown that for a number of transformation products of unsymmetrical dimethylhydrazine, whose structure was established using this prediction, it could be very unreliable.