Graph Neural Network Prediction of Antioxidant Activity of Polyphenols from Grape Pomace

Abstract

Grape pomace is the main solid by-product of winemaking and a cheap source of polyphenols. However, the antioxidant strength of most compounds in grape pomace has never been measured on its own. Mixtures do not behave additively, so extract results cannot be separated into compound-level values. This paper reviews the chemistry and the modelling problem, sets out a full pipeline from data curation to molecular graphs to a trained network, and then runs it on measured data. Three results concern the data. Unit conversion matters, since epicatechin measured in two laboratories in two-unit systems agrees to 0.021 pIC50 log units once converted. Donor count does not order activity, since gallic acid and trans-resveratrol carry the same three phenolic hydroxyls yet differ by 0.436 log units, which is the argument for a representation that encodes adjacency. And protocols cannot be pooled, since three compounds measured in a second laboratory differ systematically by 1.152 log units, more than the 1.477 log unit range of the training corpus. For training, 25 phenolics measured in one laboratory under one protocol reduced to 16 usable records once non-numeric entries were removed and dose-response fit quality was filtered on, a filter that published curation protocols do not apply and that removed a compound whose tabulated IC50 came from a regression explaining 19 percent of its variance. On those 16 compounds a message-passing network reached a leave-one-out R2 of 0.744 against 0.138 for a mean predictor, with a training fit of 0.696, and a label-scrambling control confirmed that no structure-activity signal was recovered at this sample size. Predictions for the five targets span 0.322 log units and all ten pairwise comparisons overlap at 95 percent. The pipeline runs correctly, and closing the gap between 16 compounds and the 1911 used by published benchmarks is the work that remains.