Combining\nGroup Contribution Method and Semisupervised\nLearning to Build Machine Learning Models for Predicting Hydroxyl\nRadical Rate Constants of Water Contaminants

Machine\nlearning is an effective tool for predicting reaction rate\nconstants for many organic compounds with the hydroxyl radical (HO<sup>•</sup>). Previously reported models have achieved relatively\ngood performance, but due to scarce data (<1400 records), the applicability\ndomain (AD) has been significantly limited. To address this limitation,\nwe curated a much larger experimental data set (Primary data set),\nwhich contains 2358 kinetic records. We then employed both the group\ncontribution method (GCM) and a semisupervised learning (SSL) strategy\nto add new data points, aiming to effectively expand the model’s\nAD while improving model performance. The results indicated that GCM\nimproved the model’s performance for chemicals outside the\nAD, while SSL expanded the model’s AD. The final model, after\nincorporating 147,168 new data points, achieved an <i>R</i><sup>2</sup> = 0.77, root-mean-square-error = 0.32, and mean-absolute-error\n= 0.24 on the test set. Importantly, the AD was expanded by 117% compared\nto the model developed solely based on the Primary data set, and the\nfinal model can be reliably applied to more than 560,000 chemicals\nfrom the DSSTox database. Further model interpretation results indicated\nthat the model made predictions based on a correct “understanding”\nof the impact of key substituents and reactive sites toward HO<sup>•</sup>. This research provides an effective method for augmenting\ndata sets, which is important in improving ML model performance and\nexpanding AD. The final model has been made widely accessible through\na free online predictor.