Influence of Data Curation and Confidence Levels on\nCompound Predictions Using Machine Learning Models

While\ndata curation principles and practices are a major topic\nin data science, they are often not explicitly considered in machine\nlearning (ML) applications in chemistry. We have been interested in\nevaluating the potential effects of data curation on the performance\nof molecular ML models. Therefore, a sequential curation scheme was\ndeveloped for compounds and activity data, and different ML classification\nmodels were generated at increasing data confidence levels and evaluated.\nSequential data curation was found to systematically increase classification\nperformance in an incremental manner due to cumulative effects of\nindividual data curation criteria. The analysis of chemical space\ndistributions of compound subsets at different data confidence levels\nrevealed that the separation of compounds with different class labels\nin chemical space generally increased during sequential activity data\ncuration, which was mostly due to subsequent elimination of singletons\nrather than compounds from analogue series. These findings provided\na rationale for increasing the classification performance of ML models\nas a consequence of increasingly stringent data curation. Taken together,\nthe results reported herein suggest that further attention should\nbe paid to varying data curation and confidence levels when deriving\nand assessing ML models for chemical applications.