Chemically unstable natural products are prone to show their reactivity in the procedures of extraction, purification, or identification and turn into contaminants as so-called "artifacts". However, identification of artifacts requires considerable investments in technical equipment, time, and human resources. For revealing these reactive natural products and their artifacts by computational approaches, we set up a virtual screening system to seek cases in a biochemical database. The screening system is based on deep learning models of predicting the two main classifications of conversion reactions from natural products to artifacts, namely solvolysis and oxidation. A set of result data was reviewed for checking validity of the screening system, and we screened out a batch of reactive natural products and their probable artifacts. This work provides some insights into the formations of natural product artifacts, and the result data may act as warnings regarding the improper handling of biological matrixes in multicomponent extraction.
Keywords: artifact; deep learning; natural product; virtual screening.