Composite Machine Learning Strategy for Natural Products Taxonomical Classification and Structural Insights

Taxonomical classification of natural products (NPs) can facilitate bioprospecting efforts, aid genomic and phylogenetic analysis of source organisms, and support sustainable biosynthesis of novel drugs using NPs as building blocks. In this work, a composite machine learning strategy merging graph convolutional neural networks (GCNN), feed forward neural networks (FFN), and support vector machines (SVM) is proposed and illustrated for the taxonomical classification of NPs within five kingdoms (animal, bacteria, chromista, fungi, and plant). Our composite model trained on 133,092 NPs from the LOTUS database demonstrated a five-fold cross validated classification accuracy of 97.1%. When employed to classify NPs beyond its training set from the NP Atlas database, accuracies of 83.2% for bacteria and 86.5% for fungi were obtained. Dimensionality reduced representations of the molecular embeddings from our composite model revealed well-separated clusters of NPs that are the basis for superior classification. The top critical substructures from the NPs of each kingdom were also identified and compared to provide insights on NP structure-taxonomy relationships. Overall, this study demonstrates the potential of composite machine learning models for NP taxonomical classification and to shed light on the underlying structural reasons for such classification assignments.