In silico methods to select biologically active anti-protozoan natural products

Abstract

Neglected diseases (ND) caused by tropical protozoan parasites threaten millions of people in the world. Despite the efforts from academia, government authorities and pharma companies, the available marketed drugs are insufficient. Natural products (NP) comprise a rich source of compounds that may lead to new drugs against ND. In 2011, the Research Network Natural Products against Neglected Diseases (ResNet NPND)1 has been established as an international research initiative to make efforts against this global threat. In this work, we describe the use of in silico methods to find NP hits against protozoan parasites. A ResNet NPND data bank with ca. 1,000 chemical structures of anti-parasitic NP and their biological data has been created2. 2D/3D structure descriptors were calculated and non-relevant variables were eliminated by correlation matrix followed by linear forward selection. The in vitro activity data of 513 chemical structures was grouped into four different classes against the following parasites: Plasmodium falciparum (A; 291 structures), Trypanosoma brucei (B; 63), T. cruzi (C; 98) and Leishmania donovani (D; 61). Random Forest (RF) and k-Nearest Neighbors (k-NN), machine-learning tools used to classify objects according to their classes, were used for a multi-class classification of the 513 chemical structures of NP according to their biological activities (A-D). The cross-validated models using 2D descriptors showed an accuracy of 70 – 90% and good statistical significance (Kappa statistic and ROC area). Additional 2,948 chemical structures of NP from Analyticon Discovery GmbH (Germany) with unknown biological activities were used as external data set. The RF and k-NNN models were suitable to predict new NP hits against multiple targets that can be further investigated by docking and biological assays.