Machine learning prediction of cyanobacterial toxin (microcystin) toxicodynamics in humans.

ABSTRACT

Microcystins (MC) represent a family of cyclic peptides with approx. 250 congeners presumed harmful to human health due to their ability to inhibit ser/thr-proteinphosphatases (PPP), albeit all hazard and risk assessments (RA) are based on data of one MC-congener (MC-LR) only. MC congener structural diversity is a challenge for the risk assessment of these toxins, especially as several different PPPs have to be included in the RA. Consequently, the inhibition of PPP1, PPP2A and PPP5 was determined with 18 structurally different MC and demonstrated MC congener dependent inhibition activity and a lower susceptibility of PPP5 to inhibition than PPP1 and PPP2A. The latter data were employed to train a machine learning algorithm that should allow prediction of PPP inhibition (toxicity) based on MCs 2D chemical structure. IC50 values were classified in toxicity classes and three machine learning models were used to predict the toxicity class, resulting in 80-90% correct predictions.