Training data

The training data was largely sourced from a HTS for βH-inhibiting antimalarials, piloted by Vanderbilt University (VU)¹⁵. Only βH inhibitors (βH IC₅₀ ≤ 100 µM) which had been tested against Pf were included; molecules were considered ‘active’ (1) if they exhibited Pf IC₅₀ ≤ 1 µM and ‘inactive’ (0) if Pf IC₅₀ ≥ 1.5 µM. The same criteria were applied to ∼100 bioactive neo- and isocryptolepine derivatives synthesised by a group at Okayama University (OU)^45,46. In addition, a number of molecules from the Tres Cantos Antimalarial Compounds Set (TCAMS)⁴⁷ were included in the training set using unpublished βH inhibition data; these molecules all exhibited ≥ 90% βH inhibition and were considered active if they inhibited ≥ 90% Pf growth, both relative to the chloroquine control drug. The resultant set contained a total of 1606 molecules, with 374 actives (Supplementary Table ^S2). For the SVM model, the training molecules were converted to their 3D energy minimised representations at pH 5 ± 0.5 in LigPrep and multiple protonation states were retained, resulting in 1707 data instances.