Phenotypic Screens, Chemical Genomics, and Antimalarial Lead Discovery

Several extensive small molecule screens against growing Plasmodium falciparum have recently been published [1–3] and thousands of hit structures are now publicly available. This represents a large majority of the drug-like chemical diversity currently available for screening and hence delineates the currently drugable target space of P. falciparum, since the ‘‘drugability’’ term includes an ‘‘availability’’ concept. From a chemical standpoint, some hits may look like bona fide drug leads while others more like chemical probes for target identification, but if the hit set is globally biased for any physicochemical property, relative to the starting screening libraries, it is the microorganism that ‘‘selected’’ for it. We can ask next what is the nature of the bias, and whether the chemical diversity identified in the screens is a reasonable representation of the chemotypes needed to inhibit the essential and potentially drugable targets in the pathogen. The usual answer to that question is ‘‘surely not’’, but why? The starting compound libraries are purposely biased to fit into the ‘‘ADME space’’ for orally bioavailable compounds [4,5] and by the practicalities of synthetic chemistry. Screening libraries at companies also reflect their interest in certain human targets, although in GlaxoSmith Kline’s case half of the starting compounds were purchased from outside vendors, and other published hit sets contain commercial compounds only [2,3]. It is difficult to estimate what coverage of target space has been achieved with the published structures. Nobody knows the total number of potentially drugable targets in Plasmodium, but as a first approximation we can use the figure of 400 predicted eukaryotic core essential genes [6]. Some gene functions will not be drugable, but others not belonging to the core set may be indispensable in the human host cell. In principle we could use the chemical families identified in the screens to roughly estimate the number of therapeutically relevant targets, meaning those that can be lethally affected by achievable concentrations of drug-like compounds. However, in the authors’ view, establishing a one-toone correlation would be unsatisfactory, as we would need to assume that each chemical family inhibits a different target. Chemoinformatic tools to classify compounds leave plenty of room to make each classification subjective. Chemists accept as a fact of life that the same compound can be classified in either one of two, or even more, related chemical families. That may carry less consequence in terms of chemical thinking, but in the context of this discussion, it means we cannot reliably establish a univocal correspondence between chemical families and individual targets. During our ongoing analysis of the Tres Cantos Antimalarial Set (TCAMS) [1] (deposited at https://www.ebi.ac.uk/ chemblntd together with similar sets from St. Jude Children’s Research Hospital and Novartis-GNF), we are finding that compounds in the same chemical family show different parasitological properties. Some inhibit an identified essential enzyme while others do not, and some exhibit a delayeddeath phenotype, but not so their fellow class members. Conversely, different chemical families are being found to kill parasites through inhibition of the same target. These findings show that a one chemical family–one target correlation cannot be reliably established when based on purely chemical criteria. Computational tools to analyze hit sets need to be biologically informed in order to be useful for generating target hypotheses. It is not helpful simply trying to define ‘‘the’’ physicochemical properties common to antimalarial hits from wholecell screens. One would not expect all binding sites for small ligands in a microorganism to have common features, and that they will differ between taxonomic divisions. So unless there is one, or very few, prevailing killing mechanism for most compounds in the set, the dominant requirements common to all hits against a given pathogen will be those broadly related to cellular transport (influx, efflux, and intracellular accumulation). Biological information must be layered on top of the chemical clustering to make it useful for investigating the target space of a pathogen. Computational exercises can estimate chemical similarities between compounds in the set and known ligands of specific proteins, or estimate the physicochemical complementarity between compounds and binding sites in proteins with known or modelled structures. Given the large gaps in the basic knowledge of Plasmodium, all these analyses require a great deal of extrapolation and lots of modelling, anchoring target predictions to very few known structures and making them highly operator dependent. The approach has recently claimed some successes [7–9], but to date most practitioners admit that truly novel insights are usually needles in a haystack of already known or strongly suspected targets. Sets of whole-cell hits should be rescreened for specific modes of action. Even with the small numbers of compounds in such sets, single target screens are probably not practical given the effort required to validate individual targets and develop robust assays, and attempts in that direction have failed to produce useful results so