Druggability — the likelihood of being able to modulate a target with a small-molecule drug — is crucial in determining whether a drug discovery project progresses from 'hit' to 'lead'. With only 10% of the human genome representing druggable targets, and only half of those being relevant to disease, it is important to be able to predict how druggable a novel target is in early drug discovery.

A target's druggability is usually estimated by classifying it with known gene families that have previously been successfully targeted with drugs. But as the targets of some marketed drugs are considered as conventionally non-druggable, this approach comes with limitations. A new approach to predicting druggability has now been reported by Cheng et al. in Nature Biotechnology. They have devised a mathematical model that uses structural information about a target's binding site to estimate druggability.

On the basis that a target binding site has a maximal achievable affinity for a drug-like molecule, the authors proposed that this affinity could be calculated by modelling desolvation — the release of water from the target and the ligand after binding. Desolvation is dependent on the curvature and surface-area hydrophobicity of the binding site, which were represented by applying recently developed computational geometry algorithms to ligand-bound crystal structures of the target. The resulting maximal affinity predicted (MAPPOD) value from the calculation can then be converted to a more commonly used druggability score (Kd value) for different protein structures.

The model was applied to 63 structures representing 27 targets, of which 23 were marketed drugs. Targets known to be druggable were found to have more favourable MAPPOD values (<70 nM). However, there were five known druggable targets that had scored in the 'undruggable' target range. Further analysis of these targets showed that the model had eliminated troublesome structures, albeit ones for which drugs were ultimately developed, as the known drugs for these targets were prodrugs or required active transporter mechanisms.

A comparison of calculated MAPPOD values for 11 targets with published affinities of the most potent known drugs for those targets also showed a correlation. The only outliers were cABL kinase, which is targeted by imatinib (Gleevec; Novartis) and might therefore be explained by a preference for selectivity over affinity, and HMG-CoA (3-hydroxy-3-methylglutaryl-coenzyme A) reductase, the target of the statins. The statins are more polar than is normally desirable but are also exceptional in that their mode of action does not require good systemic bioavailability.

But how does this model fare against conventional high-throughput screening (HTS)? Using two novel targets, fungal homoserine dehydrogenase (HSD) and haematopoietic prostaglandin D synthase (H-PGDS), the authors showed that HTS results (16 hits for HSD and 200 hits for H-PGDS) correlated well with the prediction from MAPPOD affinities that HSD would be a difficult target. Indeed, follow-up lead optimization efforts from HTS against HSD did not find any drug-like leads, whereas 11 sub-micromolar leads were identified for H-PGDS.

As the authors acknowledge, it takes more than druggability alone to make a good target, and this model does not account for conformational changes, flexible binding sites or allosteric pockets. However, despite these constraints, such calculations might offer an explanation for unsuccessful screening campaigns and provide useful information when undertaking discovery efforts against therapeutically relevant, yet challenging, targets.