Drug discovery

Chemical beauty contest

Most drug candidates fail clinical trials, in many cases because the compounds have less than optimal physico-chemical properties. A new method for assessing the 'drug-likeness' of compounds might help to remedy the situation.

Experienced medicinal chemists develop a sense of chemical aesthetics — a feel for how drug-like any particular molecule is. But is it possible to measure such chemical beauty? Reporting this week in Nature Chemistry, Hopkins and colleagues1 provide a quantitative estimate of drug-likeness that assesses a combination of a molecule's physical properties. Unlike the commonly used descriptions of drug-likeness, their approach allows a single, continuous scale to be defined, so that molecules can be ranked in order of desirability.

Drugs are developed from the optimization of 'lead' molecules, which are frequently found through the biological screening of compound collections. Before being finally accepted into use by regulatory and paying bodies, an optimized drug candidate must undergo years of intensive toxicological and clinical-efficacy studies. Most orally active drugs that survive these arduous developmental pressures have a set of physico-chemical properties that fall within a certain range of values — they are said to lie in a defined physical and chemical 'drug-like space'2,3,4,5. Until now, this drug-like space has been defined using cut-off values for permissible physical properties, perhaps most notably the values defined by the medicinal chemist Christopher Lipinski and his colleagues in the 'rule of five'2 (Box 1).

Hopkins and co-workers1 point out that Lipinski's rule can be misleading, because undesirable compounds could pass the drug-likeness test by only just meeting all four cut-off criteria, whereas better compounds could fail because they just miss one of the cut-offs. The application of the rule in this unintended way may help to explain why the compounds in current patents from pharmaceutical companies are, on average, significantly less drug-like than marketed drugs6,7,8.

Taking a cue from a study9 that used mathematical 'desirability functions' to assess how suitable a range of compounds would be as drugs that act in the central nervous system, Hopkins and co-workers1 used a similar approach to analyse the drug-likeness of a set of 771 oral drugs approved by the US Food and Drug Administration. The authors defined desirability functions for eight physical properties proposed to be important for oral drugs, including the four Lipinksi properties. They also took into account the number of aromatic rings and rotatable bonds in a molecule, the polar surface area (a measure of how hydrophilic a molecule is) and the number of groups in the molecule known to cause toxicity. The functions captured the full distribution of each physical property and provided a continuous quantitative estimate of drug-likeness (QED) on a scale from most to least drug-like.

Because the bulk physical properties of compounds are known to correlate with each other to some extent, Hopkins and colleagues used different property weightings to maximize the overall information content of the combined QED values for each drug. This weighting system will be controversial to some, because it may not reflect the importance of each property to drug-likeness. For example, there is a strong case to be made for lipophilicity as the dominant drug-like property, because it is important for a molecule's absorption, metabolism, promiscuity (binding to unwanted targets), toxicity and survival in the drug-development pipeline3,4,6,7. Nevertheless, the benefits of the authors' strategy are clear: their method not only computes drug-likeness on a single quantitative scale, but, more importantly, it reveals that drugs that fail the Lipinski criteria have distributions of drug-likeness that overlap with drugs that pass the criteria (Fig. 1).

Figure 1: Assessments of 'drug-likeness'.

Physico-chemical criteria, such as those defined by the Lipinski rule of five, are typically used to predict whether a compound is drug-like or not. a, The bar chart shows the number of oral drugs that fail or pass the Lipinski rule, based on a set of 771 drugs approved by the US Food and Drug Administration. b, Hopkins and colleagues1 report a method for predicting the drug-likeness of compounds on a scale of 0 (not drug-like) to 1 (drug-like). The chart shows the distribution of drug-likeness calculated using this method for the same drugs depicted in a. The analysis shows an overlap of Lipinski passes and failures for a range of drug-likenesses. Notably, some very drug-like molecules fail the Lipinski rule, whereas some very un-drug-like compounds pass it.

Hopkins and colleagues went on to show that their QED approach is better at differentiating drugs from non-drugs than the Lipinksi rule and other schemes based on cut-off values. They further validated the discriminatory power of their method by comparing QED scoring with the results of a study1 in which 79 medicinal chemists decided which of 17,117 molecules were attractive starting points for optimization as drugs, on the basis of only visual inspection of the molecular structures. Impressively, the QED scores for molecules considered to be attractive by the chemists were significantly higher than those for molecules considered unattractive. This suggests that the QED method may, at least in part, capture a sense of the chemical aesthetics that medicinal chemists develop through knowledge, experience and intuition10.

Finally, Hopkins and co-workers used their method to predict the drug-likeness of 167,045 ligand compounds that bind to 1,729 proteins (on the basis of binding information from ChEMBL, a database of biologically active compounds11). This allowed them to determine which proteins had the most drug-like set of ligands. Proteins whose ligands had the highest QED scores should be the most chemically tractable targets for drug discovery, because their known ligands are the most drug-like.

QED is not the final word in understanding the underlying features of chemical beauty and drug-likeness, but it does provide a holistic, more balanced assessment than previous approaches. It is also customizable, highly flexible and seems to be straightforward to implement. Any combination of physical properties can be chosen to define desirability functions, and users can set the relative weightings for properties as desired. QED should therefore find immediate use in replacing cut-off rules for the selection of oral-drug-like compounds for screening. The method can also be applied to control sets other than oral drugs, such as lead-like molecules12, compounds that belong to specific target and therapeutic classes, and drugs that are administered non-orally.

The widest long-term impact of Hopkins and colleagues' method1 should be on the optimization of lead compounds, where it will help medicinal chemists to prioritize which drug-like compounds to prepare. It is to be hoped that the implementation of improved guidelines for drug-likeness, such as QED, at this stage of drug discovery will improve the quality7 of candidate drug molecules, and eventually help to reduce the 96% attrition rate of compounds that enter clinical trials13.


  1. 1

    Bickerton, G. R., Paolini, G. V., Besnard, J., Muresan, S. & Hopkins, A. L. Nature Chem. 4, 90–98 (2012).10.1038/nchem.1243

    ADS  CAS  Article  Google Scholar 

  2. 2

    Lipinski, C. A., Lombardo, F., Dominy, B. W. & Feeney, P. J. Adv. Drug Delivery Rev. 23, 3–25 (1997).

    CAS  Article  Google Scholar 

  3. 3

    Wenlock, M. C., Austin, R. P., Barton, P., Davis, A. M. & Leeson, P. D. J. Med. Chem. 46, 1250–1256 (2003).

    CAS  Article  Google Scholar 

  4. 4

    Hughes, J. D. et al. Bioorg. Med. Chem. Lett. 18, 4872–4875 (2008).

    CAS  Article  Google Scholar 

  5. 5

    Luker, T. et al. Bioorg. Med. Chem. Lett. 21, 5673–5679 (2011).

    CAS  Article  Google Scholar 

  6. 6

    Leeson, P. D. & Springthorpe, B. Nature Rev. Drug Discov. 6, 881–890 (2007).

    CAS  Article  Google Scholar 

  7. 7

    Leeson. P. D. & St-Gallay, S. A. Nature Rev. Drug Discov. 10, 749–765 (2011).

    CAS  Article  Google Scholar 

  8. 8

    Hann, M. M. Med. Chem. Commun. 2, 349–355 (2011).

    CAS  Article  Google Scholar 

  9. 9

    Wager, T. T. et al. ACS Chem. Neurosci. 1, 435–449 (2010).

    CAS  Article  Google Scholar 

  10. 10

    Lajiness, M. S., Maggiora, G. M. & Shanmugasundaram, V. J. Med. Chem. 47, 4891–4896 (2004).

    CAS  Article  Google Scholar 

  11. 11


  12. 12

    Teague, S. J., Davis, A. M., Leeson, P. D. & Oprea, T. Angew. Chem. Int. Edn 38, 3743–3748 (1999).

    CAS  Article  Google Scholar 

  13. 13

    Bunnage, M. E. Nature Chem. Biol. 7, 335–339 (2011).

    CAS  Article  Google Scholar 

Download references

Author information



Corresponding author

Correspondence to Paul Leeson.

Rights and permissions

Reprints and Permissions

About this article

Cite this article

Leeson, P. Chemical beauty contest. Nature 481, 455–456 (2012). https://doi.org/10.1038/481455a

Download citation

Further reading


By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.