Drug-likeness, multiple optimization and Alerts

As mentioned in the last blog I have been reading and digesting the fascinating paper by Bickerton et al., entitled “quantifying the chemical beauty of drugs”. Part of their paper came up with quantitative estimate of drug likeness (QED) based on multicriteria optimization. Had a busy week so I am a bit slower than Derek over at in the pipeline at reading this. What was interesting was that this study used molecular mass, logP, number of hydrogen bond acceptors, number of hydrogen bond donors, polar surface area, number of rotatable bonds, number of aromatic rings and structural alerts. Basically these “usual suspects” descriptors have been used in some format by some of the previous methods for developing rules for drug-likeness. Then the authors went through various benchmarking steps to compare QED vs many of the prior rules like the rule of five etc. So far so good. And for the purposes of the blog I am going to omit a good part of the paper on drug targets etc.

Something hit me in the methods section, they evaluated their method using 771 oral compounds, 554 were identical to those used to essentially build their QED method. So perhaps it is not unexpected that the method does slightly better than the other qualitative methods. Because their training and testing sets overlap to such a large extent. Surely an additional graphic (in addition to fig 2a) would have been to focus on the non overlapping ~220 or so compounds.

One point not discussed by the authors is the general tendency for scientists to convert quantitative measures to binary outcomes anyway by simply implying a cut off for drug likeness with QED score = x. I think this is human nature. As compounds with just one breach of the Rule of five would not be classed as failures, this scores comparably to the QED method on Fig2a showing the receiver operator curve. So if I had a choice, yes a quantitive score like QED would may be of interest but the ease and interpretability of the rule of 5 is very compelling. Yes it has its limitations but..I am still none the wiser on what is a beautiful drug.

This paper also gave me strong Deja Vu feelings and brought back some work I did a little over a year ago looking at the Alerts rules from different drug companies and ran a set of drugs through them as well as comparing to the rule of 5. I put this out as a white paper on slideshare in late 2010 and also used these images below in a paper with with Dr. Joel Freundlich focussed on TB. To quote from my white paper (bold added for emphasis) “One common observation looking at hits and approved drugs for neglected diseases is that to the experienced chemist many of the molecules appear ugly. As beauty is in the eye of the beholder it is hard to define ‘ugly’ but the incorporation of rules for chemical reactivity or structural alerts [24-28] can help. These filters in particular pick up a range of undesirable chemical substructures such as thiol traps and redox-active compounds, epoxides, anhydrides, and Michael acceptors. Reactivity can be defined as the ability to covalently modify a cysteine moiety in a surrogate protein [26-28]. Older rules such as the Lipinski rule of 5 [29] have been more widely used. For example if you look at the FDA approved drugs nearly 90% pass this rule (Figure 1). However the more Lipinski violations a compound has also correlates with the increase in the failure using various pharmaceutical filtering methods for reactive groups (Figure 2) [30]. So this suggests some undesirable or ugly molecules may have additional risks such as undesirable promiscuity or toxicity [31]. 

Sure I did not filter all the drugs (obtained from the CDD database) for just those that are oral, but there is clearly a correlation between rule of 5 failures and Alerts failures..and heck over 90% of drugs pass the rule of 5 and therefore are “drug-like”. Give me a rule anyday that gets ~90% right..A phone interview with a journalist today reminded of this work in the context of the Malaria dataset from GSK and an analysis done on this that showed though most of the compounds passed the GSK rules a large percentage failed the Pfizer and Abbott alerts. So what may pass in one company as a good compound is not the same across the board. Maybe we are in the same situation with drug-likeness. Everyone has an opinion but no one has the answer. The rule of 5 may guide you 90% of the way but after 15yrs the other 10% is still out of reach. The methods for predicting drug-likeness depend so much on what is a drug today that they may have a harder time telling us what will be a drug in 15-20yrs. Which is part of the challenge. Do not get me started on multiple optimisation as I have done my bit over the years to suggest we should think beyond a single property or rule or even measure of drug-likeness for making decisions!



Leave a Reply

Your email address will not be published.

You may use these HTML tags and attributes: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>