It's "well known" that we all need to find the stronger inhibitors against any specific therapeutic target as possible. The motivation is that the strongest inhibitors are most probably less toxic. It's so obvious and thus needs to be checked.

To do so we took drug cards from the DrugBank and attempted to correlate the activity against the specific therapeutic targets (in log units) with LD50 (in log units as well).

The results are impressive on its own. In any activity range the toxicity of the compounds is very uniform. Let me repeat: there is the same probability to hit a highly toxic compound in pKI range, say 8-10 and 6-8. Once again: the toxicity is the ability of a compound to interact with a number of allimportant targets and has nothing to do with the ability of a compound to bind to a specific target of a therapeutic interest. What matters is of course the so called therapeutic window, which is the ratio of activity vs. toxicity.

We use a few chemical compounds vendors in our drug discovery programs (let us call them "Provider I", "Provider II", and "Provider III"). Nowadays chemicals providers normally claim about 1M of drug-like compounds readily available for shipping at a very reasonable cost. A few dozen of providers, with up to 10 of them very large overall could give up to a 10M of distinct compounds. How many compounds are out there?

To optimize our in-silico operations we use clustering to collect molecules of similar structure. The cluster centroids represent the clusters in the lead identification process and hence the number of clusters is a good measure of chemical diversity of a library. To cluster  the compounds we use Tanimoto similarity criterion, since the metrics lets us use fingerprint sorting to avoid much of pair-distance calculations. The number of common structures is a measure of (dis)similarity between a two chemical libraries. With this in mind we performed a co-clusterisation of the three vendor libaries (see the Graph on the left). It appears that the vendors II and III are roughly of the same "size" (diversity), whereas the vendor I has the most diverse collection of the compounds. What's remarkable, is that number of the compounds is the largest in the collection II and the smallest in the collection III.

Now we are ready to see what the claimed "drug likeness" of the compounds might mean. There are two great online libraries containing all the (small molecules) drugs (Drug Bank) and a large collection of the compounds with identified activity against specific molecular (proteins) targets. To see how the compound libraries and the biologically active compounds relate, we co-clustered each of the vendor libraries with those obtained from the DrugBank and BindingDB databases:

The results are remarkable in a few ways. First of all, all the three vendor libraries are very similar in their properties. Each of them contains roughly the half of the similarity classes representing the known drugs. This means that half of the current drugs is not "drug like enough" to be picked up by the modern "drug like" compound selection algorithms. Still, the number of stable compounds of reasonable size is about one or two orders of magnitude larger than the size of the modern vendor libraries. This means that we still have a long way to go on a chemical synthesis progress road to cover the chemical diversity enough at least to "rediscover" just the already known drugs!

There is also another remarkable conclusion: about 30% of the compounds classes overlap with the biologically active compounds from the BindingDB and therefore up to a one third of the compounds is biologically active! This may be a signature of a major library construction flaw: the compounds where selected to be "drug like", meaning rather similarity to compounds with known biological activity. In practice such promiscuity could mean a lot of side effects and toxicity.

Druglikeness is a qualitative concept used in drug design for an estimate on how "druglike" a prospective compound is. Usually it is estimated from the molecular structure, often even before the substance is synthesized and tested.

A good drug should show good availability, low toxicity and high potency. The quantitative measures of such properties are bioavailability (BA, measured in %), Maximum Recomended Daily Dose (MRDD, mmol/L) and IC50 against a drug's target.

The product of toxicity and availability, MRDD*BA, gives an upper bound on target IC50 and hence is an indication of a drug quality. The Figure above represents the distribution of such product for slightly over 100 drugs. As it can be seen from the Graph, most of drug compounds have the product small, roughly below 2*10^-5mol/L. Hence, small value of MRDD*BA product may be regarded as an indication of druglikeness.

In fact the situation gets even more interesting if the same druglikeness parameter is plotted in log-scale (see the Figure on the right). Since MRDD*BA limits drug's IC50 against its target, we can deduce that most drugs are centered around pIC50 = 5 (which means that the target pIC50 should exceed 5).