Virtually screening 11 billion compounds – no problem!

Three years ago we highlighted virtual screens of roughly 100 million molecules which led to numerous high-affinity ligands against two targets. Those efforts made use of the Enamine “readily available for synthesis” (REAL) library, a virtual catalog of molecules that can be rapidly made and delivered. Enamine is continuing to grow this resource, which as of last year stood at 11 billion compounds. This is an impressive number, but how do you make use of it? In a just-published paper in Nature, Vsevolod Katritch (University of Southern California, Los Angeles) and a large group of collaborators provide a promising fragment-based solution.

 

Molecules in the Enamine REAL collection can be made using one-pot parallel synthesis from two or three reagents; for example, an amide could be made from an amine and a carboxylic acid. Enamine built a set of 75,000 reagents and 121 different reactions which collectively could produce 11 billion molecules (it’s even larger now). However, docking all of these could take thousands of years on a single CPU or cost hundreds of thousands of dollars on a computing cloud.

 

Rather than docking all the Enamine REAL compounds, the researchers developed an approach called virtual synthon hierarchical enumeration screening, or V-SYNTHES. The first step is to create a library of scaffolds with molecular weights in the 250-350 Da range. Taking the amide example above, imagine linking a set of 1000 amines to benzoic acid and a set of 1000 carboxylic acids to methylamine. This 2000 compound minimal enumeration library, or MEL, could be considered a subset of the full 1000 x 1000 = 1,000,000 virtual amide library. The numbers are even more dramatic for a three-component reaction: a MEL of just 1500 compounds could represent 125,000,000 fully elaborated molecules.

 

The MEL is docked against a protein of interest, and a diverse set of the top-scoring compounds chosen for fragment growing. In our example, the benzoic acid “cap” on the best compounds would be replaced by the full set of 1000 carboxylic acids. These would then be virtually screened, and the top compounds synthesized and tested.

 

The researchers applied V-SYNTHES to two targets. The first was a cannabinoid receptor bound to an antagonist. A total of 1.5 million molecules were docked against CB₂, representing 11 billion fully enumerated compounds. After filtering the best hits to remove PAINS and molecules similar to known CB₂ ligands, 80 diverse compounds were chosen for actual synthesis and testing, of which Enamine was able to deliver 60 in less than 5 weeks. One-third of these turned out to be antagonists with K_i values < 10 µM in biological assays.

 

How does this compare to a brute-force approach? Screening all 11 billion molecules wasn’t feasible, so the researchers screened a representative subset of the Enamine REAL library consisting of 115 million molecules – two orders of magnitude larger than the libraries screened in V-SYNTHES. Of 97 compounds synthesized and tested, only 5 turned out to be antagonists of CB₂ with K_i values < 10 µM.

 

A nice feature of V-SYNTHES is that it is well-suited to SAR-by-catalog. This was demonstrated by looking for analogs of the three best hits within Enamine REAL space. Of 104 compounds synthesized and tested, more than half had K_i values < 10 µM, and 23 were submicromolar antagonists. In fact, several turned out to be low nanomolar and selective not just against the related CB₁ receptor but against a panel of 300 other GPCRs.

 

V-SYNTHES was also applied to the kinase ROCK1 and achieved similarly impressive results: six of 21 compounds synthesized and tested had K_d < 10 µM in a binding assay, and one was a low nanomolar inhibitor.

 

This is a lovely and practical application of fragment concepts. Importantly, because the computational cost only increases linearly with the number of synthetic components while the library size increases with the square (for two-component molecules), it is very scalable; the researchers suggest that “terascale and petascale libraries” should be “easily” accommodated. These are numbers beyond even what DNA-encoded libraries can promise.

 

Currently V-SYNTHES relies on a good structural model for docking, but as computational predictions of protein structures become ever more accurate, perhaps even this will cease to be a limitation. Our SkyFragNet post from 2019 is looking ever more prophetic, in a good way.