Computationally-enabled fragment growing without a structure

Advancing fragments without high-resolution structural information remains a challenge scientists often choose not to take on, according to our poll last year. But for many appealing targets, such as membrane proteins, structural information is difficult to obtain. In a new paper in J. Med. Chem., Peter Kolb and collaborators at Philipps-University Marburg and Vrije Universiteit Brussel describe a computational strategy.

The approach, called “growing via merging”, starts with a core fragment that binds to a target, in this case the β₂-adrenergic receptor (β₂AR). Ideally this interaction is structurally characterized, but if not a model can suffice. Here, the researchers started with five fragments they had previously discovered. All of these had in common a lipophilic core with a primary or secondary amine appendage; this is a known pharmacophore for β₂AR, so modeling could be used to orient the fragments.

Next, this core fragment is derivatized in silico with other fragments using a selection of 58 common reactions. Since all five fragments contained an amine, reductive amination was used here. A set of nearly 19,000 fragment-sized aldehydes and ketones was extracted from the ZINC database and computationally transformed into amines – as if they were reacted with one of the core fragments. These were then docked into the receptor, and those that did not overlap with the core fragments and also placed the amine near the amine of the core fragment were kept for further analysis.

The top 500-scoring fragments were then “reacted” – again in silico – with the core fragments and again docked. Eight of these were actually synthesized and tested for binding, of which four had higher affinity than the initial fragments. The best, compound 11, showed a 40-fold boost in affinity over its starting fragment.

This is an appealing approach, and it will be interesting to see how generalizable it proves. The β₂AR is a somewhat forgiving test case due to prior work on the target and the fact that the ligand’s amine interaction with a critical asparate residue helps to orient the core fragment. Laudably though, the computational toolbox (called PINGUI, for Pyton in silico de novo growing utilities) is open access. Please leave a comment and share your experiences if you’ve tried it.

More hits from a complex library?

One of the cornerstones underpinning fragment-based lead discovery is molecular complexity: fragments are less complex than larger molecules, and are thus likely to bind to more sites on more proteins. In theory, then, you want relatively simple fragments, and in fact Astex has actually formalized this with the concept of the “minimal pharmacophore”, in which each fragment contains a single pharmacophore (such as a hydrogen bond donor next to a hydrogen bond acceptor). But this is not the only way to build a fragment library; in 2016 we noted a paper out of the University of Dundee describing fragment libraries built with “caps” for easy derivatization. In a new paper in ChemMedChem, Paul Wyatt, Peter Ray, and collaborators at the University of Dundee and GlaxoSmithKline describe a screen with this “functional group complexity” (FGC) library.

The researchers were interested in the protein InhA, a drug target for Mycobacterium tuberculosis, the organism causing the eponymous disease. A relatively small library of 1360 fragments was assembled from six different sources, loosely defined by the authors:

• 573 commercial fragments
• 170 “3D” fragments from the 3DFrag consortium
• 326 of the designed FGC fragments
• 46 commercial fragments chosen based on known InhA inhibitors
• 124 “inventory” fragments
• 121 “project” fragments

These were screened against InhA in pools of 8, with each fragment present at 0.5 mM, using STD NMR, resulting in a fairly high hit rate of 11% (149 fragments). The commercial fragments and FGC fragments both gave a marginally higher hit rate (12.6%, 72 fragments and 13.2%, or 46 fragments respectively) while the 3D fragments gave a considerably lower hit rate (5.9%, or 10 fragments).

Previous work had suggested that more potent molecules seemed to reduce the STD signals for the NADH cofactor, so these molecules (32 fragments) were prioritized. The 13 FGC fragments represented a hit rate of 4%, nearly double the 2.4% for the library as a whole.

All 149 of the initial fragments were tested in a biochemical assay at 0.5 mM, but only 4 gave measurable inhibition – too few to draw conclusions. Five compounds were characterized crystallographically bound to InhA, including two of the FGC fragments. This information was used to merge two fragments, compound 24 (an FGC fragment) and compound 12 (a commercial fragment), yielding a mid-micromolar inhibitor. Adding a “magic methyl” gave a satisfactory ten-fold boost in potency. Fragment 24 was also merged with a previously reported molecule, compound 3a, to produce compound 42.

These results suggest that more heavily functionalized fragments don’t necessarily have a lower hit rate, albeit for a small library and a single target. And as we noted last year, molecular complexity is difficult to define; it is not immediately obvious that FGC fragment 24 is actually more complex than commercial fragment 12. The old cliché still holds: more data are needed.

Fragments in the clinic: ABBV-075 / Mivebresib

Bromodomains bind to acetylated lysine residues in proteins to control gene transcription. These epigenetic regulators have received considerable attention as drug targets, particularly for oncology. Last year we highlighted work out of AbbVie in which fragments found in an NMR screen were advanced to two series of molecules that potently inhibit the four members of the BET family of bromodomains. A more recent publication in J. Med. Chem. by Keith McDaniel and his colleagues at the company describes how one of the fragments was transformed into the clinical compound ABBV-075, or mivebresib.

Compound 6 was not the most potent fragment identified, but crystallography confirmed that it binds in the acetyl lysine binding pocket. The earlier work described how the pyridazinone moiety was replaced with a pyridone and another phenyl ring was added to make molecules such as compound 9, with sub-micromolar activity.

Further modification of the pyridone led to compound 19, with a nearly 20-fold boost in affinity. Crystallography revealed that the pyrrolopyridone makes a bidentate interaction with a critical asparagine residue in BRD4, and also displaces a “high-energy” water molecule.

Next, the researchers sought to pick up additional interactions, and it turned out that introducing a nitrogen off the central ring was synthetically straightforward and would point substituents towards a pocket in the protein. This led to low nanomolar inhibitors such as compound 25, and crystallography revealed that one of the sulfonamide oxygen atoms makes a hydrogen bond with a backbone amide. Happily, the improvement in potency was also accompanied by an improvement in stability in liver microsome assays.

Unfortunately, although the pharmacokinetics in mice were reasonable, these compounds showed high clearance in rats. Analysis of the metabolites revealed that this was largely due to oxidation of the unsubstituted phenyl ring, so the researchers took the classic route of introducing halogen atoms to both deactivate the ring and block metabolism sites. This ultimately led to ABBV-075.

In addition to excellent potency in biochemical, biophysical, and cell-based assays, ABBV-075 showed excellent antitumor effects in a mouse xenograft assay when dosed orally at the low concentration of just 1 mg/kg. In addition to BRD4, the compound binds tightly to the other BET family members but is selective against most of the other bromodomains. It also demonstrates good pharmacokinetic properties in mice, rats, dogs, monkeys, and humans. ClinicalTrials.gov lists a Phase 1 study currently recruiting.

This is a lovely, textbook example of how structurally-enabled fragment growing combined with careful pharmacokinetic-based optimization can lead to a clinical candidate. Obviously there is a long and uncertain road ahead for the molecule prior to approval, but getting this far is a victory in itself.

Pointless stereochemistry

Designing fragments to be more “three dimensional” than the flatter aromatic molecules that dominate most libraries is a topic often discussed in fragment library design. One way to make fragments more shapely is to introduce a stereocenter, but doing so often complicates the synthesis. In fact, new methods for efficient enantioselective synthesis constitute a major theme of organic chemistry research. In a recent paper in Angew. Chem. Int. Ed., Niklaas Buurma (Cardiff University), Andrew Leach (Liverpool John Moores University) and collaborators at Hawler Medical University Erbil and AstraZeneca demonstrate that the effort is sometimes not worthwhile.

Because proteins are chiral, different enantiomers can have profoundly different activities. The classic case is thalidomide, the racemic mixture of which was sold as a sedative in the 1950s, leading to the birth of thousands of babies with profound birth defects. Only one enantiomer appears to be responsible for the teratogenic effects, and many people are taught that had the manufacturer sold just one enantiomer, the disaster would have been averted. Unfortunately, biology is not so simple: the hydrogen atom attached to the chiral center is slightly acidic, and thalidomide rapidly racemizes at physiological pH.

Such racemization is more common than generally appreciated. The researchers experimentally measured the racemization of a couple dozen compounds using either circular dichroism (CD) spectroscopy or NMR (in the latter case, this involved dissolving the molecule in deuterated buffers and measuring the rate of deuterium incorporation, which occurs through an achiral intermediate).

The experimental results were then compared with those obtained through computational methods. Initially these were intensive quantum mechanical calculations, but the researchers also developed a rapid and effective approach by considering each of the attached substituents around the stereocenter independently. Importantly, the details for doing this are provided in the supporting information.

How much of a problem is this? The researchers provide four examples of what they call “potentially pointless stereoselective syntheses,” all published in high profile journals in 2016 (interestingly, three are fragment sized).

According to calculations, all of these molecules would undergo 19 to 70% racemization in 24 hours under physiological conditions.

So before embarking on any onerous stereoselective synthesis, it would be worth running a quick calculation. If the molecule goes forward you’ll still need experimental evidence for stability, but at least you’re less likely to be unpleasantly surprised by the answer.