If you haven't had enough computational fragment papers, here is one more. In this paper, Zhao et al. set out to find potent inhibitors of EphB4, a receptor tyrosine kinase. This is not novel target space; the inhibitor dasatinib is already on the market. This paper is an extension of the groups previous computational method, ALTA (Anchor-based Library Tailoring). In ALTA, 1. small, mainly rigid fragments are docked. 2. Compounds with the most favorable binding energy are used to select compounds which contain that fragment. 3. Fragment is then flexibly docked. In this paper, they add explicit solvent molecular dynamics to this process.
The figure below shows the approach to selecting fragments for screening. Most of the 563,000 fragments are 150-300 Da, have fewer than 5 rotatable bonds, and no formal charge. They then selected a kinase focused collection by retaining only those with molecular weight smaller than 300 Da, a maximum of three rotatable bonds, more than one ring [Emphasis mine], and the capability to form two hydrogen bonds with the backbone polar groups of the so-called hinge region. For the latter criterion acidic CH groups (e.g., in aromatic rings) were also considered as donors. The one ring criterion is because single ring anchors don't give enough energy of binding AND to open up IP space.
I find it strange that they consider an aromatic hydrogen capable of H-bonding for the purpose of calculating free energies. I would like to know what impact this additional factor had in picking the compounds; it is not explained in the paper, nor is it explained in the SI.
This led to the three active compounds shown in the table below. Previous work by them showed that the hydroxy at position 5 of compound 1 (Compound 7) would generate a significant increase in binding energy, through two additional hydrogen bonds. [And excuse my pedantry here, but there is no position 5, is there? Aren't position 3 and 5 here the same and indistinguishable? I am not the worlds best chemist, but I do know an equivalent position when I see one.]
Modeling confirmed this (prior to the initiation of chemistry). This was confirmed by compound 7 being 50x more potent than the parent compound 1. It is nice to see that this potency correlates with 2.5 kcal/mol or the addition of 2 additional hydrogen bonds, as was predicted. They then co-crystallized the compound with the target EphA3, despite EphB4 being the actual target. 32 of 36 residues in the actives site are the same, including the 100% identity for those involved in binding 7. The 1.7A structure confirmed the predicted mode of binding.
It was then tested against related Y-kinases: 0.338 μM for Src, 0.864 μM for Abl1, 1.38 μM for Lck, 1.62μM for EGFR, while no inhibition was observed for IGF1R.Thus compound 7 has higher affinity for EphB4 than for these five tyrosine kinases. It also showed cellular activity.
To sum up the state of the art of computational FBDD:
You can find fragments that fit in an active site, even if you have to model the active site.