Covalent fragment linking on the SARS-CoV-2 Main Protease

Fragment linking is a compelling approach to advancing fragments, but, as our poll from several years ago showed, it often works better in theory than practice. Also, most examples of fragment linking involve non-covalent molecules. In a new open-access ChemMedChem paper, György Keserű and collaborators at HUN-REN Research Centre for Natural Sciences, University of Ljubljana, and Diamond Light Source apply linking to a covalent fragment.

 

The main protease (3CL^pro or MPro) of SARS-CoV-2 was the first target pursued in the fight against COVID-19. A crystallographic fragment screen was conducted just weeks into the pandemic, yielding dozens of hits. Some of these, such as compound 1, formed covalent bonds to the active-site cysteine. Compound 2 bound noncovalently nearby. In the new paper the researchers used the program DeLinker, a “graph-based deep generative method,” to try to link them.

 

Just three linked molecules were made and tested, each with a different amide connectivity. While two were quite weak, compound 5 showed low micromolar activity. This is especially impressive given that both compounds 1 and 2  themselves were inactive. SAR studies revealed that the nitrile warhead is essential for activity, and a crystal structure of compound 5 bound to 3CL^pro confirmed covalent binding. Surprisingly though, the structure showed that the moiety derived from compound 2 binds in a “flipped” position relative to the non-covalent complex. Further fragment-growing led to compound 28, the most potent compound described in the paper.

 

The researchers also tried warhead swapping, replacing the reversible nitrile warhead with a chloride for irreversible aromatic nucleophilic substitution. Although some of these molecules showed activity, the best was about 30-fold less potent. This observation makes sense, as the cysteine sulfur atom must take different trajectories to react with the different warheads.

 

As the researchers point out, more potent inhibitors of 3CL^pro have been reported; the approved drug nirmatrelvir targets this enzyme. However, the paper is still a rare example of fragment linking starting from a covalent fragment; see here for fragment merging on a different target.

 

Personally what I find most interesting is the fact that it took just three molecules to go from inactive fragments to a low micromolar inhibitor. But given that this  molecule binds in a somewhat different manner than predicted, perhaps a high-throughput approach, such as direct-to-biology, would generate even more potent molecules.

(Not) getting misled by crystal structures part 6: low ligand occupancies

It’s been a while since our last “getting misled by crystal structures” post. That one described unrecognized conformational heterogeneity of ligands. A more basic issue is ligand occupancy. It’s normally assumed that every protein in the crystal lattice has a bound ligand. A new paper in Structure by Timothy Stachowski and Marcus Fischer at St. Jude Children’s Research Hospital reveals that this is not always the case.

 

The researchers selected roughly 10,000 protein-ligand structures from the Protein Data Bank (PDB) and did a simple re-refinement of the ligand occupancies and B-factors (measures of conformational heterogeneity and modeling errors). 10% of the structures already presumed ligand occupancies at or below 0.9, but re-refinement saw the fraction jump to 35%, more than three-fold higher. There were no overall differences between covalent and non-covalent ligands, but 37% of fragments (defined as having MW <300 Da) saw decreased occupancy in re-refinement compared to just 22% of larger ligands. A few structures even saw occupancy drop completely. The authors wrote that “manual reviewing these revealed that ligands were built into spurious electron-density.”

 

Crystallographers use several metrics to assess the quality of their structures. In addition to B-factors, unique to each residue or atom, real-space correlation coefficients (RSCC) and real-space R values are commonly used, and the researchers compared these parameters before and after re-refinement. In many cases the metrics improved with decreased occupancy, but not for all metrics, and not always meaningfully. This means that standard assessments do not always flag partially occupied ligands.

 

OK, so a ligand binds at only 80% occupancy rather than the 100% assumed: does this matter? The researchers describe three categories where the answer may be yes. In the first, correcting ligand occupancy reveals alternative conformations of protein side chains, which could be informative for understanding the mechanism of binding. In the second, correcting ligand occupancy can reveal water molecules that interact with the ligand and/or protein. As we noted more than a decade ago, water is an essential player in protein-ligand interactions, and a single water molecule can make the difference between a binder and a non-binder. Finally, correcting ligand occupancy can reveal alternative binding modes for the ligand, and even ligands binding to other sites.

 

Importantly, this analysis was done on individually refined structures in the PDB, and it seems likely that the issues would be even more severe for structures batch-refined in high-throughput crystallographic fragment screens. As we wrote last year, the community needs to figure out how to deal with the increasing number of these structures.

 

The fact that roughly a third of PDB structures have less than 100% ligand occupancy has implications for training AI models. It also has implications for individual targets. As the researchers note, “non-crystallographers who rely on the PDB often assume that deposition itself is an implicit stamp of approval. Structural biologists, however, know that this is not always the case.” Before using a structure, it would be wise to re-refine it yourself.

FTO revisited: fragment linking this time

Four years ago we highlighted a fragment-merging approach targeting fat mass and obesity-associated protein (FTO), an RNA demethylase implicated in acute myeloid leukemia (AML). In a new J. Med. Chem. paper, Ze Dong, Cai-Guang Yang, and collaborators at University of Chinese Academy of Sciences Beijing describe a fragment linking approach that arrives at a similar outcome.

 

As we noted in 2021, previous research had revealed that meclofenamic acid (MA) binds in the substrate-binding site, near the binding site for the 2-oxoglutarate (2-OG) cofactor. In the new paper, the researchers synthesized an analog of MA which they then linked to analogs of 2-OG through a variety of linkers. Among the best of these was compound 8a, with low micromolar activity in an assay using PAGE (polyacrylamide gel electrophoresis) as well as mid-nanomolar activity in a different type of assay. This is an improvement over the MA analog itself (which was only tested in the PAGE-based assay, results shown in the figure). Compound 8a was also selective for FTO over two related proteins.

 

 

As is often seen in fragment linking, the SAR is quite sharp. The two-carbon linker was critical; lengthening the linker by one methylene or adding a methyl group abolished activity. The double bond in the 2-OG mimetic was also important; the saturated version of this molecule was inactive.

 

Surprisingly, a crystal structure of FTO bound to a molecule closely related to 8a revealed that while the MA moiety bound as expected, the 2-OG analog adopted a different conformation. But inexplicably, according to the experimental section, 2-OG was added to the crystallization solution, which would compete with compound 8a. Indeed, the structure deposited in the protein data bank shows the 2-OG analog N-oxalylglycine bound to the catalytic metal ion.

 

With two carboxylic acid moieties, it is no surprise that compound 8a showed no antiproliferative activity in AML cell lines. However, ester prodrugs did show low micromolar activity. Further characterization of one of these showed changes in protein levels consistent with FTO inhibition. This molecule also caused tumor growth inhibition after intraperitoneal dosing in a mouse xenograft model.

 

Superficially, compound 8a resembles compound 11b from the 2021 paper. Like that molecule it is probably too weak to serve as an ideal chemical probe. That said, with one fewer aromatic ring, compound 8a may be better suited for further optimization.

Fragment merging without crystallography for CGRP receptor antagonists

Migraines are the third leading cause of disability worldwide. Although the pathology is complex, blocking the interaction of calcitonin gene-related peptide (CGRP) with its receptor, thereby decreasing vasodilation, has proven successful in the clinic. However, some of the early small molecule antagonists were discontinued due to hepatotoxicity. In a recent J. Med. Chem. paper, Naohide Morita, Isao Azumaya, and collaborators at Kissei Pharmaceutical and Toho University describe a new class of inhibitors.

 

CGRP binds at the interface of a heterodimeric receptor comprised of the calcitonin receptor-like receptor (CLR) and receptor activity-modifying protein 1 (RAMP1). To find hits, the researchers screened a library of 2500 fragments (which could be up to 350 Da) at 500 µM against the extracellular CLR/RAMP1 domains using SPR. This yielded 565 hits, which were clustered based on similarity, and 250 were chosen for dose-response studies, leading to 38 confirmed hits. Competition studies with a known CGRP antagonist whittled this number down to just four, with compound 1 being chosen for further study due to ease of analog synthesis.

 

Compound 1 was confirmed as a binder using isothermal titration calorimetry (ITC). Unfortunately, co-crystallography with CLR/RAMP1 was unsuccessful, so the researchers turned to docking using information from known small molecule inhibitors. This work suggested that compound 1 binds to the CGRP receptor but does not interact with RAMP1, a conclusion further supported by mutagenesis studies.

 

To find fragments that bind RAMP1, the researchers performed a second fragment screen, again using SPR. This time the fragments were chosen from those in the first set that had not been tested in dose-response studies, supplemented with several hundred more selected based on structures of known CGRP antagonists. Of 784 fragments screened, 114 were taken into dose-response studies, leading to 8 hits. Compound 2 was the most potent, and mutagenesis studies suggested it interacted with RAMP1.

 

Crystallography of compound 2 was also unsuccessful, but docking, supported by NMR studies, suggested a possible binding mode. Compounds 1 and 2 were merged to yield compound 3, which had a satisfying 2000-fold improvement in potency compared to compound 1. Compound 3 also showed cell activity.

 

Compound 3 contains three stereocenters, so the researchers sought to simplify the molecule. They also needed to improve potency and metabolic stability. Multiparameter optimization ultimately led to compound 15, with picomolar(!) affinity for the receptor, subnanomolar activity in cells, and good pharmacokinetic properties. A standard model for migraine is inhibition of facial blood flow in marmosets, and compound 15 was active. The compound was also clean in tests for hepatotoxicity.

 

Although no further development of compound 15 is reported, this is a nice case study in fragment merging. As the researchers note, it is also one of just a handful of examples that succeeded in the absence of crystallographic data (we wrote about another one here). Hopefully this will further embolden researchers to pursue fragment merging and linking without direct structural information.

Solving protein-ligand NMR structures without isotopic labeling

Last week we highlighted a protein-detected NMR method that does not require expensive and sometimes difficult isotopic labeling of proteins. However, while that approach is able to provide affinity information, it does not provide structural information. A new (open-access) paper in J. Am. Chem. Soc. by Roland Riek, Julien Orts, and collaborators at the Institute for Molecular Physical Science and the University of Vienna tackles this challenge.

 

The approach builds on NMR Molecular Replacement (NMR²), which we last wrote about here. In NMR², brute force calculations obviate the need for assigning individual NMR peaks to specific protein residues, thereby sidestepping considerable up-front effort. Most of the new paper focuses on applying NMR² to ligand discovery for the oncogenic G12V mutant of KRAS, which I’ll briefly summarize.

 

The researchers start by screening the 890-membered DSI-poised fragment library (in pools of six, with each fragment at 0.6 mM) against KRAS using ligand-detected STD NMR. This produced 133 hits, which were then retested at 1 mM each using [¹⁵N,¹H]-HSQC two-dimensional protein-observed NMR, invalidating about 30% of them. Dose-response titrations were performed on the top 13 hits; all of them were found to be weak binders, with at best low millimolar affinity. NMR² was then used to determine protein-ligand structures for some of these hits. That information guided the design of additional ligands, which had slightly higher affinities.

 

This thorough description of the NMR² workflow should be useful if you’re trying to do this at home. But what really caught my eye was a bit at the very end of the paper describing a new relaxation-filtered NOESY pulse sequence. Specifically, “an inversion recovery pulse block serves as a T₁ filter, followed by a perfect echo sequence and a CPMG without J-modulation, as a T₂ filter.” In essence, the experiment takes advantage of the fact that proteins relax more rapidly than small molecules, so NMR peaks coming from the protein are filtered out. But NMR peaks from protons in the ligand that are in close proximity to protons on methyl groups of the protein are observed, and the intensity of these peaks correlates with the distance between ligand and protein protons. Feeding these distance constraints into NMR² generates a three-dimensional structural model. The researchers compare models generated using NMR² on unlabeled KRAS to those generated using NMR² on labeled KRAS and show that they are roughly similar.

 

This is a neat approach, and it will be interesting to see whether it catches on. According to our poll last year ligand-detected NMR has fallen to fourth place among fragment-finding methods, and protein-detected NMR is in seventh place. Perhaps approaches like this and that described last week will usher in a new era of NMR for FBLD.

Berotralstat: an overlooked fragment-derived drug

At the end of 2023 I mentioned that a paper by Dean Brown listed berotralstat as a fragment-derived drug. Readers will notice this molecule does not appear on our “fragments in the clinic” list. Did we miss it? After reading a (2021!) J. Med. Chem. paper by Pravin Kotian and colleagues at BioCryst, I believe the answer is yes.

 

Hereditary angioedema (HAE) is a rare genetic disease caused primarily by deficiencies in a protein that inhibits a serine protease called plasma kallikrein, or PKal. Drugs had already been developed to replace the inhibitor protein, but these need to be injected or infused. Since PKal is an enzyme, the researchers sought to make a small molecule inhibitor that could be taken as a pill.

 

BioCryst had developed an earlier drug called BCX4161, which is potent but has poor oral bioavailability. To find a better molecule, the researchers turned to the rich literature around serine protease inhibitors, which led them to make compound 2, a fragment of previously reported inhibitors of other serine proteases. The protonated benzylamine was expected to bind in the S1 pocket of the enzyme, and indeed the molecule did show weak but measurable activity.

 

Fragment growing led to compound 4, with double-digit micromolar activity. Building off the new phenyl ring led to more potent molecules such as compound 13, with low micromolar activity. Further structure-based design eventually led to BCX7353, or berotralstat. The paper provides good descriptions of the design rationale. For example, the fluorine was added to improve permeability, and the nitrile was added to improve the ADME profile. Modeling was used both to improve potency as well as to gain selectivity over other serine proteases. This proved to be successful: berotralstat is a subnanomolar inhibitor of PKal and at least several thousand-fold selective over trypsin and other serine proteases such as thrombin and FXa.

 

The pharmacokinetic properties of berotralstat in rats and monkeys were also good, and according to clinicaltrials.gov the molecule first entered the clinic in 2015. In December of 2020 the FDA approved berotralstat for prophylactic treatment of HAE attacks.

 

This is a nice story, and I agree with Dean that the discovery of berotralstat was “based on a legacy clinical candidate and fragment approaches.” The earlier molecule BCX4161 contained a benzamidine moiety, which was in part responsible for the poor oral bioavailability. Replacing this with a benzylamine fragment from the literature is a classic fragment strategy, and compound 2 is fully compliant with the rule of three.

 

So how was it missed? The abstract only states that berotralstat was discovered “using a structure-guided drug design strategy.” Indeed, the word “fragment” appears precisely once in the paper, albeit in a very telling sentence: “We evaluated these fragments in our PKal^pur inhibitor assay…”

 

From a timeline perspective, the approval of berotralstat makes it the fifth approved fragment-derived drug, after pexidartinib and before sotorasib. I’ll include it in the next update of clinical compounds, along with my standard disclosure that “the list is almost certainly incomplete.” What else are we missing?

The limits of published data

Machine learning (or, for investors, artificial intelligence) has received plenty of attention. To be successful you need lots of data. If you’re trying to, say, train a large language model to write limericks you’ve got oodles of text to draw from. But to train a model that can predict binding affinities you need lots of measurements, and they must be accurate. Among public repositories, ChEMBL is one of the most prominent due to its size (>2.4 million compounds) and quality, achieved through manual curation. But even here you need to be cautious, as illustrated in a recent open-access J. Chem. Inf. Model. paper by Gregory Landrum and Sereina Riniker (both at ETH Zurich).

 

The researchers were interested in the consistency of IC₅₀ or K_i values for the same compound against the same target. They downloaded data for >50,000 compounds run at least twice against the same target. Activity data were compared either directly or after “maximal curation,” which entailed removing duplicate measurements from the same paper, removing data against mutant proteins, separating binding vs functional data, and several other quality checks. They used a variety of statistical tests (R², Kendall τ, Cohen’s κ, Matthew’s correlation coefficient) to evaluate how well the data agreed, the simplest being the fraction of pairs where the difference was more than 0.3 or 1 log units, roughly two-fold or ten-fold differences.

 

The results were not encouraging. Looking at IC₅₀ values, 64% of pairs differed by >0.3 log units, and 27% differed by more than 1 log unit. In other words, for more than a quarter of measurements a molecule might test as 100 nM in one assay and >1 µM in another.

 

Of course, as the researchers note, “it is generally not scientifically valid to combine values from different IC₅₀ assays without knowledge of the assay conditions.” For example, the concentration of ATP in a kinase assay can have dramatic effects on the IC₅₀ values for an inhibitor. Surely K_i values should be more comparable. But no, 67% of pairs differed by >0.3 log units and 30% differed by >1!

 

The situation improved for IC₅₀ values using maximal curation, with the fraction of pairs differing by >0.3 and >1 log units dropping to 48% and 13%. However, this came at the expense of throwing away 99% of the data.

 

Surprisingly, using maximal curation data for K_i data actually made the situation worse. Digging into the data, the researchers found that 32 assays reporting K_i values for human carbonic anhydrase I, all from the same corresponding author, include “a significant number of overlapping compounds, with results that are sometimes inconsistent.” Scrubbing these improved the situation, but 38% of pairs still differed by >0.3 log units, and 21% differed by >1 log unit.

 

This is all rather sobering, and suggests there are limits to the quality of available data. As we noted in January there are all kinds of reasons assays can differ even within the same lab. Add in subtle variations in protein activity or buffer conditions and perhaps we should not be too surprised at log-order differences in experimental measurements. And this assumes everyone is trying to do good science: I’m sure sloppy and fraudulent data only make the situation worse. No matter how well we build our computational tools, noisy data will ensure they often differ from reality, whatever that may be.

Detailing hot spots with atomic consensus sites

Practical Fragments has written frequently about hot spots, regions on proteins that are predisposed to bind ligands such as drugs. Determining whether a protein has a hot spot can help prioritize a target for screening, and one of the more established computational approaches to do so is FTMap, which we wrote about most recently just a couple months ago.

 

While FTMap can tell you whether a protein has one or more hot spots, it provides few further details, such as which regions might prefer a hydrogen bond donor or acceptor. This has now been addressed in a new J. Chem. Inf. Mod. paper by Sandor Vajda and collaborators at Boston University, Stony Brook University, and Acpharis. (Diane Joseph-McCarthy presented some of this work at the CHI DDC conference earlier this month.)

 

The original version of FTMap started with a collection of 16 very small molecule probes: these were docked all over a protein, with hot spots being identified as consensus sites where many probes bound. To get more information about each hot spot, the researchers have extended the method – now called E-FTMap – by increasing the number of probes to 119 covering key functional groups. For example, whereas FTMap included dimethyl ether as a probe, E-FTMap also includes 2-methoxypropane, 2-methoxy-2-methylpropane, and tetrahydropyran. If all these probes bind with the oxygen in the same part of the hot spot, this suggests a predilection for a hydrogen bond acceptor, and also provides information about nearby hydrophobic contacts.

 

By using a sufficiently diverse group of virtual probes, E-FTMap is able to more finely detail hot spots, tallying the “atomic consensus sites” within them. This is reminiscent of an approach we wrote about several years ago, though that method used just three different probes.

 

To benchmark E-FTMap, the researchers took 109 fragment-to-lead pairs with published crystallographic information and assessed whether the program could identify interactions that had been experimentally observed. The results were encouraging and far superior to the original version of FTMap. The highest ranked atomic consensus sites generally overlapped with appropriate atoms in fragments and leads. Interestingly, the results for fragments were better than those for leads, and the researchers suggest this is because the fragment “core is responsible for the bulk of the binding free energy in a ligand and that larger ligands bind by forming additional interactions at weaker hot spots that surround the fragment binding site.”

 

Next, E-FTMap was tested against five proteins for which between 31 and 353 fragment-bound crystal structures were available. Here too the program was broadly successful, though some fragments bound regions of the protein that E-FTMap overlooked, particularly in cases where there were conformational changes. This is not surprising given that the program assumes the protein remains rigid. (Other computational approaches such as SWISH, which we wrote about here, are starting to account for protein flexibility.)

 

E-FTMap looks qualitatively at specific atomic interactions, and one question I had was how well the atomic consensus sites matched up with binding affinities of known fragments; perhaps some crystallographically identified fragments bind so weakly one would not expect to find them computationally, as we discussed here and here. This hypothesis might be tested by focusing on comparisons with experimentally characterized fragments with the highest ligand efficiencies.

 

Also, I was struck by the fact that the virtual probes in E-FTMap are roughly the size of MiniFrags or MicroFrags, and I couldn’t help but wonder how well the atomic consensus sites from the virtual screens would correlate with the binding modes of these tiniest of fragments.

 

One nice feature of E-FTMap is that it can be accessed through a simple web server, so if you’re interested in these and other questions you can test it for yourself. If you do, please share your experiences.

Hot spots real and imagined

Practical Fragments has written several times about “hot spots”: regions on proteins where small molecules and fragments readily bind. Knowing whether your target protein has a hot spot can help you decide whether to pursue the target in the first place. A variety of computational approaches have been developed for finding hot spots, most of which start with a crystallographically determined structure. In a new J. Chem. Inf. Mod. paper, Sandor Vajda and collaborators at Boston University and Stony Brook University ask whether computational models of proteins can also be used for one of the more popular methods, FTMap.

 

The researchers started with a set of 62 proteins, each of which had a published crystal structure bound to a fragment (MW < 200 Da) as well as to a larger molecule. The predicted structures of these proteins were then downloaded from the AlphaFold2 (AF2) site, and these models were truncated to correspond to the residues seen in the crystal structures to facilitate comparisons. The computational models were quite similar to the experimental models, particularly when comparing the positions of the peptide backbone atoms which define the overall shape of the proteins.

 

Next, the researchers applied the program FTMap, which computationally explores the surface of proteins using a set of 16 very small probes such as ethanol. Hot spots are regions where lots of probes bind, and the “hotness” of these spots correlates with the number of bound probes. FTMap assessed hotness on the AF2 structures and the crystallographicaly determined structures. (Before running FTMap, the bound ligands in the crystal structures were computationally removed.) Additionally, the researchers ran FTMap on unliganded crystal structures for the 47 proteins where these had been reported.

 

FTMap was broadly successful at finding the hotspots defined by bound fragments, succeeding 77% of the time starting with either the fragment-bound or unliganded structures and 71% starting with the AF2 models. Implementing stricter criteria (demanding the experimental fragment binding site be the top hot spot, for example) reduced the success to 56% for the crystallographic starting points and 47% for the AF2 models.

 

The paper discusses several examples in detail, in particular the two where the AF2 models were most different from the experimental models. Both of these were large, multidomain proteins. When AF2 models of just the ligand-binding domains were used, the models were significantly improved. This seems to be a generally useful hack: generating truncated AF2 models for other proteins also improved the performance of FTMap.

 

The utility of AF2 models for docking has been the subject of some debate, with some arguing that even though the overall protein folds may be accurate, local side chain conformations may be wrong, and a single side chain rotation may make the difference between ligand binding or not. This paper suggests that hot spots are not too sensitive to these subtleties, and that AF2 models can be used for finding hot spots.

Fragments vs IP6K1 without using structural information

The three members of the inositol hexakisphophate kinase family are potential targets for a wide range of diseases, from Alzheimer’s to cancer to metabolic disease. However, current inhibitors are not specific for individual isoforms. Also, the most potent compounds contain a carboxylic acid moiety, which is usually at odds with brain penetration. In a new ACS Med. Chem. Lett. paper, James Barrow and collaborators at Johns Hopkins School of Medicine, the Lieber Institute for Brain Development, and AstraZeneca describe neutral, selective inhibitors.

 

The researchers started with a high-throughput biochemical screen of 17,000 fragments, each at 100 µM, against IP6K1. The library itself is available here. After dose-response follow-up studies, 90 hits confirmed, with IC₅₀ values as good as 2 µM. Most of the hits contained carboxylic acids, but compound 5 did not, and also had good ligand efficiency.

 

No crystal structures of IP6K1 have been reported, so the researchers used an AlphaFold model for docking compound 5. This suggested a fragment growing approach. A variety of replacements for the pyrrolidine were attempted, and while some of these had improved activity many also proved to be chemically unstable. Removing the nitrogen and growing led to compound 24, which was both chemically stable and had sub-micromolar activity.

 

The quinazolinone core itself was associated with poor solubility, and the researchers made multiple attempts to modify it, such as introducing additional nitrogen atoms or methylating to remove a hydrogen-bond donor. Unfortunately, all these modifications led to significant losses in potency.

 

Compound 24 is highly selective for IP6K1 over IP6K2 and somewhat selective over IP6K3. Unfortunately, it showed no cellular activity, possibly due to modest biochemical potency and solubility. Nonetheless, this brief paper illustrates that starting with a larger than normal fragment library can lead to new chemotypes. Screening the larger library gave the researchers more chances to find fragments that did not contain carboxylic acids. Indeed, the difficulty of modifying the quinazolinone moiety demonstrates the utility of screening more molecules. Had the closely related molecules been in the library, they might not have turned up as hits, but their presence would suggest that relevant chemical space had been interrogated.

 

The paper is also a nice example of optimizing hits in the absence of structural information. Although much needs to be done to turn compound 24 into a chemical probe, the fact that it is still so small (almost rule-of-three compliant) provides hope that this can be done.

Solving structures with selective labeling and NMR2

Protein-detected NMR first enabled fragment-based lead discovery way back in 1996, but improvements in crystallography have now allowed synchrotrons to surpass big magnets as preeminent tools to determine how fragments bind to proteins. One of the major challenges in NMR is assigning the chemical shift values of atoms in all the individual amino acid residues. A technique called NMR Molecular Replacement (NMR²) sidesteps the need for this tedious, time-consuming process. A refinement to this technique, making it more broadly applicable, has just been published (open-access) in Sci. Reports by Julien Orts (University of Vienna), Martin Scanlon (Monash University) and collaborators.

 

As we discussed previously, NMR² relies on intensive calculations using experimental intermolecular NOEs between a protein and a ligand to generate a model. Although the method does not require assignment of backbone or side chain chemical shifts, it does require high-quality spectra. For example, if the spectra of several amino acid residues overlap it is impossible to distinguish them (this applies to conventional NMR methods too). The researchers realized that one way to simplify the spectra is through selective labeling, in which the methyl groups of the amino acid residues alanine, isoleucine, leucine, valine, and threonine are isotopically labeled with ¹³C. Going one step further, the entire protein can be deuterated (rendering most of the protein invisible to NMR), while these methyl groups retain ordinary hydrogen atoms.

 

For the present study, the researchers focused on the protein EcDsbA, an antibacterial target we’ve written about previously. They selectively labeled methyl groups so that, in isoleucine, leucine, and valine, only one of the two methyl groups was labeled. That reduced the total number of protons to just 6% of the unlabeled protein.

 

The researchers then solved the structure of EcDsbA with a previously identified ligand. At 23 heavy atoms the ligand is on the large side, though with an affinity of just 0.9 mM it presents a difficult test case. A total of twelve intermolecular NOEs were used in NMR² to build a model of the complex. One challenge with NMR² is that there may not be a single solution. For example, if two methionine methyl groups are both near a ligand, it may be impossible to determine a unique binding mode. This turned out to be the case, and the top two structures had different positions for a carboxylic acid group and a phenyl in the ligand.

 

To benchmark NMR², the protein-ligand complex was also determined using conventional two-dimensional techniques (HADDOCK and CYANA, which made use of assigned chemical shifts) as well as X-ray crystallography. These all agreed with the NMR² model in placing a phenylpropyl moiety from the ligand in a hydrophobic groove, but they differed in the placement of the carboxylic acid and the other phenyl moiety: the top scoring NMR² model agreed with the crystal structure and the CYANA NMR structure but differed from the HADDOCK structure, which was similar to the second-best NMR² model. Before assuming that the crystallographic structure is correct, though, it is worth noting that the ligand makes crystal contacts with a neighboring protein, and the electron density around the ambiguous phenyl is weak.

 

This is a nice demonstration of the utility of NMR². It seems to provide similar information as classic NMR methods, but the time taken is “orders of magnitude” less. And selective labeling should make NMR² applicable to even larger proteins. I look forward to seeing more people use this strategy.

Ensembles of fragment structures guide selectivity

Scientists generally want structural information when a project begins, and ideally that structural information comes from crystallography. Most of us who have been doing drug discovery for a while can remember seeing the first structure of a favorite molecule bound to a target protein and being inspired, reassured, or sometimes confused. But as crystallography becomes increasingly high throughput, it is now not uncommon to obtain dozens or even hundreds of structures. What to do with all this bounty? In a recent open-access J. Chem. Inf. Model. paper, Mihaela Smilova, Brian Marsden, and collaborators at University of Oxford, the Cambridge Crystallographic Data Centre, and Exscientia describe one application.

 

Back in 2016 we wrote about a computational approach called hotspot mapping, which uses three small fragment probes (aniline, cyclohexa-2,5-dien-1-one, and toluene) to virtually explore potential binding sites and map hydrogen bond acceptors, donors, and apolar interactions. The idea was to predict binding sites and the key interactions likely to drive affinity. The new paper focuses not just on affinity, but on selectivity.

 

The approach starts by taking multiple structures of the same protein bound to various ligands, especially fragments. Ligands and water molecules are then removed, and hotspot mapping is conducted for each structure. Then, all the hotspot maps are combined to generate an “ensemble” hotspot map, which in theory should give a more complete picture of potential attractive and repulsive interactions than a single structure.

 

To assess selectivity, the ensemble hotspot map of one protein is “subtracted” from that of another. If the proteins are very closely related, this “selectivity map” might be empty: all the interactions for one protein would be present in the other. But if there are differences, they become very apparent.

 

Several retrospective case studies are provided. In the first, ensemble hotspot maps were generated from the closely related bromodomains BRD1 and BRPF1, using 23 and 26 fragment-bound structures, respectively. The selectivity map clearly shows the potential for a hydrogen bond donor on a ligand to bind to the backbone amide of a serine in BRD1; the corresponding residue in BRPF1 is a proline, incapable of making this interaction. And indeed, an examination of the literature revealed that this interaction had previously been used to generate inhibitors of BRD1 that were 15-fold selective over BRPF1.

 

The kinases p38α and ERK2 are also closely related, but selectivity maps generated from five p38α structures and 17 ERK2 structures revealed a hydrophobic pocket in the former but not in the latter. This pocket had previously been used to generate selective inhibitors of p38α. Similarly, 28 structures of CK2α and 32 structures of PIM1 were used to generate a selectivity map that also revealed a hydrophobic pocket that can form in the former protein and had been used to generate selective inhibitors.

 

Generally, the more structures available, the more informative the selectivity maps are. The researchers note that though they only used five p38α structures, the fragments were chosen to be diverse (and interestingly all of them made interactions in the hydrophobic pocket). Also, while some protein flexibility can improve the maps, too much is a problem. (For the kinases, only DFG-in structures were used, for example.)

 

This method is a nice synthesis of experimental and computational techniques. A skeptic might argue that it doesn’t provide fundamentally new information: in the examples provided, the selectivity features had already been found and exploited by medicinal chemists. But the automated process and the clear output may speed things up, especially for newer targets, and indeed the researchers note that it is being applied in-house at Exscientia.

 

Perhaps most importantly, if you’d like to try it yourself, the code is freely available here. Happy mapping!

Fragments vs AXL, with help from an ELF

The receptor tyrosine kinase AXL has been implicated in multiple cancers, and more recently as a possible target for COVID-19. The protein has been pursued by many groups, mostly starting from high-throughput screens or existing inhibitors. In a recent Bioorg. Med. Chem. paper Pearly Ng, Alvin Hung, and collaborators at Experimental Drug Development Centre Singapore, University of Sussex, and LigatureTX describe a fragment-based approach.

 

The researchers started with a biochemical screen of 1700 fragments at 500 µM. Subsequent dose-response experiments yielded 16 fragment hits with IC₅₀ < 400 µM, all of which are shown in the paper. Four of these were based on 7-azaindole, a kinase hinge-binding motif that has previously been used in AXL inhibitors.

 

Seeking novelty, the researchers focused on the remaining 12 fragments and turned to their “expanded library of fragments (ELF),” which they have been building for several years. Though not described in depth, this collection was designed such that near-neighbors of their primary library could be rapidly screened to establish SAR and identify productive growth vectors. Compound 11, an indazole, led to several analogs from the ELF such as compound 24. This was docked into a publicly available crystal structure of AXL, and structure-based design led to compound 32. Further optimization ultimately led to molecules such as compound 52, the most potent in the paper.

 

Although compounds in this series were potent in the biochemical assay, they were typically 100-1000 less potent in a cell-based NanoBRET target-engagement assay, an annoying but not uncommon phenomenon. Several of the compounds showed high nanomolar activity in an ovarian cancer cell line with AXL overexpression. Interestingly, two compounds showed considerably more potent antiproliferative activity than target-engagement, and the researchers speculate they may be hitting other kinases. In support of this hypothesis, the researchers found that one of their compounds inhibited 12 of 97 kinases by >90% at 1 µM. This is not surprising given that indazoles are privileged fragments for kinases. Mouse pharmacokinetic studies revealed poor oral exposure for most of the compounds, though this could be improved by decreasing the basicity of the molecules.

 

A nice aspect of this paper is that most of the work was done without the benefit of structural information. Crystallography attempts with AXL and any of the ligands were unsuccessful, though the researchers were eventually able to obtain a structure of one of their more potent molecules bound to a related kinase. And just in time for the holidays, the paper also illustrates the utility of having an ELF on your shelf.

Selective fragments vs GPCRs, guided by modeling

Earlier this year we highlighted a fragment optimization success story against a G protein-coupled receptor (GPCR) which made no use of structural information. Due to the difficulty of crystallizing these membrane-bound proteins, structures have been rare for this large class of drug targets. Advances in crystallography are starting to change that. In a recent open-access Chem. Commun. paper, Jens Carlsson and collaborators at Uppsala University and the US National Institutes of Health make use of the increasing availability of such structures to develop potent, selective inhibitors.

 

The researchers were interested in A₁ and A_2A adenosine receptors (A₁AR and A_2AAR), targets for a variety of ailments from cancer to cardiovascular diseases. (A_2AAR was the subject of this blog post a few months ago.) In the current study, the researchers wanted to know whether structures and molecular dynamics (MD) simulations could guide production of selective inhibitors.

 

Previous computational and experimental work from the authors had yielded compound 1, with low micromolar activity against A₁AR and 7-fold selectivity over A_2AAR. Crystal structures of both these proteins are available, though not bound to the small molecule. Docking studies suggested that the ligand would make similar interactions to both proteins, but that there might be an opportunity for increased selectivity towards A₁AR due to the presence of a smaller threonine residue compared with a methionine in A_2AAR. Nine analogs were designed to grow into this lipophilic pocket, and free energy perturbation and MD simulations suggested that they would have improved affinity for A₁AR. This turned out to be the case when the molecules were made and tested in radioligand binding assays.

 

Although compounds 5 and 9 were more potent, selectivity was not improved. MD simulations suggested this might be due to the small size of the fragments, which could be accommodated in A_2AAR by slight shifts in the binding modes. To try to anchor compounds within the pocket, the researchers grew off the phenyl ring, leading to molecules such as compound 15. Borrowing from this molecule and compound 9 led to compound 22, the most potent and selective molecule in the series. (A separate effort led to a somewhat weaker but A_2AAR-selective ligand.) Both molecules were found to be antagonists when tested in cells, which was expected given that the crystal structures used for modeling were in the inactive conformation.

 

The correlation between predicted and measured binding energies was respectable, with a mean unsigned error (MUE) of 1.08 kcal/mol and Spearman’s rank correlation coefficient (ρ) of 0.8 for 24 compounds. Selectivity predictions were also impressive at MUE = 0.48 kcal/mol and ρ = 0.85.

 

This is a nice illustration of using computational methods to improve the affinity of a fragment by more than three orders of magnitude while also increasing selectivity. This particular system is probably on the easier side; we blogged about previous research from this group on A_2AAR back in 2013. The researchers note that proteins with larger binding sites and weaker ligands are likely to be more challenging. It will be fun to see efforts towards Class B GPCRs, for example.

A fragment library designed for merging: application to PKCζ

Advancing fragments in the absence of structural information has a reputation for being so challenging that some people do not even attempt it. Modeling can help, but what if you could improve your odds by designing your library strategically? This approach has been demonstrated in a recent J. Med. Chem. paper by Masakazu Atobe and colleagues at Asahi Kasei Pharma.

To facilitate fragment merging, the researchers synthesized a library of 5000 substituted isoquinoline fragments. As illustrated by the drug fasudil, isoquinoline is a privileged pharmacophore for binding to the hinge region of kinases. Importantly, isoquinoline has 7 different positions from which to grow: screening monosubstituted versions would potentially allow rapid merging of hits. This approach is conceptually similar to that used to discover vemurafenib.

The target of interest was protein kinase C ζ (PKCζ – that’s a zeta, by the way), one of the 11 members of the PKC family that has been implicated in diseases ranging from diabetes to cancer. Previously reported inhibitors are insufficiently potent or selective, in part because no crystal structure of the kinase has been reported. The researchers were interested in developing a chemical probe to better understand the biology.

A biochemical screen of the 5000-member isoquinoline library at 100 µM yielded just a dozen hits, with IC₅₀ values ranging from low to mid-micromolar. Importantly, substituents were found at four different positions, thus facilitating fragment merging. The researchers first merged fragment 6 with fragment 8, resulting in mid nanomolar inhibitor 10. Further optimization yielded compound 21, which is highly selective for PKCζ in a panel of 216 kinases and also has good pharmacokinetic properties in mice. However, cell potency is relatively modest.

Next, the researchers merged fragment 7 with fragment 9 to generate sub-nanomolar compound 26. This molecule also inhibited protein kinase A, but further optimization led to compound 37, which showed excellent selectivity in a panel of 381 kinases as well as good mouse pharmacokinetic properties and mid-nanomolar activity in a cellular assay. Encouragingly, the compound also showed good activity in a collagen-induced arthritis mouse model. The aniline – which adds about ten-fold to the affinity – may ultimately need to be removed, but clearly this molecule is well-suited for further optimization. 

This paper provides two lovely examples of fragment merging by design, but how general is the approach? One of the key advantages of fragment-based screening is the ability to survey huge swaths of chemical space. Building an entire library around a single fragment obviously constricts this. The fact that the hit rate (0.24%) was so low perhaps illustrates this point; it would be interesting to know how ligandable PKCζ is, or whether a library built around a different privileged pharmacophore would yield a higher hit rate. Lower expected hit rates necessitate larger libraries; 5000 fragments is already more than average according to our poll. And of course, if you are going to build a library of thousands of similar fragments, you had better be certain you choose one that has good pharmaceutical properties, further limiting your choices. Despite all these cavaeats, clearly the investment paid off for PKCζ. It will be fun to see what else comes out of this effort.

Fragments vs MNK1 and MNK2: take three

Mitogen-activating protein kinase-interacting kinases 1 and 2, or MNK1 and MNK2, are implicated in several cancers while seeming to be dispensable for normal cells, making them attractive oncology targets. Indeed, we’ve previously written about two FBDD-derived clinical compounds against these kinases, eFT508 and ETC-206. In a J. Med. Chem. paper published last month, Alvin Hung and colleagues at A*STAR describe a third series.

The researchers started by screening 1700 fragments in a biochemical assay, resulting in 11 molecules that inhibited both MNK1 and MNK2 with decent ligand efficiency. Four of these had a meta-substituted pyridine, as in compound 6. Making a few analogs led to compound 13, with low micromolar potency against both enzymes.

Crystallography proved unsuccessful, but making the reasonable assumption that the pyridyl nitrogen binds to the kinase hinge region led to two models, only one of which was consistent with the SAR. Decoration of the phenyl ring led to compound 21, with submicromolar activity. Although still early, the researchers collected both in vitro and in vivo ADME data on this molecule, which turned out to be quite promising.

Next the researchers turned to the pyridyl ring and found that appending small (5-membered) heterocycles could also boost potency, as in compound 36. At this point, after multiple attempts with previous compounds, crystallography finally yielded a structure that confirmed the proposed binding mode. Combining elements from compounds 21 and 36 directly led to only a slight boost in activity, but further tweaking ultimately led to compound 47, the most potent member of the series. Unfortunately this molecule was unstable in mouse liver microsomes, but a related compound showed good mouse pharmacokinetics as well as impressive selectivity in a panel of 104 kinases.

No pharmacodynamic studies are described, and perhaps this series was deprioritized to focus on ETC-206, which was also developed at A*STAR. Indeed, the later compounds in this paper reveal a frustrating struggle between potency and stability often seen in medicinal chemistry. This is captured in a nice timeline that shows a rapid improvement in potency over about 8 months, followed by a slight drop as the researchers tried to improve exposure. Although no Goldilocks molecule is reported, this paper is nonetheless a lovely example of fragment to lead optimization done for the most part without the aid of crystallography.

New chemistries for covalent fragments

Good things come in threes, and since our last two posts have covered covalent fragments, we thought we’d continue the theme with two more papers on the topic.

The first, in MedChemComm by György Keserű and colleagues at the Hungarian Academy of Sciences, is actually a companion to a paper we covered at the end of last year. The researchers were interested in electrophilic heterocycles, and assembled a library of 84, of which 57 were commercial and the rest were synthesized. In their previous paper, the researchers focused on reactivity against proteins. This paper is focused more on aqueous stability and intrinsic reactivity with glutathione, a biologically important thiol. The paper includes handy figures and tables summarizing reactivity rate constants and half-lives (at pH 7.4). These should be useful for selecting warheads that are sufficiently reactive as to be able to label proteins, but not so reactive as to label many proteins nonselectively.

The researchers also did computational studies to try to understand different trends in reactivity. And if you’re interested in testing the compounds yourself, they note that the “library is available for screening against relevant targets upon request from the authors.”

The second paper, by Philippe Roche, Xavier Morelli, and collaborators at Aix-Marseille University, was published in J. Chem. Inf. Model. Last year we highlighted their diversity-oriented target-focused synthesis (DOTS) approach, which combines virtual screening with automated synthesis to rapidly generate new compounds for testing. They have now expanded this approach (called CovaDOTS) to focus on covalent modifiers.

Conceptually, CovaDOTS is akin to fragment linking, in which one of the “fragments” is the nucleophilic residue in the target protein. The process starts with a known noncovalent ligand which is computationally grown by attaching it to commercially available building blocks that contain reactive warheads. These new molecules are then linked with the side chain of an amino acid (cysteine and serine in the paper), and the assemblies are docked against the protein to find molecules that fit well. Ultimately, the best would be resynthesized and tested.

The researchers applied CovaDOTS to three proteins for which covalent and non-covalent ligands had been previously characterized crystallographically. In the case of two kinases, EGFR and ERK2, the program performed well, with the “correct” (i.e., published) ligand observed in the top 14% and 4.4% of hits. For the serine peptidase PREP, the published ligand was the top hit of 303 molecules scored. In all three cases, the predicted binding mode also closely resembled the experimentally determined structure.

One limitation of CovaDOTS is that, as currently implemented, it considers only commercially available building blocks that also “possess both a warhead and an activated function compatible with the selected chemical reaction(s).” It would be interesting to combine the approach with the “make-on-demand” molecules we discussed a few months ago. And of course, it will be interesting to see real-world examples of how the program performs, in addition to these retrospective case studies.

This ends the June 2019 covalent fragment trilogy, but I think it’s fair to say that covalent fragments have a bright future. Look forward to many sequels!

Fragments vs PKC-ι: A*STAR’s second series

Just over a year ago we highlighted work out of A*STAR describing a series of inhibitors for the cancer target protein kinase C iota (PKC-ι). We ended by mentioning that the group had a second undisclosed series. This has now been described in ACS Med. Chem. Lett. by Jacek Kwiatkowski, Alvin Hung, and colleagues.

Compound 1 was among the fragment hits from the high-concentration biochemical screen previously mentioned. Although the researchers did not have a crystal structure, they assumed that the aminopyridine moiety was acting as a hinge binder, which helped them produce a computational model. A simple replacement of the phenyl ring with a pyridyl ring led to compound 2, with a satisfying improvement in potency and ligand efficiency.

As it turned out lots of diverse moieties could be substituted in place of the phenyl, including indoles and phenols. This promiscuity led the researchers to propose that the added heteroatom was making a water-mediated hydrogen bond to the protein; the water could rotate to either accept or donate a hydrogen bond to the ligand. Unfortunately, further growing from this ring did not improve potency.

Returning to their model, the researchers sought to grow from the aminopyridine ring towards a hydrophobic region of the protein. Adding a phenyl group (compound 16) was tolerated, though did not improve the affinity. However, the model suggested that an aspartic acid might be accessible from the phenyl ring, and indeed adding a positively charged piperazine as in compound 19 led to a nearly 100-fold boost in affinity. Unfortunately, the compound’s permeability (measured in a Caco-2 assay) was low, and perhaps because of this it showed only weak antiproliferative activity against hepatocellular carcinoma cells.

Ultimately the researchers were able to solve the cocrystal structure of compound 19 with PKC-ι, which mostly confirmed the model: the aminopyridine interacts with the hinge region, and the second pyridyl moiety likely makes a water-mediated hydrogen-bond with the protein, although the low resolution of the structure makes this somewhat ambiguous. The added piperazine appears to interact with a different aspartic acid than the one targeted.

Although there is more work to be done, it is notable that the researchers were able to optimize a fairly weak fragment to a sub-micromolar compound in the absence of experimental structural information. As they note, “while the empirical SAR remained our ultimate guide in fragment optimization, the model aided the successful design of potent inhibitors.” This is another nice example supporting our 2017 poll results, and recent review, that drug hunters can successfully advance fragments without NMR or crystallography.

Fragments vs PI3Kδ via deconstruction and regrowth

Ligand deconstruction, in which a larger molecule is dissected into component fragments that are subsequently optimized, can be useful for developing new chemical series. This is nicely illustrated in a paper recently published in J. Med. Chem. by Kenneth Down and colleagues at GlaxoSmithKline.

The researchers were interested in phosphoinositide 3-kinase δ (PI3Kδ), a popular target for a variety of indications from oncology to inflammation. They had already developed GSK2292767 as a clinical candidate, but they wanted a backup with a different chemotype. Crystallography revealed that the indazole moiety was interacting with the hinge region of the protein. Trimming off the top of the molecule (compound 4) led to a loss of both potency and specificity against three related members of the lipid kinase family, not surprising given the fact that indazole is a privileged fragment for kinases in general.

To generate a new series, the researchers sought to replace the indazole hinge binder using modeling and previously published information. Starting with a selection of more than 30 possible hinge binders, they synthesized 324 molecules and found that compound 11 was more potent and ligand efficient than compound 4, as well as reasonably selective against other PI3K isoforms. Growing this fragment-sized molecule led to compound 16, with low nanomolar potency against PI3Kδ, greater than 100-fold selectivity against three related PI3K isoforms and 29 additional kinases, good permeability, and activity in a cellular assay.

The careful observer will note that the dihydropyran hinge binder in compound 11 is shorter than the indazole in compound 4, and indeed crystal structures of compounds 11 and 16 complexed to PI3Kδ revealed that the pyridine sulfonamide fragment is shifted in the active site compared to the original drug molecule, accommodated by various conformational shifts in the protein.

This paper is a good illustration of what has been called fragment-assisted drug discovery. Nowhere in the article do the researchers use the phrase “fragment-based,” though they do refer to the pyridine sulfonamide as a “privileged fragment.” In the end, the proof of practicality is in the chemical matter, so we’ll need to wait until more is revealed about this series.

(Not) getting misled by crystal structures: part 5 – conformational heterogeneity

It’s been a while since the last installment in our “getting misled” series. One of the key issues with crystallography is that ligands are almost always modeled as binding in a single conformation. This does not necessarily reflect reality, as we discussed here. Indeed, as described here and here, subtle changes can cause ligands to dramatically change their binding modes, which could reflect the fact that the initial ligand itself had multiple binding modes, and the change simply shifts the equilibrium. In an effort to proactively seek out disparate binding conformations, Henry van den Bedem and a group of collaborators from Stanford, UCSF, Schrödinger, and Université Paris-Saclay have created a new program, which they describe in J. Med. Chem. (See here for In The Pipeline’s discussion.)

The open-source program, called qFit-ligand, starts with an existing protein-ligand structure and an electron density map. It first breaks the ligand into rigid fragments (such as rings) and rotatable bonds. Each rigid group is then allowed to move around and rotate to fit the density. Of course, this might entail the rest of the molecule moving as well to avoid bumping into the protein; up to five positions are stored for each rigid group. Combinations that best match the electron density are retained: for a ligand with three rigid groups, 15 conformations would be considered. Importantly, the entire process is automated.

The researchers validated qFit-ligand against a set of 73 reasonably high-resolution structures from the protein data bank (PDB) that had included two different binding conformations; they started with just one of the reported conformations and used their program to find the second. qFit-ligand was very effective at identifying cases where a terminal portion of the molecule had flipped or rotated, though less so for more difficult cases such as displacement of the entire ligand.

Next, the researchers turned to the D3R dataset of 145 high-quality, manually curated crystal structures, where qFit-ligand correctly identified 7 of the 10 structures with alternate conformations, and even identified an alternative conformation for a ligand that had not previously been detected.

The researchers then examined a large set of crystal structures that had been flagged as potentially dubious, and found several could be improved by including alternate conformations. Similarly, an examination of all 126 crystal structures of BRD2-4 bromodomain-ligand complexes in the PDB revealed that 12 almost certainly had previously undetected alternate binding conformations; another 24 likely did.

qFit-ligand strikes me as a powerful tool for getting beyond the static picture usually presented by crystallography. Because the program is automated, the researchers note, it should be complementary to high-throughput approaches such as PanDDa (which we described here). Of course, using qFit-ligand effectively assumes that everyone is aware of the potential for both false positives and negatives. As the researchers conclude, “communication between structural biologists, computational chemists, and medicinal chemists remains a requisite for successful, rational design.”