Alternative side-chain structures from methyl CPMG

As I have mentioned before on this blog, the use of tools like CS-ROSETTA holds the promise of determining protein structures using only the chemical shifts of its backbone atoms. In addition to potentially making NOEs and RDCs redundant, this technology allows biologists to determine the conformations of minor members of the structural ensemble, which are very difficult to obtain using conventional approaches in population-dominated techniques like NMR and X-ray crystallography. There are two limitations here, however. First, we only gain insight into the backbone, and as we know, the positions of side chains in minor states can be critical for function. In addition, backbone chemical shifts are not always available due to relaxation problems. Both weaknesses could, in principle, be addressed by extracting conformational information from the chemical shifts of methyl groups, which report on side-chain behavior and continue to give good signal even in very large proteins. This is the rationale behind a series of recent papers from the Kay lab [1-3] intended to determine changes in side-chain rotameric state from methyl relaxation-dispersion data.

Read the rest...

Dynamic origins of PBX1 homeodomain allostery

In the Monod-Wyman-Changeux model for cooperative binding, proteins exist in an equilibrium of low-affinity and high-affinity states in solution, absent any ligand. In this view, although it may appear that the binding of a ligand causes a conformational transition, it actually stabilizes one conformation from a pre-existing equilibrium. In the past several years, advanced NMR techniques have yielded increasing evidence that these structural equilibria exist for a number of proteins, suggesting that this model for linkage between conformational change and binding may be quite general. An upcoming paper in the Journal of Molecular Biology (1) is typical of such findings.

Read the rest...

The M2 channel controversy rides again

Most people never learn about an actual scientific controversy. Almost every "controversy" that bubbles into the public eye is manufactured, often reflecting social or ethical differences rather than genuine disagreements between experts about how different models fit to reality. Actual scientific controversies tend to be highly technical, and often concern points that lay people find to be esoteric. That doesn't mean that the issues involved aren't important, or that they're even difficult to understand. One controversy that has unfolded over the past few years and now may be over relates to a seemingly simple question. Where do adamantane drugs bind to the influenza A M2 channel?

Read the rest...

Do metamorphic proteins mediate evolutionary structural transitions?

On several previous occasions on this blog I've discussed proteins that undergo significant changes in structure without drastic changes in their primary sequence or solution conditions. In some cases, a few mutations can take a protein to a novel fold, as with Philip Bryan's protein G work. In others, closely related sequences within a whole family populate different kinds of folds, as Matt Cordes illustrated for the case of Cro proteins. In addition, there are some cases such as lymphotactin, where interconversion between two very different structures takes place at equilibrium, as illustrated by Brian Volkman's research. Each time stories like this come up I have mentioned that this kind of behavior (termed "metamorphism" in a 2008 commentary by Alexey Murzin) suggests a means by which proteins could evolve from one structure to another without losing foldedness or function. Recently, a group from the Weizmann Institute published results in PNAS that speak to this possibility.

Read the rest...

Dynamics conservation in the Ras superfamily

The proposition that general fold architecture is preserved within a family of evolutionarily-related proteins is not controversial. The amino acid sequence of a protein determines its structure, and countless studies have substantiated the idea that proteins with similar sequences will adopt similar folded conformations. Because structure and dynamics are intrinsically linked, one could reasonably assume that many features of a protein's dynamics get conserved along with the fold. A growing number of experiments show that this is indeed the case, including a recent paper in Structure (1).

Read the rest...

A single residue dictates a fold

Anfinsen's dogma — that the amino acid sequence of a protein uniquely determines its structure — naturally leads one to the idea that identity between amino acid sequences means identity between structures. This has proven to be a successful paradigm: sequence similarity reliably predicts structural and functional similarity. Evidence accruing in recent years, however, suggests that for small proteins, at least, this assumption may not be entirely safe. Adding to this view, in an article in PNAS this week (it's open access, so open it up), a research team led by Philip Bryan reports that they are able to generate significantly different folds with divergent functions from sequences that differ by a single amino acid.

Read the rest...

Non-native hydrogen bonds mediate structural transitions

A few weeks ago, I wrote that the goal of a structural biology research program ought to be to "characterize the conformation and energy of key, functionally-relevant members of the protein's structural ensemble and identify the pathways between them." The Nature paper last week, among other examples I mentioned in the preceding post, described functionally significant minor members of the native-state ensemble, and this is certainly an area where structural studies are making a lot of progress. But what about the other part of that statement, the transition pathways? How are we to study them, and what can we learn about them? Experiments alone are unlikely to tell us everything we want to know about the intermediates between different native structures. We can, however, use simulations validated by experiments to investigate the mechanisms of structural change. Today in Cell, research primarily performed by my coworkers Alexandra Gardino and Janice Velos demonstrates that the bacterial signaling protein NtrC  rapidly samples its active conformation even when it is not phosphorylated. Moreover, they confirm predictions that the intermediates between these two states are stabilized by hydrogen bonds not present in either one.

Read the rest...

Alternate structures and catalysis in cyclophilin

Of all the sources of structural variability in proteins, the hardest to pin down is side-chain conformational heterogeneity. Side chains aren't always easy to model into their primary conformation in the first place — you need excellent crystal diffraction or NMR data to do it. Even if you pulled that off, it's not always clear how (or if) side-chain fluctuations relate to a protein's activity. If we carefully examine our data for the faintest signals, however, we can sometimes find evidence of secondary conformations that play an important role in function. In an article in this week's Nature, James Fraser, myself, and others show that in the case of the proline isomerase cyclophilin A (CypA), alternative conformations of side chains play a key role in catalysis.

Previous experiments on CypA had established that the backbone amide groups of many residues were sensitive to a conformational fluctuation on the millisecond timescale. Under conditions where this enzyme is saturated with a peptide substrate, the fluctuation rate for some of these residues is very similar to the catalytic rate, suggesting that the dynamics and catalysis are linked in some way (2). Later experiments also showed that this fluctuation was an intrinsic property of the enzyme, continuing even in the absence of substrate (3). What we didn't know, however, was how the dynamics of cyclophilin were related to catalysis. We couldn't know, because we had no idea what the motion we were detecting was.

In the case of enzymes like adenylate kinase, there is a dramatic rearrangement of structural elements, and the population of conformations corresponding to the "end points" of that motion can be significantly enriched by altering the amount of substrate present in solution. In the case of CypA, neither of these things seems to be true. Supplementary Fig. 1a (freely accessible from the article page) neatly encapsulates the problem. For this figure, 48 structures of CypA, some with ligand and some without, were aligned, and the variation between them was determined. While there is some variability in the chain conformation, it is primarily limited to a group of residues known to undergo fluctuations that are not related with catalysis (blue chain). The residues involved in the catalysis-related dynamics don't seem to have much variability, even across this fairly large group. So we can't trap the unknown, minor state of CypA by adding substrate, and there's no evidence of an alternate state that explains the NMR data.

Knowing this, we suspected that some kind of side-chain motion accounted for the observed dynamics, probably involving an aromatic group of some kind. Our efforts to gather evidence for this, however, ran into some typical NMR problems — resonance overlap and poor sensitivity exacerbated by chemical exchange. Fortunately, the crystallographers came to our rescue, in the form of Tom Alber and his super-talented grad student Jaime Fraser. Jaime had determined a crystal structure of CypA at cryogenic temperature and analyzed the data using their algorithm RINGER, which examines electron density below the threshold typically considered "noise" in order to identify possible alternative rotameric states of side chains. He found evidence of multiple conformations for a few residues, but nothing that would explain the NMR results. Jaime had the bright idea to redo the experiment at room temperature, which Tom was convinced would result in nothing more than a radiation-damaged crystal and bad diffraction data.

What actually happened was that when Jaime examined the electron density from that experiment he could identify a group of side chains that had more than one conformation in the crystal, which you can see in Fig. 1. These residues included serine 99, methionine 61, and the catalytic arginine 55. Right in the middle of this group was phenylalanine 113, a residue with an aromatic side chain capable of causing changes in chemical shift at relatively long range. For context, the image to the left shows a structure of cyclophilin (PDB code: 1RMH) in complex with the model substrate we used in our own experiments (succinyl-Ala-Ala-Pro-Phe-p-nitroaniline), with the side chains of S99, F113, M61, and R55 in red. As you can see, F113 and M61 form part of the floor of the binding pocket, with S99 rather remote.

So here we have an alternative structure of CypA, hidden below the threshold typically considered when determining a crystal structure. It was certainly plausible that fluctuations in this ensemble of side chains could give rise to the NMR observations, but plausibility isn't proof. One way to address this would be to try and force CypA to adopt the less-populated conformation. If you look at Fig. 1d you can see that the two conformations of S99 lie at the standard rotameric positions, and that the less-populated rotamer of S99 would run into the more-populated rotamer of F113. So, if you replaced one of the side-chain hydrogens of S99 with a methyl group (i.e. mutated the serine to threonine), that might push the other residues of this group into their minor conformational state. So, that's what we did.

To the right you can see an overlay of structures for wild-type (WT) CypA (red) and S99T (green), aligned using structural elements on the opposite side of the protein from the active site. As you can see, the backbone traces match very closely, except for the helix and loop on the right. These elements are involved in crystal contacts in the S99T structure, but not the WT; a lower-resolution structure of S99T shows no differences here. Another key difference between these structures, of course, is the position of the side chains (thick neon); as shown here (and more clearly in Fig. 2c) they seem to have adopted the minor conformation from the WT structures. Although this mutation inspires widespread chemical shift changes (Fig. 3a) consistent with the hypothesis that this concerted side-chain rotation gives rise to the NMR observations, the structures seem very similar. Yet, S99T CypA differs from WT in two important ways.

The first difference is that the conformational fluctuations are dramatically slower, but only for residues that showed catalysis-related dynamics in WT (Fig 3d). In fact, this rate is now so slow that due to a quirk of NMR we can only determine the slowest rate of the process. At 10 °C, this fluctuation in S99T is about 60 times slower than the slowest process in WT.

The second key difference between the mutant and the WT is that catalysis is dramatically slowed. Because CypA does not consume its substrate (it acts on both cis- and trans- proline bonds) its activity can be assayed by NMR, as you can see in Fig. 4. As with any enzymatic assay, the net activity is proportional to the amount of enzyme added, so just glancing at these spectra (and knowing the enzyme concentration) you can estimate that S99T has at least 40-fold lower activity than WT enzyme. If you actually perform the fits, it turns out that the reaction velocity for S99T is about 240 times lower than that for normal CypA, but this includes a contribution due to the fact that S99T does not bind its substrate as tightly either. If you correct for this, it turns out that S99T has about 70-fold less activity than the normal enzyme. Not only is this similar to the change in dynamics, it's also quite comparable to another mutation, R55K, that removes a group that performs some of the chemistry.

These results indicate that a conformational change in a group of side chains including F113 is primarily responsible for the chemical exchange behavior observed in WT. The S99T mutation stabilizing the minor conformation dramatically and similarly reduces both the conformational fluctuation rate and the catalytic rate. This suggests that dynamics and catalysis are linked not by happenstance but by some direct relationship. Unfortunately, these experiments do not provide any direct insight into the mechanism by which dynamics contribute to catalysis. They do establish, however, that in CypA coherent fluctuations of side chains, barely detectable in protein crystals, nonetheless make a critical contribution to function.

1) Fraser, J.S., Clarkson, M.W., Degnan, S.C., Erion, R., Kern, D., & Alber, T. (2009). Hidden alternative structures of proline isomerase essential for catalysis Nature, 462 (7273), 669-673 DOI: 10.1038/nature08615

2) Eisenmesser, E.Z., Bosco, D.A., Akke, M., & Kern, D. (2002). Enzyme Dynamics During Catalysis Science, 295 (5559), 1520-1523 DOI: 10.1126/science.1066176

3) Eisenmesser, E., Millet, O., Labeikovsky, W., Korzhnev, D., Wolf-Watz, M., Bosco, D., Skalicky, J., Kay, L., & Kern, D. (2005). Intrinsic dynamics of an enzyme underlies catalysis Nature, 438 (7064), 117-121 DOI: 10.1038/nature04105

Read the rest...

Don't look for "the" structure

It surprises me how often I hear students, postdocs, and even professors talk about determining the structure of a protein. A singular structure has the advantage of being relatively easy to interpret, but the cost of this is often the loss of functional data. It's easy to understand how this terminology emerges from the discipline of crystallography, which after all only works when the protein molecules adopt only a small number of conformations. Yet even when it comes to NMR, a technique that should be very sensitive to the fact of structural multiplicity, the language of researchers and the structural tools available to them are too often oriented towards the idea of a singular structure. But any representation of a protein as a single conformation is a simplification — every protein exists in multiple structural states.

Trivially, we are aware that a given polypeptide chain can adopt a number of different conformations — the "folded state" of any given polypeptide chain covers only a tiny sliver of the possible conformational space. A protein that is "unfolded" occupies not a single, well-defined state but a vast multiplicity of states, and this kind of statement is not controversial because we tend to imagine unfoldedness as a messy chaotic jumble of conformations. The reality is less cut-and-dried: although unfolded proteins may have no regular structure, many still have a propensity to form particular secondary structures or interactions. The reality of denatured proteins is that they have a complex and varied energy landscape, not an array of possible structures that all have roughly equivalent energy. The flipside of the popular view is that the a protein's native state draws down to a sharp energy well, and this conception is also misguided.

The most dramatic counterexamples to the idea of a neat, punctate energy well come from proteins that adopt several different folds in the native state. One relevant case is lymphotactin, which freely interconverts between an α/β monomer and an all-β dimer under physiological conditions. Lymphotactin may be unusual, but the principal message from that study is one that ought be paid attention to in others, particularly when the protein in question has functional conformational diversity. Consider α-synuclein, a protein implicated in Parkinson's disease. In the presence of some detergent micelles this protein is known to take on an α-helical hairpin structure, with two helices laying down on the charged surface of the lipid headgroup. In solution, however, it seems to take on a number of different forms, and may interact with true lipid bilayers in a completely different way than it interacts with micelles. For proteins that interconvert between several different physiologically-relevant folds, one is never pursuing the structure, but rather a structure.

Of course, we don't expect most proteins or domains to regularly adopt alternate overall folds. However, reorientations of domains or monomers is a relatively common behavior, and one that poses a sticky challenge for structural biologists because incidental properties of a particular arrangement may bias our experiments towards observing it. A minor member of the ensemble, if it has favorable packing geometry, may exclusively populate a crystal. Similarly, NMR experiments to determine domain arrangement via residual dipolar couplings must always be undertaken with an eye to ensuring that interactions with the aligning media do not bias the results. No single structure of adenylate kinase can instruct us about its catalytic cycle, and structures of the unbound state do not capture the reality that the protein continues to open and close in the absence of ligand. Single structures do not capture motions of domains or monomers relative to each other and that often means an incomplete understanding of function.

Domain motions are also an overly dramatic example, because simpler rearrangements of the backbone take place in many proteins, even when regular secondary structures are evident. Fluctuations of the main chain play a functional role in several proteins — as, for instance, in the flaps of the HIV protease. Additionally, rearrangements of the backbone have a significant role in signaling, as in NtrC, which I'll talk about more in two weeks. Proteins where the main chain rearranges in response to ligand binding or post-translational modification generally cannot be described by a single structure.

Even if the backbone is rigid, every protein will have flexibility in the side chains of its amino acids. One of course expects to see this kind of behavior in side chains on the surface of a protein, where it is usually dismissed as irrelevant. However, we also know that side chains can rotate and move in the core of a protein, and that on some protein surfaces they can undergo coherent rearrangements. I'll talk a bit more about the functional relevance of side-chain motions next Thursday. For now, suffice to say that side chain rotations cannot be so easily ignored and sometimes have functional effects. Structural studies that do not capture these rotations may be missing something important.

My point here is not that single structures are stupid or useless. A structure can be very informative about about a protein's function, and often has great power to explain the effects of mutations and ligands. However, we should not mislead ourselves into thinking that any single structure will have all the answers, or indeed any of them. Every protein is a constantly interconverting ensemble of structures, and there are many layers of structural diversity within that ensemble, reaching from whole fold rearrangements to "mere" side-chain adjustments. Determining the structure of a protein is not a coherent goal for a research program. The successful structural biology study will characterize the conformation and energy of key, functionally-relevant members of the protein's structural ensemble and identify the pathways between them.

Read the rest...

The role of dynamics in catalysis

For some enzymes, dynamics on the millisecond timescale play a critical role in catalysis. I don't think this is a particularly controversial or unclear statement, but then, I know what I mean by it. In the process of communication, however, the intended meaning sometimes gets lost or transformed. A statement that addresses an entire catalytic cycle, for instance, might be interpreted as addressing only the chemical step. This seems to have happened in a pair of papers that concern the transfer of energy from conformational rearrangements to a chemical reaction.

Consider a reaction scheme in which an enzyme loosely associates with substrates (E.S), then "closes" to form a tight, catalytically-competent complex that then undergoes a reaction with the rate k_chem:

Pisliakov et al. (1) ask whether the closing process can accelerate k_chem. They ask this question primarily because a group from Harvard University proposed that this was possible in a paper printed last year in J. Phys. Chem. B (2). In that paper, Min et al. performed some simulations suggesting that such an acceleration was at least possible, and consistent with some enzymatic data. Pisliakov et al. approach the question with simulations of the reaction of the phosphotransfer enzyme Adk with 2 ADP molecules to form ATP and AMP. As part of the catalytic cycle, the enzyme goes from an open state (PDB: 4AKE) where the ATP and AMP binding sites are exposed to solvent, to a closed state (PDB: 1ANK) where the substrates are shielded from the surrounding solution by ATP and AMP "lids" that close down over the active site.


One can, perhaps, imagine that when the enzyme closes around the substrates, some motion will occur that promotes the transfer of a phosphate group from one molecule to another. Pisliakov et al. use a three-tiered system of simulations to address the question, as a way of trying to get around the difficulty of dealing with the long timescales required. Their simulations allow them to adjust the energy barrier to match the experimental rates or accelerate the reaction so that the whole pathway can be simulated. In general, they find that conformational fluctuations do not enhance the chemical reaction rate in this system.

I have two main concerns about the science that was performed here. The first is that the energy barriers in the long-timescale experiment appear to be improperly paramaterized. In estimating these barriers for the phosphotransfer reaction in Adk, Pisliakov et al. used 260 /s as k_chem. However, although the actual reaction carried out by Adk follows an extremely complex scheme, the analysis performed by Wolf-Watz et al. utilized a simplified scheme that combined all post-association steps into a single k_cat. This is why the concordance between k_cat and k_open justifies the conclusion that lid-opening is rate-limiting. In principle, the experiments used for that paper are incapable of separating the opening and closing steps from the chemical step. Therefore we have no experimental knowledge of the phosphotransfer rate, except that it is greater than 260 /s. This perplexing error appears to have originated with Min et al., but I am surprised Warshel's group did not catch it.

This is not a major problem because the bulk of the conclusions of the experiment were drawn from a different simulation in which the energy barriers were lower, but this leads to my second concern. If the structural transition involves a very smooth and coherent rearrangement of the protein, then simply manipulating energy barriers should not result in a serious error of analysis. In reality, however, ensemble motions of protein elements are not going to be so directed or uniform. Structural rearrangements are not highly singular steps, but involve a large number of intermediates and transition states. Motions in the late stages of the structural transition that promote catalysis may well be missed by simplified models, or accelerated beyond productivity by lowering the energy barrier.

That said, I'm not particularly surprised that Pisliakov et al. find that energy from the conformational coordinate does not transfer to the chemical coordinate, nor do I disagree with the finding. Despite what Pisliakov et al. appear to believe, the papers that have come out of Dorothee's group don't argue that the millisecond motions contribute directly to the chemistry. Doro doesn't believe that for a second. Neither do I. The importance of dynamics has little to do with shoving the reaction along the chemistry coordinate, but everything to do with getting substrates bound and into a state where chemistry is possible.

Dynamics allow an enzyme to reconcile incompatible functional requirements. To efficiently function as a phosphotransfer enzyme (as opposed to a hydrolytic phosphatase), Adk must expel water from the active site during catalysis. If the active site is inaccessible to solution, however, there is no way for the substrates to diffuse into it. It is difficult to create a single, rigid fold that can accommodate both these demands, but by fluctuating between two states the problem is resolved quite easily. So yes, the dynamics are essential to catalysis, but that does not imply that the conformational and chemical energy coordinates are coupled.

More perplexing is the discussion of the hierarchy of motion, which Pisliakov et al. take to mean that nanosecond motions somehow contribute to the chemical coordinate. As I discussed when that paper was initially published, the question being addressed was whether and how motions on the fast timescale (ps-ns) in Adk were related to the slower (ms) motions of the lids. In a hierarchy of motion, fast timescale fluctuations enable or promote slow timescale dynamics. In the case of Adk, this means that nanosecond flexibility at structural hinges allow the millisecond motions of the ATP and AMP lids. It was not implied, then or since, that the nanosecond motions in question make a direct contribution to movement along the chemical coordinate. This is not to say that there are no researchers who believe that ns motions contribute to catalysis — I've previously mentioned some work on hydrogen tunneling that makes precisely this argument. In the specific case of Adk, however, the contribution of ns motions to catalysis consists entirely in their enabling of the slower ensemble motions of the nucleotide binding domains, and nobody but the Warshel group has suggested otherwise.

There is an ongoing disconnect in the literature concerning the role of dynamics in catalysis. While it is true that in many cases rates of structural transitions correlate with rates of catalysis, this does not imply that the conformational transition coordinate is linked to the chemical reaction coordinate by direct transfer of energy. It is more likely that the dynamics of the enzyme contribute to catalysis by generating reaction-competent states from reaction-incompetent states. This is not to say that dynamics cannot possibly make a contribution to phenomena such as hydrogen tunneling, but it strikes me as unlikely that motions on the millisecond timescale will contribute to a chemical coordinate. Experiments, rather than simulations, will be the ultimate test of the idea. However, in principle, this hypothesis can only be tested experimentally on enzymes where the conformational changes do not limit the chemical reaction rate. Because the rate of the chemical step is unknown in Adk, it may not be an appropriate model system for addressing this question.

1. Pisliakov, A., Cao, J., Kamerlin, S., & Warshel, A. (2009). Enzyme millisecond conformational dynamics do not catalyze the chemical step Proceedings of the National Academy of Sciences, 106 (41), 17359-17364 DOI: 10.1073/pnas.0909150106

2. Min, W., Xie, X., & Bagchi, B. (2008). Two-Dimensional Reaction Free Energy Surfaces of Catalytic Reaction: Effects of Protein Conformational Dynamics on Enzyme Catalysis The Journal of Physical Chemistry B, 112 (2), 454-466 DOI: 10.1021/jp076533c

3. Wolf-Watz, M., Thai, V., Henzler-Wildman, K., Hadjipavlou, G., Eisenmesser, E., & Kern, D. (2004). Linkage between dynamics and catalysis in a thermophilic-mesophilic enzyme pair Nature Structural & Molecular Biology, 11 (10), 945-949 DOI: 10.1038/nsmb821

Read the rest...

Filling the donut hole in dynamics

This is the last of my series of posts about the dynamics-focused topical issue of JBNMR. There are plenty of other excellent papers in it, and I encourage you to at least glance over all of them, especially if you're an NMR person.

Standard NMR dynamics experiments on isotropically tumbling proteins cover a wide, but not comprehensive, swath of fluctuation timescales. Limited information about motions that take milliseconds or more can be obtained from hydrogen exchange data; the AMORE-HX experiment is meant to obtain this kind of information. Fluctuations with time constants in the range of μs-ms can be measured by relaxation-dispersion experiments, as were used in the Peng lab's paper. Motions that are faster than the rotational correlation time of the protein can be characterized using dipolar relaxation data of the kind I collected for the field-cycling experiment. There are additional kinds of data that also cover these areas, but the glaring hole lies between the correlation time of the protein (several ns) and the low end of the chemical exchange regime (several μs). Several teams, including that of Christian Griesinger, propose to fill this gap using data derived from residual dipolar couplings (RDCs). In the topical dynamics issue of the Journal of Biomolecular NMR, his group demonstrates the use of this technique in measuring the dynamics of side chains in the small protein ubiquitin. The article is open access, so feel free to open it up and read along.

The strength of the dipolar coupling between two nuclei depends on their magnetic properties (specifically their gyromagnetic ratio), the distance between them, and the angle between the internuclear vector and the vector describing the external magnetic field. For solution NMR this last component is typically not important because most proteins tumble randomly with respect to the magnetic field, causing this interaction to be averaged away. However, if you were to somehow introduce a tiny amount of bias into the tumbling, very slightly aligning the protein parallel or perpendicular to the magnetic field, a residual portion of this coupling could be recovered. The effect is considerable: a net alignment of less than a fraction of a percent generates couplings on the order of 30 Hz or more.

There are many ways of inducing this alignment. Large charged particles such as phage or DNA nanorods have been used, as have assemblies such as charged or polar lipid bicelles. In addition, proteins can be labeled with paramagnetic metals to induce fractional orientation. Even mechanically manipulated media, such as acrylamide gels, can be used to achieve alignment if they are stretched or compressed along the field axis. When a new method of alignment is introduced, it is typically tried out on a small, abundant protein with good relaxation characteristics, most often the regulatory protein ubiquitin. The upshot of this is that there is a fantastic amount of RDC data on this protein, and Griesinger's group uses this data to model the motions of its methyl-bearing side-chains.

This may sound strange, because RDCs are typically employed for structure determination. The angle defined by the measured coupling results from the overall tumbling bias of the protein (an alignment tensor) that is the same for each coupled pair, and their angle within that frame of reference, which can be used to uniquely define a structure. However, the dipolar coupling cannot be measured in an instantaneous fashion. It must evolve over time, just like chemical shift or a J coupling. As such, the dipolar coupling reflects an averaged orientation over the evolution period. In principle, the degree of averaging can be modeled as some kind of order parameter, similar to the S² of the Lipari-Szabo system, reflecting all these motions. This should encompass not only the fast dynamics that determine dipolar relaxation rates, but also motions slower than the global correlation time, up to near the ms range.

In order to derive dynamics data from their set of experiments in 13 different alignment media, the authors first scaled the RDCs from C-H methyl bonds. The reason they did this is that the three hydrogens in a methyl group rotate constantly around an axis passing through the adjacent C-C bond (cyan in the isoleucine side chain depicted at right). Because this rotation is typically very fast, simple to model mathematically, and pretty uninteresting, it can be deconvoluted from the dynamics data to give us what we're really interested in, the behavior of the C-C bond. Using an alignment tensor derived from a separate dataset of N-H RDCs in 36 different alignment media, the calculated C-C RDCs are combined into a matrix, from which simplified parameters describing motion can be derived. Fig. 1 depicts this schematically (note, the legend has the variables m and i reversed in meaning).

The order parameter (S²_rdc) reflects the rigidity of the bond and ranges from 0 (highly flexible) to 1 (perfectly rigid). The values measured for the methyl groups of ubiquitin cover almost this entire range (Fig. 2a), which is typical of side chains which tend to have less constrained motions than the backbone. It's also evident from this figure that S²_rdc is roughly anticorrelated with the number of dihedral angles between a given methyl group and the peptide backbone. The anisotropy of motion (η_rdc) is generally low, and appears to be roughly correlated with the number of intervening dihedrals. Both of these observations agree with previous data from other dynamics experiments, as well as reasonable expectations about the movements of these groups.

Farès et al. compare their S²_rdc to order parameters determined using other approaches. As one would expect, for most methyls the S²_rdc, which encompasses motions from a wide array of timescales, is lower than the S² determined using a Lipari-Szabo model-free approach that is only sensitive to ps-ns motions. The most notable exceptions are three residues for which the RDC method fit order parameters greater than 1. Probably these represent some unanticipated failure of the model, although these violations occur at groups that are expected to be rigid (two are alanines) and which have high S² from the Lipari-Szabo model. One potential culprit is the previously discussed scaling, which may need to be adjusted if the axial rotation is unusually slow. The best agreement with existing data comes from an approach that combined J-coupling data with less-specific RDC information, but only when those data are corrected for very fast motions using backbone order parameters.

Because the RDC fundamentally contains data about the orientation of a given bond, it should be possible, with a minimal degree of modeling, to extract specific information about the kinds of motions being made, data that cannot be obtained from methods such as the Lipari-Szabo model-free interpretation. The authors modeled side-chain motions around the dihedral angles χ₁ and χ₂ using the MFA data. The agreement between these results and the existing ensembles is not particularly good, but it's not completely obvious where the fault for this lies. Similarly there was limited agreement between these order parameters and those derived from existing ensembles. Agreement with the EROS ensemble structure determined last year was good, but it is difficult to judge what this means, as that ensemble was calculated using the same or similar data to what was used in this paper.

It's fair to ask whether this sort of data analysis will be possible in systems with less comprehensive data. Even this extremely rich dataset proved problematic, for instance in the cases of the alanines and the disagreement concerning rotameric states. It remains to be seen whether this approach will be practical in cases where significantly fewer alignment media can be used. However, it is also true that methyl groups have the best relaxation properties of any group in a protein, and that the experiments for the determination of RDCs are simple and sensitive. This means that if this approach can be made to work, it can provide important dynamic data even in the largest proteins studied by NMR. Farès et al. have performed an impressively comprehensive analysis of dynamics in a time regime that NMR has had difficulty accessing. Hopefully they will next bend their considerable talents towards reproducing as much of this analysis as possible in a more difficult system with sparser data.

Farès, C., Lakomek, N., Walter, K., Frank, B., Meiler, J., Becker, S., & Griesinger, C. (2009). Accessing ns–μs side chain dynamics in ubiquitin with methyl RDCs Journal of Biomolecular NMR, 45 (1-2), 23-44 DOI: 10.1007/s10858-009-9354-7 OPEN ACCESS

Read the rest...

A step towards incorporating dynamics data into drug design

My field-cycling article (previous post) is part of a dynamics-focused topical issue of JBNMR. In my next few science posts I'll describe some of the other contributions.

Research into the interplay between protein structural dynamics and function is a window into important fundamental knowledge about biochemistry, but the general justification for public funding of these studies by medical agencies is that they will have the ultimate effect of improving our ability to design and optimize drugs. However, even though our ability to characterize macromolecular dynamics has increased dramatically in the past few decades, there are few, if any, cases in which this knowledge has been applied successfully to the design of therapeutic agents. In part this is because incorporating data on fluctuations into the design algorithms poses a significant challenge. It's also true, though, that we understand only part of each system, i.e. the dynamics of the protein target, not the small molecules it binds. If dynamics studies are to make the maximum possible contribution to pharmaceutical sciences, the motions of the ligand must be characterized. In their article in Journal of Biomolecular NMR, Jeffrey Peng and students from Notre Dame attempt to address this shortcoming in the case of a substrate for the phosphorylation-directed prolyl isomerase Pin1.

Pin1 is implicated in a number of regulatory and signaling pathways, which seems strange because it doesn't possess any intrinsic transcriptional regulation ability, nor does it covalently add or remove phosphate groups. Instead, Pin1 has an enzymatic activity that accelerates, generally without altering the relative populations, the conversion of prolines from their cis- to trans- state and vice versa. This activity is specifically targeted to prolines that are adjacent to phosphorylated serines or threonines. In addition to the catalytic domain that does this work, Pin1 possesses a WW domain that has identical specificity. Because Pin1 does not alter the balance between cis- and trans- Pro, only the rate at which one changes to the other, its role in signaling has been difficult to ascertain, although there is intense interest in this area.

You don't need to understand an entire pathway to design an effective inhibitor. What you do need to understand is the relationship between specific chemical groups and binding affinity. Getting that knowledge can be very difficult if the proposed drug is flexible. In that case, refinement methods that focus only on the particular chemical groups rather than their dynamic properties could go badly astray. Unfortunately, the dynamics of protein-bound drug molecules are difficult to measure. Their proton signals are likely to be swamped by the protein, and small molecules are often difficult to label with isotopes convenient for NMR. Peng et al. propose to address this by studying ¹³C relaxation at natural abundance.

A little less than 99% of the world's carbon is in the form of NMR-inactive ¹²C, which is a problem for NMR because carbon is very abundant in proteins and drugs. Of the rest, most (about 1% of all carbon) is dipolar, NMR-detectable ¹³C, which is usually not enough to accomplish anything in terms of protein NMR. As a result, NMR researchers typically adopt the strategy of expressing their proteins using bacteria grown in media containing ¹³C₆ D-glucose. Such enrichment of drug molecules probably could not be carried out for pharmaceutical research due to the cost and the limited availability of properly labeled reagents. Fortunately, advances such as magnets stronger than 17 T and cryoprobes make sensitive detection of natural-abundance ¹³C a plausible approach. Because natural-abundance measurements also simplify the experiments and analysis considerably, Peng et al. adopt this approach in their study.

Peng et al. measure μs-ms fluctuations in a 10-residue peptide in the presence and absence of Pin1. Keeping in mind that such motions can only be detected when they are associated with a change of chemical shift, it is reasonable that no such motions are detected when the peptide is all by itself. In the presence of Pin1, however, methyl groups on phospho-Thr 5 and Val 7 experience some kind of chemical exchange process on the order of several 100 /s (at 278 K), as does a methylene group in Pro 6.

Peng et al. rationalize their observations with reference to a previously-determined structure of the Pin1 WW domain in complex with this peptide (explore it at the PDB). As you can see from the lowest-energy member of this NMR ensemble (left), the residues where they detect these fluctuations in the methyls and methylenes are those that are most involved in the binding interaction. The WW domain is represented as blue ribbons, while the peptide is shown as sticks down at the bottom. That the Pro and pThr form part of the interface is unsurprising, as they constitute the specific binding sequence, while the Val side chain appears to be in position to make some hydrophobic contacts. Everything makes sense, but that doesn't mean it's telling us what we want to know.

The structure above shows us the interaction of the peptide with the WW domain, while what we're really interested in getting at is the catalytic domain. Using an exchange spectroscopy experiment, Peng et al. determined that the ms dynamics they were observing probably reflected the binding of the peptide to the WW domain. To avoid this interaction, they created an artificial Pin1 that contained only the catalytic domain, and found that this also caused chemical exchange in the methyls and methylenes. Cross-checking against the exchange spectroscopy rates suggested that the ms dynamics in this case reflect the result of Pin1 catalytic activity, namely the interconversion from cis to trans and vice versa.

Unfortunately, this experiment did not report the most desired data, i.e. the dynamics of the ligand on the enzyme. On-enzyme fluctuations certainly contribute to the exchange experienced by the ligand, but because the on-enzyme population is so small (at most 2.5% of ligand) this would only be detectable in the case of an extremely large change in chemical shift. In principle one could deconvolute the dynamics from a partial-occupancy system where 50% or more of the ligand is bound to enzyme, but reliably fitting all the parameters for a two-state chemical exchange system from CPMG data is an already-dicey proposition. Fitting a four-state process from data like these is unlikely to be practical. So, in order to observe on-enzyme dynamics the drug of interest will need be saturated with its target protein, which would require millimolar protein concentrations for most ligands. Under those conditions, the spectra will also contain significant signal from the protein. The 70% deuteration used in this experiment, combined with ¹³C depletion, will probably be enough to suppress this, although these isotopes will increase the cost of the technique (and diminish protein yields).

Nevertheless, this paper establishes that the natural-abundance approach to measuring ligand dynamics on the µs-ms timescale is feasible. Because methylenes and methyls are common moieties in drugs and small molecules this technique may have broad applicability. Investigating the motions of small molecules bound to large proteins poses a unique problem because these systems don't have the advantages of either small molecules (low R₂) or proteins (exotic labeling schemes). The ongoing work of Peng et al. suggests that this problem is tractable, which may have positive consequences for our ability to design and optimize drugs.

Peng, J., Wilson, B., & Namanja, A. (2009). Mapping the dynamics of ligand reorganization via ¹³CH₃ and ¹³CH₂ relaxation dispersion at natural abundance Journal of Biomolecular NMR DOI: 10.1007/s10858-009-9349-4

Read the rest...

Mesodynamics, field cycling, and SARS: an explanation

Part of the motivation for my previous post about the spectral density was the recent appearance online (and upcoming appearance in print) of my paper in the Journal of Biomolecular NMR, which is open access, so you can open it up from home and read along as I tell you about it. The obscure-sounding title "Mesodynamics in the SARS nucleocapsid measured by NMR field cycling" means that we were able to characterize an interesting fluctuation in a protein from the SARS coronavirus, and that we used a cool technique to do it.

In the previous post, I mentioned that NMR dynamics studies ought to use data collected at multiple static magnetic field strengths. This is typically accomplished by increasing the strength, because of clear advantages in sensitivity and resolution at high and ultra-high field. Corresponding author Alfred Redfield, however, created a device (left) to capture information about relaxation at lower magnetic field strengths while retaining the advantages of, say, a 500 MHz magnet. This is accomplished by field-cycling, which in this case means physically moving the sample from the center of the magnet's superconducting coil to a spot several centimeters away. If one has carefully measured the magnetic field gradient with respect to distance, one can reproducibly measure relaxation at a desired (lower) field within the bore of the 500 MHz magnet.

As you might surmise from the photograph, Al built the field-cycling device himself, often jury-rigged from whatever parts were convenient. For instance, as you can see at right, the push-rod that connects the sample in the tube to the motor assembly was made from an arrow purchased at a sporting goods store. I'll count myself lucky if I'm half as creative and active in my 50s as Al is in his 70s. Al has used this device to investigate the dynamics of nucleic acids and lipids, but he was interested to see what we could learn about proteins by examining relaxation at low magnetic fields. Relatively low, at any rate — the weakest magnetic field I used is higher than you would typically encounter in, say, a clinical MRI. In his ³¹P research, however, Al has gone near zero field during the relaxation period.

Elan Eisenmesser, now a professor at the University of Colorado Health Sciences Center, did some initial investigations using this technique on cyclophilin A, and edited the pulse sequences so they could control the field-cycling device. Unfortunately, the results in CypA were kind of boring because for that protein the dynamics on the ps-ns timescale are relatively homogeneous. At this time, Elan was also working on the N-terminal domain of the SARS nucleocapsid protein (henceforth SARSN). As you can see from the structure at left (explore it at the PDB), SARSN has a long β-hairpin (sticking out to the right) which is known to be flexible. The hairpin is thought to interact with RNA as part of the viral assembly process, as well as binding to several host proteins during the process of infection. As Elan prepared to move on, he passed the project to me, and with Wladimir Labeikovsky assisting for the first couple of months, I took a bunch of spectra under various conditions.

You can see what a low-field spectrum looks like at right: this is an HSQC from an R₁ experiment where excitation and acquisition were performed at 50.7 MHz (¹⁵N) and the relaxation period took place at ~17 MHz (blue). I've also overlayed a spectrum collected entirely at 50.7 MHz (red). The peaks are all in the same place and the sensitivity is good, but the signal/noise ratio is clearly lower for the 17 MHz spectrum, and we get some sidebands from the water on the right side of the spectrum. Getting the water signal to behave was a significant challenge for these experiments and took several tries to get right.

Besides the experiments I performed personally, spectra were collected by Elan and Geoffrey Armstrong at the Rocky Mountain magnet facility (the 900 MHz R₁ and NOE) and Karl Koshlap at the UNC Pharmacy School (500 MHz data). Karl's involvement was necessitated by a change in sample conditions and the unfortunate incident our spectrometer had with an HDTV channel (chronicled here, here, and here).

In the end I managed to gather relaxation data from four high fields (using standard equipment) and two low fields (using the field cycler). The R₁ data are shown in Figure 2, and if you read the previous post then they shouldn't surprise you very much. For most of the protein, R₁ decreases steeply as the strength of the static magnetic field is increased, but for a subset of amides this field dependence is substantially reduced. Most of these residues fall into a continuous stretch encompassing the β-hairpin of SARSN and an adjacent loop (shown on the structure in Figure 1). In addition, the heteronuclear NOE measurements for these residues show a very large R_NOE/R₁ ratio at 50 MHz that decreases substantially as the field increases (Figure 3). As I discussed in the last post, these patterns of field dependence are characteristic of flexible regions in a protein, but more specifically they indicate flexibility with an internal correlation time of around a nanosecond or so.

One might expect a large, relatively unconstrained feature like the hairpin to have flexibility on multiple timescales. In particular, it seems like the sort of structural element that might move with a time constant of microseconds or milliseconds. These slower motions can't be fit with great accuracy using the experiments performed here, but evidence of their absence can be found in the R₂ experiments performed at 500 and 600 MHz (Fig. 4). Assuming that they are correlated with changes in chemical shift, we would expect motions on this timescale to increase the R₂, but in the hairpin this relaxation rate is substantially reduced, consistent with high flexibility on the nanosecond timescale (low S²).

In order to gain a more complete picture of the dynamics, I fit the relaxation data to model-free formulations of the spectral density. For most residues, the classic Lipari-Szabo formalism worked quite well, although the S² are generally higher than I like. An analysis of the fits, however, indicated that many residues needed to be fit to a more complex model, called extended model-free or model 5. In this model the spectral density is given as:

where S²_f and S²_s are order parameters for a fast and slow internal motion, respectively, and τ_s is the internal correlation time for the slow motion (τ_f is assumed to be ~0). The residues that were fit to this alternative model happened to be those with anomalous R₁ and NOE dispersions, meaning they mostly belonged to the β-hairpin and the loop incorporating residues 60-65.

Ultimately I didn't include the low-field data in the quantitative fits. The large random errors in these rates (error bars in Fig. 2) meant that the more precise high-field data would dominate the fits, for one thing. For another, the low-field data were not entirely consistent with the high-field results. Although the general features of the relaxation at 17 and 30 MHz agree with predictions from high field, the observed low-field R₁ differ substantially from predictions. This could be due to a number of error sources, the two biggest being positioning error and interference between the CSA and dipolar relaxation mechanisms (because we cannot suppress this interference in the fringe field). Al also thinks some of the error may be due to the influence of a low-amplitude fluctuation in the globular portion of SARSN. Qualitatively the R₁ behave much as we would expect, but bringing them into line quantitatively will take more work.

The upshot of all of this effort to fit the dynamics is that the residues in the hairpin have an interesting duality. On very short timescales (< 10 ps or so) they are quite rigid, much like the rest of the protein. On a slightly longer timescale, however, they are very flexible, with S²_s of around 0.6, and similar internal correlation times across the entire feature in the range of 600-800 ps (Fig. 5). Because the correlation time of this fluctuation is significantly faster than molecular tumbling but much slower than typical backbone fluctuations, Al called them "mesodynamic", a word Dorothee seems to like. At any rate, these observations led us to propose that the hairpin fluctuates widely (based on S²_s and τ_s) as a coherent structural unit (based on S²_f), rather than having its strands fall apart and flop around randomly. The hairpin is both ordered and disordered, depending on the timescale of analysis and frame of reference.

The physical plausibility of this dynamic model was assessed using a pair of 15 ns all-atom molecular dynamics simulations performed by Ming Lei. What these found, shown in Fig. 7, was that the hairpin maintained its internal structure while moving freely with respect to the globular portion of the protein. In addition, the simulations suggested a reason the 60-65 loop had similar dynamics to the hairpin — transient hydrogen bonds formed between side chains in the hairpin and residues in the loop, causing their motions to be correlated.

The qualitative agreement between the low-field and high-field data supports our contention that this technique can be made to work and to give valuable data about certain kinds of fluctuations. Future work on proteins with this technique will require a rigorous approach to control for the systematic bias we observed. Additionally, this study re-emphasizes the value of taking relaxation data at many fields in order to fully characterize biomolecular dynamics.

As for the dynamics of SARSN, the finding is interesting but doesn't yet provide any specific insight. Disordered regions of a protein are often associated with promiscuous binding activity, and this hairpin is no exception. However, the existence of multiple binding sites in one of these regions is usually attributed to a significant ability to restructure itself. Here, that possibility would seem to be limited by the apparent persistence of the hairpin's intrinsic structure. The ability of the hairpin to move freely while maintaining a particular internal arrangement may have advantages in capsid construction, an idea that could potentially be tested by inserting prolines or glycines in the β-strands, which should disrupt the hydrogen bonding that preserves the hairpin.

Al and his collaborator Mary Roberts are currently continuing their investigations of ³¹P dynamics in nucleic acids and lipids using low field. They're even advertising:


Clarkson, M., Lei, M., Eisenmesser, E., Labeikovsky, W., Redfield, A., & Kern, D. (2009). Mesodynamics in the SARS nucleocapsid measured by NMR field cycling Journal of Biomolecular NMR DOI: 10.1007/s10858-009-9347-6 OPEN ACCESS

Read the rest...

How do adamantane drugs block M2?

Vaccination plays such an important role in our seasonal influenza strategy in part because we don't have many medicines that can be brought to bear on the disease. The neuraminidase inhibitors (specifically Tamiflu) are widely stockpiled, and continue to work for now, but the specter of resistance is already lurking. If these drugs are too widely or too improperly used, there is a good chance that resistance mutations will eventually render these drugs ineffective. Universal drug resistance may already be the fate of the drugs amantadine and rimantadine, built on an adamantane backbone (1). The adamantane drugs inhibit the M2 proton channel from influenza A, a tiny tetrameric protein that equalizes pH between the virus and the endosome of the cell that has swallowed it. This process releases the virus contents so that they can do their damage to the cell, so these medicines can significantly retard the infection process. Or rather, they could, if so many influenza strains didn't harbor the S31N mutation that almost completely nullifies their effect. If we are to develop new drugs to attack the M2 channel, it would be helpful to know how this mutation causes drug resistance. Over the past few years a great deal of structural evidence has accumulated showing how adamantane drugs work on the older, non-resistant channels. The problem is that the evidence supports two different models of M2 inhibition, and so far it has proven difficult to determine which of them is probably correct.

How the question arose

The controversy is the result of two structures published in Nature early in 2008 (2,3). The first of these is a crystal structure of a tetramer of peptides encompassing the transmembrane (TM) region of the M2 channel reported by the DeGrado group at UPenn, which you can see at right (explore this structure at the PDB, noting that the numbering is off by 21). In the detergent used for crystallization, the peptides form a tetramer with a roughly conical pore, which amantadine (purple in these models) physically occludes, giving rise to the pore-blocking model (PBM). This model is consistent with previous results indicating that a single amantadine molecule is sufficient to inhibit the proton channel. In addition, in this model the drug binding site is adjacent to S31 (blue side chain), which is what we'd expect given that an S31N mutation is responsible for most amantadine resistance. The authors propose, given the position of the S31 side chain, that the mutant asparagines form a hydrogen-bonded network that is too constricted for amantadine to bind. Click on the picture for a larger view.

An alternative model was proposed by Schnell and Chou from Harvard University (3). They produced an NMR structure (left) of a 42 amino-acid peptide from M2 encompassing the TM region and an additional C-terminal helix (explore this structure at the PDB). In their structure, taken at pH 7.5 in detergent micelles, the tetramer forms a roughly cylindrical pore that is blocked by the side chains of the known gating residues W41 and H37 (light green in these models). Their structure shows rimantadine bound at four sites near the base of the helix but not in the pore. Using pH-dependent conformational exchange experiments, Schnell and Chou showed that a decrease in pH caused rapid structural changes in the channel, motions that rimantadine slowed. On the basis of this evidence, they proposed a mechanism in which protonation of the gating histidines destabilizes the packing of the TM helices and allows the conductance of protons. Rimantadine blocks the channel by stabilizing the helices, thus this is a dynamic quenching model (DQM). The position of S31 in this model is also somewhat different than the crystal structure, although these models were made at different pH conditions and so this may represent a difference between the closed and open states of the channel.

The distinction here is important. If Stouffer et al. are correct, then drug development should abandon the adamantane backbone altogether and start with a set of significantly different leads to address the resistance problem. The PBM implies that any molecule large enough to occlude the pore will be too large to fit in there following the S31N mutation that induces amantadine resistance. If the DQM is correct, however, then it is conceivable that further refinements to the adamantane base, or similar molecules, could improve affinity enough to overwhelm the mutational effect.

Unfortunately, neither result is unimpeachable. Although it agrees with a great deal of experimental evidence, the low resolution of the crystal structure means that the electron density called amantadine cannot be assigned unambiguously. It is also curious that a hydrophobic molecule like amantadine would bind tightly in the hydrophilic pore. In addition, the crystal form with amantadine bound contains a mutation, G34A (black side chain), which is near the drug binding site and could conceivably have altered the binding specificity of the protein.

The NMR structure has the advantage that it directly includes distance information in the form of NOEs. However, the authors used 40 mM rimantadine to obtain these results, meaning that there were as many rimantadine molecules in the solution as phosphate buffer molecules. Under these conditions, it is possible that the drug bound to a secondary, low-affinity site. Even if this is what happened, it is strange that the rimantadine never bound to the high-affinity site indicated by the crystal structure.

Both experiments use significantly truncated constructs and highly artificial systems to mimic a membrane environment. The structure of any membrane protein depends in often unexpected ways on the composition of the lipid bilayer in which it is embedded and on the structure of that bilayer. The intense curvature of the micelles may have distorted the structure in the NMR experiment, and possibly inappropriate lipids may have had effects on both structures. We know these considerations are relevant for this system, because Schnell and Chou report that the construct used for the crystal structure would not form stable tetramers in the micelle system. Also, as Chris Miller notes in his commentary on these papers (4), there were questions about both constructs with respect to their proton conductivity. Lacking significant stretches of the protein and placed in these environments, it is possible that both structures deviate from in vivo reality in significant ways.

Because the conditions diverge so much, it is difficult to weigh the mechanisms based on these structures alone. The binding site identified by Schnell and Chou is only at the very end of the construct used by Stouffer et al.. In addition, the inhibited crystal structure comes from a low-pH condition while the NMR structure exclusively represents a high-pH condition. Given these differences in conditions, it is not impossible that both models, in whole or in part, are correct. We must turn to additional experiments and alternative evidence to choose between them, specifically data on the stoichiometry of binding and the effects of mutations.

Binding stoichiometry

The crystal structure shows a single binding site for the drug, while the NMR structure implies four, and this is at odds with existing results that indicate that a single molecule of drug is sufficient to inhibit a single channel. Given the homotetrameric nature of the M2 channel, it is in principle not possible for the NMR experiment to distinguish between a single rimantadine binding event and four. That is, the NMR experiment cannot tell us whether the rimantadine-M2 inhibition occurs with a single binding event or requires four drug molecules to bind. Therefore, to argue that DQM is inconsistent with 1:1 stoichiometry overstates the case somewhat.

It may also be somewhat overstating the case to say that there is only one amantadine binding site on M2. Washing amantadine out of your buffer does not reverse inhibition, in part because of slow kinetics of leaving the binding site and in part because these drugs, being very greasy, preferentially partition into the lipid membranes and are therefore not readily removed from a system when its aqueous phase is replaced. It is difficult to measure a binding constant for the drugs because the equilibria under consideration will be quite complex. The studies often cited on the 1:1 stoichiometry (5,6) use structural and kinetic evidence to get at this question.

Czabotar et al. (5) measured tryptophan fluorescence in M2 as a function of pH and rimantadine concentration. They found that fluorescence from W41 was quenched by decreased pH, but recovered when 1 equivalent rimantadine per tetramer was added. This result implies that structural or dynamic changes caused by histidine protonation are reversed by rimantadine inhibition, but this is so general that it cannot be taken to support either the PBM or DQM.

Wang et al. (6) measured the reduction of surface currents in X. laevis oocytes after addition of various concentrations of amantadine. From these results they are able to construct a Hill plot with a coefficient of 1, showing that binding of amantadine is not cooperative. In further results, Wang et al. find that amantadine inhibits M2 channels slightly better at high pH (when the pore is closed) than at low pH, and that amantadine inhibits proton conductance in either direction (rather than favoring one). Both these outcomes are unexpected for PBM, but can be easily explained by DQM. However, the differences in the binding constants are relatively minor and the linearity of the current-voltage relationship may result from some other idiosyncratic feature of the M2 channel, so these results are not unequivocal.

Neither experiment refutes DQM because they do not measure the number of binding sites, but rather the number of efficacious binding sites. If there are four binding sites, but 95% or more of the inhibitory or structural effect is caused by the first drug molecule bound, then these experiments would be unable to distinguish DQM from PBM. Overall, the evidence on the question of binding stoichiometry does not eliminate the possibility of four binding sites existing, but it does place a requirement on DQM that the inhibitory effect of amantadine on the tetramer result from a single binding event. Because the proposed DQM binding site for rimantidine lies between monomers and is linked to the gating tryptophan, this is not unbelievable. Other evidence from these experiments is equivocal, but can be seen as somewhat more problematic for PBM than DQM.

Functional effects of mutations

A serious problem for DQM is that the mutations known to give M2 resistance to adamantane drugs are all located near the PBM binding site. In particular, S31 is adjacent to the drug in the crystal structure and quite distant in the NMR structure. As Miller notes in his commentary, mutational studies are substantially more difficult to interpret than is typically suggested, so this isn't absolutely probative. In general, however, one predicts mutations to have short-range rather than long-range effects, so at least some resistance mutations ought to evolve at the binding site. However, many of the residues surrounding the DQM site are almost absolutely conserved, presumably because they are essential to the function of the channel. As a result, it would be very difficult to interpret any studies on point mutants in this area. What would be ideal, however, would be to find a set of mutations that produced a functional protein and abrogated amantadine inhibition.

This is the basis for an interesting experiment conducted by the lab of Robert Lamb and reported last year in PNAS (7). In this case, the authors took advantage of the fact that the M2 protein from influenza B virus is not sensitive to adamantane drugs. They constructed a chimeric protein containing about a dozen residues from influenza A M2 — specifically, the dozen or so residues surrounding the PBM site. If PBM is correct, then we would expect that these residues, which define that site, would impart amantadine susceptibility to the influenza B channel. This is what happens, sort of. For your benefit, I have shamelessly stolen their figure (right), but you can check out this paper yourself because it is open access. In this assay, again involving X. laevis oocytes, the hybrid channel is sensitive to amantadine (bottom trace), but only half as sensitive as the wild-type influenza A channel (second from top). This result suggests that there is important context conferring susceptibility outside the PBM site. However, this could be something as simple as helix orientation, so the result does not necessarily imply that there is an external binding site.

Additionally, the authors made point mutations at residues (L38, D44, and R45) that were presumed to be important in the DQM mechanism or have long-range effects on amantadine binding. None of these mutations appeared to affect amantadine resistance. In contrast, experiments in liposomes reported by the Chou group this May showed that a D44A mutation prevented rimantidine from having an effect (8). This conflict in results is difficult to reconcile, but may result from the different constructs used (the Chou group used a truncated form of M2 while the Lamb group used the full-length protein) or from changes in ion specificity caused by the D44A mutation. It might be of value to repeat these experiments with the alternative construct: truncated in oocytes, full-length in liposomes. Because the D44A mutant does not appear to conduct protons as efficiently as WT, the proposition that the function of this mutant is too deranged to provide trustworthy information should also be considered.

Additional experiments in the Chou paper are meant to address the relationship between the DQM site and the mutations at the PBM site. They show that the S31N mutation prevents rimantidine binding to the remote site, and also that this mutation makes the protein generally more dynamic. From this evidence they propose that this mutation, at least, disrupts amantadine binding by destabilizing the helical packing of the channel and thus interfering with the organization of the lipid-facing pocket.

They also examine an S31A mutation and find that it is not rimantadine-resistant or destabilizing to the packing. This supports their dynamic model in a limited way, because it demonstrates that only certain mutations at the S31 site will generate resistance. It does not cast any doubt on PBM, however, because in that model resistance in the S31N mutant is explained by the idea that its side chain will partially obstruct the pore so that rimantadine will not fit. I do not think it was ever proposed that specific contacts between S31 and the drug stabilize the binding; in fact, the general absence of such contacts strikes me as a concern about PBM.

Chou et al. also examine the effect of rimantadine on the shorter construct used for the crystal studies. They find that the inhibition of this construct is substantially weaker. However, it also conducts protons at a much slower rate in this assay, suggesting that there may be additional serious problems with the function of this construct. It may be that it simply is not appropriate to use this construct for studies in solution or living membranes. That doesn't necessarily imply that this peptide will give incorrect information in the stabilizing environment of a crystal.

What do we know, and what do we need?

Very little of this evidence unequivocally prefers one model to the other. We know that a single adamantane molecule is sufficient to inhibit M2, and while this is most obviously compatible with PBM it need not be inconsistent with DQM. It is also apparent that various constructs of the M2 channel retain adamantane susceptibility after ablation of the DQM site, either by truncation, mutation, or the construction of a chimeric protein. In all assays, however, the adamantane drugs lose a considerable amount of inhibitory power, so these results are not entirely consistent with PBM either. And, at least in the Chou lab's assays, interference with the DQM site also reduces adamantane susceptibility and deranges function. Moreover, the NMR data from the Chou lab shows that mutations at PBM site have a long range effect on the DQM site, which mitigates the probative power of the S31N mutation.

How do we address this question? One important step would be to start comparing like to like. We are considering evidence from a plethora of constructs and conditions, and the evidence in conflict is often collected in very divergent experiments. Ideally we would like to have structures of the wild-type channel at low and high pH in a lipid environment that closely mimics the composition and curvature of a mature influenza virion. As this is unlikely in the near term, we must hope for NMR and crystal structures that at least use the same construct, minimally mutated, under similar conditions. NMR studies at low pH would be of value in assessing whether these studies in fact contradict one another. Additionally, it would be useful to make adamantane derivatives labeled with a free radical or other paramagnetic label; this would presumably allow the identification of a binding site at lower drug concentrations in an NMR experiment. Labeling the drug with a metal might also allow its identification in a crystal structure without any need to push the resolution significantly higher. Finally, actual structures of the S31N mutant, positively identifying the disposition of this side chain, would be of great value in judging the question.

Structural experiments can take a great deal of time and careful tuning, a requirement exacerbated by the often-fickle behavior of membrane proteins. As such, additional mutational studies could prove useful. Inverting the chimera experiment of Jing et al. to create a chimeric protein with the upper channel from influenza B and the lower channel from influenza A may be a helpful supplement to the existing experiments. If the C-terminal portion of the channel makes a contribution to adamantane inhibition this chimera will also be rimantidine-sensitive. In addition, new mutations at S31 could help distinguish the possibilities. The PBM supposes that the N31 side chains stick into the pore and form hydrogen bonds, while the DQM supposes that they stick into the interface of the TM helices and destabilize them. An S31L mutant should disrupt the helical packing but not form hydrogen bonds or extend the L31 side chains into the pore. If functional, such a mutant ought to be rimantidine resistant if the DQM is correct, but not if PBM is correct. Assuming the geometry of the longer side chain is wrong for formation of a hydrogen bonding network, an S31Q mutation might be useful as well. Similarly, mutations that increase the size of the L40, I42, or L43 side chains could prevent adamantane binding to the DQM site without degrading the channel's transport capabilities; the drug sensitivity of such a mutant would be a powerful argument either way. Any experiments of this kind would likely be easier to perform than to interpret, but could provide valuable insight. Obviously, it would also be important to establish that each mutant was competent at transporting protons.

The unspoken assumption of the debate so far is that these mechanisms are mutually exclusive, but there is no particular reason to believe this must be so. The structural experiments definitively show that binding to both sites is at least possible — even if one clings tenaciously to the idea that the density observed in the crystal is not in fact amantadine, that structure at least shows that the PBM site is capable of accommodating the drug. It might therefore be plausible that adamantane drugs inhibit M2 using both mechanisms simultaneously, or that DQM predominates at high pH and PBM at low pH. Redundancy in inhibitory mechanisms may explain the curious features of amantadine inhibition noted by Wang et al., and the inability of experiments specific to a single site to completely account for adamantane inhibition. In addition, the fact that S31N interferes with both mechanisms may explain why it is the primary resistance mutation.

An experiment with the alternate chimera mentioned above could test this possibility. In addition, if the mechanisms switch off in a pH-dependent fashion, then this should be testable with the hybrids: specifically, the A/B M2 used by Jing et al. should have lower susceptibility to adamantane drugs at high pH than at low pH. Similarly, the B/A M2 chimeric protein, if inhibited by amantadine, would be more resistant at low pH.

Doubtless these suggestions are nothing new to the members of the labs working on this perhaps unexpectedly hairy question. Membrane protein structure and function is one of the most difficult experimental subjects in biochemistry, and constitutes a critically important frontier in scientific efforts to improve human health. It is infinitely easier to propose most of these experiments than it is to perform them, and I would be remiss if I did not temper the persistently critical tone of this post with some praise for the efforts of all the scientists involved in this research, and for their commitment to getting the right answer. These papers represent years of work by incredibly talented people using some of mankind's most advanced scientific techniques. The lack of clarity on the question of adamantane drugs binding to M2, even in the face of this amazing effort, is a testament to the enormous difficulty of researching these critical systems.

---

1. Deyde, V., Xu, X., Bright, R., Shaw, M., Smith, C., Zhang, Y., Shu, Y., Gubareva, L., Cox, N., & Klimov, A. (2007). Surveillance of Resistance to Adamantanes among Influenza A(H3N2) and A(H1N1) Viruses Isolated Worldwide The Journal of Infectious Diseases, 196 (2), 249-257 DOI: 10.1086/518936 OPEN ACCESS

2. Stouffer, A., Acharya, R., Salom, D., Levine, A., Di Costanzo, L., Soto, C., Tereshko, V., Nanda, V., Stayrook, S., & DeGrado, W. (2008). Structural basis for the function and inhibition of an influenza virus proton channel Nature, 451 (7178), 596-599 DOI: 10.1038/nature06528

3. Schnell, J., & Chou, J. (2008). Structure and mechanism of the M2 proton channel of influenza A virus Nature, 451 (7178), 591-595 DOI: 10.1038/nature06531

4. Miller, C. (2008). Ion channels: Coughing up flu's proton channels Nature, 451 (7178), 532-533 DOI: 10.1038/451532a

5. Czabotar, P., Martin, S.R., & Hay, A.J. (2004). Studies of structural changes in the M2 proton channel of influenza A virus by tryptophan fluorescence Virus Research, 99 (1), 57-61 DOI: 10.1016/j.virusres.2003.10.004

6. Wang, C., Takeuchi, K., Pinto, L.H., & Lamb, R. (1993) Ion Channel Activity of the Influenza A Virus M2 Protein: Characterization of the Amantadine Block J. Virol. 67 (9) 5585-5594 Available free from PubMed Central

7. Jing, X., Ma, C., Ohigashi, Y., Oliveira, F., Jardetzky, T., Pinto, L., & Lamb, R. (2008). Functional studies indicate amantadine binds to the pore of the influenza A virus M2 proton-selective ion channel Proceedings of the National Academy of Sciences of the United States of America, 105 (31), 10967-10972 DOI: 10.1073/pnas.0804958105 OPEN ACCESS

8. Pielak, R., Schnell, J., & Chou, J. (2009). Mechanism of drug inhibition and drug resistance of influenza A M2 channel Proceedings of the National Academy of Sciences of the United States of America, 106 (18), 7379-7384 DOI: 10.1073/pnas.0902548106

The Scientific Activist and Discovering Biology in a Digital World also have interesting posts on this subject.

Read the rest...