In a recent Science paper, Sarel Fleishman et al. report the de-novo computational design of a protein interface to specifically target and tightly bind a surface patch of the flu hemaglutinin protein. We interview Sarel to get the insights from behind the scenes and the outlook for this exciting approach.
Fleishman SJ, Whitehead TA, Ekiert DC, Dreyfus C, Corn JE, Strauch EM, Wilson IA, & Baker D (2011). Computational design of proteins targeting the conserved stem region of influenza hemagglutinin. Science (New York, N.Y.), 332 (6031), 816-21 PMID: 21566186
Fleishman et al. describe a general computational method for designing proteins that bind a surface patch of interest on a target macromolecule. They first identify good binding sites for ‘anchor’ residues on the target surface and utilize these to anchor de novo designed interfaces. The method was used to design proteins that bind a conserved surface patch on the stem of the influenza hemagglutinin (HA) from the 1918 H1N1 pandemic virus. Two of several tens of designs showed very nice binding, reaching (after affinity maturation) low nanomolar affinity. One of the designs was shown to inhibit the HA fusogenic conformational changes induced at low pH. The crystal structure of the designed protein in complex with HA revealed an actual binding interface nearly identical to that in the computational prediction. Such designed binding proteins may be useful in the future for both diagnostics and therapeutics.
To further understand what made this approach work, and what can we learn from it on the nature of protein-protein interactions, we interviewed Sarel, a senior post-doctoral (HFSP) fellow at David Baker’s group and first co-author of this work.
You mention in the paper that this is a general design strategy to target an arbitrarily selected protein surface. How easily can you apply it to another system? Is it restricted to certain surface topologies, or can you target also e.g. flat surfaces?
The method is general to the extent that any surface where a cluster of amino acid sidechains can form favorable interactions would be a plausible target. In unpublished work we have designs that bind surfaces that are much flatter and less hydrophobic than flu hemagglutinin’s. In fact, the depth of the hemagglutinin surface was a huge challenge for our approach as it restricted the type of protein that can approach the surface. Our success rate (2 out of 71) was accordingly quite low. In other projects we have somewhat higher success rates. The major push now is to extend the method and see what sort of target surfaces can be reasonably tackled and also what sort of scaffolds can be designed effectively. We have some exciting projects on other viral targets, such as Ebola and measles as well as a host of biologically interesting targets. I hope that within the next few years we’ll have this question mapped out more precisely.
Realistically, I don’t think that at this point we can target very polar surfaces as sometimes seen in natural protein complexes. Also, this method would clearly not work if extensive backbone contacts are to be made, as seen in beta-sheet extensions across the interface. Such targets would require more and likely different approaches.
You describe in the paper the methodology, but in your mind what was the key to making it work? Many others have tried before with less success.
This is a subtle point. In my mind the main aspect that made this method successful was the use of proxies to negative design. Negative design has been known to be important for binding (and folding) for a long time. The basic premise is that in order to bind effectively a protein need not only optimize its sequence and conformation for binding the particular target; it must also preclude binding of the myriad other off-target molecules. Obviously tackling this negative design problem head on is computationally impossible as there is a vast number of surfaces to be considered. What we found in previous work was that natural protein binders cluster residues that are crucial for binding in spatially small regions (known as interaction hotspots in the field) to effect such negative design. By clustering the residues together and promoting energetically favorable interactions among them protein binders ensure that the surface won’t reconfigure and bind to off-target molecules. Our design strategy in effect centers around this strategy. We start by designing a small cluster (2-3) of amino acid residues that form high-affinity interactions both with one another and the target surface and then identify protein scaffolds that can incorporate those residues without introducing strain. A third step further stabilizes the monomer by designing surrounding residues. Without this implicit negative design component, I’m fairly convinced that we would get non-specific binders that would interact with other surfaces on hemagglutinin or other hydrophobic proteins.
Indeed, in this case your designed protein struck the bulls-eye. Can you elaborate though on what ensures specificity? Some studies have shown that for protein-protein interactions just designing for affinity is not enough.
That’s absolutely right, and in my opinion this point is tightly related to the question of what made the strategy work. Specificity is ensured in our design strategy by the rigidity of the designed binding surface. Since the core residues are designed to form favorable interactions among themselves they don’t reconfigure, and that precludes alternative binding modes and targets. It’s important to keep in mind though that negative design and specificity is impossible to fully compute because of the vast number of potential binding surfaces available for binding in an experiment. This is why experimental validation is crucial for any design, and the gold standard is a molecular structure by x-ray crystallography. In the case of the hemagglutinin binder the molecular structure showed very accurate modeling. This is very encouraging, but we need structures of many more designs to probe this point. We also need structures of failed designs to understand what it is that we’re missing in our computations. The long-term goal is to have reliable computational methods that allow one to go from a target surface to a binder without experimental trial-and-error. There’s still quite a distance to go for that.
Speaking of failed designs, would you like to comment on the success ratio of the designs? Although you only need one, 2/80 is rather low. Also, how did the prize designs ranked in comparison to the other 78?
That’s a crucial point. Our success rate is very low at present. In my mind this point is also connected to the question of specificity and negative design. We recently carried out an analysis with the help of 28 research teams from around the world that are interested in computational prediction of protein binding. The question in this analysis was what discriminates the inactive designs from a set of natural protein complexes. One of the most intriguing points to have come out of the analysis was that many of the surfaces that we designed are predicted to be unstable, because the backbone is poorly connected to the remainder of the designed protein. By contrast, many of the designed surfaces that are true binders have very stable backbones. This is a fascinating point as it suggests that designed surfaces are very likely to form the modeled interaction, if (and this is the crucial point) the binding surfaces actually form in experiment. Our concepts of designability and binding are constantly evolving as we’re learning what works and what doesn’t in de novo design. The exciting aspect is that this is shedding new light on evolution, binding, and has many potential applications in helping us reliably generate binders and inhibitors.
Have you gained new insights on the nature of protein-protein interactions along the way? Or was everything already there and just needed to be assembled correctly?
Yes. As I started work on this project my main focus was on elements of positive design: generate a protein sequence and conformation that best binds to the target protein. The emphasis on excluding alternatives came by thinking about what it is that makes designs that came out of early design strategies inferior to natural protein binders. Even now, as we’re tackling different protein targets and retrospectively analyzing why some designs worked and others failed our ideas about what shapes natural protein binding sites are changing. I now think that the way that binding surfaces avoid off-target binding is a crucial force in the evolution protein-protein interactions. Think about it this way: how does evolution ensure that a given protein binder would not bind to the vast number of surfaces of proteins in the crowded cell environment. There has to be a general solution to this problem, or else, each time a surface mutation arises in evolution the entire cell-repertoire of proteins would need to reconfigure in an impossibly long cascade. We know that in certain cases, mutations that do not affect the function of a the protein but form alternative binding modes cause disease. The most famous case is of sickle-cell anemia, where a surface mutation causes hemoglobin to form huge oligomers that cripple red-blood cells. I find it extremely exciting that protein design can now be used to think about issues such as the evolutionary pressures on protein-interaction networks, the thermodynamic principles of binding, as well as disease.
What are the future plans for this design strategy? where are you going to take it too?
To best understand all the outstanding issues we need a lot more data on what works and what doesn’t. For that, we need to design proteins that bind to a variety of protein surfaces. One of the nice features of this project is that it is so multifaceted. By designing inhibitors of flu hemagglutinin, we made a molecule with potential uses as a therapeutic and diagnostic, and we also shed light on the biophysics of protein binding. Likewise, we’re now designing binders of other viral target proteins as well as biomedically interesting cellular targets. I’m expecting more insights to come from these projects.
However, the strategy is currently limited to working on sidechain-dominated interactions. One of the fascinating phenomena in protein binding is the rich world of backbone flexibility in binding. Consider how potent antibodies are as general binders. Most of this potency comes from the use of backbone flexibility. We’re currently very far away from being able to accurately capture the energetics of this phenomenon, but I’m excited to pursue it. My guess is that the same questions of how to incorporate negative design and specificity will come out as strongly as ever in this research.
Lastly, what would you say is currently the biggest challenge in computational interface design?
My impression is that we need different strategies to tackle the many different types of interactions observed in nature. We have only addressed the issue of sidechain-dominated interactions. How to design backbone-mediated interactions, as well as highly polar and charged interactions remains unanswered. Other issues are what is it that makes a protein bind to its target, which is intimately related to the question of our low success rate in de novo design. There are numerous questions that this research has raised. I consider this an especially exciting time for protein design, when questions of protein biophysics have such high stakes for our understanding of evolution, protein function, and our ability to treat and diagnose disease.