Here’s a very interesting computational paper that’s taking on one of the major challenges in the field: given a particular protein surface, can you predict de novo what other proteins might bind to it? That probably sounds like a significant challenge just in a combinatorial sense, and that it is. The approach taken here is to chop the problem down to a size that can actually be computed, and in such a way that you’re spending your resources sorting through the possibilities that are more likely to give you good answers.
Easy to say! But in practice, here’s the sort of thing this team did: they picked a particular protein surface, and then docked a huge number of disembodied amino acid structures against it, looking for good interactions. A “rotamer interaction field” method sped this up, allowing useful approximations of these interaction energies without computing everything down to the details (they’re just concerned with the protein backbone at this point rather than all the side chains). That’s a huge boost, because it reduces the algorithmic complexity – in “Big O” notation, you’re normally looking at a problem like this in O(N) or O(N2). That is, the time/number of computational steps needed will either scale linearly as the number of cases examined, or (God help you) as the square of such cases. But the RIF approach knocks that down to O(1), which is just a constant that doesn’t depend on the number of protein-sidechain interactions being considered. In real world terms, that boosts the number of things you can study by a factor of millions. The interactions themselves were stored in a 6-D hash table, which allows for quick lookup and summing of each docking example by just looking up the individual amino acids. (See below, though – even this choked up the process later on).
One of the things I like most about this paper is that it moves out of the computational world to the physical one of actual proteins: the team was prepared to test things out on a large scale. The proteins involved were over 84,000 calculated miniproteins (50 to 65 amino acids long, in this case), with variations on five different sorts of predicted surface shapes. These were also designed to have definite hydrophobic cores to help to ensure folding to present these surfaces, but in reality, about 34,000 of these were found to be stable once they were expressed. That itself is a bit of a commentary on the limits of what we can currently accomplish with protein design, but it did provide a good set of test proteins in the end and using this set surely made the final results much more meaningful.
There was a lot of work to do in docking these protein surfaces against that big pile of RIF data, and you can tell that the team ran into some difficulties there. It’s the old story: you can get crappy numbers quickly, or useful ones much more slowly. They were using the RosettaFold forcefields (which have, along with AlphaFold, famously performed very impressively with protein structures), but going “Full Rosetta” on every case was still far too computationally intensive – and that was true even after cutting the RIF data down to size with a low-resolution shape-matching step. But they found that if they only calculated the hydrophobic amino acid side chains (and using stripped down energy function for speed), that this still gave useful results and allowed for a lot more RIF data to be shoveled into the hopper than they otherwise would have been able to handle. There were other changes made to the standard Rosetta approach (for example, dealing with a persistent tendency to bury polar groups and some problems with the final computed packing quality), but I won’t go into all those details, partly because I’m not qualified to and partly because I’m not sure how many people will be left reading if I try.
As the authors note, the size of the space sampled is completely insane (my term, not theirs, but I think they’d agree): tens of thousands of protein backbones (as mentioned), times one billion disembodied side chain interactions (as above), times about ten to the sixteenth interfaces per scaffold placement. The paper says at this point, with what I hope is extremely dry wit, that “Sampling of spaces of this size is necessarily incomplete”. They developed several other wheat-from-chaff steps to try to cut this down to size and to find the motifs quickly with unusually dense numbers of favorable side chain interactions. These were then sent back around for another docking-and-design cycle, and finally, the best of the best were selected for actual experimental checking.
So how did that work out? At this point, they selected thirteen proteins. From human cells: TrkA, FGFR2, EGFR, PDGFR, the insulin receptor as well as IGF1R, the Tie2 angiopoietin receptor, IL-7 receptor alpha, the CD3 delta chain, and TGF-beta. And from pathogen organisms they chose influenza A H3 hemagglutinin, VirB8-like protein from Rickettsia typhi, and everyone’s favorite, the SARS-CoV-2 spike protein. I have to add at this point that this is a very solid list of protein-protein interaction targets, and I have to congratulate the team on choosing these and being willing to tackle such pertinent examples.
They designed tens of thousands of possible protein binders for 13 sites across the dozen proteins (two sites for EGFR. To see how robust the projections were to changes in sequence, they tested these computationally against a full-saturation mutagenesis set at the protein surfaces themselves (each residue run through all 20 amino acids). The most optimized binders for each target after this round were then actually expressed in E. coli, and their affinities were measured in biolayer interferometry assays.
The best hits were the binder for the coronavirus spike protein (0.15 nM KD), IL-7R (0.3 nM) and VirB8 (0.5 nM). The EGFR and TrkA candidates had single-digit nanomolar potency, but most of the others were hundreds of nanomolar, up to IGF1R at 860 nM. The spike protein example was already described in a separate manuscript while the rest of these were being worked on, I should note). That’s not going to put us all out of business immediately, but it’s pretty damn good for pulling protein structures right out of the air, isn’t it? Functionally, the receptor ligands were all antagonists, and testing each of them against all the other targets yielded very little cross-reactivity (which was expected, since their own structures are rather diverse). Most of the computed proteins turn out to be three-helix bundles, and the resemblance to antibodies and other designed antibody substitutes is something that many will note immediately.
The team got X-ray structures of the bound complexes in several cases, and the actual interactions matched the predicted ones quite well indeed. Interestingly, the success rates for these designed binders varied a great deal across the different targets – FGFR2 and PDGFR had hundreds of good candidates , while Tie2 and CD3 had fewer than ten hits from libraries of 100,000 potential designs. These success rates correlated very strongly with the hydrophobicity of the targeted regions on the proteins, which suggests that attempting to design in hydrogen bonds is still very much an unsolved problem. I would be interested to know how much of this hydrophobic-interaction space was “greasy side chain stuff” and how much was more directional things like pi interactions (stacking or edge-to-face). My guess is more of the former, for similar reasons to the lack of hydrogen bonding success.
But this tells us that the state of the art (which is certainly is) has gotten pretty interesting. No one has ever seen a native protein that binds to the targeted region of VirB8, for example, but this technique yielded a sub-nanomolar one. I think a very interesting experiment would be to produce a large library of 50-mers (as much as experimental conditions would stand – perhaps up to DNA-encoded library levels) and screen those against these targets to see what you might come up with. In other words, what’s currently more work: going through these calculations or just banging out peptides and hoping for the best? If you wanted to get real mano a mano with that idea, you’d run a big set of the optimized protein candidates identified here to see if the later steps missed anything good. And as a small molecule guy, I would be very interested to see something like this applied to a large set of plausible non-peptidic compounds, although that would certainly be a lot more work.
Either way, though, it’s important to remember that the computational route, as always, is improving faster than the bench route. As you can tell, this paper had to chop, chop, chop this huge problem down to size at many points – we, as in we the human species, do not have the computational resources to go after such things without such filters in place. But we’ll not only get more computing power in general as time goes on, but we’ll also get better at narrowing down things like this at the same time. Never bet against things that are driven by software and hardware improvements – that’s a lesson that has been driven home many a time over the last fifty years or so, and it will be driven home some more in the decades to come.
