Scientist Stories: Joe Marsh, Deep Mutational Scanning

We're going to jump to Joe Marsh, who was scheduled after lunch, but we're just going to keep going now with Joe, who was kind of José's successor in a sense, and worked on protein complexes. And he still does that today up in the MRC Human Genetics Unit, where he's a professor. Over to you, Joe.

Thank you, Sarah. Can everyone hear me, see my screen?

Yep, everything's great.

Excellent. So I changed the title of my talk a little bit. I decided to go with something a little bit more general to talk about some of the research that we're doing now up in Edinburgh and how it relates specifically to work that I did as a postdoc in Sarah's group. So my decision to join Sarah's lab as a postdoc was probably the most important scientific decision that I ever made in terms of my overall direction of my career. I did a PhD in Toronto in starting experimental NMR and moved increasingly computational over the course of the PhD. And I had the thought that, you know, I loved the computational work, I wanted to do a postdoc in structural bioinformatics, really excited about the idea of, you know, taking advantage of all these protein and protein complex structures that were now available and doing large-scale analyses. So I got in touch with Sarah. We had some great conversations. I came out and visited Cambridge. It went really well. And I remember Sarah said to me— I was finishing up my PhD and she said, you know, you can start basically as soon as possible as long as this ERC starter grant on protein complex evolution is funded that she was shortlisted for. And of course, given Sarah's track record at the time, it seemed, you know, of course they were going to give it to her. But it turns out that grant wasn't successful, so I had to wait for a fellowship. And it's interesting, one year later, Sarah tried again with a completely different project on T-cell differentiation that was funded for a starter grant. And interesting to think— obviously, she's been incredibly successful moving in that direction. So I ended up getting an EMBO fellowship and an HFSP fellowship, and I was here for 4 years total. And it was a great time with a great group of people. In the lab, and we got a lot of exciting stuff done. So we published quite a number of papers investigating various aspects of protein complex assembly and protein complex evolution. And I think that in combination with the work Tina did around the same time, we really showed that ERC made a huge mistake by not funding that grant, because there was certainly a lot of interesting stuff to do there. So In 2014, I moved to Edinburgh to start my own group at the MRC Human Genetics Unit. And so part of our research here is sort of continuing along these lines, thinking about sort of fundamental evolutionary and biophysical questions about complexes. So I just wanted to share this paper, for example, that's up as a preprint now. We just resubmitted it to a journal after major revisions called Interface Size Drives Co-Translational Assembly of Protein Complexes. And I thought I just wanted to share this because it sounds exactly like the kind of paper that we would have written when I was there. And the basic take-home message is protein complexes that assemble co-translationally tend to have big interfaces, which is basically exactly what you would have expected if you had read all the research on this coming out of the Type-ME group. So it's been fun to kind of continue to dabble in these kinds of questions. But our main focus now is really transitioned towards studying human genetic disease, which of course makes sense given that I'm at the MRC Human Genetics Unit now. But thinking about protein complexes can actually be very useful for this. So this is another preprint that we just put up last month looking at the properties of human disease mutations at protein interfaces. And basically it's fairly well known that disease mutations tend to be enriched in protein interfaces, but what we showed is, you know, different types of interfaces— different homomeric and heteromeric interfaces and the symmetry of protein complexes— can be very important determinants of the likelihood of mutations being pathogenic. And this has significant implications for the performance of different computational variant effect predictors at different types of mutations. So in addition to these kind of large-scale analyses where we're looking at patterns across thousands of structures and many thousands of mutations. We also do a lot of studies of individual proteins and protein families, in particular in, in collaboration with other sort of clinical geneticists here in Edinburgh and around the world. And one that we've worked on a lot is one that's been studied a huge amount at the MRC Human Genetics Unit where I'm located is PAC6. And I saw Veronica was in the audience earlier, who was, you know, originally showed that loss-of-function mutations in PAC6 were associated with aniridia. PAC6 is called the master regulator of eye development. And it's well known that if you get protein null mutations, so premature stop codons, frame shifts, that cause, you know, one copy of the protein not to be produced, you get this aniridia phenotype due to haploinsufficiency. It's very much known to be a dosage phenomenon where you have less PAC6, you get the eye phenotype. But in addition to these null mutants that cause aniridia, there's been a large number of missense mutations that have been identified with a, with a much broader range of eye phenotypes. And a number of these have been identified at the Human Genetics Unit over the last couple of decades. And the interesting thing about the disease-causing missense variants in PAC6 is that they is this much broader spectrum of phenotypic variants. They range from, you know, much milder phenotypes to very severe, so-called worse-than-null mutants. So certain single amino acid changes within PAC6 are much worse than actually knocking out your whole copy of the protein, suggesting that some kind of, you know, gain of function or dominant negative effect must be going on. So what we wondered is whether thinking about this from a protein structural perspective could help better understand what was going on with the PAC6 missense mutations. And so we did sort of extensive compilation of known disease mutations and classifying them in terms of their phenotypic properties. But to make it simple, what we found is that the most severe variants, the sort of so-called MAC-associated variants— microthymia and ophthalmia and coloboma— tended to— using a combination of analysis and mutation location and molecular modeling, they're far more likely to perturb protein protein-DNA interactions. So actually, if you look at the severe versus less severe mutations on the structure, there's not much of an obvious pattern. But if you actually model the effects of mutations on DNA binding and protein stability, you get quite a clear difference where the most severe mutations tend to sort of occur near DNA binding sites and don't destabilize the protein but seem to interfere in some way with DNA binding. And in a way, this, this explains the sort of worse than null effects that we observe. Whereas, you know, if you have a mutation that totally destabilizes the folding of PAX6, the protein is going to get degraded. It's essentially, you know, a phenotype of a complete knockout, heterozygous knockout, because the protein's not there, it's not doing its job. But if you get a mutation where the protein is still totally stable, but it's interfering with DNA binding somehow, that protein is still there and it's going to change the way that it's binding to DNA. And the important thing about PAX6 is it's not just one DNA binding domain, it's but it's got sort of two halves of the so-called paired domain, and then another homeodomain. So even if you interfere with binding of one, it's still going to bind in other ways. And it's probably changing its patterns of binding in this, driving a different transcriptional program, which can therefore lead to these worse-than-null phenotypic effects. So we published this paper a couple of years ago now, the sort of, you know, the structural bioinformatics analysis of the existing mutations. But we've still been really interested in what's actually going on at a molecular level with these mutations. And so, so to get at this in more detail, we're moving to a deep mutational scanning type of strategy to probe the effects of PAX6 mutants in the, in the paired domain. And so for that, we're using a kind of yeast-one hybrid-like screen where we've put the PAX6 enhancer into the yeast genome next to an antibiotic resistance gene. And so binding of PAX6, the paired domain, will drive the expression of antibiotics resistance protein. So what we can do is we can use saturation mutagenesis to build a library of all possible single amino acid mutants of, of the paired domain, transfer them into the yeast, and grow them together in the presence of antibiotic in a massively parallel competition assay. And so by performing deep sequencing at various time points, we can basically get the growth rate for all of these variants, which is then coupled to PAC6 binding to DNA. So the ones that grow fast will tend to be good binders, whereas those that ablate binding are going to grow slowly. And so we've spent a, you know, a couple of years getting this working. There's a lot of sort of technical difficulties to work out, but in the last few months now we're at an exciting point where we're starting to get some, some good measurements. And so this is our PAC6 sequence structure map where we've measured the effect on growth rate for almost all single amino acid substitutions. Substitutions in the paired domain. And so we can start to use this to see if we can better understand what's going on. So when we map the mutations to the structure of the protein, it makes a lot of sense. You get, you know, much higher effects on fitness and the mutations that are close to the DNA binding site. Almost all the mutations tend to disrupt binding. Or grow slower than wild type. But I've highlighted in yellow here on this isoleucine 71 position, which is really interesting in that almost every mutation at that position substantially increases the growth rate of the protein. So it seems to enhance DNA binding. So there's no— we don't know of any disease mutations or even human population variants occurring at that position, but it's, it's very interesting to think about what specifically is going on there. And the other thing that we can do then is we can use this data to try and see, you know, how well does it explain the human PAC6 disease. So we can compare the experimentally measured growth rates for reputatively benign variants observed in the human population taken from the NOMAD database and the known pathogenic missense variants. And you can see that the experimentally measured fitness score does a pretty good job at distinguishing between the between the two groups, and most of the disease mutations tend to have very strong effects on fitness. But it's certainly not perfect. You know, some of the disease mutations, which are— we're quite certain are real disease mutations, are, you know, even similar to the nomad variants or even appear to have enhanced— a little bit enhanced binding. So again, the question is what exactly is going on here. And so one of the issues potentially is that, you know, we're only looking at one enhancer, but PAX6 is sort of very quite broad in terms of the different enhancer sequences that it can bind to. So we're going to try and repeat the experiments next using at least one different enhancer sequence and see how similar the results are. And it could be just the other effects on the proteins that are something other than simply disrupting on DNA binding that can lead to disease. But the crucial thing here is we can compare these effects on growth rate to existing computational predictors. So we use sort of a ROC area under the curve to compare our DMS measurements to a bunch of different computational variant effect predictors and see how good they distinguish between these two groups of mutations. And so the DMS data is here in red, and the rest are all a bunch of different computational predictors. And what you can see is that the DMS data ranks third in terms of its ability to distinguish between the pathogenic and benign variants. There's two that perform better, but the crucial thing to emphasize is that most of these methods and all of the, the best performing methods are based upon supervised learning approaches where they've been trained on, you know, datasets of known disease mutations. So almost certainly these predictors have seen some of the PAC6 disease mutations before, so there's a bit of overfitting going on, which is always such a huge issue when it comes to the assessment of any computational predictors and especially variant effect predictors. So if we compare the experimental data to the best predictor based upon an unsupervised learning approach that shouldn't be biased in this way, which is deep sequence here, you can see that it does do substantially better. So it does strongly suggest that these kind of deep mutational scanning measurements, while they may be coupled to just one specific experimental phenotype, really have huge potential and are quite possibly better than any existing computational approaches for identifying new pathogenic variants. And so if you're interested in this, we actually published a whole paper last year on this issue of using deep mutational scanning data to benchmark variant effect predictors and to directly identify disease mutations. So that's it. I'd like to thank my group members, in particular Alex and Marcin, who did the deep mutational scanning work. Rolando Hernandez, who helped analyze some of that DMS data. Mihai and Ben, who wrote the papers that I showed. And my very close collaborators at the Human Genetics Unit, Greg Kudla. Our groups are incredibly well integrated now, where we have several shared members between us, and he's really enabling us to do all this kind of deep mutational scanning work. And David Fitzpatrick, who's a clinical geneticist who I've worked with extensively and in particular on the PAC6 stuff here. It's really been a big supporter within the unit for the sort of power of my structure-based approaches to help us understand human genetic disease. And also to my girls: Hannah, who was born when we were in Cambridge, but since moving to Edinburgh we've had Sophie, and just earlier this year Isabelle. So 3 girls just like Sarah, so I can only hope that they turn out as well as the, the Tyckman girls. Quick add, we've got a program leader track position in computational biology open at the Human Genetics Unit, great place to work, a core-funded position. So if you're interested, send me an email or contact Chris Ponting, who's the head of biomedical genomics, who's leading the search. So thank you very much.

Thank you, Joe.