Scientist Stories: Nicole Soranzo, The genetic basis of common traits and diseases

Thank you. Thank you for the introduction. I'm Nicole Soranzo. I'm a senior group leader at the Sanger Institute, and I'm a professor of human genetics at the University of Cambridge, and I've also taken a recent appointment at Human Technopole in Italy to lead the genomics center. So today I'd like to tell you a little bit about our studies working on the genetics of common complex traits. And diseases. I don't think I need to tell you too much about the importance of common diseases. So there are the diseases including cardiovascular diseases, cancer, and respiratory diseases with the highest burden in both mortality and morbidity, measured here as disability-adjusted life years in, in the world. And indeed, we know a lot about these diseases and also about the underlying We can also quantify the burden of underlying risk factors, such as, for instance, in this case, blood pressure. So over the course of the last 20 years or so, we made tremendous progress in understanding the genetic basis of common diseases, and really well explained in this review on the left-hand side. And this has particularly been fostered through the rapid technological advancement which have enabled the development of really breathtaking resources that in turn have fostered key findings and knowledge in genetics. And despite this tremendous progress, however, there's a lot more to be done and understood. In particular, we're still trying to measure to what extent we have discovered all the genetic variation that affects human traits, disease risk, and patient outcomes, particularly for those variants at the left-hand side of the allelic spectrum, which we know have been poorly represented in SNP-based imputation arrays that we've used until recently. The second really big challenge is that to assign a molecular interpretation and function to the genetic variants discovered and to in fact identify effective genes. And although we have available a growing and really sophisticated array of genomic technologies, our ability to use this toolbox to really advance understanding of mechanistic basis of genetic variation is still relatively limited because these processes are highly complex and painstaking. The third big challenge is that to harness the wealth of genetic information to really improve prognosis, diagnosis, and treatment of disease. So we know, for instance, that genetic variation can identify drug targets that have a higher success rate in drug development pipelines. But for the most part, really, the wealth of information in genetic studies has not yet been harnessed for therapeutic purposes. So my group uses a number of approaches shown here. We mostly rely on the analysis of large-scale genetic and genomic data. Datasets, including different types of genetic variants, to try and discover and characterize genetic variants affecting susceptibility to common diseases and their underlying risk factors. And we also use a growing number of molecular phenotypes and also experimental screens of regulatory activity to try and identify and validate targets for drug development. In terms of discoveries, we are moving very rapidly from SNP arrays to whole genome sequencing approaches, as many of the community are doing. And these typically rely on the analysis of large-scale cohorts where participants are predominantly healthy. They can be i-selected or partially selected and ideally population, but not always population representative. Participants have typically been characterized at a range of traits and phenotypes, including risk factors for disease, candidate biomarkers, but progressively also measurement of environmental and disease risk. And one of the first studies that we did was the UK10K study, was published now a few years ago, where we applied whole genome sequencing to about 4,000 individuals from two population-based cohorts. And as shown in the plot on the right, we were able to identify a handful of genetic variants, mostly linked to lipid metabolism. But by all means, it was very clear already that the sample sizes are inadequate for discoveries of complex phenotypes. Opportunities come, and we use them to study disease risk by also studying many of the underlying risk factors that capture indeed intermediate layers of variation linking a genetic variant to a disease of interest, including, for instance, phenotypes related to transcription, to translation, and to function. And we began to really use these phenotypes at scale to try and get a better sense for how genetic variants affect disease traits through specific pathways of molecular processes. Indeed, we were among the early adopters of these technologies. At scale in population cohorts. And using quantitative trait loci approaches, we were able to show that indeed phenotypes such as gene regulation phenotypes, metabolites, proteins, or indeed cellular composition phenotypes can really provide us with a much more resolved view of the effect of individual genetic variants. So today I'd like to tell you a little bit about the 3 examples of work that we've done exploiting, particularly exploiting some of these approaches. The first study has been a collaboration with John Danesh, Adam Butterworth, and many colleagues in the Cardiovascular Epidemiology Unit at the University of Cambridge. And here we really started to apply the power of genome sequencing to the power of multi-omic phenotypes. Phenotyping in cohort patients. John and colleagues have collected this wonderful resource called the INTERVAL study, which is indeed a randomized clinical trial of blood donation involving over 50,000 participants that were recruited between years 2012 and 2014. Importantly, these participants also volunteered to create one of the richest resources that includes extensive profiling of thousands of phenotypes, including multiple measurement of blood counts, metabolites, proteins, as well as traditional risk factors. So in partnership with them, we decided to genome sequence nearly 17,000 donors of the cohort, including a combination of whole genome and exome sequencing technologies. And we have now collected a large set of large catalog of genetic variation from these data. Our analysis focused on 3 classes of genetic variants, including small SNVs and indels, structural variants, and mitochondrial DNA variation. And it is really about the latter that I want to tell you a little bit about. So this is unpublished work that was collected and done in collaboration with wonderful colleagues Patrick Shinery and Oliver Steggall. And it was really done very much by two really talented postdocs. Na Kai was jointly supervised by myself and Oliver, and Aurora Gómez Durán was supervised by Patrick Shinery. And indeed, both of them have now moved to independent faculty positions in Germany and in Spain, respectively. So mitochondrial DNA is a small molecule, is only about 17 kb in size. It encodes for a small number of genes, about 13 protein-coding genes, which are mostly encoding for protein of the electron transport chain, and also about 24 tRNAs that are involved in the translation of mitochondrial encoded proteins. So although this genome is very small, it encodes for a hugely important function, particularly because, as you know, mitochondria are the energy powerhouses of the cell. And indeed, it has been shown that the combination of environmental factors, of genetic variation encoded by the mitochondrial genome itself or by the nuclear genome, but affecting mitochondrial function, have really the potential to create dysfunction, particularly to the main function of the mitochondria, which is oxidative phosphorylation, and this leading to progressive bioenergetic decline and the onset of severe diseases such as cancers, metabolic diseases, degenerative diseases, as well as the normal processes of aging. So we used the whole genome or exome sequencing data to call genetic variants, and we validated that through data collected from Affymetrix arrays. Ultimately, we decided to settle on a set of 184 common mitochondrial DNA variants that were robustly called in both the whole genome and whole exome sequencing data, and we tested association with over 5,000 phenotypes. Here on the right, you can see that the mitochondrial variants fully resolve the haplotypic structure or the haplogroup structure of the mitochondria in the population., as shown here on the panel on the left, which is obviously very distinct from the genetic structure in the same individuals that we can glean from nuclear variation. So as I said, we tested association of over 5,000 phenotypes, and here on the left is the Manhattan plot showing the location of the SNPs along the mitochondrial genome with their encoding genes, and p-values of association for all 5,000 metabolites— 5,000 phenotypes. As you can see, the top associations are few, and their surpassing genomic significance are few. And indeed, they are all observed with a single metabolite called N-formylmethionine, or fMet for short. So although we were a little disappointed that it was a single association, this turned out to be extremely interesting. Now, she applied a number of statistical and genetic analysis to fully resolve the haplogroup structure of these variants, and she was able to show that there are indeed 3 independent association signals driven by 3 variants in dark blue, light blue, and red that independently drive the FMET associations. And these 3 variants are in fact sitting on 2 on the early branch of the U haplogroup, at the end of the UK1 supergroup as part of the U haplogroup, and also on the H4 haplogroup, mitochondrial haplogroup. So overall, there are 3 independent associations linking N-formylmethionine with mitochondrial variants, and all of these have been replicated in an independent data set. So what was really intriguing about this association is that N-formyl methionine is known to be the amino acid that is responsible, or it's critical, for the initiation of protein synthesis in the mitochondria. And indeed, both mitochondria and bacteria use the N-formylated version of methionine, but indeed in this, in the cytoplasm, the non-formulated methionine version is used. So first of all, we thought to validate this signal was really truly driven by mitochondrial variation. So we did a separate association study in nuclear genomes and found that indeed there was a single independent association in the nuclear genome that accounted for about 14% of variation in Fmet levels, mostly involving the homolog of a yeast tRNA transferase gene. However, the mitochondrial association was independent and not associated with that and accounted for an additional 6% variation in Fmet levels. The rest of Fmet variation was driven by a combination of either unexplained factors or noise, or indeed by variation that is captured through metabolites and blood cell phenotypes that is potentially capturing an unmeasured environment. So because of the potential role of N-formylmethionine in protein synthesis, we decided to look at additional datasets for human samples to try and identify, to see whether we could detect an association of this mitochondrial variant also with gene expression and protein phenotypes. So on the bottom panels are associations with GTEx. Gene expression data. As you see on the left, there are indeed 5 different eQTLs, all driven by fMET-associated variants involving 5 genes. And what we see is that indeed there is association. In this case, we report the association for one of these genes, for MTND3, which is a member of the respiratory chain. We see that indeed there is a consistent downregulation of gene expression phenotype associated with the fMET-associated variant, whereas, sorry, other variants do not show an effect. We saw a similar but not really conclusive pattern with protein levels, and this obviously requires validation in independent cohorts. So having seen this association, particularly tantalizing hypothesis that this genetic variant might affect the protein protestasis in mitochondria, we set up a partnership with Patrick Chinnery at the Mitochondrial Biology Unit at the University of Cambridge. And really, Aurora, who is a genius of mitochondrial DNA experimental work, has used an enormous variety of techniques to really try and drill down on the effect of these genetic variants and to build a potential model for how fMET contributes to variation in human phenotype. She used a model system called the cybrid where the different mitochondrial variants, in this case the H4 and the UK haplogroups, are implanted into this same cytoplasm of a single reference cell. And this allows us to really look at the effect of mitochondrial variation without the confounding of the genome, nuclear genome background. She used the assays including different type of assays including protein synthesis assays, both of mitochondrial cytoplasmic proteins. She looked at mitochondrial complexes through electrophoretic analysis both in native and denatured condition. And she did various other techniques including cell signalling immunoblotting, bioenergetic studies, and ubiquitination. Tination quantification of cytoplasmic proteins. So, the model that she came up with is shown here on the right, is quite complex. This data is in advanced stage of review. But basically, what we can show is that there is clearly an effect of fmet in the mitochondria in controlling the rate of mitochondrial protein synthesis. And there are parallel efforts, effects also in correlating fmet levels correlated to Apo groups to mitochondrial stress and oxygen consumption by the proteins of the complex I and IV of the mitochondria. She was also able to show that there is a parallel effect on the cytoplasm, and indeed where she can observe again a decrease in protein synthesis and increase in integrated stress response. All pointing to really a role of Fmet in controlling proteostasis. What was really intriguing is that this effect on protein synthesis didn't appear to be mediated by the usual rapamycin-mediated pathway, but indeed seemed to more involve a role of Fmet as an endogran involving the— basically controlling proteasomal degradation through ubiquitination. So overall, our model really shows that fmet levels coordinate mitochondrial and cytoplasmic proteases both in synthesis and in degradation, which in turn controls cellular processes well beyond the well-studied mitochondrial bioenergetics, oxphosphatase and ATP synthesis. So given this potential mechanism highlighted that could be mediated by fMET, we also sought to understand whether fMET levels could be implicated in human diseases. And this is partly because it's been already shown that the haplogroup UK, which is the haplogroup that we associate with fMET, has been associated with the risk of late onset diseases such as Parkinson's disease and and ischemic stroke. So we set up another collaboration, this time with Claudia Langenberg and her team. This analysis has been done by Mike Pitzner to test whether, in a prospective manner, whether variation in N-formylmethionine would be associated with a host of severe health outcomes, including many late onset diseases. So, what this plot shows is the odds ratio of f-methyl levels measured at baseline against the 24 different outcomes across a 20-year follow-up period in 11,900 EPIC-North participants. And the line represents a significance threshold and an association where— sorry, in black, if it surpassed the significant threshold adjusted for the experiment.. So, what you see there is a number of diseases including renal disease, coronary disease, heart failure, peripheral artery disease, COPD, and indeed even mortality that are significantly associated with Fmet levels, suggesting that Fmet could be a biomarker at best and potentially also mediator of differences in disease phenotypes. In the other parts of the talk, I would like to switch gears a little bit and tell you a little bit more of the work that we've been doing studying genetics in a whole biological system. This is really work that I've done and my group has been doing in the course of the last 10 years or so at the Sanger Institute together with many collaborators. So what we typically do is that we use clinical blood cell reports to derive cellular phenotypes, including the counts, ratios, volume proportion, and physiological parameters that are associated with the function of blood cells. In particular, mature blood cells such as red cells, platelets, and white cells. So, we know these cells are obviously important for a number of physiological processes including oxygen transport and immunity and are tightly controlled under normal homeostatic conditions.. We also know that these parameters are altered in diseases that result in hematopoietic dysfunction, and these are many common rare diseases as well as cancers. So over the course of the last 10 years or so, we've done a number of progressively wider genome-wide association studies taking genetic variation and linking them to these phenotypes to test, to discover genetic variants predisposing to different potential set counts and function measures. The last of this was published about a year and a half ago and really looked at nearly 800,000 participants and 29 different traits and discovered nearly 11,000 independent variants associated with these 29 phenotypes. So this is really the largest set of correlated phenotypes that have been discovered to date, and what was a really useful is that we discovered a large number of low frequency and rare variants, providing us with a new opportunity to study the property of these variants at scale. As we expect for genome-wide association discovered variants, many of these variants are non-coding, and this was particularly true for common variants. But indeed, what was surprising is that many rare variants, some of which have really high effect sizes, so effect sizes comparable to coding variants still had a presumed regulatory annotation. So we really took advantage of this unprecedented catalog of genetic variation to ask a number of questions related to the function of genetic variants across the spectrum of allelic predisposition. I'm just going to show you a couple of, of these examples. So, the main question that we really wanted to address is this dual function of blood cells, both in normal physiology, where, as I said, their counts are really controlled in tighter ranges under normal hemostasis. And on the other side, the potential role of these associations also in dysregulated diseases such as common diseases, rare diseases, and even cancers. So we wanted to first of all test whether the contribution of genomide variants that represent effectively the background genome might also contribute to the risk or indeed modify the risk of these diseases. And at best also to understand whether the same genes might be implicated in both processes under potentially disease liability model. So to capture the effect of the overall genetic variation affecting these phenotypes, we built what we call polygenic risk scores. They are effectively just weighted sums of the number of alleles that each person in population carries with respect to the alleles associated with any given phenotype. And so if you see in this plot on the left-hand side, the majority of the population will have average level of polygenic risk course, but people at the extreme might be particularly low or high risk of any given, you know, carrying an extreme phenotype. And so one could ask, for instance, if people at the lower end of the spectrum might be protected against the risk of given diseases, and people on the right might indeed be at increased risk over the normal background of the population. So we developed the polygenic risk factors, risk scores in UK Biobank, and then we validated them in two populations in the INTERPOS study and also in a Canadian study led by Guillaume Lettre, among others. What we can see is that these polygenic risk scores that are derived from hundreds of sentinel variants are both highly predictive with our Pearson correlation measure between 0.22 and 55, and also highly portable across European populations. We can then ask whether these polygenic risk scores modify or influence the risk of diseases. To do that, we first of all characterized the genetic landscape of, in the case of rare disorders, of classical blood disorders in UK Biobank. This is really work driven by Dragan Avuković together with Willem Overhand. And so we classified the UK Biobank study participants by effectively scoring genetic variation coming from SNP arrays and intersecting them with the main disease databases including ClinVar, the HEMD human gene mutation database, and also a recently curated list of variants from causative for rare blood disorders from the rare disease project of the 100,000 Genome Project. So what we found out of all our sentinel variants, 101 were included in one or more of these databases, of which 80% were coding. And 16 out of 101 were annotated to be pathogenic in either database using very strict criteria, and those involved 11 different genes. So if you look at the pattern of homozygote mutation for recessive diseases and indeed heterozygote for dominant diseases, we see that there are several instances of apparently healthy UK Biobank participants that were carriers for validated pathogenic variants for autosomal recessive disorders. And so this really suggests that, you know, there is some level of misclassification in Biobank that could be indeed caused by many different, by many different reasons. One could be low penetrance of the variant itself. It could be missing health record data that means that the person has a disease that has not been reported in Biobank. Misannotation of the pathogenicity of the variant, meaning that the variants are not really pathogenic, although we use the strict validation criteria., or indeed other undiscovered compensatory effects that could be driven by either other rare variants that are not captured by the SNP arrays, or indeed polygenic variation. So while obviously this analysis needs to be repeated in the presence of a complete genome sequence, we here sought to explore whether polygenic variation might be a contributor to, in addition to these variants. To modifying the disease phenotype. So for UK Biobank, we retrieve ICD codes for 423 different diseases of blood using Chapter 3 definition, and that defined 29,080 patients and controls. And then for each of these, we estimated a weighted genetic genetic risk score or polygenic risk score for each patient and fit logistic regression models to test the association of the GTS with the rare disorders of the blood. So the results for association surpassing our experiment-wide significance threshold are shown here. And so we can show that there are indeed polygenic effects on these diseases. For instance, we see a protective effect of platelet count and PCT and EO on thrombocytopenias. We see an increased risk for secondary polycythemia, which is a disorder of red cells associated with HCT, HGB, and RBC. We see, for instance, a risk effect linked to variants associated with eosinophil counts for hypergammaglobulinemia, and indeed a protective effect for platelet volume for hereditary factor with ACE deficiency. So while these, as I said, these effects need to be validated in the presence of whole genome sequencing data and ideally also in additional validation cohorts, we have managed to replicate some of these effects in a cohort of bleeding, thrombotic, and pleural disorders. And so we are pretty confident that some of these results will hold up.. And so overall, this really suggests that the common variation of small effects, which are, you know, what we call the background polygenic genome, drive or at least modify the risk of rare blood disorders independent of Mendelian mutations. What will be intriguing is also to see if similar effects are also observed for cancers and particularly where the polygenic risk scores might modify the risk of cancer manifestation or indeed outcome in the presence of somatic mutations. And there's already some evidence for that coming from work of Jyoti Nangalia that has shown some of the SNPs that we previously associated with red cell and platelet phenotypes are indeed also associated with a risk of diagnosis of ET versus Bv. The last few minutes, I'd like to talk a little bit about a third strand of work, again focusing on full blood count interpretation, where we exploit multi-omic regulatory phenotypes to study disease mechanisms. So clearly, as I said, this is a big challenge and a challenge that is particularly difficult in case of poorly accessible tissues, for instance, in in brain. In blood, we are relatively lucky that we are able to isolate to a high degree of purity not only mature cells but indeed many of the blood cell progenitors. And in this work done by our collaborator Vijay Sankaran, he was able to beautifully characterize at single-cell resolution patterns of chromatin accessibility in all the major blood cells, including, as shown here, for instance, the lineage leading to the maturation of mature red blood cells from hematopoietic stem cell progenitors. And what we are able to do is then to intersect our genetic data and our pharma mapping data to derive what are the more likely regulatory elements that underpin some of these associations. And as you see in this plot, these are 75 different red blood cell phenotypes. In some cases, the chromatin accessibility is only available in the late stages of differentiation, potentially indicating an answer that opens late in differentiation. In other cases, it matches to answers that are indeed open throughout the differentiation process. So we can intersect genetic data to derive hypotheses on the identity and location of many regulatory elements. In this case, this is a genetic variant in BCL2 that has been associated with decreasing red blood cell count using off-the-counter monocyte count. And we see indeed that this region maps to peaks of accessible chromatin that are only specific to myeloid cells, whereas they are basically not present in lymphoid cells. This is another example where we see association of a well-found map and now validated variant interleukin-1 and interleukin-2 associated with type 1 diabetes and multiple sclerosis. And we see again that the association maps to open chromatin regions in lymphoid cells and is also associated with highly deleterious alteration of transcription factor binding, again in lymphoid cells. So we can use this wealth of information to begin to drive some of the experimental approaches to validate the function of these associations. In vitro. And this is just a very beautiful summary of some of the main approaches available for blood in a recent review. And in this case, through work particularly of Manuel Tredaglia, who was a postdoc in my team, together with Eve Coomber and Dominique von Schiller, we were able to develop a multiparallel reporter assay to begin to probe some of these associations in reference cell types that are effectively hematopoietic stem cell types. And what we begin to see is really that a subset of the variants that we discovered through association study show positive function, both as enhancers and also with allele-specific effects linking effectively the genetic variant that we identified to regulatory effects. So we can also use this information to begin to map common disease associations. And this is work again of Manuel together with Koichi Kondo. And this is just one example of a locus where it's a predisposition locus for IPD, which also intersects regulatory annotation for eQTLs for expression and histone modification quantitative trait loss. So what we can see that out of all the variants that map in find mapping credible sites. The variants that are identified through gene expression and histone modification QTLs mostly map to histone-bound regions, and indeed the top variant maps to a transcription factor binding site. Whereas the variants associated with the IBD risk are find mapped predominantly to a region of the genome that has no regulatory hallmark. So we put some of these variants into the mPRADesign itself, and we're able to show that out of the credible set of 11 variants discovered in IBD GWAS, indeed a single variant which corresponds to the CBP-beta-bound variant has significant regulatory NLE-specific activity, suggesting that this is potentially a prime candidate causal variant. So this effort of resolving association will be painstaking. Both the rationale and the approaches and the priorities for this work have been really nicely formalized by a working group of the International Common Disease Alliance, which is really an international group of scientists that is attempting to really define framework by which we can take genetic information and continue to develop both genetic maps to map their genetic mechanism to regulatory and other type of approaches, and ultimately to harness this data through— in medical applications. And in addition to these efforts that I described So obviously, also a lot of work geared towards defining both the data infrastructure components, but also components of ethics and global equity to ensure that the benefit of these studies benefits all the world population and not just focusing on individuals of Northern European ancestry, as it's mostly been done so far. So with this, I really like to thank all the people that contributed to this work have been highlighted over the course of the talk. This is my team on the left, or a previous version of it. And these are some of the really wonderful postdocs, PhDs, and collaborators that have been working on these projects. And I've acknowledged most of them previously, so I will not repeat all of the names. So thank you very much for your attention.