Friday, April 27, 2018

New T2DKP release adds individual-level data for interactive analysis

With the April release of the Type 2 Diabetes Knowledge Portal, we are increasing the number of datasets and samples available for interactive analysis via the LocusZoom and GAIT tools. These tools now access individual-level data from three additional datasets, all of which were quality controlled and analyzed at the Accelerating Medicines Partnership in Type 2 Diabetes (AMP T2D) Data Coordinating Center (DCC):
  • CAMP GWAS: 3,628 multi-ancestry samples from the MGH Cardiology and Metabolic Patient cohort, generated by a public-private partnership between Pfizer Inc. and Massachusetts General Hospital;
  • METSIM GWAS: 8,791 European ancestry samples from the Metabolic Syndrome in Men study.
These individual-level data are available as "dynamic" datasets, powered by Hail software, in LocusZoom on Gene pages and Variant pages of the T2DKP, for the following phenotypes: 
  • BioMe AMP T2D GWAS: type 2 diabetes, BMI, diastolic blood pressure, fasting glucose, HbA1c, HDL cholesterol, LDL cholesterol, systolic blood pressure
  • CAMP GWAS: type 2 diabetes, BMI, fasting glucose, fasting insulin
  • METSIM GWAS: type 2 diabetes, BMI, diastolic blood pressure, fasting glucose, fasting insulin, HbA1c, HDL cholesterol, LDL cholesterol, systolic blood pressure
To perform interactive analyses on these data in LocusZoom, select one of the available phenotypes in step 1 and then choose a "dynamic" dataset in step 2.

When you click on a variant in the resulting LocusZoom plot, the option to condition on that variant appears in the tooltip:

Clicking on that link starts on-the-fly association analysis for the region while conditioning on that variant, which can reveal whether association signals are independent of each other. You can choose to condition on multiple variants. The variants of your choice are listed in the upper left-hand corner of the plot, and the list may be edited:

Individual-level data from these three datasets are also available for interactive analysis via the Genetic Association Interactive Tool (GAIT) on Variant Pages. After selecting one of the datasets, you will be able to choose a phenotype for association analysis, filter the sample pool by specifying a range of values for one or more phenotypes, choose custom covariates, and then run on-the-fly association analysis for your chosen subset of samples. Find all of the details about how to use this tool in our GAIT guide.

We hope that the increased ability to interact with individual-level data in the T2DKP will be helpful to your research. As always, we are happy to answer any questions about these or other data and tools; please contact us for help.

Tuesday, April 17, 2018

Developing a model for collaborative science: a mid-term perspective on the AMP T2D Partnership

In 2011, Dr. Francis Collins, Director of the National Institutes of Health (NIH), met with leaders in biomedical research to discuss a frustrating problem. Continual improvements in molecular biological and genomic techniques were generating an avalanche of data relevant to complex diseases, yet the translation of these data into insights about disease mechanisms and drug targets was unacceptably slow. It was clear that an entirely new paradigm for collaborative research would be needed to speed up the extraction of knowledge from data.

The result of these discussions was the creation of the Accelerating Medicines Partnership (AMP), one branch of which focuses on type 2 diabetes (T2D)—a life-threatening disease that affects hundreds of millions of people worldwide, whose incidence is growing, and whose progression cannot yet be effectively stopped or reversed. AMP T2D, a five-year project, includes the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK); the pharmaceutical companies Janssen Pharmaceuticals, Eli Lilly and Company, Merck, Pfizer, and Sanofi; the University of Michigan; the University of Oxford; the Broad Institute; and other researchers around the globe. The Foundation for the National Institutes of Health (FNIH) also provides funding and coordination for the project.

Drawing on the strengths of both academia and industry, this public-private partnership brings together all stakeholders in a pre-competitive space to share data and combine resources, with the goal of validating new drug targets faster. Now in Spring 2018, roughly mid-way through the funding period, it is evident that this collaboration has resulted in remarkable progress on both scientific and collaborative fronts.

Genetic association data: the foundation of AMP T2D

Genetic association studies interrogate the genomes of individuals at millions of specific genomic positions to discover sequence variants that are correlated with the incidence of disease. From the outset, AMP T2D aimed to support the generation of unprecedented amounts of new genome-wide association study (GWAS), exome sequencing, and whole-genome sequencing data within the project as well as their aggregation with all relevant publicly available data. Originally, 5 sites were funded by the NIDDK to generate new data and deposit them into the AMP T2D Data Coordinating Center (DCC) at the Broad Institute. As the project evolved, another site was funded by the NIDDK and 8 more sites were funded by the FNIH. Additionally, an Opportunity Pool of funds from the NIDDK was created, allowing the AMP T2D Steering Committee to award smaller grants for complementary research projects in a flexible, science-driven manner.  Currently 10 Opportunity Pool projects are in progress, and more awards will be given in the future.

Not only has the number of genetic association studies increased since the inception of AMP T2D, but also the number of samples surveyed in each has grown dramatically, from typically under 100,000 to approaching 1 million today. The increased statistical power conferred by these large sample sizes has led to a huge increase in the number of loci found to be significantly associated with T2D, from about 70 at the start of the project to nearly 430.

Improvements in genomic technologies in the past few years have allowed AMP T2D collaborators to generate increasing amounts of sequencing data, which make it possible to comprehensively interrogate all alleles and to uncover rare variation. At the project’s start, T2D associations with exome sequences (covering the protein-coding regions of the genome) were available for about 13,000 samples, and no whole-genome sequencing studies had been published. Now, more than 2,600 whole genomes are available, and analysis of a set of 50,000 exomes—the largest disease-specific aggregation of exome sequencing data to date—is nearly complete. Importantly, many of the associations that have been newly discovered in sequencing studies involve relatively rare variants that affect protein-coding regions. It is often more straightforward to develop hypotheses about the impact of such variants than it is for variants outside of coding regions.

As the AMP T2D partnership has grown in prominence in the diabetes field, the DCC has been approached by investigators outside the project who want to contribute their data in order to aggregate and display them in the context of AMP T2D data. In early 2017, researchers in the 70kforT2D project, which found novel T2D associations by re-analyzing existing GWAS data, offered their results for integration into the DCC and display in the Type 2 Diabetes Knowledge Portal (T2DKP; see below) before publication.

70kforT2D GWAS was first pre-publication dataset to be added to the T2DKP from outside the AMP T2D partnership, and it was particularly appropriate that these scientists, whose results illustrate the value of data sharing, themselves chose to freely share their results. Incorporation of datasets into the AMP T2D DCC and T2DKP offers investigators the chance to take advantage of the expertise of the AMP DCC analysis team, apply cutting-edge analysis tools to their data, and display their results broadly to the T2D research community in the context of multiple datasets. The AMP T2D DCC is open to incorporating T2D-relevant datasets from all investigators (find details on contributing data here).

In addition to the datasets generated by AMP T2D partners and other T2D researchers, which focus on associations with T2D, glycemic measures, and T2D complications, the AMP T2D DCC also collects publicly available genetic association datasets for traits relevant to T2D, such as anthropometric measures, blood pressure and lipid levels, and heart and kidney disease.

Orthogonal data types to help identify and prioritize causal variants and genes

Finding genetic variants that are associated with T2D risk is critically important to understanding the genetics of T2D, but it is only a first step. The most significantly associated variant in a genomic region may not be the causal variant that is responsible for altered T2D risk. Researchers perform fine mapping to analyze genetic associations in specific regions of the genome and generate credible sets—that is, sets of variants that are predicted to include the causal variant. Mid-way through the AMP T2D funding period, emphasis among the data-generating partners is beginning to shift from simply generating association data to performing fine mapping and credible set analysis.

But even after predicting which sequence variations are responsible for altered risk, finding clues about how they affect risk requires integration with additional data types. Information about the functional importance of the genomic region where a variant is located—its relevance to gene expression, protein function, networks and pathways, metabolite levels, and more, all determined on a tissue-specific basis—can help prioritize genes and pathways for in-depth experimental investigation. These kinds of research were built into AMP T2D from the beginning, and as the importance of these data types became even clearer, several Opportunity Pool awards were given to projects focusing on complementary data types that shed light on the significance of genetic associations.

Several of these projects focus on generating tissue-specific epigenomic data: histone modifications, DNA methylation, chromatin conformation, transcription factor binding, 3-dimensional chromosome structure, and other data types. Epigenomic data can provide important clues about the mechanisms by which sequence variation affects T2D risk, particularly for variants that lie outside of protein-coding regions. For example, if a risk-associated variant is seen to disrupt a transcription factor binding site, this would support the hypothesis that the transcription factor and its target genes are relevant to T2D.

To make these data accessible to researchers, one Opportunity Pool award supports the creation of the Diabetes Epigenome Atlas, which collects and displays epigenomic datasets relevant to T2D. In the near future, these data will be fully integrated with genetic association data in the Type 2 Diabetes Knowledge Portal (see below).

Other Opportunity Pool projects are concerned with processes downstream of gene expression. Discovering interactions between proteins implicated in T2D risk, for example, could help to uncover all of the players in pathways important for the development of T2D, increasing the number of potential drug targets. Determining the effects of variants on the levels of key metabolites can illuminate the metabolic pathways that change during the development of T2D. 

In addition to generating all of these orthogonal data types, AMP T2D partners are developing algorithms and using machine learning to classify and prioritize variants on the basis of the functional annotations that accompany them. Finally, other Opportunity Pool projects will use model organisms to test and validate drug targets that are suggested by these analyses.

Tools and methods to speed analysis and interpretation

At the inception of AMP T2D it was also clear that the development of new methods and tools would need to accompany the generation of data, and support for these activities was built into the program. One major technical effort has addressed an obstacle to global data aggregation: because of institutional and national privacy regulations, some datasets may not leave their site of origin to be aggregated with other datasets at the AMP T2D DCC. A group at the European Bioinformatics Institute has built a technical replicate of the DCC and knowledgebase, such that data stored there are equally as accessible for browsing, searching, and interactive analysis as are the data stored at the AMP T2D DCC at the Broad Institute. This federation mechanism allows global data accessibility even when data aggregation is not permitted.

Other efforts supported by AMP T2D are aimed at improving the speed and efficiency at which data can be taken in and analyzed. In one project, a data intake system is being developed that will streamline the process for both data submitters and for the DCC team, and will be applicable to data submission both at the Broad DCC and at other federated sites. Another project has created a software pipeline, LoamStream, that will largely automate quality control and association analysis of incoming data. Currently, LoamStream is in use for quality control of genotype data, and this has already greatly reduced the time required to process new datasets. Future work will extend the pipeline to association analysis and will also allow it to take in sequence data as well as genotype data.

A genetic association of a variant with T2D gains credibility if multiple independent studies replicate the association. Thus, it is important for researchers to be able to evaluate the weight of available evidence. But currently this is difficult to assess from the association datasets in the AMP T2D DCC, because many are based on overlapping sets of subjects. AMP T2D partners at the University of Michigan and University of Oxford are working on a method to take these overlaps into account and synthesize associations from multiple datasets into a “bottom-line” significance for association of a variant with T2D, which will aid in prioritizing variants for future work.

Multiple AMP T2D projects for analysis, interpretation, and custom interactive analysis of variant-phenotype associations are ongoing at the Universities of Michigan, Chicago, and Oxford, Vanderbilt University, and the Broad Institute. These projects are aimed at facilitating, in various ways, the path from variant associations to functional knowledge, and all have been or will be integrated into the T2D Knowledge Portal (see below).

Hail software offers a pipeline that speeds up the analysis of huge genomic datasets, while the gnomAD resource aggregates and harmonizes exome and genome sequences to provide a catalog of genetic diversity, in more than 100,000 humans, that aids in interpretation of variant associations with disease. A tool under development in the gnomAD project will display the effects of variants on protein structures as another way to deduce their potential impact.

Other analysis modules include gene-based association methods for using expression data to predict genes that may impact a phenotype (PrediXcan and MetaXcan), and a phenome-wide association study (PheWAS) method for visualization of the associations of a variant across multiple phenotypes, which is a crucial consideration during drug development. 

The interactive visualization tool LocusZoom will integrate many of these methods to display variant associations and credible sets, epigenomic and functional annotations, and phenotype associations across a genomic region as well as offering custom association analysis.

An example LocusZoom plot

AMP T2D Knowledge Portal: democratizing T2D genetic results for researchers world-wide

AMP T2D was founded on the idea that in order to truly accelerate progress, genomic information must be freely accessible to all scientists and presented in a way that is understandable by a broad range of researchers working on T2D biology, not only by human geneticists and bioinformaticians with special computational skills. So the roadmap for the project included not only data generation and analysis, but also the production of a publicly available web resource that would integrate data types, interpret the evidence, and present of all these results. 

While it is under continuous development, mid-way through the initial funding period the T2D Knowledge Portal (T2DKP) is already a well-established resource. Other web resources collect genetic association data, but the T2DKP is unusual in providing harmonized datasets to which a consistent analysis pipeline has been applied. Rather than simply cataloging datasets, it offers distilled and synthesized results along with their interpretation, to guide more detailed exploration of the evidence. And, unlike any other extant resource, it offers researchers the ability to perform interactive queries on protected individual-level data. 

T2DKP home page

The Gene page of the T2DKP (see an example) illustrates the presentation of immediately understandable summary information along with the opportunity to drill down to the details. An algorithm considers the associations of all variants across a gene, for all phenotypes and in all datasets aggregated at the DCC, and calculates from them a “traffic light” signal for the gene: green to indicate that there is a significant association for at least one phenotype; yellow to indicate suggestive, if not highly significant, associations; and red to indicate that there is no evidence for association for any of the phenotypes considered in the T2DKP. Below this, tables and graphics invite users to explore all variants across the gene, their impacts on the encoded protein, and their associations, as well as their positions relative to epigenomic marks across the region in multiple tissues.

The T2DKP currently offers the ability to run custom, interactive association analyses using two different tools. In the LocusZoom visualization, users may choose one or more variants as covariates before performing association analysis. The Genetic Association Interactive Tool (GAIT) for single variant associations, which also powers the custom burden test for gene-level associations, is even more versatile, presenting the distributions of different characteristics of the sample set (age, sex, BMI, glycemic measures, blood lipid levels, and many more) and allowing users to filter the set by multiple criteria and to choose custom covariates before performing association analysis. Both of these tools allow analytical access to the individual-level data, whether housed at the Broad DCC or at the EBI federated node, in a secure environment so that data privacy is always protected.

Evolution of a collaborative environment

AMP T2D organization

The AMP T2D partnership is a multifaceted project (illustrated above) that embraces several aspects of basic research and combines them with building a product, the T2DKP. In connecting scientists both within and outside of consortia, in academia and in industry, working on genetic associations or functional studies, it is becoming the nexus of the T2D genetics community. Researchers are finding the T2DKP helpful for accessing even their own results and for viewing them in the context of multiple phenotypic associations and other complementary data types. Pharmaceutical partners are finding help via the Target Prioritization project, in which the tools and methods developed within AMP T2D are being used to prioritize a list of genes of mutual interest for further investigation.

Perhaps most importantly, AMP T2D has made researchers—both within and outside of the project—aware of the value of sharing data for representation in the context of all other relevant data. Only by compiling and interpreting all available information will we be able to make the best hypotheses about genes and pathways that are possible drug targets and prioritize them for in-depth functional investigation.

AMP T2D and beyond

In the remainder of the initial AMP T2D funding period, we expect continued progress in each of the areas discussed above. The data intake and analysis pipelines will be improved, and new data will be incorporated at an increasing pace—including data from the UK Biobank, which has generated association results for 500,000 genotyped subjects and more than 2,500 traits. Associations will be added for many more phenotypes related to T2D, including diabetic complications and longitudinal phenotype data that connect the development of various traits to the timeline of incident T2D.  Much more T2D-relevant epigenomic data will be available for query as well as for browsing, via dynamic connection with the Diabetes Epigenome Atlas. And entirely new data types (for example, metabolomic and proteomic data) arising from Opportunity Pool projects will be added to the T2DKP.

Ongoing work on tools and methods will result in the addition of many more interactive modules to the T2DKP. Researchers will be able to view PheWAS data; prune lists of variants by their linkage disequilibrium relationships; calculate credible sets and genetic risk scores with custom parameters; perform more versatile interactive burden tests; prioritize genes by pre-calculated association scores; overlay the positions of coding variants on protein structures to help assess their impact; and perform enrichment analysis on sets of loci to suggest pathways implicated in disease processes.

The Knowledge Portal platform developed for AMP T2D has already proved extensible to other complex diseases: in 2017, both the Cerebrovascular Disease and Cardiovascular Disease Knowledge Portals were launched. In the future, connections within the ecosystem formed by the T2D, Cerebrovascular, and Cardiovascular Portals will be improved, so that researchers can easily assess the impact of a variant or involvement of a gene for all of these related diseases. If funding and collaboration considerations allow, perhaps one day these Portals will merge into a single cardiometabolic disease genetics Knowledge Portal to accelerate the development of new therapeutics in this broader area.

Finally, the ultimate goal of this funding period is that by its end, the data generation, analysis, and interpretation will have facilitated the validation of multiple promising drug targets for further investigation. Given the rate of progress on multiple fronts, this seems a realistic goal. We hope that this unique collaborative environment will continue to accelerate T2D genetic research and will become a paradigm for other research communities.

Monday, April 9, 2018

Those hoofbeats just might come from zebras

Image by Eric Dietrich via Wikimedia Commons
A physician in the 1940s wanted to convey to his students that the most obvious diagnosis is most likely to be the correct one, so he coined a saying that has become famous: “When you hear hoofbeats, think of horses not zebras.” Applying this concept to complex disease genetics, if a risk-associated variant causes a non-synonymous mutation in a coding sequence, the first hypothesis to consider is that it affects disease risk by altering the protein. But although this is often the case, one of the lessons we can learn from a large new study, published today and now available for browsing and searching in the T2D Knowledge Portal, is that we should not forget about zebras.

The new study, from a global coalition of scientists (Mahajan et al., Nature Genetics 2018), is an exome-wide association study that surveyed the T2D associations of variants within the protein-coding regions of the genome. Including more than 81,000 T2D cases, over 370,000 controls, and multiple ancestries, this study has a three-fold larger effective sample size than any previous study. Using p-value < 2.2 x 10-7 as a threshold for significance across the exome, the authors found 69 significantly associated coding variants representing 40 distinct association signals in 38 loci—16 of which had not been previously associated with T2D risk.

To get a better idea of which variants in these loci were causal for T2D risk, the researchers performed fine mapping for 37 of the 40 significant signals. They meta-analyzed T2D associations for over 500,000 individuals of European descent, performed imputation, and then generated 99% credible sets for each signal—that is, sets of variants that are 99% likely to include the causal variant. To calculate the credible sets, they used an “annotation-informed prior” model of causality that took into account the distribution of associations for different variant impact classes and also the overlap of variants with putative enhancer elements.

The 37 association signals for which the authors generated credible sets were all due to coding variants that would cause changes in the sequence of the encoded protein. But surprisingly, the fine mapping analysis found that coding variants were likely to be causal for T2D risk at fewer than half of these loci.

One of these surprising results involves a gene that is well-known to be relevant to T2D: PPARG. Involvement of the PPARG protein in T2D is beyond doubt, since this ligand-inducible transcription factor is the target of thiazolidinedione drugs that are used to treat T2D. A common variant in PPARG, rs1801282, that causes a p.Pro12Ala change in the protein has been assumed to account for the T2D association, but there is little experimental evidence that this change affects PPARG function.

In the credible set generated in this study, the probability that rs1801282 is causal was not found to be particularly high. Included in this credible set along with rs1801282 are 19 non-coding variants. One of these was previously shown to affect a binding site for the transcription factor PRRX1 and to affect expression of PPARG2, a PPARG isoform. This suggests the intriguing possibility that the T2D risk in this locus is caused, partly or wholly, by variants affecting regulation rather than protein sequence.

A similar pattern, with partial causality due to non-coding variants, was seen at an additional 7 loci. And in 13 other loci, even though these loci were discovered via coding variant signals, non-coding variants had the highest probability of causing risk.

According to Professor Mark McCarthy of the University of Oxford, one of the principal investigators of the study, “Our study shows that we should not jump to conclusions when we see that one of our association signals includes a variant around which we can base an attractive mechanistic narrative. The “average” coding variant is more likely to be causal than the “average” noncoding variant, but even at the set of loci where we detect a significant coding variant association, it is as likely as not that the signal is driven instead by one of the non-coding variants nearby. By bringing together genetic and genomic data, we can improve our prospects for finding the causal variants at GWAS loci, but these should be the starting points for empirical studies not a destination in themselves.” Dr. McCarthy has written a commentary on this study; read it here.

So, in investigating complex disease genetics, it is still a good bet that a coding variant affects disease risk via altered protein sequence: at least in some parts of the world, hoofbeats are very often due to horses. But this study reminds us that it is always a good idea to look beyond the obvious hypothesis, and remember the zebras.

This paper includes many other discoveries, and we recommend that you read the paper to get the full story. We are pleased to announce that in addition to publishing the paper, the authors have made their results available to the T2D research community immediately upon publication, in the T2D Knowledge Portal.

The dataset in the T2DKP is named ExTexT2D (ExTended exome array genotyping for T2D) and includes associations for T2D, both unadjusted and adjusted for BMI. A description of the dataset along with a table listing the cohorts of the study subjects can be found on the Data page, and you can browse and search the ExTexT2D exome chip analysis dataset at these locations in the T2DKP:

On Gene pages (see an example) on the Common variants and High-impact variants tabs
On Variant pages (see an example) in the Associations at a glance section and the Association statistics across traits table
Via the Variant Finder search
View a Manhattan plot of associations across the genome by selecting “type 2 diabetes” or “type 2 diabetes adj BMI” in the View full genetic association results for a phenotype menu on the home page.

This dataset offers by far the largest sample size for exploring associations of low-frequency and common coding variants with T2D. The size of the study enabled evaluation of which coding variants mediate GWAS signals and which are simply "proxies" to the true causal variant, as revealed in the credible set analysis. With the addition of this dataset, the T2DKP offers in-depth information on two aspects of exome associations: common and low-frequency variant associations in ExTexT2D, and comprehensive coding variant associations in the 19K exome sequence analysis dataset (soon to include 50,000 exomes).

We are pleased to provide access to these important new results. Please contact us with any questions or comments about these new data or the T2DKP in general!