arXiv subjects

Matthew Stephens

Publications and source records attributed to Matthew Stephens.

34 records · Page 2Linked to original sources

arXiv2018

Empirical Bayes Shrinkage and False Discovery Rate Estimation, Allowing For Unwanted Variation

We combine two important ideas in the analysis of large-scale genomics experiments (e.g. experiments that aim to identify genes that are differentially expressed between two conditions). The first is use of Empirical Bayes (EB) methods to handle the large number of potentially-sparse effects, and estimate false discovery rates and related quantities. The second is use of factor analysis methods to deal with sources of unwanted variation such as batch effects and unmeasured confounders. We describe a simple modular fitting procedure that combines key ideas from both these lines of research. This yields new, powerful EB methods for analyzing genomics experiments that account for both sparse effects and unwanted variation. In realistic simulations, these new methods provide significant gains in power and calibration over competing methods. In real data analysis we find that different methods, while often conceptually similar, can vary widely in their assessments of statistical significance. This highlights the need for care in both choice of methods and interpretation of results. All methods introduced in this paper are implemented in the R package vicar available at https://github.com/dcgerard/vicar .

stat.ME↗

arXiv2017

varbvs: Fast Variable Selection for Large-scale Regression

We introduce varbvs, a suite of functions written in R and MATLAB for regression analysis of large-scale data sets using Bayesian variable selection methods. We have developed numerical optimization algorithms based on variational approximation methods that make it feasible to apply Bayesian variable selection to very large data sets. With a focus on examples from genome-wide association studies, we demonstrate that varbvs scales well to data sets with hundreds of thousands of variables and thousands of samples, and has features that facilitate rapid data analyses. Moreover, varbvs allows for extensive model customization, which can be used to incorporate external information into the analysis. We expect that the combination of an easy-to-use interface and robust, scalable algorithms for posterior computation will encourage broader use of Bayesian variable selection in areas of applied statistics and computational biology. The most recent R and MATLAB source code is available for download at Github (https://github.com/pcarbo/varbvs), and the R package can be installed from CRAN (https://cran.r-project.org/package=varbvs).

stat.CO↗

arXiv2017

Unifying and Generalizing Methods for Removing Unwanted Variation Based on Negative Controls

Unwanted variation, including hidden confounding, is a well-known problem in many fields, particularly large-scale gene expression studies. Recent proposals to use control genes --- genes assumed to be unassociated with the covariates of interest --- have led to new methods to deal with this problem. Going by the moniker Removing Unwanted Variation (RUV), there are many versions --- RUV1, RUV2, RUV4, RUVinv, RUVrinv, RUVfun. In this paper, we introduce a general framework, RUV*, that both unites and generalizes these approaches. This unifying framework helps clarify connections between existing methods. In particular we provide conditions under which RUV2 and RUV4 are equivalent. The RUV* framework also preserves an advantage of RUV approaches --- their modularity --- which facilitates the development of novel methods based on existing matrix imputation algorithms. We illustrate this by implementing RUVB, a version of RUV* based on Bayesian factor analysis. In realistic simulations based on real data we found that RUVB is competitive with existing methods in terms of both power and calibration, although we also highlight the challenges of providing consistently reliable calibration among data sets.

stat.ME↗

arXiv2014

An estimate of the average number of recessive lethal mutations carried by humans

The effects of inbreeding on human health depend critically on the number and severity of recessive, deleterious mutations carried by individuals. In humans, existing estimates of these quantities are based on comparisons between consanguineous and non-consanguineous couples, an approach that confounds socioeconomic and genetic effects of inbreeding. To circumvent this limitation, we focused on a founder population with almost complete Mendelian disease ascertainment and a known pedigree. By considering all recessive lethal diseases reported in the pedigree and simulating allele transmissions, we estimated that each haploid set of human autosomes carries on average 0.29 (95% credible interval [0.10, 0.83]) autosomal, recessive alleles that lead to complete sterility or severe disorders at birth or before reproductive age when homozygous. Comparison to existing estimates of the deleterious effects of all recessive alleles suggests that a substantial fraction of the burden of autosomal, recessive variants is due to single mutations that lead to death between birth and reproductive age. In turn, the comparison to estimates from other eukaryotes points to a surprising constancy of the average number of recessive lethal mutations across organisms with markedly different genome sizes.

q-bio.PE↗

arXiv2014

Bayesian methods for genetic association analysis with heterogeneous subgroups: From meta-analyses to gene-environment interactions

Genetic association analyses often involve data from multiple potentially-heterogeneous subgroups. The expected amount of heterogeneity can vary from modest (e.g., a typical meta-analysis) to large (e.g., a strong gene--environment interaction). However, existing statistical tools are limited in their ability to address such heterogeneity. Indeed, most genetic association meta-analyses use a "fixed effects" analysis, which assumes no heterogeneity. Here we develop and apply Bayesian association methods to address this problem. These methods are easy to apply (in the simplest case, requiring only a point estimate for the genetic effect and its standard error, from each subgroup) and effectively include standard frequentist meta-analysis methods, including the usual "fixed effects" analysis, as special cases. We apply these tools to two large genetic association studies: one a meta-analysis of genome-wide association studies from the Global Lipids consortium, and the second a cross-population analysis for expression quantitative trait loci (eQTLs). In the Global Lipids data we find, perhaps surprisingly, that effects are generally quite homogeneous across studies. In the eQTL study we find that eQTLs are generally shared among different continental groups, and discuss consequences of this for study design.

stat.ME↗

arXiv2013

Efficient Algorithms for Multivariate Linear Mixed Models in Genome-wide Association Studies

Multivariate linear mixed models (mvLMMs) have been widely used in many areas of genetics, and have attracted considerable recent interest in genome-wide association studies (GWASs). However, fitting mvLMMs is computationally non-trivial, and no existing method is computationally practical for performing the likelihood ratio test (LRT) for mvLMMs in GWAS settings with moderate sample size n. The existing software MTMM perform an approximate LRT for two phenotypes, and as we find, its p values can substantially understate the significance of associations. Here, we present novel computationally-efficient algorithms for fitting mvLMMs, and computing the LRT in GWAS settings. After a single initial eigen-decomposition (with complexity O(n^3)) the algorithms i) reduce computational complexity (per iteration of the optimizer) from cubic to linear in n; and ii) in GWAS analyses, reduces per-marker complexity from cubic to quadratic in n. These innovations make it practical to compute the LRT for mvLMMs in GWASs for tens of thousands of samples and a moderate number of phenotypes (~2-10). With simulations, we show that the LRT provides correct control for type I error. With both simulations and real data we find that the LRT is more powerful than the approximate LRT from MTMM, and illustrate the benefits of analyzing more than two phenotypes. The method is implemented in the GEMMA software package, freely available at http://stephenslab.uchicago.edu/software.html

q-bio.QM↗

arXiv2013

Wavelet-based genetic association analysis of functional phenotypes arising from high-throughput sequencing assays

Understanding how genetic variants influence cellular-level processes is an important step towards understanding how they influence important organismal-level traits, or "phenotypes", including human disease susceptibility. To this end scientists are undertaking large-scale genetic association studies that aim to identify genetic variants associated with molecular and cellular phenotypes, such as gene expression, transcription factor binding, or chromatin accessibility. These studies use high-throughput sequencing assays (e.g. RNA-seq, ChIP-seq, DNase-seq) to obtain high-resolution data on how the traits vary along the genome in each sample. However, typical association analyses fail to exploit these high-resolution measurements, instead aggregating the data at coarser resolutions, such as genes, or windows of fixed length. Here we develop and apply statistical methods that better exploit the high-resolution data. The key idea is to treat the sequence data as measuring an underlying "function" that varies along the genome, and then, building on wavelet-based methods for functional data analysis, test for association between genetic variants and the underlying function. Applying these methods to identify genetic variants associated with chromatin accessibility (dsQTLs) we find that they identify substantially more associations than a simpler window-based analysis, and in total we identify 772 novel dsQTLs not identified by the original analysis.

stat.ME↗

arXiv2013

A likelihood-based scoring method for peptide identification using mass spectrometry

Mass spectrometry provides a high-throughput approach to identify proteins in biological samples. A key step in the analysis of mass spectrometry data is to identify the peptide sequence that, most probably, gave rise to each observed spectrum. This is often tackled using a database search: each observed spectrum is compared against a large number of theoretical "expected" spectra predicted from candidate peptide sequences in a database, and the best match is identified using some heuristic scoring criterion. Here we provide a more principled, likelihood-based, scoring criterion for this problem. Specifically, we introduce a probabilistic model that allows one to assess, for each theoretical spectrum, the probability that it would produce the observed spectrum. This probabilistic model takes account of peak locations and intensities, in both observed and theoretical spectra, which enables incorporation of detailed knowledge of chemical plausibility in peptide identification. Besides placing peptide scoring on a sounder theoretical footing, the likelihood-based score also has important practical benefits: it provides natural measures for assessing the uncertainty of each identification, and in comparisons on benchmark data it produced more accurate peptide identifications than other methods, including SEQUEST. Although we focus here on peptide identification, our scoring rule could easily be integrated into any downstream analyses that require peptide-spectrum match scores.

stat.AP↗

arXiv2012

A statistical framework for joint eQTL analysis in multiple tissues

Mapping expression Quantitative Trait Loci (eQTLs) represents a powerful and widely-adopted approach to identifying putative regulatory variants and linking them to specific genes. Up to now eQTL studies have been conducted in a relatively narrow range of tissues or cell types. However, understanding the biology of organismal phenotypes will involve understanding regulation in multiple tissues, and ongoing studies are collecting eQTL data in dozens of cell types. Here we present a statistical framework for powerfully detecting eQTLs in multiple tissues or cell types (or, more generally, multiple subgroups). The framework explicitly models the potential for each eQTL to be active in some tissues and inactive in others. By modeling the sharing of active eQTLs among tissues this framework increases power to detect eQTLs that are present in more than one tissue compared with "tissue-by-tissue" analyses that examine each tissue separately. Conversely, by modeling the inactivity of eQTLs in some tissues, the framework allows the proportion of eQTLs shared across different tissues to be formally estimated as parameters of a model, addressing the difficulties of accounting for incomplete power when comparing overlaps of eQTLs identified by tissue-by-tissue analyses. Applying our framework to re-analyze data from transformed B cells, T cells and fibroblasts we find that it substantially increases power compared with tissue-by-tissue analysis, identifying 63% more genes with eQTLs (at FDR=0.05). Further the results suggest that, in contrast to previous analyses of the same data, the majority of eQTLs detectable in these data are shared among all three tissues.

q-bio.QM↗

arXiv2012

Small World MCMC with Tempering: Ergodicity and Spectral Gap

When sampling a multi-modal distribution $π(x)$, $x\in \rr^d$, a Markov chain with local proposals is often slowly mixing; while a Small-World sampler \citep{guankrone} -- a Markov chain that uses a mixture of local and long-range proposals -- is fast mixing. However, a Small-World sampler suffers from the curse of dimensionality because its spectral gap depends on the volume of each mode. We present a new sampler that combines tempering, Small-World sampling, and producing long-range proposals from samples in companion chains (e.g. Equi-Energy sampler). In its simplest form the sampler employs two Small-World chains: an exploring chain and a sampling chain. The exploring chain samples $π_t(x) \propto π(x)^{1/t}$, $t\in [1,\infty)$, and builds up an empirical distribution. Using this empirical distribution as its long-range proposal, the sampling chain is designed to have a stationary distribution $π(x)$. We prove ergodicity of the algorithm and study its convergence rate. We show that the spectral gap of the exploring chain is enlarged by a factor of $t^{d}$ and that of the sampling chain is shrunk by a factor of $t^{-d}$. Importantly, the spectral gap of the exploring chain depends on the "size" of $π_t(x)$ while that of sampling chain does not. Overall, the sampler enlarges a severe bottleneck at the cost of shrinking a mild one, hence achieves faster mixing. The penalty on the spectral gap of the sampling chain can be significantly alleviated when extending the algorithm to multiple chains whose temperatures $\{t_k\}$ follow a geometric progression. If we allow $t_k \rightarrow 0$, the sampler becomes a global optimizer.

stat.ME↗

arXiv2012

Polygenic Modeling with Bayesian Sparse Linear Mixed Models

Both linear mixed models (LMMs) and sparse regression models are widely used in genetics applications, including, recently, polygenic modeling in genome-wide association studies. These two approaches make very different assumptions, so are expected to perform well in different situations. However, in practice, for a given data set one typically does not know which assumptions will be more accurate. Motivated by this, we consider a hybrid of the two, which we refer to as a "Bayesian sparse linear mixed model" (BSLMM) that includes both these models as special cases. We address several key computational and statistical issues that arise when applying BSLMM, including appropriate prior specification for the hyper-parameters, and a novel Markov chain Monte Carlo algorithm for posterior inference. We apply BSLMM and compare it with other methods for two polygenic modeling applications: estimating the proportion of variance in phenotypes explained (PVE) by available genotypes, and phenotype (or breeding value) prediction. For PVE estimation, we demonstrate that BSLMM combines the advantages of both standard LMMs and sparse regression modeling. For phenotype prediction it considerably outperforms either of the other two methods, as well as several other large-scale regression methods previously suggested for this problem. Software implementing our method is freely available from http://stephenslab.uchicago.edu/software.html

q-bio.QM↗

arXiv2012

Integrated analysis of variants and pathways in genome-wide association studies using polygenic models of disease

Many common diseases are highly polygenic, modulated by a large number genetic factors with small effects on susceptibility to disease. These small effects are difficult to map reliably in genetic association studies. To address this problem, researchers have developed methods that aggregate information over sets of related genes, such as biological pathways, to identify gene sets that are enriched for genetic variants associated with disease. However, these methods fail to answer a key question: which genes and genetic variants are associated with disease risk? We develop a method based on sparse multiple regression that simultaneously identifies enriched pathways, and prioritizes the variants within these pathways, to locate additional variants associated with disease susceptibility. A central feature of our approach is an estimate of the strength of enrichment, which yields a coherent way to prioritize variants in enriched pathways. We illustrate the benefits of our approach in a genome-wide association study of Crohn's disease with ~440,000 genetic variants genotyped for ~4700 study subjects. We obtain strong support for enrichment of IL-12, IL-23 and other cytokine signaling pathways. Furthermore, prioritizing variants in these enriched pathways yields support for additional disease-association variants, all of which have been independently reported in other case-control studies for Crohn's disease.

q-bio.QM↗

arXiv2011

Bayesian variable selection regression for genome-wide association studies and other large-scale problems

We consider applying Bayesian Variable Selection Regression, or BVSR, to genome-wide association studies and similar large-scale regression problems. Currently, typical genome-wide association studies measure hundreds of thousands, or millions, of genetic variants (SNPs), in thousands or tens of thousands of individuals, and attempt to identify regions harboring SNPs that affect some phenotype or outcome of interest. This goal can naturally be cast as a variable selection regression problem, with the SNPs as the covariates in the regression. Characteristic features of genome-wide association studies include the following: (i) a focus primarily on identifying relevant variables, rather than on prediction; and (ii) many relevant covariates may have tiny effects, making it effectively impossible to confidently identify the complete "correct" subset of variables. Taken together, these factors put a premium on having interpretable measures of confidence for individual covariates being included in the model, which we argue is a strength of BVSR compared with alternatives such as penalized regression methods. Here we focus primarily on analysis of quantitative phenotypes, and on appropriate prior specification for BVSR in this setting, emphasizing the idea of considering what the priors imply about the total proportion of variance in outcome explained by relevant covariates. We also emphasize the potential for BVSR to estimate this proportion of variance explained, and hence shed light on the issue of "missing heritability" in genome-wide association studies.

stat.AP↗

arXiv2010

Using linear predictors to impute allele frequencies from summary or pooled genotype data

Recently-developed genotype imputation methods are a powerful tool for detecting untyped genetic variants that affect disease susceptibility in genetic association studies. However, existing imputation methods require individual-level genotype data, whereas, in practice, it is often the case that only summary data are available. For example, this may occur because, for reasons of privacy or politics, only summary data are made available to the research community at large; or because only summary data are collected, as in DNA pooling experiments. In this article we introduce a new statistical method that can accurately infer the frequencies of untyped genetic variants in these settings, and indeed substantially improve frequency estimates at typed variants in pooling experiments where observations are noisy. Our approach, which predicts each allele frequency using a linear combination of observed frequencies, is statistically straightforward, and related to a long history of the use of linear methods for estimating missing values (e.g., Kriging). The main statistical novelty is our approach to regularizing the covariance matrix estimates, and the resulting linear predictors, which is based on methods from population genetics. We find that, besides being both fast and flexible---allowing new problems to be tackled that cannot be handled by existing imputation approaches purpose-built for the genetic context---these linear methods are also very accurate. Indeed, imputation accuracy using this approach is similar to that obtained by state-of-the-art imputation methods that use individual-level data, but at a fraction of the computational cost.

stat.AP↗

arXiv2010

Statistical inference of transmission fidelity of DNA methylation patterns over somatic cell divisions in mammals

We develop Bayesian inference methods for a recently-emerging type of epigenetic data to study the transmission fidelity of DNA methylation patterns over cell divisions. The data consist of parent-daughter double-stranded DNA methylation patterns with each pattern coming from a single cell and represented as an unordered pair of binary strings. The data are technically difficult and time-consuming to collect, putting a premium on an efficient inference method. Our aim is to estimate rates for the maintenance and de novo methylation events that gave rise to the observed patterns, while accounting for measurement error. We model data at multiple sites jointly, thus using whole-strand information, and considerably reduce confounding between parameters. We also adopt a hierarchical structure that allows for variation in rates across sites without an explosion in the effective number of parameters. Our context-specific priors capture the expected stationarity, or near-stationarity, of the stochastic process that generated the data analyzed here. This expected stationarity is shown to greatly increase the precision of the estimation. Applying our model to a data set collected at the human FMR1 locus, we find that measurement errors, generally ignored in similar studies, occur at a nontrivial rate (inappropriate bisulfite conversion error: 1.6$%$ with 80$%$ CI: 0.9--2.3$%$). Accounting for these errors has a substantial impact on estimates of key biological parameters. The estimated average failure of maintenance rate and daughter de novo rate decline from 0.04 to 0.024 and from 0.14 to 0.07, respectively, when errors are accounted for. Our results also provide evidence that de novo events may occur on both parent and daughter strands: the median parent and daughter de novo rates are 0.08 (80$%$ CI: 0.04--0.13) and 0.07 (80$%$ CI: 0.04--0.11), respectively.

stat.AP↗

arXiv2010

A nested mixture model for protein identification using mass spectrometry

Mass spectrometry provides a high-throughput way to identify proteins in biological samples. In a typical experiment, proteins in a sample are first broken into their constituent peptides. The resulting mixture of peptides is then subjected to mass spectrometry, which generates thousands of spectra, each characteristic of its generating peptide. Here we consider the problem of inferring, from these spectra, which proteins and peptides are present in the sample. We develop a statistical approach to the problem, based on a nested mixture model. In contrast to commonly used two-stage approaches, this model provides a one-stage solution that simultaneously identifies which proteins are present, and which peptides are correctly identified. In this way our model incorporates the evidence feedback between proteins and their constituent peptides. Using simulated data and a yeast data set, we compare and contrast our method with existing widely used approaches (PeptideProphet/ProteinProphet) and with a recently published new approach, HSM. For peptide identification, our single-stage approach yields consistently more accurate results. For protein identification the methods have similar accuracy in most settings, although we exhibit some scenarios in which the existing methods perform poorly.

stat.AP↗