Athina Spiliopoulou

Application of high-dimensional feature selection: evaluation for genomic prediction in man

In this study, we investigated the effect of five feature selection approaches on the performance of a mixed model (G-BLUP) and a Bayesian (Bayes C) prediction method. We predicted height, high density lipoprotein cholesterol (HDL) and body mass index (BMI) within 2,186 Croatian and into 810 UK individuals using genome-wide SNP data. Using all SNP information Bayes C and G-BLUP had similar predictive performance across all traits within the Croatian data, and for the highly polygenic traits height and BMI when predicting into the UK data. Bayes C outperformed G-BLUP in the prediction of HDL, which is influenced by loci of moderate size, in the UK data. Supervised feature selection of a SNP subset in the G-BLUP framework provided a flexible, generalisable and computationally efficient alternative to Bayes C; but careful evaluation of predictive performance is required when supervised feature selection has been used.

Comparing probabilistic models for melodic sequences

Modelling the real world complexity of music is a challenge for machine learning. We address the task of modeling melodic sequences fromthe same music genre. We perform a comparative analysis of two probabilistic models; a Dirichlet Variable Length Markov Model (Dirichlet-VMM) and a Time Convolutional Restricted Boltzmann Machine (TC-RBM). We show that the TC-RBM learns descriptive music features, such as underlying chords and typical melody transitions and dynamics. We assess the models for future prediction and compare their performance to a VMM, which is the current state of the art in melody generation. We show that both models perform significantly better than the VMM, with the Dirichlet-VMMmarginally outperforming the TC-RBM. Finally, we evaluate the short order statistics of the models, using the Kullback-Leibler divergence between test sequences and model samples, and show that our proposed methods match the statistics of the music genre significantly better than the VMM.

GeneImp: Fast Imputation to Large Reference Panels Using Genotype Likelihoods from Ultra-Low Coverage Sequencing

We address the task of genotype imputation to a dense reference panel given genotype likelihoods computed from ultralow coverage sequencing as inputs. In this setting, the data have a high-level of missingness or uncertainty, and are thus more amenable to a probabilistic representation. Most existing imputation algorithms are not well suited for this situation, as they rely on prephasing for computational efficiency, and, without definite genotype calls, the prephasing task becomes computationally expensive. We describe GeneImp, a program for genotype imputation that does not require prephasing and is computationally tractable for whole-genome imputation. GeneImp does not explicitly model recombination, instead it capitalizes on the existence of large reference panels—comprising thousands of reference haplotypes—and assumes that the reference haplotypes can adequately represent the target haplotypes over short regions unaltered. We validate GeneImp based on data from ultralow coverage sequencing (0.5×), and compare its performance to the most recent version of BEAGLE that can perform this task. We show that GeneImp achieves imputation quality very close to that of BEAGLE, using one to two orders of magnitude less time, without an increase in memory complexity. Therefore, GeneImp is the first practical choice for whole-genome imputation to a dense reference panel when prephasing cannot be applied, for instance, in datasets produced via ultralow coverage sequencing. A related future application for GeneImp is whole-genome imputation based on the off-target reads from deep whole-exome sequencing.

Genome-wide association study of response to methotrexate in early rheumatoid arthritis patients

Methotrexate (MTX) monotherapy is a common first treatment for rheumatoid arthritis (RA), but many patients do not respond adequately. In order to identify genetic predictors of response, we have combined data from two consortia to carry out a genome-wide study of response to MTX in 1424 early RA patients of European ancestry. Clinical endpoints were change from baseline to 6 months after starting treatment in swollen 28-joint count, tender 28-joint count, C-reactive protein and the overall 3-component disease activity score (DAS28). No single nucleotide polymorphism (SNP) reached genome-wide statistical significance for any outcome measure. The strongest evidence for association was with rs168201 in NRG3 (p = 10−7 for change in DAS28). Some support was also seen for association with ZMIZ1, previously highlighted in a study of response to MTX in juvenile idiopathic arthritis. Follow-up in two smaller cohorts of 429 and 177 RA patients did not support these findings, although these cohorts were more heterogeneous.

MR-PheWAS: exploring the causal effect of SUA level on multiple disease outcomes by using genetic instruments in UK Biobank

Objectives We aimed to investigate the role of serum uric acid (SUA) level in a broad spectrum of disease outcomes using data for 120 091 individuals from UK Biobank. Methods We performed a phenome-wide association study (PheWAS) to identify disease outcomes associated with SUA genetic risk loci. We then implemented conventional Mendelianrandomisation (MR) analysis to investigate the causal relevance between SUA level and disease outcomes identified from PheWAS. We next applied MR Egger analysis to detect and account for potential pleiotropy, which conventional MR analysis might mistake for causality, and used the HEIDI (heterogeneity in dependent instruments) test to remove cross-phenotype associations that were likely due to genetic linkage. Results Our PheWAS identified 25 disease groups/outcomes associated with SUA genetic risk loci after multiple testing correction (P<8.57e-05). Our conventional MR analysis implicated a causal role of SUA level in three disease groups: inflammatory polyarthropathies (OR=1.22, 95% CI 1.11 to 1.34), hypertensive disease (OR=1.08, 95% CI 1.03 to 1.14) and disorders of metabolism (OR=1.07, 95% CI 1.01 to 1.14); and four disease outcomes: gout (OR=4.88, 95% CI 3.91 to 6.09), essential hypertension (OR=1.08, 95% CI 1.03 to 1.14), myocardial infarction (OR=1.16, 95% CI 1.03 to 1.30) and coeliac disease (OR=1.41, 95% CI 1.05 to 1.89). After balancing pleiotropic effects in MR Egger analysis, only gout and its encompassing disease group of inflammatory polyarthropathies were considered to be causally associated with SUA level. Our analysis highlighted a locus (ATXN2/S2HB3) that may influence SUA level and multiple cardiovascular and autoimmune diseases via pleiotropy. Conclusions Elevated SUA level is convincing to cause gout and inflammatory polyarthropathies, and might act as a marker for the wider range of diseases with which it associates. Our findings support further investigation on the clinical relevance of SUA level with cardiovascular, metabolic, autoimmune and respiratory diseases.

Genome-wide association study of response to tumour necrosis factor inhibitor therapy in rheumatoid arthritis

Rheumatoid arthritis (RA) is characterised by chronic synovial joint inflammation. Treatment has been revolutionised by tumour necrosis factor alpha inhibitors (TNFi) but each available drug shows a significant non-response rate. We conducted a genome-wide association study of 1752 UK RA TNFi-treated patients to identify predictors of change in the Disease Activity Score 28 (DAS28) and subcomponents over 3–6 months. The rs7195994 variant at the FTO gene locus was associated with infliximab response when looking at a change in the swollen joint count (SJC28) subcomponent (p = 9.74 × 10−9). Capture Hi-C data show chromatin interactions in GM12878 cells between rs2540767, in high linkage disequilibrium with rs7195994 (R2 = 0.9) and IRX3, a neighbouring gene of FTO. IRX3 encodes a transcription factor involved in adipocyte remodelling and is regarded as the obesity gene at the FTO locus. Importantly, the rs7195994 association remained significantly associated following adjustment for BMI. In addition, using capture Hi-C data we showed interactions between TNFi-response associated variants and 16 RA susceptibility variants.

Reply to ‘Misestimation of heritability and prediction accuracy of male-pattern baldness’

Yap et al.1 present two criticisms of our recent analysis2 of the genetic architecture of male pattern baldness (MPB)2. First they note our earlier study in ref.2 overestimated SNP heritability (hereafter heritability) by excluding people in category two on the UK Biobank scale. We agree, heritability should have been reported as 0.64 not 0.94. This arose for a natural, if mistaken for this purpose, desire to categorize subjects clearly on a binary scale, and thus exclude indeterminate subjects. Their second criticism is that we overestimated the proportion of heritability explained by the 71 loci that have been identified. We stand by the broad conclusion that about one-third of the genetic effects (on a baldness trait dichotomized as category 1 versus 2, 3, or 4) are explained by the 71-locus SNP score. In principle, the proportion of polygenic variance explained by the SNP score can be evaluated in the following three possible ways: (1) as the ratio of the phenotypic variance explained by the SNP score to the variance explained by polygenic effects; (2) as the ratio between the heritability due to the SNPs and the baseline heritability estimate; or (3) as ratio of the reduction in polygenic variance in a model that includes the SNP score to the polygenic variance in a model that does not include this fixed effect (see Supplementary Method for details). These three methods should give the same result if the residual variance and the phenotypic variance do not change between the models with and without the SNPs. Yap et al. have used the first method, and estimate that the 107 SNPs from 71 loci explain about 15–20% of variation in total liability. Our own estimate using the same method is 20% on the liability scale, close to theirs, implying that about 31% of the total heritability of 0.61 is explained by the SNP score. Our article, however, reported an estimate by method (2), in which the ratio of the difference in heritability in models including and excluding the SNP to the baseline heritability was 38%. Including category two did not change this estimate (Supplementary Table 1 method (2)). Of course, to evaluate predictive performance requires an independent test dataset, beyond the scope of both our original study2 and the correspondence1. GCTA implements a mixed linear model and therefore estimates phenotypic variance from the variances of the random effects in the model. Therefore, the estimated phenotypic variances from models with different fixed effects (i.e., with and without the SNP predictor) are different. We thus applied method (3) which is not affected by the same issue as it works on the absolute and not relative scale (see Supplementary Method) and gives an even greater estimate of 45%. Given the limitations of fitting mixed linear models to a binary trait to estimate the parameter of interest, it is not easy to be certain which is the best one. However, irrespective of which is used, our conclusion that we can explain a relatively large proportion of heritability using SNPs from only 71 loci is still valid. Having now corrected the error in the estimation of heritability, we thus believe that the remainder of the results and conclusions are still valid, including in particular that we can explain a large proportion of the genetic variance using a relatively small number of SNPs. Furthermore, our identification and replication of several new loci for MBP remains accurate and DOI: 10.1038/s41467-018-04808-2 OPEN

Author Correction: GWAS for male-pattern baldness identifies 71 susceptibility loci explaining 38% of the risk

We have been alerted that in our recent Article the calculations used to transform the heritability from the observed scale to the liability scale did not take into account the individuals in category 2 of the baldness scale, who were removed in our original analysis. This led to an overestimation of the heritability on the liability scale, which should have been 0.62 instead of 0.94. Moreover, in the Title and in the Abstract, we report that we can explain 38% of the risk, while in fact that is the proportion of heritability explained by the loci we discovered. These errors do not substantially change the paper or its conclusions apart from the statement MBP is therefore probably one of the most heritable complex traits. Genome-wide significant associations and pathway analyses are not affected in any way and male-pattern baldness remains less genetically complex than other complex traits. We wish to thank Yap et al. for bringing this to our attention.

Prediction of treatment response in rheumatoid arthritis patients using genome-wide SNP data

Although a number of treatments are available for rheumatoid arthritis (RA), each of them shows a significant nonresponse rate in patients. Therefore, predicting a priori the likelihood of treatment response would be of great patient benefit. Here, we conducted a comparison of a variety of statistical methods for predicting three measures of treatment response, between baseline and 3 or 6 months, using genome‐wide SNP data from RA patients available from the MAximising Therapeutic Utility in Rheumatoid Arthritis (MATURA) consortium. Two different treatments and 11 different statistical methods were evaluated. We used 10‐fold cross validation to assess predictive performance, with nested 10‐fold cross validation used to tune the model hyperparameters when required. Overall, we found that SNPs added very little prediction information to that obtained using clinical characteristics only, such as baseline trait value. This observation can be explained by the lack of strong genetic effects and the relatively small sample sizes available; in analysis of simulated and real data, with larger effects and/or larger sample sizes, prediction performance was much improved. Overall, methods that were consistent with the genetic architecture of the trait were able to achieve better predictive ability than methods that were not. For treatment response in RA, methods that assumed a complex underlying genetic architecture achieved slightly better prediction performance than methods that assumed a simplified genetic architecture.