Volume 21, Issue 1

RNA-protein interactions have long being recognised as crucial regulators of gene expression. Recently, the development of scalable experimental techniques to measure these interactions has revolutionised the field, leading to the production of large-scale datasets which offer both opportunities and challenges for machine learning techniques. In this brief note, we will discuss some of the major stumbling blocks towards the use of machine learning in computational RNA biology, focusing specifically on the problem of predicting RNA-protein interactions from next-generation sequencing data.

Correlation analysis is widely used in biological studies to infer molecular relationships within biological networks. Recently, single-cell analysis has drawn tremendous interests, for its ability to obtain high-resolution molecular phenotypes. It turns out that there is little overlap of co-expressed genes identified in single-cell level investigations with that of population level investigations. However, the nature of the relationship of correlations between single-cell and population levels remains unclear. In this manuscript, we aimed to unveil the origin of the differences between the correlation coefficients at the single-cell level and that at the population level, and bridge the gap between them. Through developing formulations to link correlations at the single-cell and the population level, we illustrated that aggregated correlations could be stronger, weaker or equal to the corresponding individual correlations, depending on the variations and the correlations within the population. When the correlation within the population is weaker than the individual correlation, the aggregated correlation is stronger than the corresponding individual correlation. Besides, our data indicated that aggregated correlation is more likely to be stronger than the corresponding individual correlation, and it was rare to find gene-pairs exclusively strongly correlated at the single-cell level. Through a bottom-up approach to model interactions between molecules in a signaling cascade or a multi-regulator-controlled gene expression, we surprisingly found that the existence of interaction between two components could not be excluded simply based on their low correlation coefficients, suggesting a reconsideration of connectivity within biological networks which was derived solely from correlation analysis. We also investigated the impact of technical random measurement errors on the correlation coefficients for the single-cell level and the population level. The results indicate that the aggregated correlation is relatively robust and less affected. Because of the heterogeneity among single cells, correlation coefficients calculated based on data of the single-cell level might be different from that of the population level. Depending on the specific question we are asking, proper sampling and normalization procedure should be done before we draw any conclusions.

B atch effect R eduction of m I croarray data with D ependent samples usin G E mpirical Bayes ( BRIDGE ) is a recently developed statistical method to address the issue of batch effect correction in batch-confounded microarray studies with dependent samples. The key component of the BRIDGE methodology is the use of samples run as technical replicates in two or more batches, “bridging samples”, to inform batch effect correction/attenuation. While previously published results indicate a relationship between the number of bridging samples, M , and the statistical power of downstream statistical testing on the batch-corrected data, there is of yet no formal statistical framework or user-friendly software, for estimating M to achieve a specific statistical power for hypothesis tests conducted on the batch-corrected data. To fill this gap, we developed pwrBRIDGE , a simulation-based approach to estimate the bridging sample size, M , in batch-confounded longitudinal microarray studies. To illustrate the use of pwrBRIDGE , we consider a hypothetical, longitudinal batch-confounded study whose goal is to identify Alzheimer’s disease (AD) progression-associated genes from amnestic mild cognitive impairment (aMCI) to AD in human blood after a 5-year follow-up. pwrBRIDGE helps researchers design and plan batch-confounded microarray studies with dependent samples to avoid over- or under-powered studies.

Gene selection is one of the key steps for gene expression data analysis. An SVM-based ensemble feature selection method is proposed in this paper. Firstly, the method builds many subsets by using Monte Carlo sampling. Secondly, ranking all the features on each of the subsets and integrating them to obtain a final ranking list. Finally, the optimum feature set is determined by a backward feature elimination strategy. This method is applied to the analysis of 4 public datasets: the Leukemia, Prostate, Colorectal, and SMK_CAN, resulting 7, 10, 13, and 32 features. The AUC obtained from independent test sets are 0.9867, 0.9796, 0.9571, and 0.9575, respectively. These results indicate that the features selected by the proposed method can improve sample classification accuracy, and thus be effective for gene selection from gene expression data.

Due to the advancement of genome sequencing techniques, a great stride has been made in exome sequencing such that the association study between disease and genetic variants has become feasible. Some powerful and well-known association tests have been proposed to test the association between a group of genes and the disease of interest. However, some challenges still remain, in particular, many factors can affect the performance of testing power, e.g., the sample size, the number of causal and non-causal variants, and direction of the effect of causal variants. Recently, a powerful test, called T REM , is derived based on a random effects model. T REM has the advantages of being less sensitive to the inclusion of non-causal rare variants or low effect common variants or the presence of missing genotypes. However, the testing power of T REM can be low when a portion of causal variants has effects in opposite directions. To improve the drawback of T REM , we propose a novel test, called T ROB , which keeps the advantages of T REM and is more robust than T REM in terms of having adequate power in the case of variants with opposite directions of effect. Simulation results show that T ROB has a stable type I error rate and outperforms T REM when the proportion of risk variants decreases to a certain level and its advantage over T REM increases as the proportion decreases. Furthermore, T ROB outperforms several other competing tests in most scenarios. The proposed methodology is illustrated using the Shanghai Breast Cancer Study.

We propose two new statistics, V ̂ $\hat{V}$ and S ̂ $\hat{S}$ , to disentangle the population history of related populations from SNP frequency data. If the populations are related by a tree, we show by theoretical means as well as by simulation that the new statistics are able to identify the root of a tree correctly, in contrast to standard statistics, such as the observed matrix of F 2 -statistics (distances between pairs of populations). The statistic V ̂ $\hat{V}$ is obtained by averaging over all SNPs (similar to standard statistics). Its expectation is the true covariance matrix of the observed population SNP frequencies, offset by a matrix with identical entries. In contrast, the statistic S ̂ $\hat{S}$ is put in a Bayesian context and is obtained by averaging over pairs of SNPs, such that each SNP is only used once. It thus makes use of the joint distribution of pairs of SNPs. In addition, we provide a number of novel mathematical results about old and new statistics, and their mutual relationship.

Due to many advantages such as higher statistical power of detecting the association of genetic variants in human disorders and cost saving, extreme phenotype sequencing (EPS) is a rapidly emerging study design in epidemiological and clinical studies investigating how genetic variations associate with complex phenotypes. However, the investigation of the mediation effect of genetic variants on phenotypes is strictly restrictive under the EPS design because existing methods cannot well accommodate the non-random extreme tails sampling process incurred by the EPS design. In this paper, we propose a likelihood approach for testing the mediation effect of genetic variants through continuous and binary mediators on a continuous phenotype under the EPS design (GMEPS). Besides implementing in EPS design, it can also be utilized as a general mediation analysis procedure. Extensive simulations and two real data applications of a genome-wide association study of benign ethnic neutropenia under EPS design and a candidate-gene study of neurocognitive performance in patients with sickle cell disease under random sampling design demonstrate the superiority of GMEPS under the EPS design over widely used mediation analysis procedures, while demonstrating compatible capabilities under the general random sampling framework.

Association of phenotypes or exposures with genomic and epigenomic data faces important statistical challenges. One of these challenges is to account for variation due to unobserved confounding factors, such as individual ancestry or cell-type composition in tissues. This issue can be addressed with penalized latent factor regression models, where penalties are introduced to cope with high dimension in the data. If a relatively small proportion of genomic or epigenomic markers correlate with the variable of interest, sparsity penalties may help to capture the relevant associations, but the improvement over non-sparse approaches has not been fully evaluated yet. Here, we present least-squares algorithms that jointly estimate effect sizes and confounding factors in sparse latent factor regression models. In simulated data, sparse latent factor regression models generally achieved higher statistical performance than other sparse methods, including the least absolute shrinkage and selection operator and a Bayesian sparse linear mixed model. In generative model simulations, statistical performance was slightly lower (while being comparable) to non-sparse methods, but in simulations based on empirical data, sparse latent factor regression models were more robust to departure from the model than the non-sparse approaches. We applied sparse latent factor regression models to a genome-wide association study of a flowering trait for the plant Arabidopsis thaliana and to an epigenome-wide association study of smoking status in pregnant women. For both applications, sparse latent factor regression models facilitated the estimation of non-null effect sizes while overcoming multiple testing issues. The results were not only consistent with previous discoveries, but they also pinpointed new genes with functional annotations relevant to each application.