Discussion on the paper ‘Statistical contributions to bioinformatics: Design, modelling, structure learning and integration’ by Jeffrey S. Morris and Veerabhadran Baladandayuthapani

Abstract

Morris and Baladandayuthapani (M and B) discuss statistical contributions to bioinformatics in four different areas: design, modelling, structure learning, and integration. The authors not only manage to highlight many relevant and important contributions of statistics to the field of bioinformatics, but also illustrate cases where proper and rigorous statistical principles are not considered. We like to congratulate the authors with this very thorough and highly relevant piece of work. To complement the paper, we wish to highlight several areas of bioinformatics in which statisticians have also made significant contributions. Moreover we like to share our own experience in working with a novel type of omics, namely glycomics. We finish with a discussion about the challenges for statisticians to work in the field of statistical bioinformatics. Data Cleaning, Filtering and the Effect on Downstream AnalysisOur experience with the statistical analyses of Glycomics datasets measured in epidemiological studies motivated our consortium MIMOmics (www.MIMOmics.eu) which develops methods for integrated analysis of multiple omics datasets. Firstly we encountered that statistical analyses of Glycomics data was hampered by our lack of knowledge about high throughput technology and preprocessing steps which are taken before the data were given to us. Secondly straightforward association analysis of one omic variable with an outcome was the state of the art methodology which fails to identify multi omics risk profiles associated with complex traits. Also M and B discuss such elementwise modeling approaches (p 43). While elementwise approaches are straightforward to implement, they ignore the correlation structure among the elements which results in suboptimal testing and inefficient estimators. In our opinion it also reduces the insight in the complex biological mechanisms underlying the traits. Thirdly we wondered what the effect of data cleaning would be on downstream analysis and whether different statistical approaches needed different data cleaning procedures.

Choosing proper normalization is essential for discovery of sparse glycan biomarkers

“…In particular, for network analysis where the analysis is based 0.00 0.00 0.00 0.00 0.00 0.00 0.00 GP17 0.00 0.00 0.00 0.00 0.00 0.00 0.00 GP18 0.00 0.00 0.00 0.00 À0.63 0.00 0.00 GP19 0.00 0.00 0.00 0.00 0.00 0.00 0.00 GP20 0.00 0.00 0.00 0.00 0.00 0.00 0.00 GP21 0.00 0.00 0.00 0.00 0.00 0.00 0.00 GP22 0.00 0.00 0.00 0.00 0.00 0.00 0.00 GP23 0.00 0.00 0.00 0.00 0.00 0.00 0.00 GP24 0.00 0.48 9.39 0.00 0.00 on correlation structure, it is not at all clear how to perform analysis based on the TA normalized data. 32 With regard to biomarker discovery, it might be a better strategy to analyse glycans jointly. New groups of a few glycans, called derived traits, can be constructed, which represent groups of glycan structures that have similar structural and chemical properties.…”

Detection of suspicious interactions of spiking covariates in methylation data

“…Hence, it is important that the researcher decides, if possible detected suspicious interaction will violate the linearity assumption of the regression analysis or if the independent assumption of the bioinformatical pipeline can not hold. We would state, that a flipped or misleading effect is much more problematic than a lower statistical power [1,16].…”

“…Scientists using a linear regression model often ignore the properties of the independent variable or covariate. Especially, if the scientist is not aware of the use of a linear regression in differential expression analysis, because the regression analysis is hidden in the depths of *Correspondence: jochen.kruppa@charite.de 1 Charité -University Medicine, corporate member of Freie Universität Berlin, Humboldt-Universität zu Berlin, and Berlin Institute of Health, Institute of Biometry and Clinical Epidemiology, Charitéplatz 1, 10117 Berlin, Germany 2 Berlin Institute of Health (BIH), Anna-Louisa-Karsch-Strane 2, 10178 Berlin, Germany Full list of author information is available at the end of the article a bioinformatical pipeline [1]. A classical at first glance unsuspicious continuous covariate might be smoking in pack years.…”

Systematic Evaluation of Normalization Methods for Glycomics Data Based on Performance of Network Inference

“…This need for a glycomics-specific evaluation is further supported by the observation that the de facto standard for large-scale glycomics data preprocessing is Total Area (TA) normalization [ 8 ], which describes each glycan intensity in a sample as a percentage of the total. Following this transformation, the normalized intensities of a sample sum up to one (or 100%) by definition, leading to the loss of one degree of freedom.…”