Clinical molecular research and biomarker development rely on a high level of data quality. Ensuring data quality extends beyond the establishment of reproducible technical processes involved in measurement of variables. Obtaining accurate clinical metadata is of utmost importance for meaningful clinical research, as they are necessary for finding clinical disease-treatment or disease-biomarkers relationships [1]. Drawing conclusions based on incorrect metadata can have detrimental consequences in short-term or long-term patient care. Typical sample annotation errors may be due to sample mix-ups, database entry errors, or subjectivity, e.g. grading of a biopsy. In pharmacogenomics analyses, unrecognized annotation errors or sample mix-ups impact any supervised statistical analysis, such as certain steps during biomarker discovery and qualification, and patient stratification. Estimates of sample mix-ups or annotation errors in clinical datasets range up to 18% [2]. Several statistical approaches have been devised to address annotation uncertainty during classifier development [3–5]. These sophisticated and extremely valuable approaches are applied as part of a later phase the analytical process. To begin addressing annotation uncertainty already at the database level, we have recently reported on a transcript marker for gender annotation which can be applied to clinical datasets of whole genome microarrays immediately after the generation of data [6]. The so-called “REDKX” gender marker is based on heterosome genes with gender-dependent expression-characteristic. Application of this marker informs about the correctness of the gender annotation of a donor. In addition, control for inter-individual sample mix-up clinical datasets with multiple, e.g. longitudinal, sampling is an absolute must, and gender annotation QC is not sufficient to that goal, since sample mix-ups between individuals of the same gender would go undetected.

Of particular interest are genes with largely unchanged expression levels in samples of a donor, which would have a significantly different expression level in another donor. At best, those probe sets would display an inter-individual bimodal “on/off”-expression characteristic. Bimodal distribution of (signal intensity) data is a deviation from the assumed normal distribution of measurements within a population, and can be recognized by the presence of two modes, each characterized by a peak. Bimodality can occur by differential expression, heterosome-specific expression, or, as observed e.g. in cancer biology, by genomic lesions in some donors but not in others [7]. Candidate genes with bimodal expression characteristics exhibit heterosome-specific and/or haplotype-specific expression. Some Affymetrix probe sets were designed in polymorphic mRNA sequences and are (unintentionally) haplotype specific. Messenger RNA (mRNA) derived from a donor with a certain haplotype or polymorphism does not hybridize with the same affinity to a locus of different haplotype or lacking the polymorphism, resulting in an extremely different expression level for the latter sample. The reason for the decreased or increased affinity is of technical nature: hybridization of a labelled mRNA fragment to a site with a single nucleotide polymorphism (SNP) causes nucleotide mispairing, leading to the creation of a virtual bubble at the site of the SNP. The result is a less stringent binding of the mRNA fragment to the probe target, and lower signal intensity. Hence, sites with SNPs could be used for indirect genotyping, or “fingerprinting” [8, 9].

In order to systematically assess deviation from unimodality of individual probe sets we have utilized the principle of the dip statistic. The dip test had been proposed as a test statistic for unimodality [10]. It estimates the maximum difference between the empirical distribution function and the unimodal distribution function that minimizes that maximum difference. In order to assess the significance of the Dip Test statistic, we ran 1 × 10⁶ simulations, where a sample of the same size as the dataset in question, was drawn from a normal distribution, and a Dip Test applied to each of the draws in order to generate 1 × 10⁶ simulated Dip Test statistics. As a result, we were able to derive an empirical p-value for each probe set. Here we report on the application of the Hartigans’ dip test to transcriptome wide clinical gene expression microarray datasets, in the search of probe sets which would help flag samples with potential donor-ID annotation mix-up.

Notwithstanding technical precision, annotation precision is a major component contributing to high data quality. In an effort to standardize annotation, metadata databases have been developed which provide user guidance by implementation of controlled vocabulary [13]. However, Quantile-quantile plot of Hartigans’ dip test statisticshuman error and subjectivity is still a common source of incorrectness in such databases [1]. Analysis errors and wrong conclusions with possibly detrimental consequences can be the result if annotation errors remain undetected [4]. Current strategies of dealing with annotation errors include statistical tests at a relatively late stage of the analysis. We hypothesized that transcript markers could aid in improving detection of annotation precision at an early stage, at best before the analysis.

Our approach was to investigate, design, develop and implement quality control tools to qualify microarray data in the context of the donor sample. As an initial step we have recently developed and qualified the so-called REDKX marker which is a transcript panel marker based on expression of genes located on heterosomes, indicating gender of sample donors in clinical microarray studies [6]. However, in studies with multiple samples from donors, correct gender annotation still leaves an uncertainty about the correct assignment of a sample to a donor, as samples with the same gender annotation may come from different subjects of that gender. Hence, we devised a second annotation transcript quality control marker, which would increase the detectability of samples with donor identification mislabels.

By applying the Hartigans’ Dip Test statistic we identified probe sets with bimodal expression pattern. Of particular interest are probe sets which show allele-selectivity, such as probe sets for histocompatibility genes HLA-DQA1 and HLA-DRB4. Both genes are located on the p21 arm of chromosome 6, and could be in a linkage disequilibrium region. The markers do not represent expression quantitative trait locus (eQTL) genes, and are expressed independent of gender. HLA genes code for cell surface proteins which are expressed by antigen presenting cells and in the immune system serve the purpose of self- vs. nonself-discrimination [14]. Hence, HLA gene expression patterns represent highly specific “fingerprinting” of individual donors. In general, our results receive support in the independent findings of Joehanes et al. who have suggested that gene expression levels, including those of HLA genes, could serve as “fingerprinting” data in microarray datasets [15]. The authors came to this conclusion as part of their assessment of gene expression analysis from different blood-derived RNA sources. Here, we identify a candidate marker consisting of HLA genes in an unsupervised analysis and qualify it by training and testing its performance on a number of publically available datasets.

HLA genes are essential contributors to susceptibility to risk or resistance to several autoimmune diseases such as multiple sclerosis, rheumatoid arthritis, and Type 1 Diabetes [16–18]. As sequencing of the HLA locus continues and more alleles are being identified, the number of allele-disease associations will grow [19–22]. For our exploratory marker analysis we applied expression data for three HLA probe sets. As a consequence of the extremely high multiplexing of parameters interrogated by a microarray it is not surprising that the behaviour of a single transcript in a large transcript population is not always 100% predictable and is the result of multiple factors such as assay and array performance. A three-digit sample identifier score yields 8 different score types, not sufficient to completely discriminate all donors in studies with 9 or more donors. However, as shown in the case of donor 32 in GSE67511, only the combination of the REDKX gender marker and the score has the potential to further reduce sample ID ambiguities by eliciting follow-up investigations should discrepancies between scores within a sample collection of a single donor arise. In the present analysis, we found that about 5% of the publically available samples used herein rewarded follow-up investigations, where half of those cases could not be resolved by interrogating the data alone. Despite our encouraging results from public datasets of 4 tissue types, we recommend the optimal intensity thresholds be developed related to each tissue of interest. The reason for this suggestion lies in the tissue dependent expression level variations as well as in technical dissimilarities, which may lead to deviations in global signal intensity, e.g. by different scaling settings.

Context-dependent biomarker qualification is driven by the application of the marker. The proposed genomic fit-for purpose biomarker approach for quality control of sample ID annotations in transcriptomics clinical datasets with multiple samples per donor could be applied immediately. As part of our approach to this end, we have implemented the haplotype-based quality control as part of our microarray quality control pipeline. Microarray data files are automatically reviewed by the software after they are produced, with the analysis results cached and made available through a web interface, where likely problems are highlighted for further investigation.

In case of detected mismatches to reported metadata, a root-cause analysis will be necessary to determine the reason for the error. If the cause cannot be determined in too many cases or is systematic, the decision could be not to use the entire dataset [1].

In conclusion, we have identified a set of transcripts, which, particularly in combination with the REDKX gender marker, is capable that can be used as a starting point to control for sample ID annotation errors in clinical datasets with multiple samples per donor. In publically available datasets we have identified about 3% of unresolvable annotation errors. Thus we recommend applying the marker broadly in transcriptomics studies and to follow-up with root cause analysis where necessary. Direct sequencing will provide further confirmation and possible expansion of the marker panel.

Hartigans’ Dip Test statistic is able to robustly identify probe sets with non-unimodal expression patterns, namely 203290_at (HLA-DQA1*0401), 213831_at (HLA-DQA1*0103), and 209728_at (HLA-DRB4). Combination of the so-called HLA-score and the REDKX marker panels provides a useful molecular quality control metric for sample annotation in clinical transcriptomics studies. Biological or clinical interpretation of gene expression data should take the haplotype specificity of these probe sets into account.