Identification of an in vitro transcription-based artifact affecting oligonucleotide microarrays

Identification of artifactual microarray probe signals constantly present in multiple sample types

The detection, identification, and quantitation of transcripts have evolved from simple Northern analysis, cDNA cloning, and sequencing to RT-PCR, microarrays, and now digital gene expression using ultra-high-throughput RNA sequencing (RNA-Seq). During the course of our studies we observed that some microarray probes show very high signal intensity values yet are discordant when compared with RNA-Seq. A total of 99 probes from approximately 30,000 were identified as consistently discordant in four human tissues or cell lines. Interestingly, this set of discordant probes appears array-dependent. Among the 99 probes identified, 70 constantly exhibited a high signal in all 713 available samples surveyed using the Illumina HumanHT-12v4 platform. Some were discordant with additional probes that annotated the same genes. Absence of a number of these transcripts was confirmed by quantitative RT-PCR (qRT-PCR). Our findings suggest that one must be cautious, as some array probes do not capture the level of the target.

Evaluation of alternative RNA labeling protocols for transcript profiling with Arabidopsis AGRONOMICS1 tiling arrays

Microarrays are routine tools for transcript profiling, and genomic tiling arrays such as the Arabidopsis AGRONOMICS1 arrays have been found to be highly suitable for such experiments because changes in genome annotation can be easily integrated at the data analysis level. In a transcript profiling experiment, RNA labeling is a critical step, most often initiated by oligo-dT-primed reverse transcription. Although this has been found to be a robust and reliable method, very long transcripts or non-polyadenylated transcripts might be labeled inefficiently. In this study, we first provide data handling methods to analyze AGRONOMICS1 tiling microarrays based on the TAIR10 genome annotation. Second, we describe methods to easily quantify antisense transcripts on such tiling arrays. Third, we test a random-primed RNA labeling method, and find that on AGRONOMICS1 arrays this method has similar general performance as the conventional oligo-dT-primed method. In contrast to the latter, however, the former works considerably better for long transcripts and for non-polyadenylated transcripts such as found in mitochondria and plastids. We propose that researchers interested in organelle function use the random-primed method to unleash the full potential of genomic tiling arrays.

Analysis of antisense expression by whole genome tiling microarrays and siRNAs suggests mis-annotation of Arabidopsis orphan protein-coding genes

Background

MicroRNAs (miRNAs) and trans-acting small-interfering RNAs (tasi-RNAs) are small (20–22 nt long) RNAs (smRNAs) generated from hairpin secondary structures or antisense transcripts, respectively, that regulate gene expression by Watson-Crick pairing to a target mRNA and altering expression by mechanisms related to RNA interference. The high sequence homology of plant miRNAs to their targets has been the mainstay of miRNA prediction algorithms, which are limited in their predictive power for other kingdoms because miRNA complementarity is less conserved yet transitive processes (production of antisense smRNAs) are active in eukaryotes. We hypothesize that antisense transcription and associated smRNAs are biomarkers which can be computationally modeled for gene discovery.

Principal Findings

We explored rice (Oryza sativa) sense and antisense gene expression in publicly available whole genome tiling array transcriptome data and sequenced smRNA libraries (as well as C. elegans) and found evidence of transitivity of MIRNA genes similar to that found in Arabidopsis. Statistical analysis of antisense transcript abundances, presence of antisense ESTs, and association with smRNAs suggests several hundred Arabidopsis ‘orphan’ hypothetical genes are non-coding RNAs. Consistent with this hypothesis, we found novel Arabidopsis homologues of some MIRNA genes on the antisense strand of previously annotated protein-coding genes. A Support Vector Machine (SVM) was applied using thermodynamic energy of binding plus novel expression features of sense/antisense transcription topology and siRNA abundances to build a prediction model of miRNA targets. The SVM when trained on targets could predict the “ancient” (deeply conserved) class of validated Arabidopsis MIRNA genes with an accuracy of 84%, and 76% for “new” rapidly-evolving MIRNA genes.

Conclusions

Antisense and smRNA expression features and computational methods may identify novel MIRNA genes and other non-coding RNAs in plants and potentially other kingdoms, which can provide insight into antisense transcription, miRNA evolution, and post-transcriptional gene regulation.

Mismatch oligonucleotides in human and yeast: guidelines for probe design on tiling microarrays

Background

Mismatched oligonucleotides are widely used on microarrays to differentiate specific from nonspecific hybridization. While many experiments rely on such oligos, the hybridization behavior of various degrees of mismatch (MM) structure has not been extensively studied. Here, we present the results of two large-scale microarray experiments on S. cerevisiae and H. sapiens genomic DNA, to explore MM oligonucleotide behavior with real sample mixtures under tiling-array conditions.

Results

We examined all possible nucleotide substitutions at the central position of 36-nucleotide probes, and found that nonspecific binding by MM oligos depends upon the individual nucleotide substitutions they incorporate: C→A, C→G and T→A (yielding purine-purine mispairs) are most disruptive, whereas A→X were least disruptive. We also quantify a marked GC skew effect: substitutions raising probe GC content exhibit higher intensity (and vice versa). This skew is small in highly-expressed regions (± 0.5% of total intensity range) and large (± 2% or more) elsewhere. Multiple mismatches per oligo are largely additive in effect: each MM added in a distributed fashion causes an additional 21% intensity drop relative to PM, three-fold more disruptive than adding adjacent mispairs (7% drop per MM).

Conclusion

We investigate several parameters for oligonucleotide design, including the effects of each central nucleotide substitution on array signal intensity and of multiple MM per oligo. To avoid GC skew, individual substitutions should not alter probe GC content. RNA sample mixture complexity may increase the amount of nonspecific hybridization, magnify GC skew and boost the intensity of MM oligos at all levels.