A Leveraged Signal-to-Noise Ratio (LSTNR) Method to Extract Differentially Expressed Genes and Multivariate Patterns of Expression From Noisy and Low-Replication RNAseq Data

Abstract

To life scientists, one important feature offered by RNAseq, a next-generation sequencing tool used to estimate changes in gene expression levels, lies in its unprecedented resolution. It can score countable differences in transcript numbers among thousands of genes and between experimental groups, all at once. However, its high cost limits experimental designs to very small sample sizes, usually N = 3, which often results in statistically underpowered analysis and poor reproducibility. All these issues are compounded by the presence of experimental noise, which is harder to distinguish from instrumental error when sample sizes are limiting (e.g., small-budget pilot tests), experimental populations exhibit biologically heterogeneous or diffuse expression phenotypes (e.g., patient samples), or when discriminating among transcriptional signatures of closely related experimental conditions (e.g., toxicological modes of action, or MOAs). Here, we present a leveraged signal-to-noise ratio (LSTNR) thresholding method, founded on generalized linear modeling (GLM) of aligned read detection limits to extract differentially expressed genes (DEGs) from noisy low-replication RNAseq data. The LSTNR method uses an agnostic independent filtering strategy to define the dynamic range of detected aggregate read counts per gene, and assigns statistical weights that prioritize genes with better sequencing resolution in differential expression analyses. To assess its performance, we implemented the LSTNR method to analyze three separate datasets: first, using a systematically noisy in silico dataset, we demonstrated that LSTNR can extract pre-designed patterns of expression and discriminate between “noise” and “true” differentially expressed pseudogenes at a 100% success rate; then, we illustrated how the LSTNR method can assign patient-derived breast cancer specimens correctly to one out of their four reported molecular subtypes (luminal A, luminal B, Her2-enriched and basal-like); and last, we showed the ability to retrieve five different modes of action (MOA) elicited in livers of rats exposed to three toxicants under three nutritional routes by using the LSTNR method. By combining differential measurements with resolving power to detect DEGs, the LSTNR method offers an alternative approach to interrogate noisy and low-replication RNAseq datasets, which handles multiple biological conditions at once, and defines benchmarks to validate RNAseq experiments with standard benchtop assays.

Postnatal Arx transcriptional activity regulates functional properties of PV interneurons

“…The basis for the incoherence between the DEG and qPCR analyses with respect to Arx could be due to a number of factors. First, low Arx read counts in the RNAseq could have diminished the accuracy of quantifying Arx expression ( Lozoya et al., 2018 ). Second, P35-40 total RNA from FACS sorted Arx null PVIs used for genome-wide expression profiling may have been contaminated with slowly degrading mRNA transcribed prior to recombination at ∼ P15.…”

Single Nucleotide Resolution Analysis Reveals Pervasive, Long-Lasting DNA Methylation Changes by Developmental Exposure to a Mitochondrial Toxicant

“…Detection of differentially expressed genes (DEG) across time × treatment groupings was performed using weighed two-way ANOVA (gene × grouping) based on log2-transformed fold-change (log2FC) measurements relative to gene RPKM grand-means over all specimens; N = 24 (3 biological replicates per time × treatment grouping). Gene-wise log2FC values were weighed by a relative metric of sequencing representation (cumulative hazard of significance scores from gene-wise RPKM rate modeling with an exponential distribution and inverse link function) ( Lozoya et al, 2018b ). Genes were retained for post hoc pairwise analysis when significance level p < 0.05 after multiple comparison adjustment ( Benjamini and Hochberg, 1995 ), then filtered against a minimum gene-wise effect size δ log2FC > 0.3 × σ log2FC and post hoc pairwise significance (Student’s t test p < 0.05) for log2FC differences at each time point against PND22 among specimens in the same treatment group, as well as log2FC differences at PND22 between treatment groups.…”

Genome-wide dynamic network analysis reveals a critical transition state of flower development in Arabidopsis

“…In general, a mathematical model based on mass action kinetics and Michaelis–Menten kinetics can describe the transcriptional regulation process [64]. However, the noise inherited in the gene expression data can decrease the performance of these models [65]. Therefore, to improve the accuracy of the network inference, NARROMI algorithm was used to reduce noisy, redundant and indirect regulations [29].…”