With the rise of RNA-seq as an essential and ubiquitous tool for biomedical research, the need for guidelines on quality control (QC) is pressing. Specifically, there remains limited data as to which technical metrics are most informative in identifying low-quality samples. Here, we addressed this issue by developing the Quality Control Diagnostic Renderer (QC-DR), software designed to simultaneously visualize a comprehensive panel of QC metrics generated by an RNA-seq pipeline and flag samples with aberrant values when compared to a reference dataset. As an example, we applied QC-DR to the Successful Clinical Response in Pneumonia Therapy (SCRIPT) dataset, a large clinical RNA-seq dataset of sequenced alveolar macrophages (n = 252). Next, we used this dataset to assess relationships between a variety of QC metrics and sample quality. Among the most highly correlated pipeline QC metrics were % and # Uniquely Aligned Reads, % rRNA reads, # Detected Genes, and our newly developed metric of Area Under the Gene Body Coverage Curve (AUC-GBC), while experimental QC metrics derived from the lab were not significantly correlated. We then trained a set of machine learning models on the SCRIPT dataset to evaluate the relative contribution of QC metrics to sample quality prediction. Our model performs well when tested on an independent dataset despite differences in the distribution of QC metrics. Our results support the conclusion that any individual QC metric is limited in its predictive value and suggests approaches based on the integration of multiple metrics with QC thresholds. In summary, our work provides new insights, practical guidance, and novel QC software which can be used to improve the methodological rigor of RNA-seq studies.
Hamilton et al. (Tue,) studied this question.