Batch effect detection and correction in RNA-seq data using machine-learning-based automated assessment of quality

Table 1 Manual exploration of the PCA plots with different corrections for all datasets

GEO series	Exp. design (group vs batch)	Design bias (P_low vs group)	Kruskal Wallis’s P-value (batch vs P_low)	Batch effect on base PCA	Batch effect removed after batch correction	Batch effect removed after Plow correction	Plow performance compared to batch	Plow performance with outlier removal	Performance of combined correction
GSE120099	Good	0.655	4.24E−03	Yes	Yes	No	Worse	Worse	Worse
GSE117970	Poor	0.608	8.41E−04	Yes	No	Yes	Better	Better	Worse
GSE163857	Poor	0.522	2.09E−02	No	–	–	Comparable	Comparable	Worse
GSE162760	Good	0.496	2.36E−12	Yes	Yes	Yes	Comparable	Better	Comparabe
GSE182440	Very good	0.495	1.06E−01	No	–	–	Comparable	Better	Better
GSE144736	Poor	0.494	3.63E−01	Yes	Yes	No	Comparable	Better	Worse
GSE82177	Very good	0.493	5.75E−01	Yes	Yes	No	Comparable	Better	Better
GSE171343	Very good	0.488	8.25E−02	Yes	Yes	No	Comparable	Comparable	Comparable
GSE173078	Very good	0.479	2.93E−07	No	–	–	Comparable	Comparable	Comparable
GSE61491	Good	0.448	2.13E−01	Yes	Yes	No	Comparable	Comparable	Better
GSE163214	Good	0.443	1.03E−02	Yes	Yes	Yes	Comparable	Comparable	Better
GSE153380	Poor	0.442	1.58E−01	No	–	–	Comparable	Better	Better

Exp. Design is a manually given label evaluating the balance of the biological groups between the batches. Design Bias evaluates the clustering of the biological groups by quality scores P_low (normalized gamma of P_low against the group; the higher, the better the clustering by quality; values from 0 to 1). The Kruskal–Wallis’s P-value is derived from a Kruskal–Wallis’s test comparing average P_low values by batch. Taken together, those metrics show the potential association between our quality metric, groups, and batches. Other columns show the manual evaluation of the batch effect and correction methods

ISSN: 1471-2105