LoDEI: a robust and sensitive tool to detect transcriptome-wide differential A-to-I editing in RNA-seq data

Abstract

RNA editing is a highly conserved process. Adenosine deaminase acting on RNA (ADAR) mediated deamination of adenosine (A-to-I editing) is associated with human disease and immune checkpoint control. Functional implications of A-to-I editing are currently of broad interest to academic and industrial research as underscored by the fast-growing number of clinical studies applying base editors as therapeutic tools. Analyzing the dynamics of A-to-I editing, in a biological or therapeutic context, requires the sensitive detection of differential A-to-I editing, a currently unmet need. We introduce the local differential editing index (LoDEI) to detect differential A-to-I editing in RNA-seq datasets using a sliding-window approach coupled with an empirical q value calculation that detects more A-to-I editing sites at the same false-discovery rate compared to existing methods. LoDEI is validated on known and novel datasets revealing that the oncogene MYCN increases and that a specific small non-coding RNA reduces A-to-I editing.

Fig. 1. Schematic overview of differential A-to-I editing index calculation.

a Two sets of samples S and $S^{\prime }$ with different conditions undergo RNA-sequencing followed by standard RNA-seq data analysis of QC checking and read alignment. b and c Next, differential A-to-I editing signals are calculated by LoDEI's sliding window approach (b) yielding the final output table (c).

Fig. 2. Observed signal differences and empirically derived q values.

Rows correspond to the ADAR1 KD (a–c), RO60 KO (d–f), MYCN-amp (g–i), and sncRNA7SL OE (j–l) datasets. The left and middle columns show Bland–Altman plots of δ^G→A values considered as noise (left column), and δ^A→G values considered as a mixture of signal and noise (middle column). The third column shows empirical q values. Orange lines show the signal cutoffs derived by LoDEI corresponding to a q value ≤ 0.1 in columns 1 and 2 and a q value of 0.1 in column 3.

Fig. 3. Performance comparison of differential A-to-I site detection.

Shown are the number of detected differentially edited A-to-I sites as a function of the q value threshold for the ADAR1 KD (a), RO60 KO (b), MYCN-amp (c), and sncRNA7SL OE (d) datasets.

Fig. 4. Genomic locations of detected differentially edited A-to-sites.

The number of differentially edited A-to-I sites found by LoDEI (red), REDIT (blue), and JACUSA2 (light blue) at a q value threshold of 0.1 at different genomic locations are shown for the ADAR1 KD (a), RO60 KO (b), MYCN-amp (c), and sncRNA7SL OE (d) datasets.

Fig. 5. Overlap of detected differentially edited A-to-I events with REDIportal A-to-I sites as a function of the q value.

The percent of overlap of detected differentially edited A-to-I events found by LoDEI (red), REDIT (blue), and JACUSA2 (light blue) with REDIportal A-to-I sites are shown for the ADAR1 KD (a), RO60 KO (b), MYCN-amp (c), and sncRNA7SL OE (d) datasets.

Fig. 6. Overlap of detected differentially edited A-to-I events with REDIportal A-to-I sites for different genomic locations.

The percent of overlap of detected differentially edited A-to-I events found by LoDEI (red), REDIT (blue) and JACUSA2 (light blue) at a q value threshold of 0.1 at different genomic locations with REDIportal A-to-I sites are shown for the ADAR1 KD (a), RO60 KO (b), MYCN-amp (c), and sncRNA7SL OE (d) datasets.