Differences in ecDNA predictions between AmpliconArchitect and CoRAL across glioblastoma cell lines #59
Description
Hi,
Thank you for developing CoRAL.
We’re working with four glioblastoma cell lines (E20, E25, E26, and E28) for which we have both:
- Illumina short-read WGS (averaging ~60x sequencing depth) with matched normals.
- ONT long-read WGS (averaging ~30x sequencing depth).
Short-read results:
We ran AmpliconArchitect using CNV calls from both CNVKit and PURPLE. The ecDNA predictions were consistent across CNV callers, and the following ecDNAs were identified:
E20: ARID2, CDK4
E25: KDR, PDGFRA; AGAP2, CDK4, MDM2
E26: EGFR
E28: EGFR
These predictions were further validated via FISH (eLife 2023). So the AA predictions look good quality given the constraints of short-read data.
Long-read results using CoRAL:
We’ve run several long-read ecDNA detection tools including CoRAL, Decoil, CReSIL, and a de novo assembly-based approach. For CoRAL specifically, we used CNV calls from both CNVKit and SPECTRE.
CoRAL + CNVKit results:
E20: No ecDNAs identified
E25: CDK4
E26: No ecDNAs identified
E28: No ecDNAs identified
CoRAL + SPECTRE results:
No ecDNAs identified for any sample.
Question:
We were surprised by the difference between AmpliconArchitect and CoRAL, particularly given that multiple ecDNAs in these samples were validated. Do you have any insight into why CoRAL might not be identifying ecDNAs in cases where AmpliconArchitect does — especially when both use the same CNV inputs (e.g., CNVKit)?
CoRAL Commands Run:
coral seed \
--cn-seg $cn_seg \
--output-prefix $prefix
coral reconstruct \
--lr-bam $bam \
--cnv-seed $seed_bed \
--output-dir . \
--cn-seg $cn_seg \
--solver-threads $task.cpus
Any guidance or suggestions would be greatly appreciated.
Thanks again for your work on CoRAL.