Our preprint on the challenges of benchmarking Structural Variants (SVs) is out in BioRxiv

From Dr. Bryant Lin, Stanford I'm living in insurance hell!!! 1) I am a Stage 4 Lung Cancer patient and advocate. 2) As a medical school Clinical Professor, I taught a class at Stanford anchored around my own case which was covered widely in the The New York Times, CBS Mornings,

Glad to present Unico, our new distribution-free method for deconvolving bulk omic to uncover cell-type-level insights (with results on recapitulating B cell profile in FL tumor expression and high consistency in cell-type-level methylation association):

Do you know ~60% of human SVs fall in ~1% of GRCh38? See our new preprint: https://t.co/ZdP2r83KI8 and the companion blog post on how we started this project and longdust: https://t.co/a9W6nXmslz. Work with @QianAlvinQin1

Now preprinted at https://t.co/ffgwN8NC11

Happy to share the latest work from the lab! We long-read sequenced single-cell derived melanoma subclones to reconstruct complete view of tumor evolution. With phylogeny, we can see parallel structural evolution and lineage-specific methylation.

@l_denti, @krannich479, Tomas Vinar, @RayanChikhi, Paola Bonizzoni, @BronaBrejova.

To streamline and standardize the SV evaluation task, we developed a framework, freely available at https://t.co/TqG9aYY45y, which allows to benchmark old and new SV callers against a diverse collection of truth sets

Interestingly, the ranking of the methods changed based on the version of GIAB callset used as ground truth highlighting that the performance of callers is not readily transferable between “ground truths” without consideration of their limitations and differences.

All tools exhibit great accuracy wrt v0.6 callset but we observed a considerable drop in their performance when evaluated against the more complete curated SV callset (especially when considering CHM13).

We also compared them to the GIAB v0.6 callset and the very recently released GIAB v1.1 callsets.

We evaluated long-read based callers against assembly-based callsets and we observed substantial variation in their performance, depending on the choice of ground truth, reference genome, and genomic regions used for evaluation.

The overall similarity among these callsets decreases as the completeness of the reference genome increases.

On confident regions of the genome, the three assembly-based callers provide callsets with similarity as high as 87% while on full genome the similarity decreases to 46%.

We compared the three assembly-based callers and we observed substantial differences in the properties of the generated SV callsets.

We identified critical and often unexpected divergence in the performance evaluations of SV callers, largely driven by commonly applied user-defined choices in analysis workflows

We extensively tested 3 assembly-based callers, 8 long-reads-based callers, 2 truvari modes (bench and refine) on the 3 major human reference genomes (GRCh37, GRCh38, and T2T-CHM13)

Many methods have been proposed for discovering SVs, however their benchmarking has been inconsistent across studies, often resulting in contradictory findings.

Amazing work and achievement!

“As a public university, we are stewards of taxpayer resources, and a payment of this scale would completely devastate our country’s greatest public university system as well as inflict great harm on our students and all Californians.” —UC President James B. Milliken