Thread by @ksamocha

Peter Robinson #GRD23 on splicing-impacting variants.

Diagnostic rates have improved over time, but still are low. Where are the other diagnoses?

Notes that bioinformatics can help increase efficiency by improving variant prioritization and interpretation.

Peter Robinson #GRD23: Field is slowly moving to focus on prioritizing noncoding variants (vs missense).

For splice sites, know about the importance of the canonical sites (e.g., donor +1/+2), but variants that fall outside of those regions are harder to interpret.

Peter Robinson #GRD23: SQUIRLS
Super QUick Information content Random-forest Learning of Splice variants

Paper on it:
pubmed.ncbi.nlm.nih.gov/34289339/

Peter Robinson #GRD23: Has a new graphical representation to show degree to which a sequence matches the donor or acceptor model.

Talks through some of the feature engineering considered when training this model, including information content change.

Peter Robinson #GRD23: Random forest means that you can extract the feature importance (phyloP is important for donors, for example).

Matched performance of SpliceAI for the most part (maybe a _bit_ worse). Notes in bioinformatics, all tools are reported as the best [too true]

Peter Robinson #GRD23: Notes that SQUIRLS is faster than SpliceAI when it comes to annotating an exome. This is freely available on github:
github.com/TheJacksonLabo…