Variant present in SpliceAI lookup absent in spliceai_scores.masked.snv.hg19.vcf.gz

[Roshanshafiha]

Hi, I have actually used the file spliceai_scores.masked.snv.hg19.vcf.gz for annotating my VCF file. I have found a variant which is absent in the spliceai_scores.masked.snv.hg19.vcf.gz but present in the SpliceAI lookup. May I know whats the reason for this, thank you.

The variant is :

[kishorejaganathan]

The lookup website uses a newer annotation file compared to the one used in the precomputed scores.

[Roshanshafiha]

Thank you for clarifying my question. May I know if there is a possibility to download the newer version of the annotation file?

[Roshanshafiha]

Hi, may I know if there is a location to access those annotation file used on the website. @kishorejaganathan

[kishorejaganathan]

You should ask the developer Ben Weisburd https://github.com/broadinstitute/SpliceAI-lookup

[benstory]

Hi this is kind of concerning... Is there no interest from Illumina's side to provide the most up-to-date possible annotations?

A clinician was complaining today that my scores were wrong...

This variant from Illumina's spliceai_scores.raw.snv.hg38.vcf.gz
6 169222192 . C G . . SpliceAI=G|THBS2|0.00|0.00|0.01|0.14|5|49|9|5

Switches to a variant of potential concern in the SpliceAI-lookup:

EDIT:
So I'm looking at this and my guess is their default Max Distance is set to 500. Was the original SpliceAI set differently? (Looks like 50 was the default). Ok that makes sense... Would you recommend that we re-run the prediction locally on the most up-to-date RefSeq/Ensembl annotations?

[kishorejaganathan]

Hi,
Yes, this issue is precisely because of the max distance. We used 50 while computing the precomputed scores, and realized much later (after publication) that we should be using 500 or even larger - so we recommended that to people who are running SpliceAI instead of using the precomputed scores.

The precomputed scores require an enormous amount of compute (around 3.5B variants), so we decided not to redo this with the updated settings. We will handle this in the next version of SpliceAI.

Regards,
Kishore