Variant ranking / prioritisation / ACMG auto-classification

[davmlaw]

We have no way to prioritise variants within an analysis node. This would be good to reduce classification work on medical scientists.

ACMG classifications have 5 levels of clinical significance, but underneath that is a score from criteria weights - so I think using the score from an an ACMG classifier is also a variant prioritiser - hence moving all issues together here.

Personally, I'd much rather use an existing tool, though that may eg only work on default VEP fields, not all the information we have (though we could extend it)

[davmlaw]

From other issue, in August 2018:

You can't do everything, so this needs manual steps to enter info you can't get from data (eg family segregation / literature etc)

InterVar - free tool.
Franklin https://franklin.genoox.com/

Trouble is, InterVar uses their expected annotation (ANNOVAR) not sure if it'll work on our stuff without lots of changes. http://wintervar.wglab.org/results.php

Franklin webpage works on a single variants, but you need to sign up to get more.

Some of these we can calculate per gene, some per variant / transcript, and some need sample etc information (eg phenotype based population frequencies etc). You can combine these together to work out the final score.

Automated classification software: QCI Interpret, Golden Helix, CharGen, InterVar, Varsome, Uni of Maryland - v. basic,

The big brains: SolveBio https://omictools.com/dann-tool

[davmlaw]

A good summary of the literature is given in the CharGen (2018) paper:

There have been several efforts to develop automated variant interpretation tools. For example, one group created a browser-based ACMG variant classifier in which the ACMG criteria can be selectively applied (Kleinberger et al., 2016). Another group developed InterVar (Li and Wang, 2017), a tool that automates the initial variant interpretation but then relies on a manual review step to adjust the classification criteria based on prior information or domain knowledge before arriving at a final interpretation. Another group extended ACMG’s 33 rules with 108 refinements, including semiqualitative aspects in classification, into a framework called Sherloc (Nykamp et al., 2017). However, its source code is not publicly available.

There’s also commercial ones: QCI Interpret, SolveBio, Varsome (has an API)

The InterVar paper goes through the 28 ACMG criteria and gives a great summary of how automatable each one is. InterVar uses AnnoVar while CharGen uses VEP.

CADD and DANN (2014) are machine learning based pathogenic prediction tools.

[davmlaw]

https://www.medrxiv.org/content/10.1101/2021.10.07.21264628v1.full-text

Used Exomiser v12.1 (data version 2102) and AMELIE v3.1.0