Try to come up with a region-based blacklisting of variants. Right now it's done per gene, which is wrong because only a part of a gene might be bad.
We should train a machine learning regression method to give an "artefact probability".
We should train separate models for SNVs and InDels, maybe also for exome/genome.
Test with thw following variants:
variants flagged as causal/artefact in NGSD
de-novo variants of monozygotic twins (W:\share\evaluations\2020_07_29_twin_denovo)
mendelian errors in trios
Test the following properties of a variant as features:
QUAL
DP (maybe)
MQM
SAP
ABP
In repeatmaskter region
Distance from next repeatmasker region border
Test the following variant locus metrics as features (with different window sizes, also only left/right of variant)
Percentage of positions with variant call in NGSD database (maybe also counts if variant calls)
Percentage of soft-clipped reads/bases in the window
Percentage of mismatches in the window
Percentage of indels in the window
Percentage/count of gnomAD variants with passing and failing quality column.
Percentage overlap with GIAB high-confidence region of NA12878
Percentage overlap with Platinum genomes high-confidence region of NA12877
Percentage overlap with RepeatMasker region extended by 5 bases - also try sub-types of repeats
Try to come up with a region-based blacklisting of variants. Right now it's done per gene, which is wrong because only a part of a gene might be bad.
We should train a machine learning regression method to give an "artefact probability". We should train separate models for SNVs and InDels, maybe also for exome/genome.
Test with thw following variants:
Test the following properties of a variant as features:
Test the following variant locus metrics as features (with different window sizes, also only left/right of variant)