Very high FREEMIX value with WES experiments

[rulixxx]

We have a series of high depth WES experiments that are showing very high contamination. Pileup file looks reasonable with values between 1 - 600. I suspect the PC calculation is failing:

[SimplePileup] Total Number Bases: 603116
[SimplePileup] Total Number Markers: 8791
NOTICE - Number of marker in Reference Matrix:9892
NOTICE - Number of marker shared with input file:8791
NOTICE - Mean Depth:68.606074
NOTICE - SD Depth:41.640111
NOTICE - 8753 SNP markers remained after sanity check.
NOTICE - Passing Marker Sanity Check...
Estimation from OptimizeHeter:
Contaminating Sample PC1:0	PC2:0
Intended Sample PC1:0	PC2:0
FREEMIX(Alpha):0.5
NOTICE - Success!

When I downsample that BAM file keeping 90% of the reads I get a sensible answer:

[SimplePileup] Total Number Bases: 542302
[SimplePileup] Total Number Markers: 8759
NOTICE - Number of marker in Reference Matrix:9892
NOTICE - Number of marker shared with input file:8759
NOTICE - Mean Depth:61.913689
NOTICE - SD Depth:37.354130
NOTICE - 8718 SNP markers remained after sanity check.
NOTICE - Passing Marker Sanity Check...
Estimation from OptimizeHeter:
Contaminating Sample PC1:-0.00338479	PC2:-0.0472513
Intended Sample PC1:-0.00776845	PC2:-0.0262016
FREEMIX(Alpha):0.000633472
NOTICE - Success!

Any idea whats happening here?

[Griffan]

Hi, Alpha with value 0.5 means the optimization procedure doesn't converge and quit after reaching maximal rounds of iterations. Are these experiments all at 70X of sequencing depth? or any other common characteristics on these failed samples showed up?

[rulixxx]

We are doing titration experiments to test library preparation so the libraries might be a bit out of the ordinary. This experiment is 85X. We have a range of libraries 85- 180 X and all of them show abnormally high alpha. Experiments with coverage below 85X look OK.

Is it possible to increase the number of iterations? I don't see any parameters to do that.

[hyunminkang]

It is possible that numerical errors are introduced with extremely high depth. We may need an example to identify the source of errors.