automated chip qc
DESCRIPTION
Automated Chip QC. Michael Elashoff. Chip QC. Transition from mostly manual/visual chip QC to mostly automated chip QC Database of passing and failing chips to serve as the training set (5K passing, 2K failing). Chip QC: Defect Classes. In order of occurrence: Dimness High Background - PowerPoint PPT PresentationTRANSCRIPT
Automated Chip QC
Michael Elashoff
Chip QC
• Transition from mostly manual/visual chip QC to mostly automated chip QC
• Database of passing and failing chips to serve as the training set (5K passing, 2K failing)
Chip QC: Defect Classes• In order of occurrence:
– Dimness– High Background– Unevenness– Spots– Haze Band– Scratches– Brightness– Crop Circle– Cracked– Snow– Grid Misalignment
• Training set of 7K chips (Human, Rat, Mouse)
0 500 1000 1500 2000Intensity
Cou
nt
0 500 1000 1500 2000Expression
Cou
ntDimness/Brightness
Passing Chips
Bright/Dim Chips
A chip Low Scan
Dimness/Brightness
Passing Chips
Bright/Dim Chips
0 500 1000 1500 2000Intensity
Cou
nt
0 500 1000 1500 2000Expression
Cou
ntA chip Low Scan
Dimness/Brightness• Each chip type has a different typical
brightness range• Typical brightness range depends on
scanner setting– tuned-up versus tuned-down– scanners must be calibrated to achieve consistency
Spots, Scratches, etc.
Spots, Scratches, etc.
Implementation of Li-Wong• With training set of 5K passing chips, apply
Li-Wong algorithm
• For each probe set, algorithm yields:– “outlier” status for each probe-pair– probe weights for non-outlier probe-pairs
),0(~
...1,...12
N
NjNi
MMPM
ij
pairsprobechips
ijjiijij
j
Implementation of Li-Wong
• For QC, new chips are screened individually• For each probe set:
– Ignore “model outlier” probes– Using training ‘s, compute– Compute residuals for each probe pair– Flag residuals that are large
j
Implementation of Li-Wong• Compare distributions of outlier count for
passing and failing chips in training set• Determine upper bound of acceptable outlier
count:
Hgu95a 3100Hgu95b 4200Hgu95c 4100Hgu95d 4700Hgu95e 4300Rgu34a 3300Rgu34b 3000Rgu34c 3100
Grid Alignment
Grid Alignment
0
40
80
120
160
Outlier Count
Limitations of Li-Wong
• Must estimate 1.8 million probe weights for human/rat chip sets
• Works poorly for rare genes• Probe weights may vary
– Tissue Type– RNA Processing– Chip Lot– Training Set
Haze Band
Vertical 10th Percentile Profile
0
100
200
300
400
500
6001 15 29 43 57 71 85 99 113
127
141
155
169
183
197
211
225
239
253
Bin Number
Inte
nsi
tyHaze Band
Crop Circles
Crop Circles
Horizontal 95th Percentile Profile
0
5000
10000
15000
20000
25000
30000
35000
400001 16 31 46 61 76 91 106
121
136
151
166
181
196
211
226
241
256
Bin Number
Inte
nsi
ty
Using Spike-Ins
10
100
1000
10000
0.1 1 10 100
Concentration
Ex
pre
ss
ion
10636 10638 10639 10640 10641 10643 10644 1064510646 10648 10649 10650 10651 10652 10653 1065410655 10656 10657 10658 10659 10660 10661 1066210663 10664 10665 10666 10667 10668 10669 1067010671 10673 10674 10675 10677 10678 10679 10680
Spike-in R2 must be >96.5%
QC Metrics• Mean of Non-control Oligo Intensity• Mean OligoB2 Intensity• Spike-in R2
• Li-Wong Outlier Count• Several measures of LiWong Outlier
“clustering”• Vertical profiles• Horizontal profiles• Thresholds differ for each chip type
QC Metrics
QC Metrics: Performance
AutoQCPass
AutoQCFail
Total
ManualPass
449 131 580
ManualFail
2 149 151
Total 451 280 731
Two week validation run
False Negative Rate = 0.4%
These will not be manually QC’d anymore
False Positive Rate = 46.8%
These are still manually QC’d
Conclusions
• Automated QC has: – reduced the number of chips in visual QC– made the process more objective
• Automated QC has not:– eliminated the need for visual QC– incorporated the impact on real world data
quality/analysis
Thanks
• Peter Lauren• Chris Alvares• John Klein• Michelle Nation• Jeff Wiser