1. Curate patient-centric multi-omics data for precision medicine
Jun Zhu, Ph. D.
Genomics and Genetic Sciences
Icahn Institute of Genomics and Multi-scale Biology
The Tisch Cancer Institute
Icahn Medical School at Mount Sinai
New York, NY
@IcahnInstitute

2. Why it is critical to curate patient-centric multi-omics data?

3. Why errors are common in large data sets?
Clustering analysis and differential analysis do not require a large number of samples. A few errors in the samples don’t affect results much. Data sets are over-powered for the types of analyses (low hanging fruits), or the data sets are under-used.

4. Sample swaps
Gene expression
Swap across genders

5. Gene expression based gender DNA methylation based gender
Gender errors in TCGA (The Cancer Genomic Atlas)
Tumor Type
Gene expression based gender
DNA methylation based gender
Microarray
RNAseq
HM27
HM450
# of tumor samples
Male vs Female
Inconsistent samples
BLCA -
408
292:116
9 (2.2%)
153
112:41
3 (2%)
BRCA
534
6:528
2 (0.4%)
1059
12:1047
0 (0%)
318
3:315
748
9:739
COAD
166
84:82
1 (0.6%)
258
141:117
14 (5.4%)
GBM
561
336:218
9 (1.6%)
169
109:60
225
138:87
125
71:54
KIRC
219
142:77
2 (0.9%)
283
183:100
8 (2.9%)
LUAD 33
15:18
7 (21.2%)
517
243:274
3 (0.6%)
127
63:64
1 (0.8%)
306
145:161
LUSC
134
99:35
227
163:64
5 (2.2%)
READ 68
37:31 96
54:42
STAD
384
242:142
11 (2.9%) 82
42:40
210
134:76

6. Gender consistency check is not sufficient!
Only ~50% sample labeling errors can be detected by gender QC.
Not only to identify problems/errors, but also to unambiguously correct errors

7. Iteratively aggregating patient-centric omics data
Multi-omics data Matcher (MODMatcher)
Yoo et al, PLoS CompBio. 2014

8. Aggregating different types of profiles based on intrinsic barcodes
Yoo et al, PLoS CompBio. 2014

9. Simulation results

10. TCGA (The Cancer Genomic Atlas)
BRCA
GBM
Data type
RNASeq
Microarray
HM27
HM450
CNV
Original tumors
1059
534
318
747
1065
170
288
142
565
Good
1056
518
314
329
1043
520
284
141
559
With other 3 16 14 4 1
Poor 18 31
Error rate
0.3%
3.0%
1.2%
2.4% 3% 0%
2.6%
1.4%
0.7%
1.1%
Cancer
LUAD
PRAD
STAD
Data type
RNASeq
Microarray
HM27
HM450
CNV
Original tumors
517 33
127
460
421
498
493
384
476 82
395
Good
515 23
110
453
420
374
369
376
446 77
371
With other 9 17 1 5
Poor 2 7
124 6 29
Error rate
0.4%
28%
13%
1.5%
0.2%
24.9%
25.2%
1.6%
6.3%
6.1%

11. Probabilistic MODmatcher
when intrinsic barcode size is small

12. Sample cross contamination
Gene expression
Sample contamination

13. Clean matching patterns

14. Possible sample contamination

15. mis-annotated clinical information
Gene expression

16. mis-annotated clinical information

17. Software Will be available in Docker’s container soon.
Curation results will be shared through our web page.

18. Aknowledgements Supported by: Zhu lab Mount Sinai Seungyeul Yoo
Eunjee Lee
Li Wang
Wenhui Wang
Yongjae Woo
Yi Zhang
Mount Sinai
Genomics Institute
Matthew Galsky
Yixuan Gong
Spiros Hiotos
Qin Wang
Supported by:
Icahn Institute of Genomics and Multiscale Biology, Icahn School of Medicine at Mount Sinai
NIH R01-AG046170
BD2K U01-HG008451