Supplementary Materials Figures

Transcription

1 Supplementary Materials Figures!"#$%&'(%)**$*$! +,-%!."$/01-0,2! +,-%!!"#$.,30-04$ 5)*3$ +$<)-04$ 5)*3$.,30-04$!"#$ +$<)-04$!"#$%$&'(!)'*$+%$&'(,-./'%6 7%,8%9-"#$%:,30-04$; 7%,8%9:"$/01-0,2%03%!; 0'1-%$&'(!)'*$+%$&'(,-./'%6 7%,8%9-"#$%2$<)-04$; 7%,8%9:"$/01-0,2%03%2,-%!; 2'3#$%$&$%4%6 7%,8%9-"#$%:,30-04$; 7%,8%9-"#$%)**$*$%03%!; 25'+$6$+$%4%6 7%,8%9-"#$%2$<)-04$; 7%,8%9-"#$%)**$*$%03%2,-%!; 7..'.'(7++/)-+4%6 7%,8%9-"#$%:,3;%=%9-"#$%2$<; <")2/%3#> Figure S1: The standard statistical quantities of prediction quality for a specific HL allele H: sensitivity, specificity, positive predictive value, negative predictive value and allele accuracy. 100 ccuracy (%) Flanking Region in KB (on the SNP intersect of Illumina platforms) HL HL B HL C HL DRB1 HL Figure S2: The relationships between the four-digit accuracies (no call threshold) and size of flanking region from 50kb to 1000kb on each side, stratified by HL loci. The HIBG models were built using the HLRES samples of European ancestry as the training data, and the imputation accuracies were assessed with the independent testing data of the British 1958 birth cohort study. SNP markers were genotyped on the intersect of Illumina platforms. 500kb flanking region is an appropriate region for predicting HL alleles. 1

2 European ancestry sian ancestry Hispanic ancestry frican ancestry ccuracy (%) Flanking Region in KB (on the SNP intersect of Illumina platforms) HL HL B HL C HL DRB1 HL DQ1 HL HL 2 Figure S3: The relationships between the four-digit accuracies (no call threshold) and size of flanking region from 50kb to 1000kb on each side, stratified by HL loci and ethnicities. STUDY Data were divided into training and validation sets with equal sizes for each ancestry and each HL gene. SNP markers were genotyped on the intersect of Illumina platforms. 500kb flanking region is an appropriate region for predicting HL alleles.

3 ccuracy European sian Hispanic frican European B sian B Hispanic B frican B European C sian C Hispanic C frican C European DRB1 sian DRB1 Hispanic DRB1 frican DRB1 (a) Posterior Probability European DQ1 sian DQ1 Hispanic DQ1 frican DQ European sian Hispanic frican European sian Hispanic frican ccuracy (b) Posterior Probability 50 individuals 10 and < 50 individuals < 10 individuals Figure S4: The relationship between posterior probability and overall accuracy. STUDY Data for each ancestry are divided into training and validation sets with equal sizes, and the accuracies are calculated from ten bins of posterior probabilities: (a) stratified by HL loci and ancestries; (b) over all HL loci and ancestries, the curve is fitted by a function y = x r with a parameter r = 0.31, and 0.5 posterior probability approximately corresponds to the prediction accuracy 80%. 3

4 ccuracy (%) Call Rate (%) B C DRB1 DQ Call Threshold Call Threshold Figure S5: The relationships among call threshold, accuracy and call rate when HLRES data for individuals of European ancestry are divided into training and validation sets with equal sizes Sensitivity B C DRB1 DQ Number of Copies of Training Haplotypes Sensitivity B C DRB1 DQ llele Frequency (%) Figure S6: The relationship between four-digit sensitivities (no call threshold) and the number of copies of training haplotypes for each HL allele when HLRES data for individuals of European ancestry were divided into training and validation sets with equal sizes. SNP markers on the intersect of Illumina platforms were used. For, C, DQ1, and, 10 copies of training haplotypes seem sufficient to attain 90% sensitivity, but B and DRB1 require many more training haplotypes. 4

5 100 B C DQ1 DRB1 90 ccuracy (%) Missing Proportion (%) (a) no call threshold: accuracy vs. missing proportion. 100 B C DQ1 DRB1 90 ccuracy (%) Missing Proportion (%) (b) 0.5 call threshold: accuracy vs. missing proportion. 100 B C DQ1 DRB1 80 Call Rate (%) Missing Proportion (%) (c) 0.5 call threshold: call rate vs. missing proportion. Figure S7: Box plots of accuracy and call rate with missing SNPs. HLRES data of European ancestry were divided into training and validation sets with equal sizes. The HIBG models were built using the training parts. For each run of simulation, a fraction of the SNP predictors used in the ensemble classifier (e.g, 10%, 20%) was removed randomly for the validation set, where every validation sample has the same missing SNPs, and repeat it 100 times. The missing SNPs do not significantly reduce the accuracies for missing fraction < 80%, but it does decrease the call rates. 5

6 European ncestry C B DRB1 DQ sian ncestry 500 C B DRB1 DQ Number of Classifiers Hispanic ncestry C B DRB1 DQ frican ncestry C B DRB1 DQ SNP Position (kilobase) B C DRB1 DQ1 Figure S8: The number of classifiers used in the published pre-fit models for each SNP predictor. Each HIBG model consists of 500 individual classifiers, and more important SNP markers tend to be used more frequently. 6

7 Supplementary Materials Tables Table S1: ssessing the prediction accuracies using different model parameter settings, when STUDY Data of European ancestry were divided into training and validation sets with equal sizes. No call threshold was executed 1. HL ccuracy (%) B C DRB1 DQ1 #ofsnps #oftrainingsamples #ofvalidationsamples The total number of classifiers K =25 m try = m try = m m try = 1 3m m try = m The total number of classifiers K =100 m try = m try = m m try = 1 3m m try = m : K is the total number of individual classifiers, m is the total number of SNP markers, and m try is the number of variables randomly sampled as candidates for eachselection. 2 :SNPmarkerscommontotheIllumina1MDuo,OmniQuad,OmniExpress, 660K and 550K platforms within a flanking region of 500kb are used. Table S2: ssessing the computational times (hour) of building a HIBG model for our published parameter estimates of European ancestry on a Linux system with Intel processor (2.27GHz) and 32 GB RM. HL B C DRB1 DQ1 #ofsnps #ofhlalleles #oftrainingsamples Building a HIBG model: computing time per individual classifier 0.86h 6.12h 0.84h 3.36h 0.58h 0.56h 0.28h 1 :SNPmarkerscommontotheIllumina1MDuo,OmniQuad,OmniExpress, 660K and 550K platforms within a flanking region of 500kb are used. 7

8 Table S3: Summary of the four-digit accuracies from HIBG and BEGLE using the same SNP sets. STUDY Data were randomly divided into training and validation sets with equal sizes for each ancestry. No call threshold was used, and the SNP markers within a 500kb flanking region on each side were used. HL ccuracy (%) B C DRB1 DQ1 European ancestry # of training samples # of validation samples # of HL alleles #ofsnps 1 (1M/intersect) 937/ / / / / / /279 BEGLE 1M HIBG 1M BEGLE Common HIBG Common sian ancestry # of training samples # of validation samples # of HL alleles #ofsnps 1 (1M/intersect) 942/ / / / / / /272 BEGLE 1M HIBG 1M BEGLE Common HIBG Common Hispanic ancestry # of training samples # of validation samples # of HL alleles #ofsnps 1 (1M/intersect) 965/ / / / / / /278 BEGLE 1M HIBG 1M BEGLE Common HIBG Common frican ancestry # of training samples # of validation samples # of HL alleles #ofsnps 1 (1M/intersect) 949/ / / / / / /269 BEGLE 1M HIBG 1M BEGLE Common HIBG Common : Illumina Human1M / Common to the Illumina 1M Duo, OmniQuad, OmniExpress, 660K and 550K platforms. 8

9 Table S4: The SNP list used by HL*IMP when Illumina 1M platform is specified. Locus (# of SNPs) Marker list HL (50) rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs915669, rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs166326, rs , rs , rs , rs , rs HL B (39) rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs HL C (27) rs , rs , rs , rs , rs , rs130075, rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs HL DRB1 (50) rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs615672, rs , rs482044, rs660895, rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs HL (34) rs , rs , rs , rs743862, rs , rs , rs , rs482044, rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs , rs

10 Table S5: The sensitivity (SEN), specificity (SPE), positive predictive value (PPV) and negative predictive value (NPV) calculated from validation samples for each four-digit HL allele with call threshold 0.5, when STUDY Data of European ancestry were divided to training and validation parts with equal sizes. The SNP markers in the intersect of Illumina platforms were used. llele 1 Num. Freq. Num. Freq. CR 2 CC 3 SEN SPE PPV NPV Miscall 4 HL : Overall accuracy: 98.7% 01: : : :05 (100) 02: : : : : : :01 (100) 26: :01 (80) 29: : : : : : : : : : :01 (100) 68: : : HL B: Overall accuracy: 97.8% 07: : : : : : : : : :01 (100) 18: :01 (100) 27: :05 (83) 27: : :02 (60) 35: : :01 (80) 35: :01 (100) Continued on next page... 10

11 Table S5 continued from previous page llele 1 Num. Freq. Num. Freq. CR 2 CC 3 SEN SPE PPV NPV Miscall 4 37: : : :01 (67) 39: : : : : : : : :03 (50) 44: : : : : : :03 (50) 50: : :05 (100) 52: : :01 (100) 55: :01 (100) 56: :01 (100) 57: : : HL C: Overall accuracy: 99.2% 01: : : : :04 (100) 03: :03 (100) 04: : :02 (100) 06: : : : : :01 (100) 08: : : :02 (100) 14: : : : Continued on next page... 11

12 Table S5 continued from previous page llele 1 Num. Freq. Num. Freq. CR 2 CC 3 SEN SPE PPV NPV Miscall 4 16: : :01 (100) 17: HL DRB1: Overall accuracy: 94.9% 01: :02 (100) 01: : :01 (100) 03: :01 (100) 04: :07 (80) 04: : :04 (65) 04: :01 (100) 04: :01 (100) 04: :01 (67) 07: :01 (100) 08: : : : : : :01 (100) 11: :01 (100) 11: :01 (93) 12: :01 (100) 12: : :02 (50) 13: : : :01 (100) 15: :01 (100) 15: : : :01 (100) HL DQ1: Overall accuracy: 97.8% 01: :04 (80) 01: :03 (100) 01: :02 (100) 01: : : : :03 (89) 03: : :01 (100) 04: : : Continued on next page... 12

13 Table S5 continued from previous page llele 1 Num. Freq. Num. Freq. CR 2 CC 3 SEN SPE PPV NPV Miscall 4 06: :01 (100) HL Overall accuracy: 99.2% 02: : : :02 (100) 03: : :01 (100) 03: :01 (100) 03: :01 (100) 04: : :03 (100) 05: : : : :04 (100) 06: :02 (100) 06: :01 (50) 06: :04 (100) HL : Overall accuracy: 94.8% 01: : :01 (95) 03: :01 (100) 04: :01 (100) 04: : : : :01 (100) 11: : : : : : : : : the HL alleles with more than one copy and non-zero sensitivity in the training are listed. 2 : CR call rate. 3 : CC allele accuracy. 4 : the most likely miscalled allele and the proportion of the most likely miscalled allele in all miscalled alleles. 13

14 Table S6: The sensitivity (SEN), specificity (SPE), positive predictive value (PPV) and negative predictive value (NPV) calculated from validation samples for each four-digit HL allele with call threshold 0.5, when STUDY Data of sian ancestry were divided to training and validation parts with equal sizes. The SNP markers in the intersect of Illumina platforms were used. llele 1 Num. Freq. Num. Freq. CR 2 CC 3 SEN SPE PPV NPV Miscall 4 HL : Overall accuracy: 93.8% 01: : :07 (67) 02: : : :01 (100) 02: : : : :02 (100) 11: : : : : : : : : : : HL B: Overall accuracy: 94.7% 07: : : : : : : :01 (100) 15: :35 (100) 15: : : : : : : : :01 (100) 35: : : Continued on next page... 14

15 Table S6 continued from previous page llele 1 Num. Freq. Num. Freq. CR 2 CC 3 SEN SPE PPV NPV Miscall 4 38: : :02 (100) 40: : :06 (100) 40: :04 (100) 44: : : :02 (100) 47: : : : :02 (100) 51: :01 (100) 51: : :01 (100) 54: :02 (100) 57: : : :02 (100) HL C: Overall accuracy: 97.8% 01: :01 (50) 02: : : :04 (100) 03: :03 (100) 04: : : : : : :01 (50) 07: : : : : : : : : HL DRB1: Overall accuracy: 95.8% 01: : : : : Continued on next page... 15

16 Table S6 continued from previous page llele 1 Num. Freq. Num. Freq. CR 2 CC 3 SEN SPE PPV NPV Miscall 4 04: :10 (50) 04: : : : : : : : : :01 (67) 12: : : : : :04 (100) 14: :01 (100) 14: : :04 (100) 15: :02 (100) 15: : HL DQ1: Overall accuracy: 90.0% 01: :04 (83) 01: :03 (67) 01: :02 (100) 01: : : : :02 (56) 03: :03 (100) 03: :01 (100) 04: : :05 (80) 05: : : :05 (100) 06: HL Overall accuracy: 98.1% 02: : :03 (100) 03: :03 (100) 03: : :01 (100) 04: :02 (100) 04: :01 (100) 05: Continued on next page... 16

17 Table S6 continued from previous page llele 1 Num. Freq. Num. Freq. CR 2 CC 3 SEN SPE PPV NPV Miscall 4 05: :01 (100) 05: : : : : : HL : Overall accuracy: 95.3% 01: :01 (100) 02: :01 (67) 02: :01 (100) 03: : :01 (50) 04: : :01 (100) 09: : :01 (50) 13: :02 (100) 14: : : : :01 (100) 1 : the HL alleles with more than one copy and non-zero sensitivity in the training are listed. 2 : CR call rate. 3 : CC allele accuracy. 4 : the most likely miscalled allele and the proportion of the most likely miscalled allele in all miscalled alleles. 17

18 Table S7: The sensitivity (SEN), specificity (SPE), positive predictive value (PPV) and negative predictive value (NPV) calculated from validation samples for each four-digit HL allele with call threshold 0.5, when STUDY Data of Hispanic ancestry were divided to training and validation parts with equal sizes. The SNP markers in the intersect of Illumina platforms were used. llele 1 Num. Freq. Num. Freq. CR 2 CC 3 SEN SPE PPV NPV Miscall 4 HL : Overall accuracy: 96.0% 01: : : : : : : : : : : : : : : : : :17 (100) 68: : HL B: Overall accuracy: 93.8% 07: : : : : : :43 (100) 15: : : : : : : : : : : : : : Continued on next page... 18

19 Table S7 continued from previous page llele 1 Num. Freq. Num. Freq. CR 2 CC 3 SEN SPE PPV NPV Miscall 4 44: :01 (100) 44: : : : : : :01 (50) 53: : : HL C: Overall accuracy: 98.4% 01: :01 (100) 02: : : : : : : : : : : : : : : : : : HL DRB1: Overall accuracy: 93.5% 01: :03 (100) 01: : : :01 (50) 04: :07 (100) 04: : : : : :02 (100) 08: : : : : Continued on next page... 19

20 Table S7 continued from previous page llele 1 Num. Freq. Num. Freq. CR 2 CC 3 SEN SPE PPV NPV Miscall 4 13: : : : :06 (100) 14: :01 (50) 15: : : : HL DQ1: Overall accuracy: 95.8% 01: : :01 (100) 01: :02 (100) 01: :01 (100) 02: : : :01 (100) 03: :01 (100) 04: : : : HL Overall accuracy: 98.9% 02: : :01 (100) 03: : : : : : : : : : HL : Overall accuracy: 97.5% 01: : :01 (100) 02: : : : : : : : Continued on next page... 20

21 Table S7 continued from previous page llele 1 Num. Freq. Num. Freq. CR 2 CC 3 SEN SPE PPV NPV Miscall 4 104: : the HL alleles with more than one copy and non-zero sensitivity in the training are listed. 2 : CR call rate. 3 : CC allele accuracy. 4 : the most likely miscalled allele and the proportion of the most likely miscalled allele in all miscalled alleles. 21

22 Table S8: The sensitivity (SEN), specificity (SPE), positive predictive value (PPV) and negative predictive value (NPV) calculated from validation samples for each four-digit HL allele with call threshold 0.5, when STUDY Data of frican ancestry were divided to training and validation parts with equal sizes. The SNP markers in the intersect of Illumina platforms were used. llele 1 Num. Freq. Num. Freq. CR 2 CC 3 SEN SPE PPV NPV Miscall 4 HL : Overall accuracy: 100% 01: : : : : : : : : : : : : : : : HL B: Overall accuracy: 96.7% 07: : : : : : : : : : : HL C: Overall accuracy: 96.5% 02: :10 (100) 02: : : : : : : : : : Continued on next page... 22

23 Table S8 continued from previous page llele 1 Num. Freq. Num. Freq. CR 2 CC 3 SEN SPE PPV NPV Miscall 4 18: HL DRB1: Overall accuracy: 100% 01: : : : : : : : : : : : : HL DQ1: Overall accuracy: 97.2% 01: : : : : : HL Overall accuracy: 97.7% 02: : : : : : : :01 (100) 06: : HL : Overall accuracy: 75.0% 01: : : the HL alleles with more than one copy and non-zero sensitivity in the training are listed. 2 : CR call rate. 3 : CC allele accuracy. 4 : the most likely miscalled allele and the proportion of the most likely miscalled allele in all miscalled alleles. 23

24 Table S9: The accuracies calculated from the ethnic-specific and multi-ethnic models. For each ethnicity, STUDY Data were divided into training and validation sets with equal sizes. The multi-ethnic model was built using all training samples from multiple ethnicities, whereas the ethnic-specific models were built using the training part of each ethnicity respectively. No call threshold was executed. HL B C DRB1 DQ1 European ancestry multi-ethnic model ethnic-specific model sian ancestry multi-ethnic model ethnic-specific model Hispanic ancestry multi-ethnic model ethnic-specific model frican ancestry multi-ethnic model ethnic-specific model