DeepVariant training data

March 5, 2026 ยท View on GitHub

WGS models

versionReplicates#examples
v0.49 HG00185,323,867
v0.59 HG001
2 HG005
78 HG001 WES
1 HG005 WES(1)
115,975,740
v0.610 HG001 PCR-free
2 HG005 PCR-free
4 HG001 PCR+
156,571,227
v0.710 HG001 PCR-free
2 HG005 PCR-free
4 HG001 PCR+
158,571,078
v0.812 HG001 PCR-free
2 HG005 PCR-free
4 HG001 PCR+
(and, more dowsample_fraction since last version)
346,505,686
v0.910 HG001 PCR-free
2 HG005 PCR-free
2 HG006 PCR-free
2 HG007 PCR-free
5 HG001 PCR+
325,202,093
v0.1010 HG001 PCR-free
2 HG005 PCR-free
2 HG006 PCR-free
2 HG007 PCR-free
5 HG001 PCR+
339,410,078
v1.011 HG001
2 HG005-HG007
2 HG002-HG004(7)
317,486,837
v1.112 HG001
3 HG002
3 HG004
3 HG005
3 HG006
3 HG007(9)
388,337,190
v1.212 HG001
6 HG002(12)
6 HG004(12)
3 HG005
3 HG006
3 HG007
518,709,296
v1.3Same model as v1.2
v1.412 HG001
6 HG002(12)
6 HG004(12)
3 HG005
3 HG006
3 HG007
517,209,566
v1.513 HG001
14 HG002
8 HG004
9 HG005
4 HG006
4 HG007
815,200,320
v1.621 HG001
17 HG002
8 HG004
9 HG005
4 HG006
4 HG007
929,199,066
v1.8Same model as v1.6
v1.911 HG001
17 HG002
7 HG004
8 HG005
3 HG006
3 HG007
942,514,071
v1.1011 HG001
17 HG002
8 HG004
8 HG005
3 HG006
3 HG007
942,514,071

--

WES models

versionReplicates#examples
v0.578 HG001
1 HG005
15,714,062
v0.678 HG001
1 HG005(2)
15,705,449
v0.778 HG001
1 HG005
15,704,197
v0.878 HG001
1 HG005(3)
18,683,247
v0.981 HG001
1 HG005(3)(4)(5)
61,953,965
v0.10Same model as v0.9
v1.032 HG001
9 HG002
6 HG003
6 HG004
12 HG005
9 HG006
9 HG007(7)
10,716,281
v1.141 HG001
9 HG002
6 HG004
12 HG005
9 HG006
9 HG007(9)
13,450,688
v1.241 HG001
9 HG002
9 HG004
12 HG005
9 HG006
9 HG007(11)
22,288,064
v1.3Same model as v1.2
v1.441 HG001
9 HG002
9 HG004
12 HG005
9 HG006
9 HG007(11)
21,212,424
v1.540 HG001
9 HG002
9 HG004
12 HG005
9 HG006
9 HG007
21,027,625
v1.657 HG001
9 HG002
9 HG004
12 HG005
9 HG006
9 HG007
21,027,614
v1.858 HG001
9 HG002
9 HG004
11 HG005
9 HG006
9 HG007
25,598,763
v1.957 HG001
9 HG002
9 HG004
11 HG005
9 HG006
9 HG007
25,598,763
v1.1057 HG001
9 HG002
9 HG004
11 HG005
9 HG006
9 HG007
25,598,763

PACBIO models

versionReplicates#examples
v0.816 HG002160,025,931
v0.949 HG002 (6)357,507,235
v0.1049 HG002, 2 HG003, 2 HG004, 1 HG002 (amplified) (6)472,711,858
v1.01 HG001
2 HG002
2 HG003
2 HG004
1 HG005 (8)
302,331,948
v1.11 HG001
9 HG002
2 HG004
1 HG005(9)
569,225,616
v1.21 HG001
19 HG002
2 HG004
1 HG005(10)
1,036,056,726
v1.31 HG001
19 HG002
3 HG004
1 HG005
1 HG006
1 HG007
1,177,109,190
v1.41 HG001
19 HG002
3 HG004
1 HG005
1 HG006
1 HG007
1,177,596,708
v1.53 HG001
29 HG002
7 HG004
2 HG005
3 HG006
2 HG007
1,729,659,396
v1.66 HG001
60 HG002
16 HG004
4 HG005
6 HG006
4 HG007
3,195,507,862
v1.83 HG001
10 HG002
4 HG004
5 HG005
0 HG006
0 HG007
416,516,418
v1.93 HG001
10 HG002
4 HG004
5 HG005
0 HG006
0 HG007
416,516,418
v1.103 HG001
17 HG002
4 HG004
2 HG005
2 HG006
2 HG007
964,634,414

ONT models

versionReplicates#examples
v1.63 HG002
1 HG004
1 HG005
534,302,654
v1.87 HG002
1 HG004
1 HG005
1,591,950,794
v1.97 HG002
1 HG004
1 HG005
1,591,950,794
v1.102 HG001
2 HG002
2 HG004
2 HG005
2 HG006
2 HG007
1,008,626,452

HYBRID models

versionReplicates#examples
v1.010 HG002
1 HG004
1 HG005
1 HG006
1 HG007
193,076,623
v1.1Same model as v1.0
v1.210 HG002
1 HG004
1 HG005
1 HG006
1 HG007
214,302,681
v1.3Same model as v1.2
v1.410 HG002
1 HG004
1 HG005
1 HG006
1 HG007
215,863,645
v1.510 HG002
1 HG004
1 HG005
1 HG006
1 HG007
215,863,664
v1.610 HG002
1 HG004
1 HG005
1 HG006
1 HG007
215,353,081
v1.8Same model as v1.6
v1.9Same model as v1.6
v1.10Same model as v1.6

(1): In v0.5, we experimented with adding whole exome sequencing data into training data. In v0.6, we took it out because it didn't improve the WGS accuracy.

(2): The training data are from the same replicates as v0.5. The number of examples changed because of the update in haplotype_labeler.

(3): In v0.8, we used the Platinum Genomes Truthset to create more training examples outside the GIAB confident regions.

(4): Previously, we split train/tune by leaving 3 WES for tuning. Starting from this release, we leave out chr1 and chr20 from training, and use chr1 for tuning.

(5): Starting from this version, we padded (100bps on both sides) of the capture BED and used that for generating training examples. We also added more downsample_fraction.

(6): (Before v1.0) PacBio is the only one we currently uses HG002 in training and tuning.

(7): In v1.0, we train on HG002-HG004 for WGS as well, but only using examples from the region of NIST truth confident region v4.2 subtracting v3.3.2.

(8): In v1.0, PacBio training data contains training examples with haplotag sorted images and unsorted images.

(9): In v1.1, we exclude HG003 from training. And we use all NIST truth confident regions for HG001-HG007 (except for HG003) for training. We've always excluded chr20-22 from training.

(10): In v1.2, we include new PacBio training data from Sequel II, Chemistry 2.2.

(11): Between v1.1 and v1.2, we fixed an issue where make_examples can generate fewer class 0 (REF) training examples than before. This is the reason for more training examples in v1.2 when number of samples didn't increase.

(12): In v1.2, we created BAM files with 100bp reads and 125bp reads by trimming to augment the training data.

Training data:

See "An Extensive Sequence Dataset of Gold-Standard Samples for Benchmarking and Development" for a publicly available set of data we released. Data download information can be found in the supplementary material.