This quarter brings lower-cost runs on standard cloud CPUs and extends the Sentieon pangenome pipeline to hybrid short- and long-read data and to more sequencing platforms. sentieon-cli 2.0 adds a DNAscope Hybrid Pangenome pipeline, and new platform-specific models bring pangenome accuracy to Element, Complete Genomics and GeneMind data.

· Runtime and cost on AWS: a whole genome in under 40 mins on the larger instances, or for about $1 of spot compute in about 1.3 hours on a balanced AWS Graviton instance (c9g.16xlarge).

· Hybrid Pangenome pipeline (sentieon-cli 2.0): combines Illumina with ONT or PacBio long reads on the pangenome graph. 30x depth PacBio reduces total errors by 50% versus Illumina alone, and even 10x depth of long reads reduces them by about 35%.

· Element Pangenome 1.0 model: errors reduced by 24% compared with our earlier Element benchmark, on data from Element’s new VITARI system.

· Complete Genomics CoolMPS SE600: under 30,000 total errors on HG002 against the NIST v5.0q benchmark, half the errors of 35x Illumina data.

· GeneMind Pangenome 1.0 model: errors reduced by 26% compared with 35x Illumina data on the NIST v5.0q benchmark, and SV F1 of 0.959.

Section 1: Runtime and Cost on AWS — Under 40 min, or About $1 per Genome

The Sentieon DNAscope Pangenome pipeline runs on standard CPUs, with no specialized hardware, so you can have both speed and cost. On the largest instances, a whole genome finishes from FASTQ to VCF in under an hour: 53 minutes on c8a.48xlarge. On a balanced instance, it costs about $1: AWS Graviton c9g.16xlarge completes a genome in about 1.3 hours for about $1 of spot compute (about $3.70 on demand).

We benchmarked HG002 whole-genome analysis on 23 AWS instance types across AWS Graviton (c8g, c9g), AMD EPYC (c6a, c7a, c8a) and Intel Xeon (c8i, m8i) processors, at spot and on-demand prices (Figure 1); the dashed line marks the best trade-offs. NUMA-aware lanes, now in development, run alignment as parallel lanes, each pinned to one processor’s cores and local memory, which avoids slow cross-socket memory access on the largest instances: on c8a.48xlarge they cut wall time from 53 to 39 minutes, and compute cost to $2.78 at spot prices and $6.77 on demand, with the same accuracy.

Section 2: DNAscope Hybrid Pangenome Pipeline — Short and Long Reads on the Pangenome Graph

Sentieon-cli 2.0 introduces the DNAscope Hybrid Pangenome pipeline. A single command takes Illumina short reads together with ONT or PacBio HiFi long reads, maps the short reads against the HPRC v2.1 pangenome graph, and calls small variants, structural variants and copy number variants on GRCh38 or CHM13. Two new models power it: Hybrid Pangenome Illumina + ONT 1.0, and Illumina + PacBio 1.0, which covers both Revio and SPRQ chemistry.

On HG003, a held-out sample, adding long reads to 35x PCR-free Illumina data cuts total SNP and indel errors from 9,415 to 4,753 with 30x PacBio (−50%) and to 5,177 with 20x ONT (−45%). SNP errors fall by about half with either technology, and PacBio also reduces indel errors by 42%, or by about 65% when the short-read library is PCR-plus. On the NIST v5.0q benchmark with the CHM13 reference, HG002 total errors fall from 60,363 with Illumina alone to 32,431 with 30x PacBio (−46%) and 47,761 with 20x ONT (−21%).

Adding just 5x of PacBio or ONT to 35x Illumina reduces total errors by about 20%, and 10x reduces them by about 35%, so a low-depth long-read run captures most of the benefit. The Hybrid Pangenome pipeline also improves on our linear reference DNAscope Hybrid: on identical 35x Illumina + 5x ONT data and the same software release, total errors fall from 12,366 to 7,749 (−37%).

Section 3: A Dedicated Pangenome Model for Element WGS

The new Element Pangenome 1.0 model is trained on 29 Element runs from 12 libraries and 3 GIAB samples, covering AVITI data from 2022 to 2024, and data from Element’s new high-throughput VITARI system (2026), PCR-free and PCR-plus libraries, and depths from 10x to 77x.

On the NIST v4.2.1 benchmark, Element VITARI HG002 (held-out library) at 30x reaches 6,187 total SNP and indel errors, with SNP F1 of 0.9993 and indel F1 of 0.9988. Compared with our earlier Element benchmark (2023 Cloudbreak HG002 data at 42x with the earlier pipeline), total errors are reduced by 24%, even at lower depth (30x vs 42x). On the NIST v5.0q benchmark with the CHM13 reference, Element VITARI HG002 at 30x reaches SNP F1 0.9972 and indel F1 0.9881, with 39,929 total errors, compared with 60,363 for 35x Illumina data on the same reference in our 2026Q2 datasheet. Most of the difference comes from indels.

The pipeline also calls SVs and CNVs from Element data. On Element HG002 data (PCR-free, 42x), SV F1 is 0.955 on the NIST v5.0q benchmark (precision 0.976, recall 0.934) and 0.899 in the challenging medically relevant genes (CMRG). CNV F1 is 0.885 overall, 0.920 for copy losses and 0.762 for copy gains.

Section 4: Pangenome Accuracy with 600-Base Single-End Reads on Complete Genomics CoolMPS 600 Chemistry

The Complete Genomics CoolMPS 600 sequencer produces SE600 reads: a single, contiguous 600-base read, compared with at most 300 bases split across a PE150 read pair. A longer read is more likely to contain a paralog-specific variant, so it maps uniquely in segmental duplications, tandem repeats and pseudogenes, where a 150-base read would be placed at random or dropped. Fewer mis-mapped reads mean fewer false SNP and indel calls.

On the NIST v4.2.1 benchmark, PCR-free SE600 data at 35–55x reaches between 3,251 and 8,025 total SNP and indel errors across all seven GIAB samples, with SNP F1 of 0.999 and indel F1 of 0.998–0.999. HG003, which was held out from model training, reaches SNP F1 0.999 and indel F1 0.998. On the NIST v5.0q benchmark with the GRCh38 reference, HG002 at 40x reaches 29,124 total SNP and indel errors, half the 58,165 errors of 35x Illumina data on the same reference in our 2026Q2 datasheet. Most of the reduction comes from indels, with 66% fewer missed indels (10,350 vs 30,005).

Section 5: A Dedicated Pangenome Model for GeneMind WGS

The new GeneMind Pangenome 1.0 model is a dedicated DNAscope Pangenome model for GeneMind whole-genome data, trained on PCR-free and PCR-plus libraries from four GIAB samples (HG001, HG002, HG004 and HG005) downsampled to 15–45x; a fifth sample, HG003, was held out for evaluation.

On the NIST v4.2.1 benchmark, PCR-free GeneMind WGS at 30x reaches between 5,636 and 6,423 total SNP and indel errors on the four training samples, with SNP F1 of 0.9992–0.9993 and indel F1 of 0.9983–0.9990. HG003, which was held out from training, reaches SNP F1 0.9985 and indel F1 0.9969. On the NIST v5.0q benchmark with the GRCh38 reference, GeneMind HG002 at 30x reaches 43,178 total SNP and indel errors, 26% fewer than the 58,165 errors of 35x Illumina data on the same reference in our 2026Q2 datasheet, with 40% fewer missed indels (18,049 vs 30,005). GeneMind data with Sentieon also reaches SNP F1 0.9966 and indel F1 0.9932 on the Chinese Quartet reference samples.

On the same GeneMind HG002 data, SV F1 is 0.959 against the NIST v5.0q benchmark (precision 0.979, recall 0.940), and CNV F1 is 0.919, with 0.936 for copy losses and 0.870 for copy gains.