What Ultima Genomics Makes, and How to Work With Its Data

Ultima Genomics is a private American company, headquartered in Newark, California and founded by Gilad Almogy, that builds high-throughput DNA sequencers: the UG 100 and its successor UG 200 series. It is not publicly traded, and it is owned by its founders, employees, and venture investors. The company had raised roughly $600 million disclosed by its 2022 emergence from stealth, with further rounds since. The reason a technical reader should care is narrower than the funding: Ultima’s instrument produces reads with a different shape and a different error profile from Illumina’s, and if you are handed an Ultima CRAM you cannot run your usual pipeline unchanged and expect correct indel calls.

How the instrument works

Conventional short-read sequencers run chemistry inside a sealed flow cell, a fluidic channel whose walls and volume set the reagent cost per base. Ultima removes the enclosure. Library molecules are clonally amplified onto beads, the beads are deposited in a dense monolayer on a circular silicon wafer, and the wafer spins while reagents are dispensed onto the open surface and spun off. Reagent volume scales with the film on the disc rather than with a channel, which is where most of the cost reduction comes from, and imaging is done on the flat wafer at high optical throughput ¹.

The chemistry is flow-based rather than reversible-terminator. Ultima calls it mostly natural sequencing by synthesis. Each flow delivers a single nucleotide species in a repeating TGCA order, as a mixture of unlabeled natural bases and a minority of labeled, blocked bases. A homopolymer run is read as a single event whose signal intensity encodes its length, exactly as in the older pyrosequencing and ion-sensing platforms, rather than as a sequence of discrete cycles. Reads are single-end, in the neighborhood of 300 bases, and a wafer run produces billions of them. Ultima’s public target has been on the order of $1 per gigabase, which translates to roughly $100 for a 30x human genome in consumables at scale, a claim that drew both attention and skepticism when the first preprints appeared without a peer-reviewed instrument description ². Ultima was one of several companies that broke the effective monopoly on high-throughput short reads in 2022 ³, and cheaper fluidics remains an active area of instrument design beyond Ultima, including roll-to-roll approaches that move the substrate continuously past reagent stations ⁴.

What the files look like

Ultima does not hand you paired FASTQ. The native output is an unaligned CRAM (or BAM) per sample, single-end, carrying flow-specific auxiliary tags. The two you will encounter are tp and t0. tp is a per-base tag describing how confident the caller is that a base belongs in its homopolymer run, where positive and negative values indicate the base would be added or removed in the next most likely hypothesis. t0 is a base-level estimate of the probability that a zero-length homopolymer was called as one base. The read group records the flow order. Standard BQ-style per-base quality still exists, but it is a projection of a homopolymer-level uncertainty onto individual bases, so treating it as an independent per-base error probability will mislead you.

Two practical consequences follow. First, because reads are single-end, you lose insert-size information. There is no fragment-length distribution and no mate-pair rescue for repetitive alignment. Structural variant calling depends entirely on split reads and depth rather than on discordant pairs. Methods that hunt structural variants in tumors have to be tuned accordingly, and the callsets that feed driver-discovery methods built on structural variant enrichment are only as good as that tuning ⁵. Second, duplicate marking is weak. Picard MarkDuplicates on single-end data keys on 5’ position and strand, so at 30x with 300 bp reads a meaningful fraction of independent fragments collide and get flagged. For germline work we simply skip duplicate marking on Ultima WGS. For anything where duplicate rate matters, such as cell-free DNA or deep panels, use a UMI library and deduplicate on the molecular barcode instead.

The pipeline we would run

Align with BWA-MEM, or with Ultima’s own UA aligner if you have it, then keep the flow tags through the entire pipeline. Any step that rewrites reads and drops tp and t0 silently degrades your indel calls downstream, and this is the most common way a first Ultima analysis goes wrong.

For germline small variants we would use DeepVariant with the Ultima-trained checkpoint (--customized_model) rather than the default WGS model, because the default model has learned an Illumina error prior that does not describe flow data. The alternative is GATK HaplotypeCaller in flow mode, which since GATK 4.3 includes a flow-based HMM likelihood engine. Run with --flow-mode STANDARD and set --flow-order-for-annotations TGCA, and expect noticeably longer runtimes than the standard pair-HMM. Sentieon’s DNAscope also ships Ultima support if you need the throughput. Whichever you choose, benchmark against a Genome in a Bottle sample with hap.py and stratify the results, because a single aggregate F1 will hide the one region class where the platform differs.

Where the errors are

Flow chemistry moves error out of substitutions and into indels. Single-nucleotide error rates are low, because a labeled base either incorporates or does not. But homopolymer length estimation degrades as runs get longer, so the dominant error mode is a one-base insertion or deletion inside a homopolymer of roughly four bases or more. Published evaluations of the UG 100 for clinical service work report SNV performance close to established short-read platforms, with the residual gap concentrated in indels and in homopolymer-rich context ⁶. A multi-modal comparison against NovaSeq on clinically relevant FFPE tissue reaches the same structure of conclusion: high concordance for substitutions, with platform-specific behavior in insertions and deletions ⁷. Reviews of recent short-read innovation place Ultima in this category generally, as a platform whose clinical adoption depends on context-aware variant filtering rather than on raw accuracy headlines ⁸.

The flip side is that a low substitution error rate is precisely what you want for detecting rare somatic variants in a background of normal DNA. Simulation work on cell-free DNA whole-genome sequencing shows that the tradeoff between depth and per-base error rate determines the detection floor for circulating tumor DNA, and that lowering error rate buys sensitivity that extra coverage cannot ⁹. Ultima’s economics and its error profile point in the same direction in this application: cheap, deep, substitution-clean genomes.

For RNA sequencing, 300 bp single-end reads are fine for expression quantification and adequate for most differential-expression work. They are less good for isoform-level inference, where paired ends help resolve junction combinations. If you are designing barcoded libraries for single-cell work on Ultima, avoid homopolymer runs in the barcode and UMI sequences, since a one-base homopolymer error is the exact failure this chemistry produces and it will scramble cell assignment.

Anything you infer from these data about your own health is a measurement. It is not a diagnosis. Variant interpretation with clinical consequences belongs with a clinical laboratory and a genetics professional, working from a CLIA-validated assay.

Questions people also ask

Is Ultima Genomics publicly traded? No. It is a private company, so there is no ticker and no public financial reporting.

How much funding has Ultima Genomics received, and who owns it? Ultima disclosed roughly $600 million raised when it emerged from stealth in 2022, with additional rounds since. Ownership sits with its founders, employees, and venture and strategic investors.

Who founded Ultima Genomics and where is it based? Gilad Almogy founded the company. It is headquartered in Newark, California, in the San Francisco Bay Area.

What does Ultima Genomics make? Sequencing instruments and their consumables, namely the UG 100 and the newer UG 200 series. They are built around a spinning silicon wafer, flow-based chemistry, and single-end reads of about 300 bases.

Is Ultima a good platform to sequence on? For high-depth whole genomes where cost per gigabase dominates and substitutions are the variant class of interest, it is a strong choice. For indel-sensitive clinical work, homopolymer-rich targets, or isoform-resolved RNA, plan on platform-specific models and validation rather than assuming your Illumina pipeline transfers.

Footnotes

1. Elizabeth Pennisi. Upstart DNA sequencers could be a ‘game changer’. Science, 2022. https://doi.org/10.1126/science.add4867 ↩

2. Julianna LeMieux. Ultima Unveiling: Big Black Box Sparks Preprint Debate. GEN Biotechnology, 2022. https://doi.org/10.1089/genbio.2022.29050.jle ↩

3. Helen Albert. Top 5 Genetic Sequencing Startups Shaking Up the Industry. Inside Precision Medicine, 2022. https://doi.org/10.1089/ipm.09.06.09 ↩

4. Yanzhe Qin, Stephan A. Koehler, Yunyan Ling, et al. Fast, cost-effective and flexible DNA sequencing by roll-to-roll fluidics. Nature Methods, 2025. https://doi.org/10.1038/s41592-025-02730-2 ↩

5. Xavi Loinaz, Johnathan Dagan, Jean-Baptiste Alberge, et al. SVelfie: A method for discovering cancer drivers based on enrichment of likely functional structural variants. 2025. https://doi.org/10.1101/2025.03.09.642284 ↩

6. Luca Santuari, Ilya Kolpakov, Anthony Blin, et al. Evaluation of the Ultima Genomics UG 100™ sequencing platform for clinical services. 2025. https://doi.org/10.64898/2025.12.11.693628 ↩

7. Quentin Bayard, Zelda Mariet, Anna Schaar, et al. Multi-modal benchmarking of the Ultima UG100 and Illumina NovaSeq sequencing platforms using clinically relevant FFPE tissues. 2026. https://doi.org/10.64898/2026.03.11.710846 ↩

8. Katarzyna Polonis, Joseph H Blommel, Andrew E O Hughes, et al. Innovations in Short-Read Sequencing Technologies and Their Applications to Clinical Genomics. Clinical Chemistry, 2025. https://doi.org/10.1093/clinchem/hvae173 ↩

9. Li-Ting Chen, Jeroen de Ridder, Myrthe Jager. Large-scale simulation of coverage and error rate tradeoffs for cancer detection in cell-free DNA whole-genome sequencing. Bioinformatics, 2026. https://doi.org/10.1093/bioinformatics/btag460 ↩