Ultima Genomics - This is how you come out of Stealth

Ultima Genomics is out of Stealth mode and announcing $600M in funding. I wrote about Ultima last year and got most things right based on their IP. So, if you’d like to hear about genomics companies than a year in advance of their public launch, you should probably consider subscribing!

Ultima’s exit from stealth is in stark contrast to Pleno’s. Ultima have posted a comprehensive description of the platform and 4 strong publications describing various applications. The white papers are easy to read I’d recommend taking a look if you’re interested in the details.

In this post I’m going to give my opinion on the three core parts of the platform. The optics, the surface and the chemistry and see where they stand in contrast to Illumina high throughput offering (the NovaSeq 6000 S4 Flowcell).

For those interested in how I think this effects Reticula. Ultima’s COGS per run is probably >$300 and Reticula’s is $1, we’re aiming for very different markets.. Feel free to contact me (nava@reticula.bio) to hear more.

Unlike both Element and Singular, Ultima are going after high-end (NovaSeq class) instruments. To do this they’ve had to address some key issues with new approaches (and not just work off expired Illumina IP).

Optics/Stage/Fluidics

Ultima use a rotating stage. As previously noted, I like this. It means that after an imaging cycle your optics are aligned to start the next cycle. I suspect you can also use line scan CCD sensors (as used in the Illumina HiSeq etc.) to do continuous imaging (TDI). TDI is one of the key improvements Illumina introduced after the original Solexa Genome Analyzer series instruments. It lets you image the flowcell far more efficiently.

The other neat thing is the fluidic approach. They release the reagents at the center of the surface and then essentially spin coat them across the surface. It sounds pretty crazy, but very neat and certainly should give a nice uniform delivery for reagents and no issues with reagents getting caught in voids and diffusing during the run.

Overall, the imaging approach used should provide NovaSeq class throughput.

Surface

Ultima are using a 200mm silicon wafer, a bare wafer costs ~$50. The wafer is patterned at micron resolution with electrostatic landing pads. Micron sized features are huge and I suspect this can be done cost effectively. Still, this likely means their density is lower than Illumina’s.

Illumina S4 flowcells output 10B reads. From images, I’d guess the active surface area of an S4 flowcell is ~3000 square mm. That’s 3.3M reads per square mm. Or 3 per square micron.

I suspect Illumina therefore have at least a 3x density advantage over Ultima. A 200mm wafer has an area of 70709 square mm. From which Ultima state also can get 10B reads. Ultima therefore seem to have a substantially lower density than Illumina, but make up for it by using a larger surface.

In the end this come down to manufacturing costs. The Ultima surface needs to be more than 3x cheaper to manufacture than Illumina’s.

Chemistry

Ultima does appear to be using the chemistry approach I previously discussed, and as I proposed there, they are indeed cleaving fluorophores.

To briefly summarize, Ultima use an unterminated single channel chemistry. This is similar to 454 and Ion Torrent. Each nucleotide is flowed in, incorporated and detected sequentially.

If a homopolymer exists in the read multiple nucleotides will be incorporated and there will be increased fluorescence. Typically this would make homopolymers error prone as it’s difficult to tell the difference between the fluorescence intensity of say AAA and AAAA. One issue is that if a strand contains multiple fluorescently labeled nucleotides the dyes can interact and the intensity increase is non-linear.

Ultima try to get round this by using a mixture of labeled and unlabeled nucleotides. So they get less overall signal, but in most cases only a single fluorophore on each strand…

But they also use machine learning:

“To accommodate for systematic variance, we take advantage of recent significant advances in machine learning as well as the large amount of data involved in each run to employ a deep convolutional neural network (CNN) to convert raw signals into sequence reads. The CNN is trained offline on a large, diverse dataset combining data from multiple runs, and it is then recalibrated on a smaller sample of genomic reads from the current run.”

This makes me uncomfortable. The Illumina basecalling process doesn’t need to use machine learning to generate accurate reads. For the most part I would prefer the sequencer to be giving an unbiased “ground truth” read out. In particular it seems concerning that they are doing fitting to genomic reads on a per-run basis. If you perform run specific calibration, this would be better approached using a spike in sequence (such as the PhiX spike-in used in Illumina sequencing).

Even so homopolymer calls seem to be poor: "homopolymer calling accuracy is at 99.5% for homopolymer lengths of 1-2 and decreases to 90% at homopolymer lengths of 8”.

This seem poor compared to Illumina, where the overall error rate is <0.5% and only about 10% of these are “post-homopolymer” errors. Homopolymer error rates at length 8 seem to be ~1% in Illumina sequencing compared to 10% for Ultima.

Overall I find the chemistry the least attractive part of the Ultima approach. I don’t see the proposed advantages over Illumina style SBS as being very compelling. Particularly not now it is off patent, and not with Chemistry X coming.

Run time

Ultima state a 20h run time for 10B reads that average 282bp. Unlike Illumina sequencing, read length will vary. An Illumina 300bp NovaSeq S4 300bp (2x150bp) run takes 44h. So they seem to have a run time advantage! This is particular impressive because they need to do more cycles (444 compared to 300).

It seems likely that Ultima’s imaging approach is helping them out here!

Costs

The whitepaper suggests a cost of $1/Gb. This would mean the run presented would cost $3000. An NovaSeq S4 flowcell costs something like $16000. Illumina don’t break out consumable margins, but they’re generally in the 70 to 90% range.

I suspect Illumina could sell a NovaSeq S4 run for around $3000 if they really wanted to.

A direct cost of goods comparison is harder. It would be great to see a detailed analysis. But I suspect the Ultima COGS is cheaper, but perhaps not by an order of magnitude cheaper.

Conclusion

Ultima have created a solid platform which seems competitive with the NovaSeq. Data quality doesn’t seem better than Illumina. This means that they will have to compete on price. But Illumina margins are high.

Last year I wrote about the next few years in sequencing. And I don’t think Ultima changes the story much. Illumina will likely be forced to reduce their margins to keep market share. But I suspect will try and avoid being drawn into a race to the bottom.

In particular I suspect Ultima’s data quality isn’t quite as good as Illumina’s. So Illumina can choose to retain slightly higher margins for those customers who really want the highest quality data, while giving up the low and mid-range market.

This is however, just the first generation of the Ultima platform and with some tweaks perhaps they will be able to match or beat Illumina on all metrics…