Roche Gives SBX Updates - and a Name!

Last week I double-dipped on conferences, going from London Calling to European Society for Human Genetics (ESHG) in Milan.  I have a raft of notes and ideas from these, which I'll try to spool out over the next week or two before jumping to a long list of more whimsical ideas I've jotted down.   First up are some updates on Roche's SBX sequencing technology, which has now been christened Axelios - which Nava Whiteford reported in his ASeq newsletter.

Roche had an ESHG corporate session mostly focused on SBX, with talks by SBX guru Mark Kokoris as well as Sean Hofherr from Broad Clinical Labs (a third talk covered the Avenio Edge robotics system but not in the context of SBX).  I didn't attend - ESHG is a constant dilemma of which simultaneous sessions to attend - but everything is on video for registrants.  One note: you can't find the corporate videos directly through the ESHG app, but must first get to the separate video site using a link for one of the non-corporate videos, then navigate there to find the corporate sessions.  Roche also had their Diagnostics Day, which is what Nava reported on and where the Axelios name was released.  The ESHG sessions were strongly focused on technical merits with little discussion of anything commercial - and the Diagnostics Day presenters shut down any questions about pricing or precise launch timing.

For the TL;DR crowd: everything is incrementally better.  Insert lengths for the fast (was 190 now 197) and duplex (was 231 now 248) libraries are now slightly longer.  F1 scores have improved for both simplex and duplex modes.  For fast and duplex, Roche is now describing accuracy as "Q40".  Speed has continued to improve; clinical sample to VCF in under five hours for a single sample and under eight hours for a trio - in part because basecalling and (for duplex) consensus generation, and mapping to a reference occur in parallel to data generation.  Kokoris also introduced data on workflows starting with FFPE and cell-free DNA.

There was also discussion of a simplex workflow (SBX-SL) optimized to produce reads of nearly a kilobase - libraries starting from 20-50 nanograms of input DNA.  One highlighted library had a mean read length (wouldn't median be a better statistic here?) of 978 basepairs, with around 1B reads per hour throughput.  These were touted first for resolving isoforms in single cell RNA-Seq data, and also for improved haplotype phasing of genomic fragments.  That latter point brings up an interesting question of whether for rare disease work, if not requiring an expedited report, will SBX labs prefer the higher baseline accuracy of duplex mode or the better phasing and SV calling offered by long simplex reads?  Roche showed data on phasing for 60X duplex vs. 30X duplex + 30X 200bp simplex vs 30X duplex plus 30X simplex >800bp, but not long duplex on its own.  Another slide showed greatly improved performance for VNTR length calling of >800bp simplex vs duplex vs an Illumina PCR-free library.

For duplex, both chemistry and bioinformatics improvements are credited with moving F1 scores upwards.  For Genome-in-a-Bottle samples HG001-HG007, the previous best and worst performers for SNV F1 score were HG001 with 99.81% and HG004/HG007 tied with 99.69; these three now have F1 scores of 99.86, 99.82 and 99.78.  For indel F1 scores, HG003 was worst before at 99.68% and is still worst at 99.71, but HG007 at 99.74 is the only one under 99.8.  Homopolymer length calling has an F1 score of  >99 up to length 21.  Roche's plotters have failed to get my memo that once things are that good, you should switch to phred  scores - it's really hard to see the subtle differences when everything is crammed up near the limit.

For FFPE, a duplex workflow with linear amplification was used after first performing enzymatic repair on the extracted DNA.  Kokoris showed a number of statistics suggesting that duplex SBX gave similar performance to an Illumina NovaSeq library prepared from the same sample.   Interestingly, for comparing to Illumina they did get the memo and plot everything as phred scores - probably so the tiny improvement over Illumina can be more easily seen.  Some of the slides suggest much better performance than Illumina for particular types of variants - such as those in high %GC regions, tandem repeats and homopolymers.

In MRD, it's a duplex protocol,  with a claimed SNP Q-score of 44 .  After aligning and grouping similar molecules, that is said to be boosted to Q46.  Ends are trimmed, only concordant duplexes kept and any variants seen in germline tossed out - resulting in a tumor-specific SNV score of ">Q50",

Both on his own and in response to audience questions, Kokoris sketched out a number of further application directions.  He said to expect updates in the fall - he didn't specify where but ASHG would be the obvious venue - to include RNA-Seq and target enrichment as well as more updates on fast mode and FFPE.  He also reiterated that any conversion-based approach to methylation detection will work on SBX, but that he still believes that a chemistry can be derived to read methylation as a distinct additional SBX signal.  

At the diagnostics day, Roche did not name any additional sites using SBX beyond Broad Clinical Labs and the Hartwig Institute, but emphasized that both had seen very simple installation processes.  

It should be clear that Roche is gunning for all the short read applications as well as trying to leverage their advantage in midi-read space.  They also reiterated the intent to play in both ultra high throughput and more modest throughput markets and strongly implied that they will come under the capital cost of existing hardware in the ultra high throughput market.  While much will hinge on the precise details of how Roche prices the instrument and consumables and even what pricing model they use for consumables, it should be clear to all other NGS vendors that Roche is going to a force to reckon with.