A quick look at Roche's SBX

Sequencing by Expansion (SBX) is Roche’s new short-read sequencing technology. Roche released the first batch of public SBX data after a webinar yesterday. This blog post provides my impression with the data. If you are interested in SBX, I highly recommend watching the recording of the webinar, as it contains a lot more information than this blog post.

Roche released duplex SBX (SBX-D) data produced in one hour by multiplexing seven human samples together. These are single-end reads, averaging ~241bp in length. The purple line in Fig. (a) shows the fraction of reads beyond a certain length. We can see more than 95% of reads are longer than 150bp but less than 10% are longer than 350bp.

With SBX-D, the sequencing of the opposite strand begins at the 3’-end of the first strand. As a result, most 3’-ends are in duplex but 5’-ends tend to be covered by one strand only. The current SBX-D basecaller assigns a quality score of 39 to bases supported by both strands, 5 to bases covered by both strands but with disagreement between the two, and 22 to bases covered by one strand only. We can use this information to estimate the fraction of duplex reads (green curve). At the 5’-end, less than 60% of read bases are in duplex and the fraction increases to 85% at ~200bp. The drop afterwards might be caused by the reduction of longer reads, though I haven’t looked into details. It might be clearer to draw the curve from the 3’-ends of all reads.

Sequencing errors in SBX-D reads are primarily due to 1bp insertions. This results from an interesting decision to always choose bases from the longer strand when the two strands disagree. In this case, the SBX-D basecaller assigns a quality score of 5 to the extra base(s). In a homopolymer run, only the first or last base(s) are lowered to 5 but the rest of the bases still receive quality 39. This quality assignment caused my previous tool to report a low empirical quality. I had to modify the source code for SBX-D.

Fig. (b) shows the empirical base quality (emQ) along the reads. The overall SBX-D emQ stays almost constant at Q21 across base positions (red curve). Bases confirmed by both strands reach a higher emQ at Q37, close to the estimated value of Q39 (dark blue curve). This is comparable to the Illumina NovaSeq results from Baid et al (2020) (purple and green curves). The two technologies are also similar when we look at recurrent errors supported by two or more Q30+ bases (orange and light blue curves). Roche further demonstrated similar SNP and indel calling accuracy to NovaSeq during the webinar.

Overall, I am really impressed by the quality of SBX-D especially given its throughput. Nevertheless, I believe the decision to always choose the longer strand is questionable. This will set a trap for many current analyses including all pileup-based ones. While DeepVariant can learn the pattern and GATK can be tuned, they are only part of the ecosystem. I have seen sequencing technologies that produced data of special structures or features, but successful ones more often generate “ordinary” data. In case of SBX-D, the basecaller doesn’t distinguish good from bad bases based on raw signals. I suspect it is possible to use raw signals to generate accurate duplex consensus such that we can analyze SBX-D data the same way as typical short reads. This would be ideal for everyone.

The elephant in the room is pricing, which Roche understandably has kept silent about. If SBX were priced similarly to NovaSeq, many would not bother to adapt their tools and workflows just for SBX; if SBX were much cheaper, we might see a shift in tool development towards SBX. We will see.

PS: Based on how base quality is estimated, we can infer the SBX basecaller gives Q22 to all simplex bases. It is not using raw signals to distinguish good and bad bases. When the duplex consensus tool does not know which strand to trust, choosing the longer strand is a simpler bet, as we can’t easily attach base quality to deleted bases. This is why the vast majority of errors are Q5 insertions. However, a simple decision is not necessarily a good decision as I explained above.

If I were to design simplex and duplex base calling, I would use a simple machine-learning model to classify each simplex base to low or high quality. If we see conflicting single-strand sequences during the duplex consensus step, choose a high-quality base over a low-quality base. When it is a tie, use a second machine-learning model based on first/second strand and two nearby bases to make the best choice. This will greatly reduce 1bp Q5 insertions, making SBX data more compatible with existing tools. Furthermore, the numbers at my hand suggest the majority of errors come from Q5 bases. If this is true (I need to modify the code to find exact numbers), the strategy above may increase the overall base quality (the red line in Fig. b) by 3 points.

Another question is how to encode base quality in homopolymer runs. A conservative strategy is to lower the base quality for the entire homopolymer/STR run if there is conflicting evidence. Given that SBX errors are rare and mostly random, this should not affect the power too much. They could also learn from Oxford Nanopore and PacBio who have the same problem.

I guess with the strategy above, the variant calling accuracy will probably stay the same as the numbers shown at the webinar. Nonetheless, the cleaner and more “ordinary” reads will make SBX more compatible with generic variant callers and other workflows. That would be a huge selling point in my opinion.