This blog post provides an initial assessment of Roche's new short-read sequencing technology, Sequencing by Expansion (SBX), specifically focusing on the duplex version (SBX-D) data released after a recent webinar. The author analyzes read length distribution, duplex coverage across reads, and the technology's error profile, noting a predominance of 1bp insertions linked to a controversial basecalling decision to always choose the longer strand when strands disagree. Empirical base quality for SBX-D is reported at an overall Q21, with duplex bases achieving Q37, comparable to Illumina NovaSeq. While impressed by the quality and throughput, the author critiques the basecalling strategy for creating data incompatible with many existing bioinformatic tools and proposes alternative machine learning-based approaches for more "ordinary" and compatible output. The crucial unaddressed factor of pricing is highlighted as the ultimate determinant of SBX's market adoption.
Summary of Heng Li's Blog: A Quick Look at Roche's SBX
Introduction to SBX-D: Roche has released initial public data for its new short-read sequencing technology, Sequencing by Expansion Duplex (SBX-D), following a recent webinar.
Read Lengths: SBX-D produces single-end reads averaging ~241bp. Over 95% of reads are longer than 150bp, but less than 10% exceed 350bp.
Duplex Coverage: Most 3'-ends are in duplex, while 5'-ends are typically covered by only one strand. Duplex coverage increases from under 60% at the 5'-end to 85% at ~200bp.
Basecaller Quality Scores: The current SBX-D basecaller assigns Q39 for bases supported by both strands with agreement, Q5 for bases covered by both strands but with disagreement, and Q22 for bases covered by one strand only (simplex bases).
Error Profile (1bp Insertions): Sequencing errors are primarily 1bp insertions. This stems from a basecaller decision to always choose the longer strand when the two strands disagree, assigning a Q5 to the extra base(s).
Empirical Base Quality (emQ): The overall empirical base quality (emQ) for SBX-D is approximately Q21 across read positions. Bases confirmed by both strands achieve a higher emQ of Q37, close to the estimated Q39.
Comparison to Illumina NovaSeq: SBX-D's emQ and recurrent error rates are comparable to Illumina NovaSeq results (Baid et al., 2020), and Roche claims similar SNP and indel calling accuracy.
Critique of Basecaller Strategy: The decision to always choose the longer strand is deemed questionable as it generates data with "special structures" that may "trap" many existing bioinformatic tools, particularly pileup-based analyses.
Proposed Basecalling Improvements: The author suggests using raw signals to generate accurate duplex consensus. For conflicts, a simple machine learning model could classify simplex bases as low or high quality, favoring high-quality bases. A second ML model could resolve ties, aiming to significantly reduce Q5 insertions and potentially increase overall emQ by 3 points.
Homopolymer Handling: A conservative strategy is suggested to lower the base quality for an entire homopolymer/STR run if conflicting evidence exists, learning from Oxford Nanopore and PacBio.
Impact of Proposed Changes: Cleaner, more "ordinary" reads would enhance SBX's compatibility with generic variant callers and workflows, making it a stronger selling point without necessarily altering variant calling accuracy.
The "Elephant in the Room" - Pricing: Roche has not disclosed pricing. The author posits that pricing similar to NovaSeq would hinder adoption due to the need for tool adaptation, while significantly cheaper pricing could drive a shift in tool development towards SBX.