This blog post by Heng Li provides a technical first look at Roche's new short-read sequencing technology, Sequencing by Expansion (SBX), based on an initial public duplex dataset (SBX-D). The analysis reveals that SBX-D produces reads of impressive quality, with duplex-confirmed bases achieving an empirical quality (Q37) comparable to the Illumina NovaSeq platform. However, the technology's primary weakness lies in its basecalling strategy. Errors are predominantly 1bp insertions, a direct result of the basecaller's decision to always favor the longer strand when the two strands disagree. This creates a non-standard error profile that poses significant compatibility challenges for existing bioinformatics tools. The author suggests alternative machine-learning-based approaches to generate more accurate and "ordinary" data, which would ease adoption. The post concludes that the ultimate success and adoption of SBX will heavily depend on its pricing relative to established competitors.
An Expert's First Look at Roche's SBX Sequencing Technology
Introduction to SBX-D Data: The post analyzes Roche's new duplex sequencing technology (SBX-D), which produces single-end reads with an average length of around 241bp. Over 95% of reads are longer than 150bp.
Unique Quality Score System: The basecaller assigns specific quality scores that reveal the data's structure: Q39 for bases confirmed by both strands, Q5 for bases where strands disagree, and Q22 for bases covered by only one strand.
High Base Quality: The data quality is impressive and comparable to the industry-standard Illumina NovaSeq. Bases that are confirmed by both DNA strands achieve a high empirical base quality of Q37.
A Critical Flaw in the Error Profile: The primary type of error is 1bp insertions. This is caused by a questionable design choice in the basecaller: when the two sequenced strands disagree on length, it always chooses the longer one, assigning the extra base a low-quality score of Q5.
Compatibility Challenges: This unique error profile sets a "trap" for many current analysis tools, especially pileup-based variant callers. The data is not "ordinary," which will hinder widespread adoption unless bioinformaticians adapt their workflows specifically for SBX.
Proposed Solutions: The author suggests Roche could use raw signal data and simple machine-learning models to create a more accurate consensus when strands conflict. This would significantly reduce the problematic 1bp insertions and make the data compatible with the existing ecosystem of tools.
The Deciding Factor is Price: The author notes that the "elephant in the room" is pricing. If SBX is priced similarly to NovaSeq, the community may not bother adapting their tools. However, if it is significantly cheaper, it could drive a major shift in the field.