Seeking input on STAR parameter optimization for projects sequenced on Ultima UG-100

[adam-cornwell]

Hello,
Recently I have been sequencing some large single-cell RNA-Seq projects on the Ultima UG-100 platform, as these projects would be cost-prohibitive for us to sequence entirely on Illumina. I have not yet seen any public comments on optimizing STAR for Ultima; I have started to look at this, but was curious if anyone else has suggestions.
There are a couple of key reasons why I think projects sequenced on Ultima instruments with the current chemistry could benefit from optimized alignment parameters- 1) variable-length (single-end) reads that are longer on average than most Illumina reads (300bp median is advertised, I tend to see around 210bp median), and 2) assessment of the number of bases in homonucleotide stretches is frequently inaccurate due to the non-terminated chemistry and model-based estimation of how many nucleotides are added per reagent flow.

In combination, I expect that this could result in alignments being called as "too short" with greater frequency than Illumina with default STAR parameters. Ultima's default read processing workflow for most scRNA-Seq applications seems to get around this by trimming reads to a substantially shorter fixed length (e.g. 90bp for the read with the cDNA insert for single-cell). However, I expected that use of the full-length reads after adequately optimizing mapping parameters would result in a net increase of usable reads due to having to drop fewer reads for multiple mapping (i.e. longer reads mean a better chance of mapping uniquely or not exceeding multi-mapping thresholds). So far, I have found that map rate improves 6-10% in my samples after moderate trimming (fastp) and then running STAR with a reduced minimum map length fraction (50% instead of the default of 66%).

I have considered devising an "Ultima alignment mode" that would consider the presence of homonucleotide stretches in the reference when scoring alignments and be more lenient for small insertions/deletions in such regions. However, this could of course also increase the risk of false positive alignments. I hope that future chemistry and base call model improvements will make this unnecessary, but it is unclear what the timeline for that will be, and that will not help projects being sequenced now. While 5% more reads seems like a small gain, at the scale of 50+ billion reads this might not be inconsequential. I still have to look more at the actual alignments being produced right now, but dealing with BAM files from alignments has been somewhat daunting.

Happy to hear if anyone has thoughts or input.
Adam Cornwell