#27 · Single-cell

Learned and accelerated methods now operate directly on raw reads and pileups

The dominant thread is computation applied to primary sequencing data, with explicit attention to throughput and memory. SomatiX applies deep learning to allele-aware pileups from matched tumor-normal long-read DNA and is reported to outperform DeepSomatic. scRM pretrains a single-cell foundation model directly on scRNA-seq reads rather than count matrices and is reported as competitive. flashscenic accelerates SCENIC GRN inference on GPU, reporting 12-27x speedups over pySCENIC with r=0.9996 equivalence. brisc runs a multithreaded CPU workflow on 10M cells in 3 minutes and is reproducible across thread counts. The arXiv workload characterization profiles WGA, long-read, and short-read mapping tools with VTune and DAMOV-SIM, identifying memory-bound functions and cache behavior. What is new is the shift toward raw-read and pileup inputs combined with systems-level measurement at scale. The items share methodological and systems concerns rather than a single biological question.