Paralome

In active development — public release pending publication

01 / Analyses

The analysis name is the verb

There is no run subcommand. One command names an analysis and hands it a cohort, and the framework decomposes that into stages, submits them, waits, validates and moves on.

paralome scRNA --rds obj.rds --project A_vs_B \
               --group-by sample,cell_type \
               --control CON:wt --treatment TRT:ko \
               --genome mouse_grcm39 --backend slurm --submit

02 / Design

What the framework guarantees

• Checkpointed, and re-running is how you resume. Every stage writes a checkpoint; a run starts at the first stage whose checkpoint is missing or invalid. There is no resume flag to remember and no restart semantics to get wrong.
• One extraction, many comparisons. Reading a large single-cell object is the expensive step, so it happens once into a shared folder that every contrast against that object reuses.
• Each stratum is its own job. Cell types are planned from the sample manifest before submission, so skips and refusals happen in the foreground and a single failed cell type is re-runnable on its own.
• Failures carry numbers. Every error has a code, a cause, an instruction and a log path — paralome explain 104 gives the long form.
• Statistical models run side by side. Donor-level and cell-level tests are reported together with their effect sizes and power, each at the replication unit it actually used, so the difference between them is visible rather than asserted.

03 / Citation

How to cite