Why use Multi‑Sample FICTURE Analysis?

CartLoader supports analyzing ≥2 samples in three ways:

• A joint run_together run (--samples): lists the samples in a sheet and trains one shared model across all of them, then packages each sample against it. This is the recommended way; internally it uses run_ficture2_multi.
• Multi‑sample FICTURE analysis with the individual commands (run_ficture2_multi): the same joint model, run step by step.
• SGE stitch + single‑sample analysis (sge_stitch → run_ficture2): stitch multiple SGEs into a single mosaic, then train one model on that mosaic.

For unrelated samples that should not share a model, use run_together --out-root instead: each sample gets its own independent model.

What to expect

• Shared factors/comparability: a joint model learns a cohort‑wide latent basis and returns per‑sample decodes for direct comparison. The stitch approach yields a single model over the merged mosaic; useful when you need a unified coordinate system (e.g., tiling adjacent sections).
• Efficiency and scale: a joint model fits once for the cohort and decodes per sample, avoiding repeated runs and post‑hoc alignment. Stitching can be simpler for mosaics but often increases I/O and memory due to very large merged files.

Recommendation

• Prefer a joint run_together run (or run_ficture2_multi) for most cohorts: clean per‑sample outputs and better computational efficiency. Use stitching when a single shared coordinate frame is required.
• If you choose stitching, plan for higher resource usage (RAM, disk, and I/O). Large mosaics can be slow to generate and train on, and may require substantially more memory and temporary storage than per‑sample runs.