Skip to content

SimPlex: expression_sigma knob (H2: high-expressing cells spread heavies across barcodes) - #23

Merged
bnemoz merged 1 commit into
simplex-v2d-recalibrationfrom
simplex-h2-expression-heterogeneity
Nov 19, 2026
Merged

bnemoz merged 1 commit into
simplex-v2d-recalibrationfrom
simplex-h2-expression-heterogeneity

Conversation

@bnemoz

@bnemoz bnemoz commented Nov 18, 2026

Copy link
Copy Markdown
Collaborator

Stacked on #21; the base is simplex-v2d-recalibration.

Summary

Adds expression_sigma, which defaults to 0. Each cell's molecule means for both chains are scaled by a mean-1 lognormal factor exp(σZ − σ²/2). The factor comes from the cell's own "expression" RNG stream. With this, a few plasmablast-like cells own most of the molecules, and so most of the ambient copies. At σ = 0 the output is bit-identical to before.

Why: of three candidate mechanisms, this is the only one supported

Real princeps memory data (donor02 and donor07 benchmark subsets) shows a signature that SimPlex lacked: a barcode's top heavy often comes from a heavy source carried by many barcodes. Each intervention ran on the full memory P50 profile at 8k cells. Sims are subsampled to 1 in 8 barcodes, like the real subsets.

real d02 / d07 base sim read jackpots ×5 (H1) ambient light 2.0 (H3) σ 2.5 (H2)
barcode's top heavy from a source in 5+ barcodes 0.16 / 0.35 0.005 0.005 0.010 0.29
top heavy is a weaker copy of a source stronger elsewhere 0.22 / 0.29 0.03 0.05 0.06 0.27
heavy units whose source is in 2+ barcodes 0.38 / 0.64 0.19 0.15 0.21 0.52
top heavy well holds a light 0.16 / 0.15 0.34 0.23 0.44 0.15
true precision, per-well / pooled real pooled < per-well 0.53 / 0.70 0.70 / 0.77 0.31 / 0.44 0.73 / 0.52
  • H1 (heavier read jackpots) raises precision and does not produce the signature.
  • H3 (more ambient light) lowers precision only by breaking statistics that already matched real data: light composition per unit, and Ig wells per barcode.
  • H2 reproduces the signature and the real ordering, with pooled worse than per-well.

Not calibrated yet

No profile ships with this PR. With molecule means lowered so that Ig UMIs per barcode stay near real (σ 2.0, 8.5/24 molecules), two statistics still drift:

σ 2.0 real
Ig wells per barcode 9 11–14
per-well purity 0.83 ~0.5

The spreader signatures stay close to real (0.14 / 0.19 / 0.38 / 0.17).

The report and scripts are in diagnosis/simplex_v2d/ (SIMPLEX_V2D.md, "Update 2026-11-18 (2)"), outside this repo.

Tests

  • test_expression_sigma_zero_identical_and_heavy_tail_same_mean: the default is identical, the mean is preserved, the tail is heavy, and both chains of a cell share its factor.
  • Config validation.
  • The full SimPlex suite passes: 93 tests.

🤖 Generated with Claude Code

…fault off)

Each cell's molecule means (both chains) are scaled by a mean-1 lognormal factor exp(sigma*Z - sigma^2/2) from
its own "expression" RNG stream, so a few plasmablast-like cells own most molecules and hence most ambient
copies. sigma=0 is bit-identical to before.

Motivation (H2, diagnosis/simplex_v2d): in real princeps memory data a barcode's top heavy often comes from a
heavy source spread over many barcodes (16-35% vs 0.5% in sim). sigma=2.5 reproduces this and the
pooled < per-well precision reversal; jackpots (H1) and extra ambient light (H3) do not.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@bnemoz
bnemoz merged commit 159c677 into simplex-v2d-recalibration Nov 19, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant