Skip to content

04_fate.md

Latest commit

 

History

History
73 lines (55 loc) · 2.62 KB

04_fate.md

File metadata and controls

73 lines (55 loc) · 2.62 KB

04 · Weinreb clonal fate (Table 3, §5.3) — zero-shot

Task. Given a d2 source population, predict each cell's clonal Neu / (Neu + Mono) fate ratio using the frozen, zero-shot Chreode backbone. No fine-tuning.

Metric. Pearson r between predicted and observed fate ratios, masked to clones whose d6 atlas neighborhood is dominated by Neu or Mono cells. Reported as $r_\text{masked}$ at $n$ clones (the count after masking).

Expected numbers

Method $r_\text{masked}$ $n$ (clones)
Static-DiT (control) 0.3825 335
WOT 0.4593 265
moscot 0.4624 265
scDiffEq 0.4627 296
Chreode (zero-shot) 0.4680 183

(kNN and PRESCIENT entries from earlier drafts were removed from the final paper — see known_issues.md.)

Prerequisites

python scripts/download_weights.py      # dynamics_dit.pt + vae.pt
python scripts/download_phase0.py       # weinreb_ortholog.h5ad + weinreb_clonal.npz

Run (≈ 10 min, 1 GPU)

export CHREODE_ROOT=$PWD
PYTHONPATH=src python scripts/paper_bench/eval_foundation_fate.py \
  --static-checkpoint  checkpoints/pretrained/static_dit.pt \
  --dynamic-checkpoint checkpoints/pretrained/dynamics_dit.pt \
  --experiment g2a_m10_wdit_time2vecu_lowfreqcurl_uncertainty_adamw \
  --space scvi128 \
  --output-dir output/reproduce/fate/

The script writes:

  • output/reproduce/fate/static.json — Static-DiT control row
  • output/reproduce/fate/dynamic.json — Chreode (zero-shot) row

Read r_masked and n_masked from each. Expected values match the table above to four decimals.

What the script does

  1. Loads frozen backbone (no fine-tune).
  2. For each d2 cell, simulates n_sim rollouts of WaddingtonDiT1D(delta=4) to produce d6 latents.
  3. Embeds those latents in the held-out d6 atlas, picks 20 nearest neighbors per simulated point, and assigns a fate ratio.
  4. Aggregates per-clone, masks clones with ambiguous neighborhoods, computes masked Pearson against ground-truth Neu / (Neu + Mono).

Why zero-shot only

The paper deliberately does NOT report fine-tuned fate results. Population $W_2$ fine-tuning destroys the clone-conditioned fate ranking: predicted cells saturate near $q \approx 1$. The fine-tuned route is shipped as the optional flag --allow-finetune but is documented as inferior — see App. C "Fate fine-tune ablation" in the paper.

Static-DiT control row

eval_foundation_fate.py produces the Static-DiT control line in the same invocation above — see output/reproduce/fate/static.json. Compare its r_masked to the Static-DiT 0.3825 entry in the table above.