03_veres.md

03 · Veres islet differentiation (Table 2, §5.2)

Task. Predict population at each of t1 … t7 from t0 source cells, fine-tuning the pretrained Chreode backbone for 5000 epochs per seed.

Metric. Sinkhorn $W_2$ in the shared scVI-128 latent (ε = 0.1, 100 iterations).

Expected numbers

Per-timepoint $W_2$ (mean over 3 seeds; bold = best per row):

t	Chreode	Scratch	PI-SDE	PRESCIENT
t1	2.4009 ± 0.0658	2.8648 ± 0.1058	3.1161 ± 0.0253	3.2033 ± 0.0107
t2	2.7711 ± 0.1854	2.8371 ± 0.1149	2.9183 ± 0.0132	3.2088 ± 0.0187
t3	2.6403 ± 0.1141	2.8759 ± 0.1815	2.7508 ± 0.0752	3.3224 ± 0.1644
t4	2.4048 ± 0.1020	2.4760 ± 0.0534	2.7555 ± 0.0422	3.1567 ± 0.0780
t5	2.4269 ± 0.0205	2.5674 ± 0.0660	2.4697 ± 0.0145	3.0537 ± 0.0468
t6	2.7621 ± 0.1143	2.9128 ± 0.0604	2.8481 ± 0.1363	3.4527 ± 0.2476
t7	2.9132 ± 0.1704	3.0892 ± 0.1144	2.9490 ± 0.0980	3.8556 ± 0.4083
avg	2.6171	2.8033	2.8296	3.3219

Three seeds (0, 1, 2).

Prerequisites

python scripts/download_weights.py
python scripts/download_downstream_weights.py
python scripts/download_phase0.py     # veres_ortholog.h5ad

Fine-tune + evaluate (≈ 2 h per seed on 1×A100)

for seed in 0 1 2; do
  PYTHONPATH=src python -m cellworldmodel.script.run_intermediate_eval \
    --method m10 \
    --dataset veres_scvi \
    --experiment g2a_m10_wdit_time2vecu_lowfreqcurl_uncertainty_adamw \
    --init-checkpoint checkpoints/pretrained/dynamics_dit.pt \
    --epochs 5000 \
    --seed ${seed} \
    --output-dir output/reproduce/veres_seed${seed}/
done

Aggregate the three seeds

python - <<'PY'
import json, glob, statistics as s
runs = [json.load(open(p)) for p in sorted(glob.glob(
    "output/reproduce/veres_seed*/results.json"))]
print(f"{'t':>4}  {'W2 mean':>9}  {'std':>6}")
ws_per_t = {}
for t in ("t=1", "t=2", "t=3", "t=4", "t=5", "t=6", "t=7"):
    ws = [r["intermediate"][t]["sinkhorn_w2_full_latent"] for r in runs]
    ws_per_t[t] = ws
    print(f"{t:>4}  {s.mean(ws):>9.4f}  {s.stdev(ws):>6.4f}")
avg = [sum(ws_per_t[t])/len(ws_per_t[t]) for t in ws_per_t]
print(f"\navg over t1..t7 mean = {sum(avg)/len(avg):.4f}")
PY

Evaluate released fine-tuned heads only

for seed in 0 1 2; do
  PYTHONPATH=src python -m cellworldmodel.script.run_intermediate_eval \
    --method m10 --dataset veres_scvi \
    --experiment g2a_m10_wdit_time2vecu_lowfreqcurl_uncertainty_adamw \
    --model-config-checkpoint checkpoints/downstream/veres_seed${seed}.pt \
    --init-checkpoint         checkpoints/downstream/veres_seed${seed}.pt \
    --epochs 0 \
    --seed ${seed} \
    --output-dir output/reproduce/veres_eval_seed${seed}/
done

Configuration touched

• config/paper_bench/branchsbm_veres_scvi128.yaml
• config/foundation_genhui_v1.yaml

Source–target setup

Source population: t0 (control). Targets: independent populations at t1, …, t7. The fine-tune uses the same multi-Δ sampling as pretrain: uniform over all ordered $(t_i, t_j)$ pairs with $i < j$, so the same checkpoint produces predictions at every elapsed time.