总览 · ← 返回运行 20261002-202908-search-t1-scr-D

节点 n3

ot_moscot + entropy-shrunk displacement: each output cell's step (kNN-smoothed position − barycentric OT ancestor) is scaled by conf^β, conf = 1 − H(normalised ancestor distribution)/log(n_prev), β=4;

方法说明?节点程序自带的 METHOD.md：这个程序做了什么、为什么。

来自 programs.git 71eceadb34:solution/METHOD.md

调研员的计划

代码改动?这个节点的程序和父节点程序的逐行差别：绿色是新增，红色是删除。

对比：父节点版本 823b43cce4。改动的文件：solution/METHOD.md +42 −81、solution/run.py +24 −7

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex ca7a3d2..c63e698 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,81 +1,42 @@-# ot_moscot — Waddington-OT / moscot coupling, one-step displacement extrapolation--Seed (2026-10-02) made from the G37 candidate `modeling/candidates/T1/ot_moscot/` (commit 227eeb2). Same method and-hyper-parameters. Changes for the seed contract only: the dev-only environment overrides (`G37_LAMBDA`, `G37_DECODE`,-`G37_GROWTH`, `G37_JAX_GPU`) and the unused `knn` decode branch are removed; JAX forced to CPU, torch threads fixed-at 8 (`EXECUTION.json {"gpu": false}`); `stage_pair` reads every input stage of the view the same way (no reference to-manifest `mode` / `source`). Output depends only on the view's data, the time differences between stages and `--seed`.--Contract: `python run.py --data <view> --out <pred.h5ad> --seed <int>`; `g37_common.py` must stay next to `run.py`.--## Method--Two input stages `prev` (time t0) and `last` (t1); target time t2 (final view: E8.5, E9.5 -> E10.5).--1. **Embedding.** Genes measured in both stages; top 2000 by variance (both stages pooled); z-score, clip at 10;-   PCA, 30 components (randomized, `random_state=seed`). Fitted on the input stages only.-2. **Growth prior (Waddington-OT).** Proliferation / apoptosis scores (`scanpy.tl.score_genes`, moscot's mouse-   gene lists) -> birth = generalised logistic(prolif; 1.7, 0.3, 0.25, 0.5), death = logistic(apopt; 1.7, 0.3, 0.1,-   0.2), per-day growth g = exp(birth - death) (Schiebinger 2019; same formula and defaults as moscot's-   `BirthDeathProblem.estimate_marginals`).-3. **Coupling.** `moscot.problems.time.TemporalProblem` prev -> last on the PCA (`joint_attr="X_pca"`),-   source marginal ∝ g^(t1-t0), target uniform; entropic unbalanced Sinkhorn, `epsilon=1e-3`, `tau_a=0.95`,-   `tau_b=1`, `scale_cost="mean"` (moscot tutorial settings). JAX on CPU unless `G37_JAX_GPU=1`.-4. **Output cells.** n = number of latest-stage cells clipped to `[min_cells, max_cells]` (as copy_last),-   drawn without replacement from the latest stage with probability ∝ g^(t2-t1) (the WOT birth-death model-   continued over the target interval).-5. **Displacement extrapolation.** For output cell j: ancestor mean a_j = Σ_i π_ij x_i / Σ_i π_ij (barycentric-   projection of the coupling, gene space, all panel genes); smoothed position s_j = mean expression of its 30-   nearest latest-stage cells in the PCA; step_j = λ (t2-t1)/(t1-t0) (s_j - a_j), λ = 1 (continue the last-   observed displacement at the same rate). The step is added **to the cell's non-zero entries only** and clipped-   at 0; genes not measured in both stages (external stages) get no step. Each cell keeps its own residual.--**One input stage (proxy view, E8.5 only):** steps 1, 3, 5 need two stages; only steps 2 + 4 run, i.e. a-growth-weighted copy of the latest stage (g^(t2-t1) resampling). This is the only thing the proxy can test.--Sources:-- Schiebinger G. et al. Optimal-transport analysis of single-cell gene expression identifies developmental-  trajectories in reprogramming. *Cell* 176, 928–943 (2019). doi:10.1016/j.cell.2019.01.006 (WOT: unbalanced-  entropic OT between snapshots, growth from proliferation/apoptosis signatures, birth-death logistic).-- Klein D., Palla G., Lange M. et al. Mapping cells through time and space with moscot. *Nature* 638, 1065–1075-  (2025). doi:10.1038/s41586-024-08453-2 (TemporalProblem; code moscot 0.5.2, BSD-3).-- Cuturi M. Sinkhorn distances. NeurIPS 2013; Chizat L. et al. Scaling algorithms for unbalanced optimal-  transport problems. *Math. Comp.* 87, 2563–2609 (2018).-- Extrapolating the barycentric displacement one more step is our use of the coupling (WOT/moscot interpolate,-  they do not extrapolate); listed in agent/knowledge/T1_methods_landscape.md §1.--## Data / knowledge used--Only the view's input stages. Generic knowledge: moscot's built-in mouse proliferation (97) and apoptosis (193)-gene lists (`moscot.utils.data`, from the WOT paper; stage-agnostic gene-function annotation). No held-out stage,-no information from (E9.5, E13.5], no pre-trained weights.--## Hyper-parameters--| Name | Value | Where it came from |-|---|---|---|-| `N_HVG`, `N_PCS`, `K_SMOOTH` | 2000, 30, 30 | a priori; not tuned |-| `EPSILON`, `TAU_A`, `TAU_B`, `scale_cost` | 1e-3, 0.95, 1, mean | a priori (moscot tutorial settings) |-| growth prior | on | a priori (WOT / moscot defaults); G37 saw it cost ~3 points on the old proxy, kept on |-| `LAMBDA` | 1 | a priori (continue the observed displacement at the same rate); G37 also ran 0.5 on X3 (50.4 vs 50.2), not changed |-| decode `addnz` | — | **chosen on the X3 ruler (G37)** against `add` (25.6) and `knn` (48.7): the dense step destroys the zero pattern. Also a first-principles choice (the scorer compares sparse log-expression), but the evidence that picked it was X3 |--No re-tuning for the seed.--## Resources (Spark, CPU)--Final view (16.8k x 17.1k coupling): 69 s, max RSS 6.7 GB; proxy 4 s / 1.6 GB; X3 ~ 8 s / 2.3 GB; proxy2 ~ 20 s / 4.3 GB.-The dense coupling (n_prev x n_last float32, ~1.1 GB on final) and the dense earlier stage (~2.2 GB) dominate memory.--## Findings (G37, local scorer: fast engine, truth half B, scorer seed = program seed)--- Dense gene-space step (`add`) is destructive: X3 25.6 at λ=1, still 32.9 at λ=0.25 (cell_state, covariation-  collapse) — the step makes every zero slightly positive. `addnz` fixes it (X3 50.2 at λ=1, 50.4 at λ=0.5);-  `knn` decode 48.7 (λ=1). Growth resampling has no effect on X3/proxy2 (all cells are kept there).-- proxy2 (E8.5 official -> Qiu E9.0 heart, then +0.5 d): every variant ≈ copy of the Qiu cells (~27.5), the-  cross-dataset step is batch effect.-- Full eval (seeds 0-2, half B): proxy 46.8 / 46.8 / 47.9 (copy_last 50.0 / 50.3 / 50.1) — the WOT growth-  resampling alone costs ~3 points on E8.5 -> E9.5 (direction, cell_state); X3 50.2 / 50.1 / 49.8 (copy_last 50.0,-  pseudobulk_shift 40.5-40.7); proxy2 27.7-27.9 (copy_last 27.4-27.6). `G37_GROWTH=0` turns the proxy into copy_last.-- Final-view prediction (`~/vec/scratch/g37/ot_moscot/final.h5ad`, seed 0): 5118 cells, 21 nearest-E9.5 types,-  composition within ±2 % of E9.5 (OFT/RV-CM -2.1 %); no new states (the method cannot create them).+ot_moscot + entropy-shrunk displacement: each output cell's step (kNN-smoothed position − barycentric OT ancestor) is scaled by conf^β, conf = 1 − H(normalised ancestor distribution)/log(n_prev), β=4; raw per-cell displacement (PLAN κ=1) rejected on X3.++## Method (family: local_ot, parent node 2 = ot_moscot, 50.65)++Unchanged from parent: joint PCA embedding (2000 HVG, 30 PCs), WOT growth prior, moscot TemporalProblem+coupling (epsilon 1e-3, tau_a 0.95), growth-weighted resampling of output cells, addnz decode, clip ≥0.++Changed (PLAN mechanism, adapted):+1. **Entropy shrinkage (mechanism ON, default):** for each output cell j, Shannon entropy H_j of its+   column-normalised ancestor distribution Pc[:,j]; conf_j = 1 − H_j/log(n_prev) clipped to [0,1];+   step_j = LAMBDA·(dt_out/dt_in)·conf_j^BETA·(smooth_j − anc_j), BETA=4 default.+2. **Per-cell displacement (PLAN item 1) tested and REJECTED:** d_j = x_j − anc_j (κ=1) scored 26.4 on X3+   (cell_state 6.3): with factor=dt_out/dt_in=2 the step becomes 3x−2anc, tripling each cell's own sparse+   noise. Hybrid κ=0.5 also failed (30.7). κ (OT_KAPPA) therefore defaults to 0 = parent's kNN-smoothed+   displacement. This is reported honestly: the PLAN's per-cell direction does not work on this data.++Env overrides (controls only; defaults = submitted behaviour): OT_BETA (0 = shrinkage off), OT_LAMBDA+(0 = mechanism-off control), OT_KAPPA, OT_KSMOOTH, OT_DEBUG_DUMP.++## Mechanism evidence (X3_qiu_heart_early, seed 0)++- shrink_j distribution (β=4): pctiles 0.46/0.61/0.74/1.00/1.00/1.00/1.00 (0/5/25/50/75/95/100) — varies,+  not constant; ~25% of cells (diffuse ancestry) get steps damped 26–54%. With ε=1e-3 Sinkhorn the coupling is+  sharp for most cells (median conf ≈ 1), so modulation acts on the diffuse-coupling tail, as the PLAN's risk 2+  anticipated.+- Per-cell step norms vary (std > 0; e.g. κ=1 run: mean 141.9, std 21.2).+- β sweep on X3 (vec-score, seed 0): β=0 (≡ parent path): 50.65*; β=1: 50.74; β=2: 50.89; β=4: 51.20;+  β=8: 51.25; K_SMOOTH=60, β=2: 51.13. All groups move together (cell_state 50.1→51.3, covariation+  52.5→53.8, de_recovery 51.4→51.0, direction 49.14→49.22 at β=4). Gains (~0.5) are below the 2-pt noise+  floor individually but monotone in β across 1→8.+- Controls: OT_LAMBDA=0 → zero steps, growth-weighted copy of last stage (copy_last level); OT_BETA=0 →+  shrink≡1, output bit-identical to the parent program (verified locally, diff = 0).++## Verified / not verified++- Verified: X3 seed 0 (queries above), vec-check pass, determinism (default rerun bit-identical to β=4 run),+  runtime ~10 s, single-input fallback path unchanged (growth-weighted copy).+- Not verified: multiple seeds per β (budget); T1 proxy/final views (no two-stage official inputs on proxy;+  code path is view-agnostic — depends only on data, time differences and seed); β between 4 and 8.+- Knowledge sources: as in parent METHOD.md (Schiebinger 2019 WOT; moscot 0.5.2; Cuturi 2013). Entropy of the+  coupling column as a branch-point/confidence measure is a standard information-theoretic construction, no+  held-out-stage information used.diff --git a/solution/run.py b/solution/run.pyindex 1411cb7..5588052 100644--- a/solution/run.py+++ b/solution/run.py@@ -7,9 +7,11 @@ Two input stages (final: E8.5, E9.5):      (proliferation / apoptosis gene scores), entropic unbalanced Sinkhorn (epsilon 1e-3, tau_a 0.95);   3. output cells = latest-stage cells resampled with weights g^dt_out (prior growth rate continued for the target      interval, Waddington-OT birth-death model);-  4. each output cell j moves by LAMBDA * dt_out / dt_in * (kNN-mean(x_j) - ancestor_mean_j) in gene space, where-     ancestor_mean_j is the coupling-weighted (barycentric) mean of its ancestors in the earlier stage: the last-     observed displacement continued for the target interval; the cell keeps its own residual. The step is applied+  4. each output cell j moves by LAMBDA * dt_out / dt_in * shrink_j * (x_j - ancestor_mean_j) in gene space, where+     ancestor_mean_j is the coupling-weighted (barycentric) mean of its ancestors in the earlier stage (the cell's+     own displacement along its coupling) and shrink_j = conf_j^BETA with conf_j = 1 - H_j / log(n_prev), H_j the+     Shannon entropy of cell j's normalised ancestor distribution: cells with a concentrated ancestry keep the full+     per-cell direction, cells with diffuse (branch-point) ancestry are shrunk toward no step. The step is applied      to the cell's non-zero entries only (DECODE "addnz": a dense step turns every zero into a small positive value      and wrecks cell_state / covariation, see METHOD.md). Clipped at 0. One input stage (proxy: E8.5 only): steps 1, 2, 4 need two stages; only the growth resampling (3) runs, i.e.@@ -45,10 +47,14 @@ N_PCS = 30 EPSILON = 1e-3 TAU_A = 0.95 TAU_B = 1.0-K_SMOOTH = 30-LAMBDA = 1.0     # 1 = continue the observed displacement at full rate+LAMBDA = float(os.environ.get("OT_LAMBDA", "1.0"))   # 1 = continue the observed displacement at full rate; 0 = control+BETA = float(os.environ.get("OT_BETA", "4.0"))       # entropy-shrinkage exponent; 0 = no modulation (control)+KAPPA = float(os.environ.get("OT_KAPPA", "0.0"))     # per-cell residual weight: 0 = kNN-smoothed displacement (parent),+                                                     # 1 = raw per-cell displacement (unstable: factor*(x-anc) ~ 3x-2anc)+K_SMOOTH = int(os.environ.get("OT_KSMOOTH", "30")) GROWTH = True    # resample output cells by g^dt_out (WOT birth-death model) N_THREADS = 8+DEBUG_DUMP = os.environ.get("OT_DEBUG_DUMP", "")   def weighted_rows(w: np.ndarray, n: int, rng: np.random.Generator) -> np.ndarray:@@ -103,16 +109,27 @@ def main() -> None:     Pc = P[:, rows]     del P     Pc /= np.maximum(Pc.sum(axis=0, keepdims=True), 1e-30)+    # coupling confidence: Shannon entropy of each output cell's normalised ancestor distribution (column j of Pc)+    pos = Pc > 0+    ent = -np.where(pos, Pc * np.log(np.where(pos, Pc, 1.0)), 0.0).sum(axis=0).astype(np.float64)+    h_max = np.log(max(Pc.shape[0], 2))+    conf = np.clip(1.0 - ent / h_max, 0.0, 1.0).astype(np.float32)+    shrink = conf ** BETA if BETA != 1.0 else conf+    qs = np.percentile(shrink, [0, 5, 25, 50, 75, 95, 100])+    print("diag: shrink pctiles " + "/".join(f"{q:.4f}" for q in qs), file=sys.stderr)     factor = LAMBDA * dt_out / dt_in     torch.set_num_threads(N_THREADS)     Xp = torch.from_numpy(prev.X.toarray())     anc = (torch.from_numpy(Pc).T @ Xp).numpy()  # barycentric ancestor mean, (n, n_genes)     del Xp, Pc     smooth = knn_mean(Zl, last.X, rows, K_SMOOTH)-    step = (smooth - anc) * factor+    X = last.X[rows].toarray()+    step = (smooth - anc) + KAPPA * (X - smooth)  # kNN displacement + shrunken per-cell residual     del smooth, anc+    step *= (factor * shrink)[:, None]+    if DEBUG_DUMP:+        np.savez(DEBUG_DUMP, step_norms=np.linalg.norm(step, axis=1), shrink=shrink, rows=rows)     step[:, ~mask] = 0.0-    X = last.X[rows].toarray()     step *= X > 0  # "addnz": move only measured (non-zero) entries, keeps each cell's zero pattern     X += step     np.maximum(X, 0.0, out=X)

调研来源?调研员查到并用到的知识条目和文献检索结果（只列标题和编号）。

用到的知识库条目

分析结果?分析员写的 ANALYSIS.json：改了什么、各组分数怎么变、假设是否成立、经验和下一步建议。

对话摘要?每个角色和大模型对话的统计：轮数、工具调用、用时、token 数和最后的回答摘录；原始记录只给路径。

只给统计和最后回答的摘录；完整对话请到原始记录位置里列出的文件看。

分析员

ANALYSIS.json written

工程师

**完成。** solution/ 最终状态：ot_moscot + 熵收缩位移（conf^β，β=4 默认），vec-check 通过，重跑位级确定。

- **做了什么**：按 PLAN 实现 per-cell 位移 + 熵收缩。Per-cell 位移（κ=1）在 X3 上灾难性失败（26.4，factor=2 使步长变成 3x−2anc，放大单细胞噪声），κ=0.5 也失败（30.7），故 κ 默认 0，保留父节点 kNN 平滑位移，只加熵收缩机制。
- **查分**（X3, seed 0）：β=1: 50.74，β=2: 50.89，**β=4: 51.20**，β=8: 51.25（父 50.65）。提升 ~0.5，低于 2 分噪声，但随 β 单调。控制：LAMBDA=0 → 生长加权复制；BETA=0 → 与父节点输出位级相同。
- **下一步最值得试**：β∈[8,16] 或把 conf 重标定（coupling 太尖锐，中位 conf≈1，收缩只作用于尾部 25% 细胞）；以及验证多种子稳定性。

调研员

ation drops > 1.5 pt. 2. With epsilon=1e-3 the coupling may already be sharp, giving low entropy for most cells (shrink ≈ 1), reducing the change to simply removing smoothing. Check and report the distribution of shrink_j. 3. Improvement may be < 2-pt noise; confirm with ≥ 2 vec-score seeds before concluding.",
  "family_id": "local_ot",
  "mechanism": "Each cell's extrapolation direction is its own coupling-derived ancestor→current vector (not a kNN average over 30 neighbours), scaled by a confidence weight from the entropy of its ancestor distribution: concentrated ancestry keeps the full per-cell direction; diffuse ancestry shrinks the step toward zero.",
  "vs_constant_shift": "A constant shift applies one vector to every cell. Here each cell gets its own direction (from its specific OT coupling) and its own magnitude (entropy-weighted). Cells at branch points move along their own branch rather than toward the average of neighbouring branches.",
  "mechanism_evidence": "1. Report the distribution of shrink_j (must vary, not be constant ≈ 1). 2. Report std of per-cell displacement norms within the output population (must be > 0, showing non-uniform steps). 3. Four-group comparison vs parent: direction should rise from 49.14; de_recovery should stay ≥ 50. 4. Identify the 50 cells whose displacement changed most vs parent; verify they are cells with high coupling entropy (diffuse ancestry).",
  "mechanism_off_control": "Same program with LAMBDA=0: all steps become zero, output is a growth-weighted copy of the latest stage (identical to the copy_last baseline except for growth resampling). Expected: direction ≈ 49.2 (copy_last level), de_recovery ≈ 44 (copy_last level). The gap between LAMBDA=0 and LAMBDA=1-with-entropy-shrinkage measures the mechanism's contribution. A second intermediate control: set shrink_j = 1 for all j (BETA=0), keeping per-cell displacement but removing entropy modulation; this isolates the shrinkage from the per-cell direction change."
}
```