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

节点 n11 在终选来历上

local_ot: PCA 空间逐细胞位移 + 残差保留解码（moscot 耦合），λ=3、kNN=30、addnz

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 016536be7c: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 8204ac3d1a。改动的文件：solution/METHOD.md +36 −66、solution/g37_common.py +2 −1、solution/run.py +82 −29

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex ca7a3d2..8c91c0c 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,81 +1,51 @@-# ot_moscot — Waddington-OT / moscot coupling, one-step displacement extrapolation+# local_ot: PCA 空间逐细胞位移 + 残差保留解码（moscot 耦合），λ=3、kNN=30、addnz -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`.+在父节点 2（ot_moscot 基因空间位移）上按 PLAN 改为：耦合重心祖先与 kNN 平滑都在拟合耦合的 30 维 PCA 空间里计算，位移经 PCA 线性解码回基因空间并保留每细胞 PCA 正交残差；λ 与 K_SMOOTH 在 X3 上网格搜索定标。 -Contract: `python run.py --data <view> --out <pred.h5ad> --seed <int>`; `g37_common.py` must stay next to `run.py`.+## 方法 -## Method+两输入阶段 prev(t0)、last(t1)，目标 t2；单输入时退化为生长加权重采样（与父节点相同，已在合成单输入视图上验证跑通 + vec-check ok）。 -Two input stages `prev` (time t0) and `last` (t1); target time t2 (final view: E8.5, E9.5 -> E10.5).+1. 嵌入：双阶段共测基因 → 2000 HVG → z-score(clip 10) → 30 PC（只在输入阶段上拟合，random_state=seed）。+2. 生长先验 + moscot TemporalProblem 耦合（ε=1e-3, τ_a=0.95, scale_cost=mean）：与父节点完全相同。+3. 输出细胞 = last 阶段按 g^dt_out 加权重采样（与父节点相同，rng 消耗顺序不变，同 seed 下 rows 与父节点一致）。+4. **PCA 空间位移（改动 1）**：anc_Z_j = 耦合重心祖先的 PCA 均值（Pc.T @ Zp，Pc 为按 rows 归一化的耦合列）；s_Z_j = Zl 中 K_SMOOTH=30 近邻（含自身）的 PCA 均值；step_Z_j = λ·(dt_out/dt_in)·(s_Z_j − anc_Z_j)。+5. **残差保留解码（改动 2）**：x_new = pca.inverse(z_j + step_Z_j) + [x_j − pca.inverse(z_j)]。残差项解析抵消，等价于 x_new = x_j + (step_Z_j @ components)·sd，只作用于 HVG 列；非 HVG 面板基因不变。随后 addnz（步长只加到该细胞非零表达项，β_zero=0）并 clip≥0。addnz 与 clip 由父节点在 X3 上定标，本节点复验：β_zero=0.3 → 51.5，β_zero=1（稠密）→ 37.9，均大幅劣于 β_zero=0 → 61.5（covariation 23.2/11.5 vs 50.3）。+6. `--lambda`、`--k-smooth`、`--beta-zero` 为 CLI 旗标，提交默认 λ=3.0、K=30、β=0；输出只依赖视图数据、时间差与 seed，不依赖绝对时间/路径/manifest 写法。 -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.+## 超参定标（X3 A 半，seed 0，13 次查分） -**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.+| 配置 | 榜分 | cell_state | covariation | de_recovery | direction |+|---|---|---|---|---|---|+| λ=0（机制关闭对照） | 50.52 | 54.10 | 50.25 | 46.49 | 50.46 |+| λ=1, k15 | 53.20 | 60.15 | 50.38 | 49.07 | 51.25 |+| λ=2, k15 | 58.84 | 79.00 | 50.77 | 48.62 | 51.33 |+| λ=2.5, k30 | 59.98 | 83.19 | 51.18 | 47.75 | 51.42 |+| λ=3, k10 / k15 / **k30** / k50 | 60.81 / 61.10 / **61.47** / 61.11 | 88.1 / 89.2 / **89.2** / 87.2 | 49.2 / 49.5 / **50.3** / 50.8 | 46.9 / 46.5 / **47.3** / 47.8 | ~51.3 |+| λ=3.5, k30 | 61.06 | 89.28 | 49.04 | 46.49 | 51.37 |+| λ=4 / λ=5, k30 | 59.09 / 53.04 | 84.1 / 66.6 | 47.6 / 44.8 | 46.1 / 45.3 | ~51.2 |+| λ=3, k30, β_zero=0.3 / 1.0 | 51.55 / 37.90 | 75.0 / 39.3 | 23.2 / 11.5 | 46.5 / 44.9 | 51.2 / 50.4 | -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.+λ=3、K=30 名义最优；K∈[10,50] 差异 < 噪声（T1 约 2 分），λ 峰值明确（3 附近，4 起下降）。 -## Data / knowledge used+## 机制生效证据（λ=3, k30, seed 0，run.py stderr 诊断） -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.+- 对照差：λ=0 → 50.52 vs λ=3 → 61.47（+10.9，主要来自 cell_state 54.1 → 89.2）；λ=0 输出与 λ>0 不同（文件与分数均不同），机制确实在跑。+- mean|step|/mean|residual| 每细胞中位数 1.23（p10 0.94, p90 1.63）> 0.5：PCA 步长相对残差是主导信号，解码不过保守（PLAN 风险 1 排除）。+- 每细胞被移动基因数（|step|>0.01）中位 1982（min 1953, max 1995），逐细胞不同。+- step 簇内弥散：8 个 KMeans 簇内逐维 std 均值 0.059 vs 全体 0.061（比值 0.97）：位移不是逐类型常数，簇内细胞间方向/幅度仍各异（每个细胞的步长取决于自身耦合重心 + 自身局部邻域）。 -## Hyper-parameters+## 验证过 / 没验证 -| 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 |+- 验证：X3 A 半 seed 0 全网格（上表）；同 seed 逐比特确定性；vec-check ok；合成单输入视图 fallback 跑通；λ=0 对照。+- seed 稳健性：seed 1（X3 A 半）57.70（cell_state 80.7, cov 45.8, de_rec 46.9, dir 50.4）vs seed 0 61.47，种子间波动 ~3.8 分，仍显著高于对照 50.5 与父节点 50.65。+- 没验证：X3 B 半与 seed 2 复跑（正式分由系统跑）；final/proxy 视图本地不可得（fallback 代码与父节点逐行相同，父节点已在 proxy 验证）；伪装视图重跑（代码不读绝对时间/路径/board 字段，时间平移下 dt_in、dt_out 不变，预测应逐比特相同）。+- 相对节点 8（同族，λ=3 PCA 位移，62.24）：本节点 covariation 50.3 略高、de_recovery 47.3 略低、cell_state 89.2 略低，A 半分差在噪声内；残差保留解码在本实现下与"直接加解码步长"数学等价，未带来 PLAN 预期的 de_recovery 修复（de_recovery 对 λ 增大单调下降，是位移外推本身的代价，不是解码丢残差造成）。 -No re-tuning for the seed.+## 知识/数据来源 -## Resources (Spark, CPU)+只用视图输入阶段 + 父节点已有组件：moscot（Klein et al., Nature 2025）、Waddington-OT 生长先验（Schiebinger et al., Cell 2019，moscot 内置小鼠增殖/凋亡基因列表，阶段无关的通用基因功能注释）。无保留阶段/保留基因型信息，无外部训练数据（X3 视图挂载的 external 未使用）。 -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.+## 资源 -## 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).+X3 视图全程 ~15 s / <2 GB（CPU，EXECUTION.json gpu:false，JAX/torch 固定 CPU、8 线程）。diff --git a/solution/g37_common.py b/solution/g37_common.pyindex 8c991b3..7676511 100644--- a/solution/g37_common.py+++ b/solution/g37_common.py@@ -59,7 +59,8 @@ def embed(mats: list, mask: np.ndarray, n_hvg: int, n_pcs: int, seed: int):     for M in mats:         out.append(Z[start:start + M.shape[0]])         start += M.shape[0]-    return out, {"hvg": hvg, "mu": mu, "sd": sd, "components": pca.components_.astype(np.float32)}+    return out, {"hvg": hvg, "mu": mu, "sd": sd, "components": pca.components_.astype(np.float32),+                 "pca_mean": pca.mean_.astype(np.float32)}   def knn_mean(Z: np.ndarray, X: sparse.csr_matrix, rows: np.ndarray, k: int) -> np.ndarray:diff --git a/solution/run.py b/solution/run.pyindex 1411cb7..971f54e 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,5 +1,7 @@ #!/usr/bin/env python3-"""ot_moscot: Waddington-OT / moscot TemporalProblem coupling, extrapolated one step past the latest input stage.+"""local_ot (node 11, from ot_moscot seed): moscot coupling, per-cell displacement computed and extrapolated in the+PCA subspace where the coupling was fitted, decoded back to gene space with each cell's PCA-orthogonal residual+preserved.  Two input stages (final: E8.5, E9.5):   1. joint PCA of both stages (HVG, z-score, 30 PCs; fitted on the inputs only);@@ -7,19 +9,18 @@ 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-     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.+  4. displacement in PCA space: anc_Z_j = coupling-weighted (barycentric) mean of cell j's ancestors' PCA+     coordinates; s_Z_j = kNN mean (K_SMOOTH nearest latest-stage cells) of its PCA coordinates;+     step_Z_j = LAMBDA * dt_out / dt_in * (s_Z_j - anc_Z_j);+  5. residual-preserving decode: x_new_j = x_j + (step_Z_j @ components) * sd on the HVG columns (equivalently+     pca.inverse(z_j + step_Z_j) + [x_j - pca.inverse(z_j)]; the residual cancels analytically), non-HVG panel+     genes unchanged. 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, 5 need two stages; only the growth resampling (3) runs, i.e. a growth-weighted copy of the latest stage. -Seed version (agent/seeds/T1__val/ot_moscot, 2026-10-02) of modeling/candidates/T1/ot_moscot (G37): same method and-hyper-parameters; the dev-only environment overrides (G37_LAMBDA / G37_DECODE / G37_GROWTH / G37_JAX_GPU) and the-`knn` decode branch are removed; JAX and torch on CPU, fixed thread count (EXECUTION.json gpu false).-Parameter provenance (METHOD.md): DECODE = addnz was chosen on the X3 ruler (G37); LAMBDA = 1 a priori, also checked-on X3 (0.5 vs 1 within 0.2).+Mechanism-off control: --lambda 0 zeroes step_Z; the output is then exactly the growth-weighted resampling of the+latest stage (steps 2, 4, 5 skipped). JAX and torch on CPU, fixed thread count (EXECUTION.json gpu false). """  from __future__ import annotations@@ -36,7 +37,7 @@ os.environ["JAX_PLATFORMS"] = "cpu"  # CPU only: deterministic, no GPU slot import numpy as np  sys.path.insert(0, str(Path(__file__).resolve().parent))-from g37_common import embed, growth_rates, knn_mean, stage_pair  # noqa: E402+from g37_common import embed, growth_rates, stage_pair  # noqa: E402  from src.task1_temporal.view_io import load_manifest, panel_genes, target_n_cells, write_prediction  # noqa: E402 @@ -45,9 +46,11 @@ 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+K_SMOOTH = 30    # kNN neighbourhood in PCA space (X3: k30 61.5 >= k15 61.1 >= k50 61.1 >= k10 60.8, within noise)+LAMBDA = 3.0     # extrapolation strength; X3 grid: 1 -> 53.2, 2 -> 58.8, 2.5 -> 60.0, 3 -> 61.5, 3.5 -> 61.1,+                 # 4 -> 59.1, 5 -> 53.0; 0 = mechanism off (growth resampling only, X3 50.5) GROWTH = True    # resample output cells by g^dt_out (WOT birth-death model)+BETA_ZERO = 0.0  # step weight on zero entries: 0 = "addnz" (keep each cell's zero pattern), 1 = dense N_THREADS = 8  @@ -61,6 +64,11 @@ def main() -> None:     parser.add_argument("--data", required=True)     parser.add_argument("--out", required=True)     parser.add_argument("--seed", type=int, default=0)+    parser.add_argument("--lambda", dest="lam", type=float, default=LAMBDA,+                        help="displacement extrapolation strength; 0 = mechanism off (growth resampling only)")+    parser.add_argument("--k-smooth", type=int, default=K_SMOOTH, help="kNN neighbourhood size in PCA space")+    parser.add_argument("--beta-zero", type=float, default=BETA_ZERO,+                        help="fraction of the step applied to zero entries (0 = addnz, 1 = dense)")     args = parser.parse_args()      manifest = load_manifest(args.data)@@ -76,13 +84,21 @@ def main() -> None:         write_prediction(last.X[rows], genes, args.out, seed=args.seed)         return +    (Zp, Zl), info = embed([prev.X, last.X], mask, N_HVG, N_PCS, args.seed)+    g_prev, g_last = growth_rates([prev, last], genes, mask, args.seed)+    rows = weighted_rows(g_last ** dt_out if GROWTH else np.ones(last.n_obs), n, rng)+    print(f"diag: growth last min/median/max {g_last.min():.3f}/{np.median(g_last):.3f}/{g_last.max():.3f}", file=sys.stderr)++    if args.lam == 0.0:+        # mechanism-off control: step_Z = 0, pure growth-weighted resampling of the latest stage+        write_prediction(last.X[rows], genes, args.out, seed=args.seed)+        return+     import anndata as ad     import pandas as pd     import torch     from moscot.problems.time import TemporalProblem--    (Zp, Zl), _ = embed([prev.X, last.X], mask, N_HVG, N_PCS, args.seed)-    g_prev, g_last = growth_rates([prev, last], genes, mask, args.seed)+    from sklearn.neighbors import NearestNeighbors      obs = pd.DataFrame({"time": np.r_[np.zeros(prev.n_obs), np.ones(last.n_obs)],                         "growth": np.r_[g_prev ** dt_in, g_last ** dt_in]},@@ -98,26 +114,63 @@ def main() -> None:     print(f"ot: converged={getattr(sol, 'converged', None)} cost={getattr(sol, 'cost', None)}", file=sys.stderr)     P = np.asarray(sol.transport_matrix, dtype=np.float32)  # (n_prev, n_last) -    rows = weighted_rows(g_last ** dt_out if GROWTH else np.ones(last.n_obs), n, rng)-    print(f"diag: growth last min/median/max {g_last.min():.3f}/{np.median(g_last):.3f}/{g_last.max():.3f}", file=sys.stderr)     Pc = P[:, rows]     del P     Pc /= np.maximum(Pc.sum(axis=0, keepdims=True), 1e-30)-    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-    del smooth, anc-    step[:, ~mask] = 0.0+    anc_Z = (torch.from_numpy(Pc).T @ torch.from_numpy(Zp)).numpy()  # barycentric ancestor mean in PCA space+    del Pc+    factor = args.lam * dt_out / dt_in++    k_smooth = min(args.k_smooth, Zl.shape[0])+    nn = NearestNeighbors(n_neighbors=k_smooth).fit(Zl)+    idx = nn.kneighbors(Zl[rows], return_distance=False)+    s_Z = Zl[idx].mean(axis=1)                     # smoothed position in PCA space+    step_Z = (s_Z - anc_Z).astype(np.float32) * factor++    # residual-preserving decode: x_new = pca.inverse(z + step_Z) + [x - pca.inverse(z)]+    #                           = x + (step_Z @ components) * sd  on the HVG columns (residual cancels exactly);+    # non-HVG panel genes keep their raw value (step 0 there).+    hvg, sd, comp = info["hvg"], info["sd"], info["components"]+    step = np.zeros((n, len(genes)), dtype=np.float32)+    step[:, hvg] = (step_Z @ comp) * sd+    del step_Z     X = last.X[rows].toarray()-    step *= X > 0  # "addnz": move only measured (non-zero) entries, keeps each cell's zero pattern+    _diagnostics(step, X, Zl, rows, info)+    if args.beta_zero <= 0.0:+        step *= X > 0  # "addnz": move only measured (non-zero) entries, keeps each cell's zero pattern+    elif args.beta_zero < 1.0:+        step *= np.where(X > 0, 1.0, np.float32(args.beta_zero))     X += step     np.maximum(X, 0.0, out=X)     write_prediction(X, genes, args.out, seed=args.seed)  +def _diagnostics(step: np.ndarray, X: np.ndarray, Zl: np.ndarray, rows: np.ndarray, info: dict) -> None:+    """Mechanism evidence (stderr only): per-cell step magnitude vs residual, within-cluster step spread,+    number of genes moved per cell. Never affects the output."""+    from sklearn.cluster import KMeans++    hvg, mu, sd, comp = info["hvg"], info["mu"], info["sd"], info["components"]+    xh = X[:, hvg]+    recon = (Zl[rows] @ comp + info["pca_mean"]) * sd + mu+    resid = np.abs(xh - recon).mean(axis=1)+    mstep = np.abs(step[:, hvg]).mean(axis=1)+    ratio = mstep / np.maximum(resid, 1e-9)+    n_moved = (np.abs(step[:, hvg]) > 0.01).sum(axis=1)+    km = KMeans(n_clusters=min(8, len(rows)), n_init=3, random_state=0).fit(Zl[rows])+    within, overall = [], step.std(axis=0)+    for c in range(km.cluster_centers_.shape[0]):+        m = km.labels_ == c+        if m.sum() > 2:+            within.append(step[m].std(axis=0).mean())+    print(f"diag: mean|step|/mean|resid| median={np.median(ratio):.3f} "+          f"p10={np.percentile(ratio, 10):.3f} p90={np.percentile(ratio, 90):.3f}", file=sys.stderr)+    print(f"diag: genes moved per cell (|step|>0.01) median={np.median(n_moved):.0f} "+          f"min={n_moved.min()} max={n_moved.max()}", file=sys.stderr)+    print(f"diag: step spread within-cluster mean={np.mean(within):.3f} overall per-dim std mean={overall.mean():.3f}"+          f" (within/overall={np.mean(within) / max(overall.mean(), 1e-9):.3f})", file=sys.stderr)++ if __name__ == "__main__":     main()

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

**Done.** Implemented PLAN on parent 2: per-cell displacement moved into the 30-dim coupling PCA space (barycentric ancestor vs kNN-smoothed position), residual-preserving linear decode to gene space (addnz, clip≥0), CLI flags `--lambda/--k-smooth/--beta-zero`.

**Scores (X3 A-half, seed 0):** grid over λ∈{0,1,2,2.5,3,3.5,4,5}×K∈{10,15,30,50}: best **λ=3, K=30 → 61.47** (cell_state 89.2, cov 50.3, de_rec 47.3, dir 51.4); mechanism-off control λ=0 → 50.5; parent 50.65. Dense decode (β>0) strongly harmful (51.5/37.9). Seed 1: 57.7 (seed variance ~3.8). Diagnostics confirm per-cell, non-constant steps (|step|/|resid| median 1.23, within/cluster spread 0.97). Deterministic, vec-check ok, single-input fallback verified.

**Next best directions:** de_recovery falls monotonically with λ — try decoupling strengths (large λ for cell_state PCs, shrinkage on PCs aligned with DE genes), or per-PC damping instead of scalar λ; also test whether dropping growth-resampling (uniform rows) recovers covariation at seed variance level.

调研员

.\n2. K_SMOOTH=15 may introduce noise in sparse regions. Check: if direction doesn't improve or cell_state drops, revert to K=30.\n3. The 30-min budget is tight if 9 configs need full runs. Mitigation: run all 9 on a 2000-cell subset first (each ~3s), pick top 2 by X3 A-half, then run full.\n4. If residual preservation fully restores covariation/de_recovery but eliminates cell_state gain, the mechanism is not working as intended — the PCA step must be the dominant signal.",
  "family_id": "local_ot",
  "mechanism": "Per-cell displacement from OT coupling, computed and extrapolated in the PCA subspace where the coupling was fitted, then decoded back to gene space while preserving each cell's PCA-orthogonal residual to maintain gene-specific expression structure.",
  "vs_constant_shift": "Each cell's displacement vector is unique: it depends on its own barycentric ancestor (from the coupling) and its own local neighborhood mean in PCA space. The residual-preserving decode further ensures gene-specific variation differs per cell. A constant shift would move all cells identically regardless of their coupling or neighborhood.",
  "mechanism_evidence": "Engineer should report: (1) std of step_Z across cells within each coarse cluster (if std≈0, displacement degenerated to per-type constant); (2) mean |step_gene| vs mean |residual| ratio per cell (confirms PCA step is non-trivial); (3) per-group score changes vs parent (cell_state should rise, covariation/de_recovery should not drop more than 1 point); (4) number of genes with |step|>0.01 per cell (should be >>0 and vary across cells).",
  "mechanism_off_control": "Run with --lambda 0 (or equivalently --no-displace flag): step_Z is zeroed, output is pure growth-weighted resampling of the last stage (identical to copy_last with growth weights). Expected: cell_state drops to ~50, direction drops to ~49, covariation returns to ~52.5. If output with λ=0 is bit-identical to λ>0, the mechanism is not running.",
  "sources": []
}

审查员

REVIEW.json written