总览 · ← 返回运行 20261003-093415-search-t1-r2-D-s0

节点 n5

n5: local_ot — damped OT displacement, mover-gene masked decoding (α=0.3, k=10)

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 922a8a6e4f: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 b4157ff52c。改动的文件：solution/EXECUTION.json +1 −0、solution/METHOD.md +45 −0、solution/README.md +0 −4、solution/run.py +150 −4

diff --git a/solution/EXECUTION.json b/solution/EXECUTION.jsonnew file mode 100644index 0000000..9d5125c--- /dev/null+++ b/solution/EXECUTION.json@@ -0,0 +1 @@+{"gpu": false}diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..78f1630--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,45 @@+# n5: local_ot — damped OT displacement, mover-gene masked decoding (α=0.3, k=10)++一句话摘要：两输入阶段间 Sinkhorn-OT 速度场（重心投影+PCA kNN 传递），只对伪批量 |Δ|≥0.15 且方向一致的基因施加 α·(Δt_out/Δt_in) 阻尼位移；单输入视图精确退化为 copy_last。++## 方法（family_id: local_ot）++1. `inputs_by_time`；单输入 → 与父节点 copy_last 逐值相同的输出（同 rng、同 sample_rows）。+2. 两输入（X3: Qiu E8.75→E9.0，target E9.5）：top-2000 HVG（两阶段合并方差）→ 25 维 PCA。+3. Sinkhorn OT（POT，`method="sinkhorn_log"`，reg=0.1，100 iters，cost 按中位数归一；每阶段最多 3000 细胞）求耦合 P。+4. 速度场：stage-1 侧重心投影 d_i = Σ_j P[i,j]·x2_j / Σ_j P[i,j] − x1_i（对每个 stage-1 细胞在基因空间求平均，抑制单细胞噪声），再经 PCA 空间 k=10 近邻平均传递给每个 stage-2 细胞。+5. 解码：shift = α · ratio · disp，ratio = (t_target − t_last)/(t_last − t_prev)（只用时间差，视图无关；X3 ratio=2，final=1）。仅当基因 |pb2−pb1| ≥ 0.15（mover）且细胞位移符号与伪批量方向一致时施加；且只改 x>0 的条目（不把零抬成小正数，避免伤 mmd/variogram），clip ≥ 0。α=0.30。+6. 稀疏激活（PLAN 步骤5）：pb delta>0.3 的基因（X3 上仅 1 个），细胞位移>0.5·α·ratio 且该基因为 0 时写入 α·ratio·disp（约 270 个条目）。作用极小，如实报告。+7. 抽样与父节点相同（先 sample_rows 再改表达），α=0 时输出与 copy_last 逐值一致（已验证 np.array_equal == True，X3 与 proxy 两视图）。++## 机制证据（T1_OT_DEBUG=1，X3 seed 0）++- 100% 细胞有 |disp|>0.01（细胞特异位移生效，非全局常数）。+- α=0→0.3 时输出 std 0.3937→0.3938（masked 解码不塌缩也不炸开；未 mask 版本 std 升至 0.46，被否决）。+- nnz 1282705→1282972：激活 268 个零条目（激活通路生效但幅度小）。+- shift 随 α 线性缩放（代码路径 α·ratio·disp）。++## 对照（mechanism_off_control，α=0 vs α=0.3，同一程序）++- α=0 输出 = copy_last（逐值相同，两视图验证）。+- X3 A 半查分：copy_last 47.92；α=0.3 masked k10 **47.35**（de_score 10.77 vs 11.02，de_dir 12.26 vs 12.30，mmd 14.64 vs 14.90，var 9.67 vs 9.69）。差距 −0.57，在 T1 ±2 噪声内。+- proxy10：单输入 → 与父节点相同路径，预期 ≈52.75（未重复查分省额度）。++## 查分记录（X3 A 半，9 次）++| 配置 | de_score | de_dir | mmd | var | 合计 |+|---|---|---|---|---|---|+| stage2 侧位移 α=0.1 | 10.77 | 11.93 | 13.89 | 8.46 | 45.05 |+| NN(k1) α=0.15 未 mask | 11.23 | 12.18 | 13.61 | 9.54 | 46.57 |+| NN(k1) α=0.3 未 mask | 11.42 | 12.13 | 12.64 | 9.76 | 45.95 |+| NN(k1) α=0.3/0.6 mask | — | — | — | — | 47.22 / 46.80 |+| NN(k10) α=0.3 未 mask | 11.33 | 12.14 | 12.68 | 9.54 | 45.69 |+| NN(k10) α=0.3 mask（提交） | 10.77 | 12.26 | 14.64 | 9.67 | **47.35** |++结论：未 mask 位移提高 de_score（−0.195→−0.143）但 mmd 大幅下降（细胞被推离 E9.5 状态流形）；mover-mask 保住 mmd/variogram 但 de_score 增益消失。**PLAN 的接受标准（>51.53）未达成；本节点在 A 半上与父节点统计打平、点估计略低**，如实报告。E8.75→E9.0 的速度方向与 E9.0→E9.5 真实变化在细胞状态空间上相关性弱，可能混入了两个 Qiu 样本间的技术差异。++## 未验证 / 下一步++- 未验证：mask 但保留未 mask 位移的 de_score 增益的折中（如按 |disp| 分位数裁剪）；用细胞类型标签分层估计速度；ratio 缩放在 final（ratio=1）上的行为。+- 生物学知识来源：无外部数据、无先验文件被使用；仅通用 OT/PCA 方法。未读任何保留阶段或禁窗数据（X3 输入 E8.75/E9.0 ≤ E9.5，合规）。+- 确定性：np.random.default_rng(seed)；PCA(random_state=0)；Sinkhorn_log、kNN、sample_rows 均确定。视图无关：分支只依赖 len(inputs) 与时间差。diff --git a/solution/README.md b/solution/README.mddeleted file mode 100644index e6cdf6f..0000000--- a/solution/README.md+++ /dev/null@@ -1,4 +0,0 @@-# copy_last--输出最新一个输入阶段（proxy：E8.5；final：E9.5），随机无放回抽到榜的细胞上限（T1 5118）。-不改表达，不改组成。这是地板：proxy 预期 ≈ 50（seed 0 实测 49.77；抽到 5118 个细胞，比整份 E8.5 的 50.00 略低）。diff --git a/solution/run.py b/solution/run.pyindex aa4bbf0..a6e65ad 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,11 +1,27 @@ #!/usr/bin/env python3-"""copy_last: the latest input stage, subsampled into the board's cell range."""+"""copy_last + damped OT displacement with sparse gene activation.++Single input stage -> exact copy_last fallback.+Two input stages -> Sinkhorn OT coupling between the last two stages in a+25-dim PCA of the top 2000 HVGs; per-cell displacement in gene space is+damped by ALPHA and applied only where it is non-negligible (zeros are not+lifted globally). Genes strongly up in pseudobulk can be sparsely activated+in cells where their local displacement is large.++Mechanism-off control: T1_OT_ALPHA=0 (output must equal copy_last exactly).+"""  from __future__ import annotations  import argparse+import os  import numpy as np+import scipy.sparse as sp+from sklearn.decomposition import PCA+from sklearn.neighbors import NearestNeighbors++import ot as pot  from src.task1_temporal.view_io import (     inputs_by_time,@@ -17,6 +33,38 @@ from src.task1_temporal.view_io import (     write_prediction, ) +ALPHA = float(os.environ.get("T1_OT_ALPHA", "0.30"))+N_HVG = 2000+N_PCA = 25+OT_MAX_CELLS = 3000+OT_EPS = 0.1+K_NN = int(os.environ.get("T1_OT_KNN", "10"))+OT_ITERS = 100+ACT_MAX_GENES = 300+ACT_MEAN_DELTA = 0.3+ACT_CELL_DISP = 0.5+ACT_SCALE = 0.1+APPLY_EPS = 0.01  # do not touch entries below this |alpha*d| (keeps sparsity)+++def hvg_indices(mats: list[sp.csr_matrix], k: int) -> np.ndarray:+    means = []+    sqs = []+    ns = []+    for m in mats:+        n = m.shape[0]+        ns.append(n)+        means.append(np.asarray(m.mean(axis=0)).ravel())+        sq = m.multiply(m)+        sqs.append(np.asarray(sq.mean(axis=0)).ravel())+    tot = float(sum(ns))+    w = [n / tot for n in ns]+    mean = sum(wi * mi for wi, mi in zip(w, means))+    sq = sum(wi * si for wi, si in zip(w, sqs))+    var = sq - mean**2+    var = np.maximum(var, 0.0)+    return np.argsort(-var)[:k]+  def main() -> None:     parser = argparse.ArgumentParser()@@ -27,10 +75,108 @@ def main() -> None:      manifest = load_manifest(args.data)     genes = panel_genes(args.data, manifest)-    last = read_stage(args.data, inputs_by_time(manifest)[-1], genes)+    inputs = inputs_by_time(manifest)+    last = read_stage(args.data, inputs[-1], genes)+     rng = np.random.default_rng(args.seed)-    rows = sample_rows(last.n_obs, target_n_cells(manifest, last.n_obs), rng)-    write_prediction(last.X[rows], genes, args.out, seed=args.seed)+    n_out = target_n_cells(manifest, last.n_obs)+    rows = sample_rows(last.n_obs, n_out, rng)++    if len(inputs) < 2 or ALPHA <= 0:+        write_prediction(last.X[rows], genes, args.out, seed=args.seed)+        return++    first = read_stage(args.data, inputs[-2], genes)+    X1, X2 = first.X.tocsr(), last.X.tocsr()++    hv = hvg_indices([X1, X2], N_HVG)+    hv = np.sort(hv)++    # subsample for OT+    def sub(n: int, rng_: np.random.Generator) -> np.ndarray:+        if n <= OT_MAX_CELLS:+            return np.arange(n)+        return np.sort(rng_.choice(n, OT_MAX_CELLS, replace=False))++    i1 = sub(X1.shape[0], rng)+    i2 = sub(X2.shape[0], rng)++    A = X1[:, hv][i1].toarray().astype(np.float32)+    B = X2[:, hv][i2].toarray().astype(np.float32)+    combined = np.vstack([A, B])+    pca = PCA(n_components=N_PCA, random_state=0)+    emb = pca.fit_transform(combined)+    E1 = emb[: A.shape[0]]+    E2 = emb[A.shape[0] :]++    cost = ((E1[:, None, :] - E2[None, :, :]) ** 2).sum(-1)+    scale = float(np.median(cost[cost > 0])) if (cost > 0).any() else 1.0+    cost = (cost / max(scale, 1e-8)).astype(np.float64)+    a = np.full(E1.shape[0], 1.0 / E1.shape[0])+    b = np.full(E2.shape[0], 1.0 / E2.shape[0])+    P = pot.sinkhorn(a, b, cost, reg=OT_EPS, numItermax=OT_ITERS, method="sinkhorn_log")+    if not np.isfinite(P).all():+        raise RuntimeError("sinkhorn produced non-finite coupling")++    A_full = X1[i1].toarray().astype(np.float32)+    B_full = X2[i2].toarray().astype(np.float32)++    # barycentric projection on the stage-1 side: smooth per-cell velocity+    # d_i = mean_j P[i,j] x2_j / sum_j P[i,j] - x1_i+    Pm = P / np.maximum(P.sum(axis=1, keepdims=True), 1e-12)+    d1 = Pm @ B_full - A_full  # displacement anchored at stage-1 cells++    # gap ratio: how far beyond the last input the target is, in units of the+    # input-to-input gap (uses only time differences -> view independent)+    times = [float(e["time"]) for e in inputs]+    gap_in = max(times[-1] - times[-2], 1e-6)+    gap_out = max(float(manifest["target"]["time"]) - times[-1], 0.0)+    ratio = gap_out / gap_in++    # transfer the field to every stage-2 cell: velocity of its nearest+    # stage-1 cell in PCA space+    nn1 = NearestNeighbors(n_neighbors=K_NN).fit(E1)+    _, idx1 = nn1.kneighbors(emb[A.shape[0] :])+    disp = d1[idx1].mean(axis=1)++    # activation gene set: strong pseudobulk increase stage1 -> stage2+    pb1 = np.asarray(X1.mean(axis=0)).ravel()+    pb2 = np.asarray(X2.mean(axis=0)).ravel()+    delta = pb2 - pb1+    act_genes = np.where(delta > ACT_MEAN_DELTA)[0]+    if len(act_genes) > ACT_MAX_GENES:+        act_genes = act_genes[np.argsort(-delta[act_genes])[:ACT_MAX_GENES]]+    act_set = np.zeros(len(genes), dtype=bool)+    act_set[act_genes] = True++    shift = (ALPHA * ratio * disp).astype(np.float32)+    delta_min = float(os.environ.get("T1_OT_DELTAMIN", "0.15"))+    if delta_min > 0:+        mover = np.abs(delta) >= delta_min+        keep = np.sign(shift) == np.sign(delta[None, :].astype(np.float32))+        shift = shift * (mover[None, :] & keep)++    Xout = X2[rows].toarray().astype(np.float32)+    s = shift[rows]+    apply_mask = (np.abs(s) >= APPLY_EPS) & (Xout > 0)+    n_activated = 0+    if ALPHA > 0:+        act_mask = act_set[None, :] & (Xout == 0) & (s > ACT_CELL_DISP * ALPHA * ratio)+        Xout = np.where(apply_mask | act_mask, np.maximum(Xout + s, 0.0), Xout)+        n_activated = int(act_mask.sum())++    if os.environ.get("T1_OT_DEBUG"):+        mag = np.linalg.norm(s, axis=1)+        print(+            f"[debug] alpha={ALPHA} disp_cells_frac(|d|>0.01)="+            f"{float((np.abs(disp) > 0.01).any(axis=1).mean()):.3f} "+            f"median|shift|={float(np.median(np.abs(s))):.4f} "+            f"activated_entries={n_activated} act_genes={len(act_genes)} "+            f"nnz_before={int((X2[rows] != 0).sum())} nnz_after={int((Xout != 0).sum())} "+            f"std_before={float(X2[rows].toarray().std()):.4f} std_after={float(Xout.std()):.4f}"+        )++    write_prediction(sp.csr_matrix(Xout), genes, args.out, seed=args.seed)   if __name__ == "__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

完成。总结：

**做了什么**：实现 PLAN 的 local_ot 家族——Sinkhorn OT（log 域，稳定）估计两输入阶段间的细胞速度场（stage-1 侧重心投影 + PCA k=10 近邻传递），按 α·(Δt_out/Δt_in) 阻尼施加到表达非零条目，仅保留伪批量 mover（|Δ|≥0.15）且方向一致的分量，含稀疏激活；单输入视图（proxy10）精确退化为 copy_last（逐值验证相同）。α=0 对照与 copy_last 逐值一致，机制证据齐全（100% 细胞有位移、268 个零条目被激活、std 不塌不炸）。

**查分结果**（X3 A 半，9/20 次）：提交配置 47.35 vs 父 copy_last 47.92（−0.57，噪声内）；预期节点分 ≈(52.75+2×47.35)/3 ≈ 49.2，**未达 PLAN 接受标准（>51.53）**。未 mask 位移提高 de_score（−0.195→−0.143）但 mmd 大跌；mask 保住 mmd 却抹掉 DE 增益——E8.75→E9.0 速度方向与 E9.0→E9.5 真实变化相关性弱，疑似混入 Qiu 样本间技术差异。proxy 与 X3 均通过 vec-check，输出对 seed 确定。

**下一步最值得试**：(1) 速度场去技术批次——先在两阶段间对齐伪批量（扣除全局均值差）再估 OT 位移；(2) 用 celltype 标签分层估计速度（同型内耦合，避免跨谱系错配）；(3) 混合解码：mask 版本 + 未 mask 位移按 |disp| 分位数裁剪，尝试同时保住 mmd 和 de_score 增益；(4) 树上的 node 3（composition_trend, proxy10 63.66）与本家族正交，组合潜力大。

调研员

ingle input, the mechanism cannot fire at all—Engineer should check inputs_by_time length on X3 first and report. 2) Damping too low → indistinguishable from copy_last (no-change protection risk, though unlikely given scoring facts). 3) Damping too high → overshoot hurts mmd_u and variogram. 4) OT computation on 2000+ cells may take >5 min; Engineer should subsample to 3000 cells per stage for OT, then map displacement to all cells. Early detection: run with alpha=0 and alpha=0.15, compare outputs; if identical, mechanism not firing.",
 "family_id": "local_ot",
 "mechanism": "OT coupling between two input stages estimates per-cell ancestry and displacement; damped extrapolation (alpha << 1) shifts the last stage toward the predicted next state, with sparse activation allowing genes to turn on.",
 "vs_constant_shift": "Constant per-type shift applies the same delta to all cells of a type regardless of their position in state space. Here, displacement is cell-specific (via OT coupling and nearest-neighbor assignment), gene activation is conditional on local displacement magnitude, and the damping factor is tuned to avoid overshooting. No type labels are used.",
 "mechanism_evidence": "Engineer should report: (1) fraction of cells with |displacement| > 0.01 (should be >50%); (2) number of genes activated from zero (should be >0 if mechanism fires); (3) per-metric change vs copy_last on both rulers; (4) correlation between displacement magnitude and alpha value (linear scaling confirms mechanism is active); (5) std of predicted expression vs input (should increase slightly, not collapse).",
 "mechanism_off_control": "Set alpha=0 in the same program. Output must be identical to copy_last (same cells, same expression, same subsampling seed). If alpha=0 output differs from copy_last, there is a bug in the pipeline. Expected difference between alpha=0 and alpha=0.15: DE metrics improve on X3, mmd_u may shift slightly, variogram approximately preserved.",
 "sources": []}
```