总览 · ← 返回运行 20261001-204440-search-t1-g21q-A

节点 n5

快照内解剖重加权：以最后一个官方输入阶段（proxy=E8.5）的真实细胞为池，按心脏谱系上调、表面外胚层/轴旁中胚层/胚外中胚层下调、去掉神经管的比例重采样 5118 个真实细胞；表达不做任何位移。

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 153ec124fb: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 001d46154c。改动的文件：solution/METHOD.md +30 −0、solution/run.py +115 −5

diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..e81b4b0--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,30 @@+快照内解剖重加权：以最后一个官方输入阶段（proxy=E8.5）的真实细胞为池，按心脏谱系上调、表面外胚层/轴旁中胚层/胚外中胚层下调、去掉神经管的比例重采样 5118 个真实细胞；表达不做任何位移。++## 方法++从 copy_last 父节点（50.04）改进。E8.5→E9.5 群体均值几乎不动（reweight 模块记录 pseudobulk r 0.994），真正变的是**解剖组成**：心脏中胚层/心内膜/JCF/心包扩大，表面外胚层、胚外中胚层、轴旁中胚层、神经管收缩。因此只改组成、不改表达，从最后一个官方输入阶段（`inputs_by_time(include_external=False)`，避免 proxy2 上误取 Qiu E9.0）重采样真实细胞：++- `HEART_TYPES`（CM 各型、aSHF/pSHF、aPHM/pPHM、内皮/心内膜/BEC、JCF、Pericardium、Proepicardium）权重 **1.6**；+- `EDGE_TYPES`（Surface Ectoderm、EXEM、Paraxial Mesoderm）权重 **0.25**；+- `DROP_TYPES`（Neural Tube）权重 0；+- 其余类型权重 1.0；+- 按权重×丰度用 largest-remainder 分配到 `target_n_cells`（proxy=5118），逐型无放回抽样（不足则有放回），seed 由 `--seed` 控制。++集合定义与 `src.task1_temporal.reweight`（run2 winner heart_jcf_peri）一致，覆盖 E8.5 与 E9.5 两套命名，故同一规则在 proxy/final 通用。++## 生物学来源++只用了不针对禁窗阶段的通用定性解剖知识：小鼠 E8.5→E9.5 心脏发生（cardiogenesis）中心脏谱系扩张、神经管/轴旁中胚层/胚外组织相对收缩，属已发表的早期发育谱系关系（≤E9.5，公开阶段）。未读取、未引用任何保留阶段（E10.5/E12.5）或禁窗（9.5<E≤13.5）的测量、类型清单、比例或标记基因；未读取 `uns.celltype_palette`。++## 验证过什么++`vec-score`（A 半，proxy 视图）：+- 默认配置（heart 1.6, edge 0.25, n=5118, 无位移）：seed0 **55.44**，seed1 **55.92**（cell_state/covariation/de_recovery/direction 均衡，direction≈59）。相比父节点 copy_last 50.04 提升约 +5.5，与 seed heart_jcf_peri（node2, 56.23）同族同量级。+- 组成扫参：heart 2.0=55.63、edge 0.10=56.2、n_cells 4000=56.19、2000=56.46(seed0) 但 2000 跨 seed 为 {56.46,55.78,54.88} 均值 55.7 —— 全部落在 T1 噪声（~2 分）内，未采纳更激进配置，保留 node2 的稳健默认（heart1.6/edge0.25/n=target）。+- **表达位移（maturation-axis shift）为净负**：均匀加多能性下降方向使 covariation 从 54 崩到 43(alpha0.15)/35(alpha0.3)，虽 direction/cell_state 略升，总分反降（54.4/53.3）。默认 shift_alpha=0，位移分支不触发。++## 没验证 / 风险++- 只在 proxy（E8.5 单输入）验证；final（E8.5+E9.5→E10.5）与 proxy2（含 Qiu E9.0 心脏外部输入）未查分。程序显式 `include_external=False` 取官方最新阶段，逻辑在两种视图都能跑通，但 final 目标阶段解剖变化更大，固定 heart1.6/edge0.25 未必最优。+- 硬编码类型名依赖官方 E8.5/E9.5 命名；外部测试题（如 X3）若用不同命名会全部落到权重 1.0，退化为 copy_last 子集（不崩，但无增益）。+- A 半调参，正式分用 B 半 + 3 seed；所有配置差异 <2 分，视为同族噪声，选择稳健默认而非最高单点。diff --git a/solution/run.py b/solution/run.pyindex aa4bbf0..e5d6a9c 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,36 +1,146 @@ #!/usr/bin/env python3-"""copy_last: the latest input stage, subsampled into the board's cell range."""+"""Anatomical composition reweighting of the latest official input stage.++E8.5 -> E9.5 the population means barely move; what changes is the dissection:+heart-field / cardiac lineages expand while surface ectoderm, extra-embryonic+mesoderm, paraxial mesoderm and neural tube shrink (published, pre-E9.5 stage+anatomy only). We therefore resample real cells from the last official input+with per-type multipliers, never shifting expression.+"""  from __future__ import annotations  import argparse  import numpy as np+from scipy import sparse  from src.task1_temporal.view_io import (     inputs_by_time,     load_manifest,     panel_genes,     read_stage,-    sample_rows,     target_n_cells,     write_prediction, ) +HEART_TYPES = {+    "OFT/RV-CM", "IFT-CM", "AVC-CM", "SV-CM", "LV-CM", "RV-CM", "V-CM",+    "Endothelium", "Endocardium", "BEC",+    "aSHF", "pSHF", "aPHM", "pPHM",+    "JCF", "Pericardium", "Proepicardium",+}+EDGE_TYPES = {"Surface Ectoderm", "EXEM", "Paraxial Mesoderm"}+DROP_TYPES = {"Neural Tube"}+# per-type overrides (tuning); values multiply the base 1.0+EXTRA_WEIGHTS: dict[str, float] = {}++# Published, non-held-out developmental biology: pluripotency factors decline as+# the embryo progresses from E8.5 toward E9.5 (gastrulation -> organogenesis).+PLURI = ["Pou5f1", "Nanog", "Sox2", "Lin28a"]+++def maturation_axis(X: sparse.csr_matrix, genes: list[str], frac: float = 0.25) -> np.ndarray:+    """Unit-free forward (developmental) direction from a single snapshot.++    Score each cell by how low its pluripotency programme is; the difference of+    the top vs bottom quartile log-expression means is the maturation direction.+    """+    idx = [genes.index(g) for g in PLURI if g in set(genes)]+    sub = np.asarray(X[:, idx].todense(), dtype=np.float32)  # (n, k)+    # per-gene z across cells, then mean; lower pluripotency -> higher score+    mu = sub.mean(0)+    sd = sub.std(0) + 1e-6+    z = (sub - mu) / sd+    score = -z.mean(1)+    n = score.shape[0]+    order = np.argsort(score)+    lo = order[: int(frac * n)]+    hi = order[-int(frac * n):]+    d = np.asarray(X[hi].mean(axis=0)).ravel() - np.asarray(X[lo].mean(axis=0)).ravel()+    return d.astype(np.float32)+++def largest_remainder(weights: np.ndarray, n: int) -> np.ndarray:+    weights = np.clip(np.asarray(weights, dtype=np.float64), 0, None)+    if n <= 0 or weights.sum() <= 0:+        return np.zeros(len(weights), dtype=int)+    raw = weights / weights.sum() * n+    out = np.floor(raw).astype(int)+    short = int(n - out.sum())+    order = np.argsort(-(raw - out))+    for i in order[:short]:+        out[i] += 1+    return out+++def type_weight(t: str, heart_w: float, edge_w: float) -> float:+    if t in EXTRA_WEIGHTS:+        return EXTRA_WEIGHTS[t]+    if t in EDGE_TYPES:+        return edge_w+    if t in HEART_TYPES:+        return heart_w+    return 1.0+  def main() -> None:     parser = argparse.ArgumentParser()     parser.add_argument("--data", required=True)     parser.add_argument("--out", required=True)     parser.add_argument("--seed", type=int, default=0)+    parser.add_argument("--heart-weight", type=float, default=1.6)+    parser.add_argument("--edge-weight", type=float, default=0.25)+    parser.add_argument("--drop", type=str, default="Neural Tube")+    parser.add_argument("--shift-alpha", type=float, default=0.0)+    parser.add_argument("--n-cells", type=int, default=0)+    parser.add_argument("--extra", type=str, default="")     args = parser.parse_args() +    global DROP_TYPES, EXTRA_WEIGHTS+    DROP_TYPES = {d for d in args.drop.split(",") if d}+    EXTRA_WEIGHTS = {}+    for tok in args.extra.split(","):+        if ":" in tok:+            k, v = tok.split(":")+            EXTRA_WEIGHTS[k.strip()] = float(v)+     manifest = load_manifest(args.data)     genes = panel_genes(args.data, manifest)-    last = read_stage(args.data, inputs_by_time(manifest)[-1], genes)+    last = read_stage(args.data, inputs_by_time(manifest, include_external=False)[-1], genes)++    labels = last.obs["celltype"].astype(str).to_numpy() if "celltype" in last.obs else np.zeros(last.n_obs, dtype=str)+    X = last.X if sparse.issparse(last.X) else sparse.csr_matrix(np.asarray(last.X))++    axis = maturation_axis(X, genes) if args.shift_alpha > 0 else None+     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_cells = target_n_cells(manifest, last.n_obs)+    if args.n_cells > 0:+        n_cells = int(np.clip(args.n_cells, manifest["min_cells"], manifest["max_cells"]))++    types = [str(t) for t in np.unique(labels) if str(t) not in DROP_TYPES]+    counts = np.array([(labels == t).sum() for t in types], dtype=np.float64)+    weights = np.array([type_weight(t, args.heart_weight, args.edge_weight) for t in types], dtype=np.float64)+    alloc = largest_remainder(counts * weights, n_cells)++    blocks = []+    for t, n in zip(types, alloc):+        if n <= 0:+            continue+        pool = np.flatnonzero(labels == t)+        choice = rng.choice(pool, size=int(n), replace=pool.size < n)+        blocks.append(X[choice])+    out = sparse.vstack(blocks, format="csr").astype(np.float32)+    out.eliminate_zeros()++    if axis is not None:+        dense = np.asarray(out.todense(), dtype=np.float32)+        dense = np.clip(dense + args.shift_alpha * axis[None, :], 0, None)+        out = sparse.csr_matrix(dense.astype(np.float32))+        out.eliminate_zeros()++    write_prediction(out, 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

计划里引用的来源

• PMID 35288709 — Qiu et al. 2022 (TOME)：小鼠胚胎相邻阶段细胞状态间存在有向的 pseudo-ancestor→pseudo-descendant 关系，仅作'单快照内可定义发育前进方向'的定性依据；未取其中任何按阶段/按簇的数值统计，尤其未使用禁窗阶段信息。

文献检索

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. Final submission in `solution/` defaults to the robust composition-reweighting config (heart 1.6, edge 0.25, drop Neural Tube, no expression shift, target cell count) — reproducing node 2's proven heart_jcf_peri family.

**What I did:** Improved parent (copy_last, 50.04) by resampling real cells from the last official input stage with anatomical composition weights (heart/cardiac lineages up, surface-ectoderm/EXEM/paraxial down, neural tube dropped), never shifting expression — E8.5→E9.5 barely moves population means, only the dissection/composition changes. Used only ≤E9.5 published lineage knowledge; no held-out stage info.

**Scoring (A-half, proxy):**
- Default config: seed0 **55.44**, seed1 **55.92** (direction≈59, balanced) — vs parent 50.04, **+5.5**, matching node 2 (56.23).
- Composition sweeps (heart 2.0, edge 0.10, n_cells 4000/3000/2000) all landed 55–56.5, within T1 noise (~2); n2000's 56.46 was a lucky seed (3-seed mean 55.7), so kept the stable node-2 defaults rather than overfitting.
- Maturation-axis expression shift was **net negative**: uniform additive shift collapsed covariation (54→43→35), outweighing small direction/cell_state gains. Left `shift_alpha=0`.

Verified `run.py` runs on the full proxy view (5118×32285, float32, finite) and passes `vec-check`; has a safe copy_last fallback when `celltype` is absent.

**Next best directions:** (1) The reweighting family plateaus ~56 — further gains need a signal beyond composition, but naive global expression shift destroys covariation; a *type-local* shift that preserves within-type covariance (needs a maturation direction estimable per lineage) is the untested lever. (2) Validate/tune weights on `final` (E8.5+E9.5→E10.5) where anatomical change is larger, and make heart/edge weights robust to non-official type names (X3 test).

调研员

 + 小幅位移'守住不掉。",
  "approach": "全部只用视图内数据；以最后一个输入阶段（proxy=E8.5，final=E9.5）为唯一快照。步骤：(1) 预处理：若输入非 log 空间则 log1p(CP10k)；选 top-2000 高变基因，PCA 取 50 维；细胞数 >40k 时先无放回抽 40k 建 kNN 图（k=30，cosine/pca 空间），其余细胞按最近邻投影到图上，控制 30 分钟时限。(2) 成熟度锚定：s_i = mean(z(器官发生/分化相关基因集)) − mean(z(多能性基因 Pou5f1, Nanog, Sox2, Lin28a))——多能性随发育下降是已发表的通用定性知识；以 s 最低的 5% 细胞为根计算扩散伪时间 p_i（或直接用图拉普拉斯特征向量按 s 定向）。(3) 前向位移：对每个细胞取其 kNN 中伪时间更高的'前向邻居'，v_i = mean(x_fwd − x_i)，再在图上做 2 轮邻居平均去噪，并对每个基因按 MAD 截断（±3 MAD），防个别基因主导。(4) 预测表达：x'_i = clip(x_i + α·v_i, 0)，α 初值 0.25，网格 {0, 0.25, 0.5}。(5) 组成倾斜：抽样权重 w_i ∝ exp(β·p̃_i)·(1 + γ·c_i)，p̃ 为标准化伪时间，c_i 为细胞周期分（Mki67, Top2a, Ccnb1, Cdk1, Pcna 均值——通用增殖标记），β 初值 0.3（网格 {0, 0.3}），γ 初值 0.5（网格 {0, 0.5}）；按 w 无放回抽 target_n_cells（5118）。(6) vec-score 筛选流程（≤20 次查分，T1 噪声 ~2 分）：第 1 查 α=β=γ=0 复现 ~50 作管线校验；随后单因子：α=0.25、α=0.5 各一查；取较优 α 后加 β=0.3 一查、再加 γ=0.5 一查；总计约 6–8 查。只有超过地板 +2 的组件才保留；两个 α 都在噪声内则固定 α=0.25（宁小勿大，防 covariation 崩）。(7) 退化自检（k016）：每基因方差比 pred/input 应在 0.8–1.2，无非有限值/负值，伪时间最低 10% 细胞在输出中占比不低于输入的 1/3（防倾斜抹掉早期/稀有状态）。(8) 单输入退路与 final 迁移：本方法主路径就是单阶段，proxy 直接可跑；final（E8.5+E9.5 两输入）时对 E9.5 快照跑同一管线，另加可选融合：KMeans(K≈30) 聚 E9.5，按质心余弦最近匹配 E8.5 簇，得 per-cluster delta_c，按 k018 取收缩系数 0.5，v_i ← (1−λ)v_i + λ·0.5·delta_c(i 所属簇)，λ=0.5；缺失/新生簇无 delta 则 λ=0 退回纯伪时间位移。proxy2（含 Qiu E9.0 心脏外部输入）不为它做任何专门设计，缺失基因按 view_io 默认 E8.5 均值补齐即可。",
  "expected_groups": ["direction", "de_recovery", "cell_state"],
  "risks": "(1) 锚定方向搞反（多能性基因不在面板或 z 化失败）→ 位移向'过去'，第一次 α=0.25 查分若 <50 且明显低于 α=0 校验值，先试翻转伪时间符号再重查一次，仍不行则退回 α=0 只留组成倾斜。(2) α 过大 → 表达离流形，covariation 与 cell_state 同时掉；靠 MAD 截断 + α≤0.5 网格 + 方差比自检尽早发现。(3) 倾斜过强 → 稀有早期状态丢失，被 ED/MMD/CSS 惩罚；用第 (7) 步占比检查，β 只允许 0.3 一档。(4) A 半过拟合：网格刻意粗糙（3×2×2），任何 <2 分的'提升'视为噪声不采纳，正式分以 B 半为准。(5) 时限：全量 PCA/kNN 超时风险 → 先 40k 子样本建图跑通、查一次分确认方向，再决定是否全量。若所有组件均无 >2 分信号，提交与父节点等价的 copy_last 变体（不亏分），并在 METHOD.md 记录'单快照伪时间外推在此尺子无效'供后续节点避坑。",
  "sources": ["PMID 35288709 — Qiu et al. 2022 (TOME)：小鼠胚胎相邻阶段细胞状态间存在有向的 pseudo-ancestor→pseudo-descendant 关系，仅作'单快照内可定义发育前进方向'的定性依据；未取其中任何按阶段/按簇的数值统计，尤其未使用禁窗阶段信息。"]
}
```