总览 · ← 返回运行 20261002-034201-search-t1-abc-r1-A-era

节点 n9 在终选来历上

类型层（增殖↓/凋亡↓）× 细胞层（OXPHOS−糖酵解↑）双权重的组成重抽样：细胞表达原样复制最新官方输入阶段，只改哪些细胞被保留。

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 7e1813b897: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：这个提交的上一版（种子程序：相对空仓库）。改动的文件：solution/METHOD.md +45 −0、solution/run.py +145 −0

diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..5be1902--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,45 @@+类型层（增殖↓/凋亡↓）× 细胞层（OXPHOS−糖酵解↑）双权重的组成重抽样：细胞表达原样复制最新官方输入阶段，只改哪些细胞被保留。++## 方法++- 基底：`view_io.inputs_by_time(manifest, include_external=False)` 的最后一个官方输入阶段（外部 Qiu E9.0 永不作基底；proxy2 与 proxy 输出一致）。没有官方输入时退回全部输入。+- 标签：`obs["celltype"]`（缺失时按 `cm_celltype/cell_type/annotation` 顺序找，都没有则在副本上 leiden 聚类，不改动写出的 X）。+- 两个权重层（全部从输入阶段现场计算，无任何硬编码统计量）：+  - 类型层：每型的增殖分（34 个周期基因每细胞均值→按型平均）与凋亡分（13 个促凋亡基因）在各型之间做 z 分（clip ±3），`w_t = exp(-0.55·z_prolif_t + (-0.25)·(-z_apopt_t))`，即 A_TP=-0.55、A_TA=0.25；细胞数 < 5 的型 w_t=1。+  - 细胞层：代谢成熟轴 `z_met_i = z(OXPHOS 13 基因均值 − 糖酵解 10 基因均值)`，系数 A_CM=+0.7；A_CP=A_CA=0（每细胞增殖轴实测被类型层取代后无增益）。+- 抽样：`w_i = w_{type(i)} · exp(A_CM·z_met_i)`，Efraimidis–Spirakis（keys = log(u)/w，取 top-n）做加权**无放回**抽样，n = `target_n_cells`（proxy/proxy2 = 5118，X3 = 2174 = 全量）。`np.random.default_rng(seed)`，同 seed 确定。+- 表达值不做任何平移/缩放（伪批量平移在三把尺子上此前均实测有害）。++## 关键参数与查分（A 半，proxy）++| 配置 | proxy |+|---|---|+| 类型层 A_TP=+0.5, A_TA=0.3 | 40.62（方向反了：de_direction −0.24） |+| A_TP=−1.0 | 49.17 |+| A_TP=−0.5, A_TA=0.3 | 54.74 |+| A_TP=−0.5 + 每细胞增殖 −0.7/凋亡 −0.2/代谢 +0.5 | 54.05 |+| A_TP=−0.5, A_CM=0.5 | 57.24 |+| A_TP=−0.6, A_TA=0.2, A_CM=0.6 | 58.15 |+| A_TP=−0.5, A_TA=0.2, A_CM=0.6 | 58.00 |+| **默认 A_TP=−0.55, A_TA=0.25, A_CM=0.7** | **57.92** |++默认取在 −0.5…−0.6 / 0.2…0.3 / 0.6…0.8 的平台中间，不是单点峰值（A/B 半与噪声约 2 分）。++## 验证过什么++- proxy 57.92、proxy2 57.92（两者同基底同输出）、X3 50.00；三视图 `vec-check` 全 ok，运行 ~2–4 s。+- 四组全部不弱于此前最佳节点 8（proxy 56.69：cell_state 56.39 / covariation 51.77 / de_recovery 52.04 / direction 56.71）→ 本节点 59.77 / 56.73 / 54.08 / 60.48。+- 类型层与细胞层符号相反于直觉但数据驱动：型水平上强增殖的型（E8.5 的神经/外胚层类 progenitor）在下一步占比下降，代谢成熟度高的细胞占比上升。++## 没验证 / 局限++- X3 上 `n_out == 池大小`，加权无放回抽样退化为恒等，所以 X3 = copy_last = 50.00；两阶段（E8.75→E9.0）信息完全没用上。+- 未在 final 视图（E8.5+E9.5 → E10.5）实测；类型名换成 E9.5 词表时，类型层仍只依赖现场计算的基因分数，不依赖名字，逻辑上可迁移，但分数未验证。+- 只测了 seed 0；未做多 seed 稳定性检验。+- 生物学知识来源：仅通用细胞周期 / 凋亡 / OXPHOS 与糖酵解基因清单（教科书级通路成员，见 run.py 顶部常量），不涉及任何保留阶段或保留基因型的测量；未读取 `uns.celltype_palette`，未使用 external/ 数据，未使用 prior/ 资源。++## 下一步最值得试++1. X3 是唯一没动的尺子：池化 E8.75+E9.0 后按同一权重层抽样（n_out < 池大小即可让权重生效），或用 E8.75→E9.0 的型比例差外推 0.5 天。+2. 类型层再加一轴：型水平的「成熟/分化标记」（如心肌 Tnnt2/Myl7、内皮 Pecam1/Cdh5、血 Hbb）z 分，可能与代谢轴互补。+3. final 视图（两官方阶段）上用 E8.5→E9.5 的型比例差与型分数差做一次收缩外推，与纯类型层重加权比较。diff --git a/solution/run.py b/solution/run.pynew file mode 100644index 0000000..c29fcb8--- /dev/null+++ b/solution/run.py@@ -0,0 +1,145 @@+"""Type-level proliferation/apoptosis weighted composition resampling.++Base pool = latest *official* input stage (external stages are never a base).+Cells are copied unchanged; only the composition (which cells are kept) is+re-weighted, using two orthogonal weight layers:++  * type layer  : w_t = exp(A_TP * z_prolif_t - A_TA * z_apopt_t)+                  z-scores taken across cell types of the base stage+  * cell layer  : w_i = exp(A_CP * z_prolif_i - A_CA * z_apopt_i + A_CM * z_met_i)++Final weight of cell i = w_{type(i)} * w_i, sampled without replacement via+Gumbel/efraimidis-spirakis keys (deterministic under --seed).+"""++from __future__ import annotations++import argparse+import os+from pathlib import Path++import numpy as np+from scipy import sparse++from src.task1_temporal import view_io++CYCLE = [+    "Mki67", "Top2a", "Pcna", "Ccna2", "Ccnb1", "Ccnb2", "Ccnd1", "Ccneg", "Ccne1",+    "Cdk1", "Cdk2", "Cdk4", "Cdk6", "Mcm2", "Mcm3", "Mcm4", "Mcm5", "Mcm6", "Mcm7",+    "Orc1", "Cdc6", "Cdt1", "Rrm1", "Rrm2", "Tyms", "Dtl", "Cenpf", "Cenpe",+    "Birc5", "Aurkb", "Plk1", "Kif20a", "Kif11", "Nusap1",+]+APOPT = [+    "Bax", "Bak1", "Bcl2l11", "Pmaip1", "Bbc3", "Bad", "Bid", "Bik", "Bmf",+    "Bnip3", "Casp8", "Casp9", "Casp3",+]+OXPHOS = [+    "Ndufa4", "Ndufb8", "Uqcrb", "Uqcrc1", "Cox5a", "Cox6b1", "Cox7a2", "Atp5a1",+    "Atp5b", "Atp5f1b", "Atp5pb", "Sdha", "Sdhb",+]+GLYC = ["Slc2a1", "Slc2a3", "Hk1", "Hk2", "Pfkp", "Pgk1", "Pgam1", "Eno1", "Ldha", "Pkma"]++DEFAULTS = dict(A_TP=-0.55, A_TA=0.25, A_CP=0.0, A_CA=0.0, A_CM=0.7, MIN_TYPE_CELLS=5)+++def gene_scores(adata, genes, groups):+    """Per-cell mean expression of each gene group (missing genes ignored)."""+    index = {g: i for i, g in enumerate(genes)}+    n = adata.n_obs+    out = []+    for grp in groups:+        cols = [index[g] for g in grp if g in index]+        if not cols:+            out.append(np.zeros(n, dtype=np.float64))+            continue+        sub = adata.X[:, cols]+        out.append(np.asarray(sub.mean(axis=1), dtype=np.float64).ravel())+    return out+++def zscore(x, w=None):+    if w is None:+        m, s = x.mean(), x.std()+    else:+        m = np.average(x, weights=w)+        s = np.sqrt(np.average((x - m) ** 2, weights=w))+    if not np.isfinite(s) or s < 1e-9:+        return np.zeros_like(x)+    return np.clip((x - m) / s, -3.0, 3.0)+++def type_labels(adata):+    for col in ("celltype", "cm_celltype", "cell_type", "annotation", "leiden"):+        if col in adata.obs.columns:+            return adata.obs[col].astype(str).to_numpy(), col+    import scanpy as sc++    tmp = adata.copy()+    sc.pp.normalize_total(tmp, target_sum=1e4)+    sc.pp.log1p(tmp)+    sc.pp.pca(tmp, n_comps=30)+    sc.pp.neighbors(tmp)+    sc.tl.leiden(tmp, resolution=1.0, key_added="leiden_auto")+    return tmp.obs["leiden_auto"].astype(str).to_numpy(), "leiden_auto"+++def main():+    ap = argparse.ArgumentParser()+    ap.add_argument("--data", required=True)+    ap.add_argument("--out", required=True)+    ap.add_argument("--seed", type=int, default=0)+    for k, v in DEFAULTS.items():+        ap.add_argument("--" + k.lower(), dest=k, type=float if k != "MIN_TYPE_CELLS" else int,+                        default=float(os.environ.get("VEC_" + k, v)))+    args = ap.parse_args()++    view = Path(args.data)+    manifest = view_io.load_manifest(view)+    genes = view_io.panel_genes(view, manifest)++    entries = view_io.inputs_by_time(manifest, include_external=False)+    if not entries:+        entries = view_io.inputs_by_time(manifest, include_external=True)+    base = entries[-1]+    adata = view_io.read_stage(view, base, genes, missing="fill")++    labels, _ = type_labels(adata)+    z_prolif_c, z_apopt_c = gene_scores(adata, genes, [CYCLE, APOPT])+    ox, gl = gene_scores(adata, genes, [OXPHOS, GLYC])+    z_met_c = ox - gl++    z_prolif_c = zscore(z_prolif_c)+    z_apopt_c = zscore(z_apopt_c)+    z_met_c = zscore(z_met_c)++    uniq, inv = np.unique(labels, return_inverse=True)+    counts = np.bincount(inv, minlength=len(uniq))+    def tmean(x):+        return np.bincount(inv, weights=x, minlength=len(uniq)) / np.maximum(counts, 1)+    z_prolif_t = zscore(tmean(z_prolif_c))+    z_apopt_t = zscore(tmean(z_apopt_c))++    w_type = np.exp(args.A_TP * z_prolif_t - args.A_TA * z_apopt_t)+    w_type[counts < args.MIN_TYPE_CELLS] = 1.0+    w = w_type[inv] * np.exp(args.A_CP * z_prolif_c - args.A_CA * z_apopt_c + args.A_CM * z_met_c)+    w = np.clip(w, 1e-6, 1e6)++    n_out = view_io.target_n_cells(manifest, adata.n_obs)+    rng = np.random.default_rng(args.seed)+    u = rng.random(adata.n_obs)+    keys = np.log(np.maximum(u, 1e-300)) / w+    if n_out >= adata.n_obs:+        idx = np.arange(adata.n_obs)+    else:+        idx = np.sort(np.argpartition(-keys, n_out - 1)[:n_out])++    X = adata.X[idx]+    coords = None+    if manifest.get("needs_coords"):+        src = adata.obsm.get("spatial_3D")+        coords = np.asarray(src)[idx] if src is not None else np.zeros((len(idx), 3), np.float32)+    view_io.write_prediction(X, genes, args.out, coords=coords, seed=args.seed)+++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

**做了什么**：draft 了「类型层 × 细胞层」双权重组成重抽样（新家族：权重在**细胞类型**层而非单细胞层）——基底取最新官方输入阶段，按型的增殖/凋亡 z 分（A_TP=−0.55, A_TA=0.25）乘每细胞代谢成熟轴 OXPHOS−糖酵解（A_CM=+0.7），Efraimidis–Spirakis 加权无放回抽样，表达值不动。符号由查分定（+0.5 型层得 40.6，方向反）。

**查分**：proxy 57.92 / proxy2 57.92 / X3 50.00（X3 池大小 = n_out，退化为恒等）；三视图 vec-check ok，运行 ~3 s。四项分组全部不低于此前最佳节点 8（cell_state 59.8 vs 56.4，covariation 56.7 vs 51.8，direction 60.5 vs 56.7，de_recovery 54.1 vs 52.0）。默认参数取在参数平台中央而非单点峰值。

**下一步最值得试**：① X3 是唯一没动的尺子——池化 E8.75+E9.0 让权重生效，或用两阶段型比例差外推；② 类型层加「成熟标记」轴（Tnnt2/Myl7、Pecam1/Cdh5、Hbb）；③ 在 final 双官方阶段视图上用 E8.5→E9.5 的型分数差做收缩外推。

调研员

ts cycling cells per-cell; but at type level, expanding types may have higher proliferation). Engineer should test α = +0.5 first (matching seed's cardiac upweight), then α = −0.5 if that fails. Step 6: Compute per-cell sampling weight: each cell inherits its type's weight w_t. Normalise to probabilities p_i = w_{type(i)} / Σ w. Sample n_out cells without replacement (if n_out ≥ n_cells, keep all; this is the X3 degenerate case where n_out == n_available). Step 7: Output sampled cells unchanged (no expression modification). Single-stage fallback (proxy, X3): method uses only the single input stage, works identically. Proxy2 two-stage: use only the last official stage as base (same as node 2/3 fix); ignore external Qiu input for base selection. If view_io.inputs_by_time returns two inputs on proxy2, take inputs[0] (E8.5 official) as the pool. vec-score fast screening: run on proxy first; if proxy score < 52 (below node 4), abort and try opposite sign of α. Then check proxy2 and X3. Total runtime target: < 30s (simple sampling, no GPU).",
  "expected_groups": ["cell_state", "direction"],
  "risks": "1) Type labels may not exist in obs or may use unexpected key names → Engineer should print adata.obs.columns first; fallback to leiden clustering. 2) The sign of α is uncertain: if proliferating types should be DOWNweighted (as node 7 suggests per-cell), α=+0.5 will fail → test α=−0.5 immediately. 3) Type-family mapping may miss rare types or split a type across families → use substring priority (longest match first). 4) If n_out ≥ n_cells (X3 case), method degenerates to copy_last (score ~50) → acceptable, matches existing nodes. 5) Proliferation at type level may be noisy for types with <10 cells → floor type count: if type has <5 cells, assign weight=1 (neutral). 6) The approach may not beat node 7's per-cell method if type composition is already well-matched; in that case the draft still provides a combinable component for future improve nodes.",
  "sources": []
}
```

审查员

REVIEW.json written