总览 · ← 返回运行 20261003-093415-search-t1-r2-D-s1

节点 n15

METHOD

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 5a1b518006: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/EXECUTION.json +1 −0、solution/METHOD.md +50 −0、solution/run.py +250 −0

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..89beaee--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,50 @@+# METHOD++lineage_branching：Isl1+→Nkx2-5+ 承诺边上，逐细胞向其 kNN 中已承诺邻居的均值做稀疏表达位移（幅度=λ·readiness），双输入时先按观测时间差校验方向，不支持则跳过。++## 方法族与机制++PLAN family_id = `lineage_branching`（最小版本）。机制（与 PLAN 一致，逐细胞、非常数位移）：++1. 取最后一个输入阶段全部细胞；承诺分数 = mean(Nkx2-5, Tnnt2, Myh6, Actc1, Tnnt1, Tbx5) − mean(Isl1, Mef2c)（log 空间）。+2. 在心脏/中胚层系标签细胞（关键词匹配 CM/SHF/Pericardium/Endocardium/JCF/Mesoderm/Blood/Endothelium/NCC/Unknown 等，对 X3 的 Unknown 也生效）中取承诺分数前 20% 且承诺标记有表达者为 committed；不足 30 个 → 不修改（support check）。+3. HVG 2500 → TruncatedSVD 25 维 → kNN k=20；readiness = 邻居中 committed 的比例；readiness > 0.15 且自身非 committed 者为 branching。+4. 每个 branching 细胞的 target = 它自己的 committed 邻居的表达均值（逐细胞不同）；位移 step = λ·readiness·(target−x)，只作用于 mask = (target>0.5) ∧ (delta>0.25) 的基因（保持稀疏结构，避免把零整体抬高），clip ≥0；λ=0.5（环境变量 `LAMBDA` 可覆盖）。非 branching 细胞原样保留。+5. 双输入方向校验（X3 / final 生效）：按共同标签加权计算前一阶段→末阶段的伪批量差 d，与平均局部位移方向 mean_shift 求余弦；cos<0.1 → 跳过位移（输出=输入复制），0.1≤cos<0.3 → λ 减半（比 PLAN 的"减半"多加了更硬的丢弃门，依据是未加门时 X3 分数下降，见下）。+6. 输出细胞数 = min(末阶段细胞数, max_cells)，`default_rng(seed)` 无放回均匀抽样。++## 关键参数++k=20，HVG=2500，PCA=25，committed_frac=0.20，readiness_thr=0.15（PLAN 风险节建议试 0.15；0.25 时 proxy 只有 461 个 branching，0.15 → 582），λ=0.5（试 λ=1.0 得 53.14 vs 53.10，噪声内，保留 0.5），mask: target>0.5, delta>0.25。++## off 对照（mechanism_off_control）++同一程序 `--off`（或 `LAMBDA=0`）→ λ=0，输出=输入的均匀子抽样（copy_last 等价）。提交版保持打开（λ=0.5，无 off 开关生效）。++## 机制生效证据（proxy E8.5→E9.5，seed 0）++- committed 2265 / pool 11323；branching（被修改）582（3.5%）。+- readiness 分布 q10–q90 = 0,0,0,0,0.9（重尾：只有心脏区细胞有 committed 邻居）。+- 逐细胞位移范数 mean=4.48, q10=2.36, q50=4.17, q90=7.09（异质，非单一常数）。+- 平均位移最大的基因：Myh6, Kcnh7, Ankrd1, Myl3, Slc2a1, Acta1 —— 心脏成熟程序方向，符合 Isl1+→Nkx2-5+ 边的生物学。+- X3（E8.75+E9.0 心脏）：branching 候选 407（18.7%），但方向校验 cos=−0.074 < 0.1 → 按规则跳过位移，on 输出与 off 逐字节相同（门在起作用，不是机制没跑）。++## 查分结果（A 半，seed 0）++| 视图 | on | off(对照) | 分项(on) de_score/de_dir/mmd/vario |+|---|---|---|---|+| proxy10 | 53.10 | 52.97 | 12.90/14.69/15.55/9.96 |+| X3 | 47.62（=off，门跳过） | 47.62 | 10.77/12.28/14.85/9.72 |++早期无门版本在 X3 上：v1（dense mask target>0.05）45.69，variogram 掉到 8.02；v2（sparse mask）46.31，variogram 8.74 —— 说明"向邻居均值位移"在 cos≈0（方向未验证）时损害共变结构，因此加了丢弃门。proxy 上 on−off = +0.13，在 ±2 噪声内：**机制在 proxy 上方向正确（top 基因是心脏成熟基因）但幅度上未证明有效**，如实报告。++## 验证过 / 没验证++- 验证过：proxy 与 X3 视图跑通、vec-check ok、seed 0 确定性（两次运行逐字节一致）、off 对照、X3 门触发后 on==off。+- 没验证：proxy2 视图（本节点无该视图目录，但代码只依赖 manifest schema，双输入路径已在 X3 走过）；伪装视图重跑（代码不读绝对时间、路径、board/mode 字段）；final（E8.5+E9.5，cos 校验可能通过并实际位移）。+- λ、readiness 阈值只在 proxy10 上粗试，未做网格。++## 知识来源++- 标记基因角色（Nkx2-5/Tnnt2/Myh6/Actc1/Tnnt1/Tbx5 = 心脏承诺/成熟；Isl1/Mef2c = SHF 祖细胞）：通用谱系知识与 PLAN motivation 引用的 k063（Jia 2018, de Soysa 2019, Ameen 2022，全部 ≤E9.5 已发表阶段）。+- 未使用任何保留阶段（E10.5/E12.5、禁窗 9.5<E≤13.5）或保留基因型的测量信息；未读外部数据文件（proxy 视图 external 为空，X3 的 external 未使用）；未读 prior/ 文件；未读 `uns.celltype_palette`（视图中已剥掉）。diff --git a/solution/run.py b/solution/run.pynew file mode 100644index 0000000..5912529--- /dev/null+++ b/solution/run.py@@ -0,0 +1,250 @@+"""lineage_branching: local-neighborhood commitment shift along Isl1+ -> Nkx2-5+ edge.++For each cell of the latest input stage, compute a commitment score+(mean of cardiac commitment markers - mean of progenitor markers). Cells in the+top fraction of this score among cardiac/mesodermal-lineage cells are+"committed". Each non-committed cell gets readiness = fraction of its k nearest+neighbors that are committed; cells above a readiness threshold are "branching"+and are shifted toward the mean expression of their own committed neighbors,+magnitude scaled by readiness and lambda. Non-branching cells are unchanged.++Off control: LAMBDA env var or --off sets lambda=0 (pure copy of subsampled+input). Mechanism uses only within-stage data + generic marker knowledge+(cardiac commitment program: Nkx2-5/Tnnt2/Myh6/Actc1 up, Isl1/Mef2c progenitor),+no reserved-stage measurements.+"""+import argparse+import json+import os+import sys++import numpy as np+import scipy.sparse as sp+++# Generic cardiac-lineage knowledge (marker roles from public developmental+# biology literature on SHF commitment, active E7.5-E9.5; not stage-specific+# measurements):+COMMITMENT_MARKERS = ["Nkx2-5", "Tnnt2", "Myh6", "Actc1", "Tnnt1", "Tbx5"]+PROGENITOR_MARKERS = ["Isl1", "Mef2c"]+# Keywords for cardiac / lateral-plate-mesoderm lineage labels (official vocab).+CARDIAC_KEYWORDS = ["CM", "SHF", "Pericardium", "Endocardium", "JCF",+                    "Mesoderm", "Blood", "Endothelium", "NCC", "Unknown",+                    "Heart", "Epicard", "Proepicardium"]++K_NEIGHBORS = 20+N_HVG = 2500+N_PCS = 25+COMMITTED_FRAC = 0.20+READINESS_THRESHOLD = 0.15+MIN_COMMITTED = 30+TARGET_HI = 0.5+DELTA_MIN = 0.25+DEFAULT_LAMBDA = 0.5+++def log(msg):+    print(f"[lineage_branching] {msg}", file=sys.stderr, flush=True)+++def gene_idx(genes, names):+    pos = {g: i for i, g in enumerate(genes)}+    return [pos[n] for n in names if n in pos]+++def is_cardiac_label(labels):+    mask = np.zeros(len(labels), dtype=bool)+    for i, lab in enumerate(labels):+        s = str(lab)+        for kw in CARDIAC_KEYWORDS:+            if kw in s:+                mask[i] = True+                break+    return mask+++def marker_score(X, genes, up_names, dn_names):+    """mean(log expr of up markers) - mean(log expr of dn markers), per cell."""+    up = gene_idx(genes, up_names)+    dn = gene_idx(genes, dn_names)+    if not up:+        return None+    Xu = X[:, up].toarray() if sp.issparse(X) else np.asarray(X)[:, up]+    s = Xu.mean(axis=1)+    if dn:+        Xd = X[:, dn].toarray() if sp.issparse(X) else np.asarray(X)[:, dn]+        s = s - Xd.mean(axis=1)+    return s.astype(np.float64)+++def main():+    ap = argparse.ArgumentParser()+    ap.add_argument("--data", required=True)+    ap.add_argument("--out", required=True)+    ap.add_argument("--seed", type=int, required=True)+    ap.add_argument("--off", action="store_true", help="off-control: lambda=0")+    args = ap.parse_args()++    from src.task1_temporal import view_io++    view = args.data+    manifest = view_io.load_manifest(view)+    genes = view_io.panel_genes(view, manifest)+    rng = np.random.default_rng(args.seed)++    lam = float(os.environ.get("LAMBDA", DEFAULT_LAMBDA))+    if args.off:+        lam = 0.0+    log(f"lambda={lam}")++    inputs = view_io.inputs_by_time(manifest)+    last = inputs[-1]+    adata = view_io.read_stage(view, last, genes)+    X = adata.X.tocsr().astype(np.float32)+    n_cells, n_genes = X.shape+    labels = view_io.labels_of(adata)+    log(f"last stage {last.get('stage')}: {n_cells} cells x {n_genes} genes")++    # ---- mechanism: local commitment shift ----+    score = marker_score(X, genes, COMMITMENT_MARKERS, PROGENITOR_MARKERS)+    n_committed = 0+    n_branching = 0+    Xout = X+    if score is not None and lam > 0:+        card = is_cardiac_label(labels)+        pool = np.flatnonzero(card)+        if len(pool) >= MIN_COMMITTED:+            thr = np.quantile(score[pool], 1.0 - COMMITTED_FRAC)+            committed = np.zeros(n_cells, dtype=bool)+            sel = pool[score[pool] >= thr]+            # require some commitment-marker signal to avoid noise-only picks+            up_idx = gene_idx(genes, COMMITMENT_MARKERS)+            if up_idx:+                m = np.asarray(X[:, up_idx].sum(axis=1)).ravel() > 0+                sel = sel[m[sel]]+            committed[sel] = True+            n_committed = int(committed.sum())+            log(f"committed cells: {n_committed} / pool {len(pool)}")++        if n_committed >= MIN_COMMITTED:+            # HVG + PCA + kNN+            Xd = X+            mu = np.asarray(Xd.mean(axis=0)).ravel()+            sq = np.asarray(Xd.multiply(Xd).mean(axis=0)).ravel()+            var = np.maximum(sq - mu * mu, 0)+            hvg = np.argsort(-var)[:N_HVG]+            hvg.sort()+            from sklearn.decomposition import TruncatedSVD+            from sklearn.neighbors import NearestNeighbors+            Z = Xd[:, hvg].toarray().astype(np.float64)+            Z -= Z.mean(axis=0)+            svd = TruncatedSVD(n_components=min(N_PCS, Z.shape[1] - 1, Z.shape[0] - 1),+                               random_state=args.seed)+            P = svd.fit_transform(Z)+            nn = NearestNeighbors(n_neighbors=min(K_NEIGHBORS + 1, n_cells))+            nn.fit(P)+            dist, idx = nn.kneighbors(P)+            idx = idx[:, 1:]  # drop self+            k = idx.shape[1]++            committed_nbrs = committed[idx]+            readiness = committed_nbrs.mean(axis=1)+            branching = (readiness > READINESS_THRESHOLD) & (~committed)+            n_branching = int(branching.sum())+            log(f"branching cells: {n_branching} ({100.0*n_branching/n_cells:.1f}%)")+            log(f"readiness quantiles: {np.quantile(readiness,[0.1,0.25,0.5,0.75,0.9]).round(3)}")++            if n_branching > 0:+                # per-cell target = mean of its committed neighbors (chunked)+                rows = np.flatnonzero(branching)+                CHUNK = 256+                tgts = np.zeros((len(rows), n_genes), dtype=np.float32)+                for s0 in range(0, len(rows), CHUNK):+                    rr = rows[s0:s0 + CHUNK]+                    nbr = idx[rr]+                    cb = committed_nbrs[rr]+                    for j in range(len(rr)):+                        cn = nbr[j][cb[j]]+                        if len(cn) == 0:+                            continue+                        tgts[s0 + j] = np.asarray(X[cn].mean(axis=0)).ravel()+                blk = X[rows].toarray().astype(np.float64)+                tgt = tgts.astype(np.float64)+                delta = tgt - blk+                mask = (tgt > TARGET_HI) & (delta > DELTA_MIN)+                masked_delta = delta * mask++                # mean local shift direction (lam-invariant; cosine is scale-free)+                mean_shift = (readiness[rows][:, None] * masked_delta).mean(axis=0)++                # ---- two-input direction validation (before applying) ----+                lam_eff = lam+                if len(inputs) >= 2:+                    prev = view_io.read_stage(view, inputs[-2], genes)+                    lab0 = view_io.labels_of(prev)+                    common = sorted(set(lab0.tolist()) & set(labels.tolist()))+                    X0 = prev.X.tocsr().astype(np.float32)+                    d = np.zeros(n_genes, dtype=np.float64)+                    wsum = 0.0+                    for c in common:+                        m0 = np.flatnonzero(lab0 == c)+                        m1 = np.flatnonzero(labels == c)+                        if len(m0) < 5 or len(m1) < 5:+                            continue+                        pb0 = np.asarray(X0[m0].mean(axis=0)).ravel()+                        pb1 = np.asarray(X[m1].mean(axis=0)).ravel()+                        w = min(len(m0), len(m1))+                        d += w * (pb1 - pb0)+                        wsum += w+                    if wsum > 0 and np.linalg.norm(d) > 0 and np.linalg.norm(mean_shift) > 0:+                        d /= wsum+                        cos = float(np.dot(d, mean_shift) /+                                    (np.linalg.norm(d) * np.linalg.norm(mean_shift)))+                        log(f"two-input direction cosine={cos:.3f}")+                        if cos < 0.1:+                            log("cosine < 0.1: local edge direction not supported by observed temporal delta -> skip shift")+                            lam_eff = 0.0+                        elif cos < 0.3:+                            log("0.1 <= cosine < 0.3: halving lambda")+                            lam_eff = lam / 2.0++                if lam_eff > 0:+                    step = lam_eff * readiness[rows][:, None] * masked_delta+                    new = np.maximum(blk + step, 0.0)+                    sn = np.linalg.norm(step, axis=1)+                    log(f"shift norms: mean={sn.mean():.3f} q10={np.quantile(sn,0.1):.3f} "+                        f"q50={np.quantile(sn,0.5):.3f} q90={np.quantile(sn,0.9):.3f}")+                    if np.linalg.norm(mean_shift) > 0:+                        log(f"genes moved (>0.01 mean): {int((mean_shift>0.01).sum())}, "+                            f"top: {[genes[i] for i in np.argsort(-mean_shift)[:8]]}")+                    Xl = Xout.tolil()+                    for j, i in enumerate(rows):+                        nz = np.flatnonzero(new[j] > 0)+                        Xl.rows[i] = nz.tolist()+                        Xl.data[i] = new[j][nz].astype(np.float32).tolist()+                    Xout = Xl.tocsr()+                    n_modified = len(rows)+                else:+                    n_modified = 0+                log(f"modified cells: {n_modified}")+    else:+        log("mechanism not applied (no markers or lambda=0)")++    # ---- subsample to allowed range ----+    n_out = view_io.target_n_cells(manifest, Xout.shape[0])+    if n_out < Xout.shape[0]:+        take = rng.choice(Xout.shape[0], size=n_out, replace=False)+        take.sort()+        Xout = Xout[take]+    log(f"output cells: {Xout.shape[0]}")++    if sp.issparse(Xout):+        Xout = Xout.tocsr()+    else:+        Xout = sp.csr_matrix(Xout)+    view_io.write_prediction(Xout, genes, args.out, seed=args.seed)+    log(f"wrote {args.out}")+++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

完成。draft 节点 15：实现 PLAN 的 `lineage_branching` 最小版——Isl1+→Nkx2-5+ 承诺边上逐细胞向 committed kNN 邻居均值做稀疏位移（λ=0.5·readiness，mask target>0.5∧delta>0.25），双输入时按观测时间差做方向校验（cos<0.1 丢弃、<0.3 减半），off 对照用 `--off`/`LAMBDA=0`。

查分（seed 0，A 半，用 9/20 次）：proxy10 on=53.10 vs off=52.97（+0.13，噪声内；机制生效：582 细胞被改，top 位移基因 Myh6/Myl3/Ankrd1）；X3 方向校验 cos=−0.07 触发丢弃门 → on=off=47.62（无门版本 X3 掉到 45.7–46.3，variogram 受损，门是必要的）。两视图 vec-check 通过、确定性验证通过。

下一步最值得试：(1) 位移幅度/覆盖太小——proxy 只有 3.5% 细胞被改，可对 committed 邻居均值做"保留残差"式插值并扩大 branching 比例；(2) X3 上门被关说明局部边方向与整体时间差不一致，值得改用型内（per-label）方向校验而不是全局加权差；(3) 与节点 7/13 的组成趋势重加权组合（本 draft 未做组成变化）。

调研员

iogram score after first run. 3) The edge may not be supported in X3 if cell types are named differently; the support check handles this but means mechanism doesn't run. 4) Readiness threshold 0.25 may be too strict/loose; Engineer should report the distribution of readiness scores and try 0.15/0.35 if too few/many cells branch. Early detection: after first vec-score, if branching cell count is 0 or <5% of input, the mechanism isn't running.",
  "family_id": "lineage_branching",
  "mechanism": "Per-cell branching along a reviewed lineage edge (Isl1+ progenitor → Nkx2-5+ committed): each cell's shift direction is toward its own committed kNN neighbors, magnitude scales with local readiness score, preserving single-cell residuals for non-branching cells.",
  "vs_constant_shift": "Direction and magnitude vary per cell: each cell moves toward its own local committed neighbors (not a global or per-type mean delta); magnitude is scaled by the cell-specific readiness score (fraction of committed neighbors); only cells with local evidence are modified. A constant per-type shift would apply the same vector to all cells of a type regardless of local context.",
  "mechanism_evidence": "Engineer should report: (1) number of committed cells detected and branching cells modified (if 0, mechanism not running); (2) distribution of readiness scores; (3) per-cell shift norm distribution (should be heterogeneous, not a single value); (4) cosine similarity between individual cell shifts and the type-level mean delta (should be <0.95 if truly per-cell varying); (5) four-group score breakdown comparing on vs off.",
  "mechanism_off_control": "Same program with λ=0 (or equivalently, skip the shift step): all cells output unchanged from input. Expected difference: off-control produces copy_last-equivalent output; on-run modifies a subset of cells. If on and off outputs are identical (branching count = 0), mechanism is not running and must be debugged before scoring.",
  "sources": []
}
```