总览 · ← 返回运行 20261001-233756-search-t1-abc-r0-B-population

节点 n21

copy_last 两级增殖重加权（β_type=-4、β_cell=-1）+ 增殖-表达斜率放大（γ=-1.6、|slope|>0.1）；PLAN 的 Reactome 分化第二表达轴实测两符号均降分，按预定判定关闭（γ_diff=0），t 统计量过滤/Reactome 细胞周期得分/γ、阈值再扫均未超对照。

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 31d48d6cfb: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 460e9acb2e。改动的文件：solution/METHOD.md +33 −26、solution/run.py +153 −51

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex 88d77db..64d854b 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,38 +1,45 @@-# copy_last 两级增殖重加权（β_type=-4、β_cell=-1）+ 增殖-表达回归斜率的逐细胞表达放大（γ=-0.8），外部/退化视图跳过表达调整。+# copy_last 两级增殖重加权（β_type=-4、β_cell=-1）+ 增殖-表达斜率放大（γ=-1.6、|slope|>0.1）；PLAN 的 Reactome 分化第二表达轴实测两符号均降分，按预定判定关闭（γ_diff=0），t 统计量过滤/Reactome 细胞周期得分/γ、阈值再扫均未超对照。 -## 方法+## 方法（最终提交 = 对照配置） -基底（继承节点 4/6，已验证）：输出 = 最新「官方」输入阶段的加权无放回抽样（Efraimidis–Spirakis）。权重两级相乘：-- 类型级 w_type = max(1e-6, 1 + β_type·(prolif_type − mean_prolif))，β_type = **-4**（节点 4 扫出）；-- 细胞级 w_cell = max(1e-6, 1 + β_cell·(prolif_cell − prolif_type))，β_cell = **-1**（节点 6 方向，本节点以乘法形式叠加）。+1. 基底（节点 4/6/18，已验证）：输出 = 最新「官方」输入阶段的加权无放回抽样（Efraimidis–Spirakis）。+   - w_type = max(1e-6, 1 + β_type·(prolif_type − mean))，β_type = -4；+   - w_cell = max(1e-6, 1 + β_cell·(prolif_cell − prolif_type))，β_cell = -1。+   - prolif 得分 = 10 个通用 cell-cycle 标记（Mki67, Top2a, Pcna, Cdk1, Ccna2, Ccnb1, Aurkb, Bub1, Cenpf, Nusc1；非阶段特异知识）log 表达均值，快照内计算，单输入阶段可用。+2. 表达调整（节点 7/18）：按类型（≥20 细胞）做基因表达对 prolif 得分的 OLS，保留 |slope|>0.1，逐细胞 x += γ·slope·(prolif_i − type_mean)（γ=-1.6，即放大类型内增殖-表达耦合），clip≥0，1024 行分块。+3. 保护：外部来源阶段或退化抽样（n_target≥池大小，如 X3）跳过表达调整，直接复制。+4. **PLAN 的第二轴（本节点新增、实测后关闭）**：分化得分 = prior/Reactome「Developmental Biology」（R-MMU-1266738，517 基因，去 PROLIF，∩面板）log 表达均值；同构 OLS 斜率 + γ_diff 调整，与增殖轴独立叠加。γ_diff 默认 0。代码保留完整实现与 `--gamma-diff` 开关。+5. 另实现但默认关闭：`--t-min`（斜率 t 统计量过滤）、`--prolif-source reactome`（Reactome Cell Cycle R-MMU-1640170 得分）。 -prolif 得分 = PROLIF 基因（Mki67, Top2a, Pcna, Cdk1, Ccna2, Ccnb1, Aurkb, Bub1, Cenpf, Nusc1，通用 cell-cycle 标记，非阶段特异知识；与面板取交集，≥3 个才启用）log 表达的逐细胞均值，完全在输入快照内计算。+## 知识来源 -新增表达调整（本节点，PLAN 机制但**符号与预期相反**）：-1. 对每个细胞类型（细胞数 ≥20），用该类型**全部**源细胞做 OLS：slope_g = cov(x_g, prolif)/var(prolif)（稀疏矩阵-向量积实现，无需稠密化全矩阵）；只保留 |slope_g| > 0.05 的基因。-2. 对每个被选中细胞 i：x_adj[i,g] = x[i,g] + γ·slope_g·(type_mean_prolif − prolif_i)，clip ≥0，按 1024 行分块稠密计算。-3. γ 扫描（proxy A 半）：0 → 56.28；+0.05/+0.15/+0.3 → 56.24/55.89/55.31（PLAN 预设的正 γ **单调降分，证伪**）；**-0.05/-0.1/-0.2/-0.4/-0.8/-1.6 → 56.40/56.58/56.68/56.83/57.20/57.37**。负 γ = 沿斜率**放大**细胞自身的增殖偏离（增强类型内增殖梯度对比），选 γ = **-0.8**（-1.6 总分仅 +0.17 且 covariation 从 53.6 掉到 51.6，判为 A 半过拟合风险，弃用）。-4. 跳过条件：所选阶段为外部来源，或抽样退化（n_target ≥ 池大小，如 X3）。X3 上未加保护时 γ=-0.8 使分数 50.0 → 43.2（covariation 28.8），加保护后恢复 50.0（已查分验证）。+- PROLIF 基因：通用 cell-cycle 标记知识（非禁窗测量）。+- Reactome「Developmental Biology」「Cell Cycle」基因集：view 内 `prior/reactome/gene_sets.gmt`（通用通路注释，非阶段特异）。 -## 查分记录（A 半）+## 查分记录（proxy A 半，seed 0，除注明外） -- proxy seed0 γ=-0.8：**57.20**（de_recovery 52.48 / direction 59.72 / cell_state 61.45 / covariation 53.56）-- proxy seed1 γ=-0.8：56.89（一致）-- proxy2 seed0：57.20（外部 Qiu E9.0 被忽略，输出与 proxy 逐字节相同，md5 已核）-- X3 seed0：50.0（保护生效，= copy_last）-- 基底 γ=0（β_cell=-1 乘法叠加）：56.28，已高于节点 4（55.23）与节点 6（55.81）的 A 半查分-- β_type 重扫（γ=0）：-3 → 55.62，-5 → 55.34，均低于 -4（56.28），维持 -4-- 预期节点分 ≈ (57.2+57.2+50)/3 ≈ 54.8 A 半（父 53.48）+| 配置 | 板分 | de_rec | dir | cell_st | cov |+|---|---|---|---|---|---|+| 对照（γ=-1.6, sm=0.1, γ_diff=0） | **57.59** | 52.48 | 59.69 | 63.15 | 53.02 |+| γ_diff=-0.4 | 57.21 | 52.48 | 60.08 | 64.84 | 48.11 |+| γ_diff=+0.4 | 55.29 | 52.48 | 60.07 | 60.52 | 45.00 |+| prolif=Reactome Cell Cycle（γ=-1.6） | 48.28 | 51.46 | 50.04 | 51.30 | 37.55 |+| t_min=2 | 57.61 | 52.48 | 59.68 | 63.15 | 53.11 |+| γ=-2.4 | 57.31 | 51.96 | 59.52 | 63.47 | 52.02 |+| slope_min=0.15 | 57.60 | 52.48 | 59.67 | 62.87 | 53.50 |+| 对照 seed1 / seed2 | 57.31 / 57.37 | | | | |++两轴斜率余弦相似度 0.40–0.53（<0.8，PLAN 风险 2 未触发），但 γ_diff 两符号均降分（负号主要砸 covariation，正号砸 cell_state+cov），触发 PLAN 中止规则（|-0.38|<0.5 且方向为负）→ γ_diff=0。  ## 验证过 -- 三视图（proxy/proxy2/X3）跑通、vec-check ok、运行 ~5 s、内存远低于 28 GB 限额。-- 确定性：仅 np.random.default_rng(seed)；proxy 与 proxy2 输出 md5 相同。-- γ=0 时与父逻辑等价（β_cell 乘法叠加版）。+- proxy/proxy2/X3 三视图跑通、vec-check 全部 ok；proxy2 输出与 proxy 逐字节相同（md5，外部 Qiu E9.0 被忽略）；X3 走保护路径纯复制。+- 默认参数输出与对照 md5 相同；γ_diff=0 时代码路径与节点 18 机制等价（对照 3 seeds A 半 57.59/57.31/57.37，均值 57.42）。+- 运行 ~4.5 s/视图，内存远低于 28 GB 限额；确定性（仅 np.random.default_rng(seed)）。  ## 没验证 / 风险 -- γ=-0.8 与 -0.4/-1.6 的差距（<0.6 分）在 T1 噪声（~2 分）内；-0.8 是折中（de_recovery +1、covariation -1.3 相对基底）。-- 负 γ 的机制解释：放大类型内增殖-表达耦合让群体在增殖轴上更分散，mmd_u 与 de_score 同升；B 半是否重现未验证。-- final 视图（E8.5+E9.5→E10.5）未测：若 E9.5 池 ≤ 目标细胞数，表达调整会被退化保护跳过，只剩 copy_last（安全但无增益）。-- 「退化即跳过」保护对 X3 有效，但若某视图池恰等于目标数而调整本可有益，也会被跳过。+- 最终配置与节点 18 机制相同，预期节点分 ≈ 54.6–54.7（B 半），非超越；本节点的贡献是证伪了 PLAN 的分化第二轴与三个替代方向，缩小后续搜索空间。+- γ=-1.6 vs -2.4 vs sm=0.15 的差距均 <0.3，在噪声内；未做更多种子区分。+- final 视图（E8.5+E9.5→E10.5）未测：若 E9.5 池 ≤ 目标细胞数则退化保护跳过表达调整，退回 copy_last（安全无增益）。+- de_recovery 对全部表达级扰动不敏感（52.48 恒定，仅 γ=-2.4 时降到 51.96），后续提升 de_recovery 需换机制（如组成/抽样层面）。diff --git a/solution/run.py b/solution/run.pyindex 71f82b5..19a133b 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,26 +1,32 @@ #!/usr/bin/env python3-"""copy_last + two-level proliferation reweighting + proliferation-gradient-expression extrapolation.+"""copy_last + two-level proliferation reweighting + two-axis expression+adjustment (proliferation gradient + Reactome developmental gradient). -Base (node 4/6): output a deterministic weighted subsample of the latest+Base (nodes 4/6/18): output a deterministic weighted subsample of the latest official input stage. Type-level weight uses the type-mean proliferation-score (beta_type=-4, tuned); cell-level weight uses the within-type-deviation (beta_cell=-1, node 6).--New (this node): per cell type, regress each gene's expression on the cell's-proliferation score (OLS slope over ALL source cells of that type). Each-selected cell's expression is nudged along that slope toward its type's mean-proliferation level: x_adj[i,g] = x[i,g] + gamma * slope_g * (tmean - p_i),-applied only where |slope_g| > SLOPE_MIN, clipped to >= 0. Types with-< MIN_TYPE_CELLS cells are left untouched. gamma=0 reproduces the base.-Proliferation markers are generic cell-cycle genes (not stage-specific-knowledge); everything is computed within the input snapshot, so the method-works with a single input stage.+score (beta_type=-4); cell-level weight uses the within-type deviation+(beta_cell=-1).++Expression adjustment axis 1 (nodes 7/18): per cell type, regress each gene's+expression on the cell's proliferation score (OLS over ALL source cells of+that type); nudge each selected cell along the slope, amplified by negative+gamma (gamma=-1.6, |slope|>0.1, node 18).++Expression adjustment axis 2 (this node): same construction but the per-cell+latent score is a DIFFERENTIATION score = mean log-expression of Reactome+"Developmental Biology" genes (from prior/, minus proliferation markers),+computed entirely within the input snapshot. gamma_diff tuned on proxy A-half;+gamma_diff=0 reproduces the node-18 base.++Both axes are applied additively (prolif first, then diff), each clipped to+>= 0, only on official non-degenerate views (external platforms like X3 and+views where the pool is not subsampled are copied untouched). """  from __future__ import annotations  import argparse+import os  import numpy as np from scipy import sparse@@ -37,15 +43,21 @@ from src.task1_temporal.view_io import (     write_prediction, ) -BETA_TYPE = -4.0   # tuned on T1 proxy A-half (node 4)-BETA_CELL = -1.0   # within-type weight (node 6)-GAMMA = -0.8       # expression-adjustment strength (negative tuned on proxy A-half)-SLOPE_MIN = 0.05   # only adjust genes with |slope| above this+BETA_TYPE = -4.0    # tuned on T1 proxy A-half (node 4)+BETA_CELL = -1.0    # within-type weight (node 6)+GAMMA = -1.6        # proliferation-axis strength (node 18)+SLOPE_MIN = 0.1     # only adjust genes with |slope| above this (node 18)+GAMMA_DIFF = 0.0    # differentiation axis: tested, harmful both signs -> off MIN_TYPE_CELLS = 20+MIN_DIFF_GENES = 10  PROLIF = ["Mki67", "Top2a", "Pcna", "Cdk1", "Ccna2", "Ccnb1",           "Aurkb", "Bub1", "Cenpf", "Nusc1"] +REACTOME_PATHWAY = "Developmental Biology"  # Reactome top-level, general+CELLCYCLE_PATHWAY = "Cell Cycle"            # Reactome, for prolif score v2+PROLIF_SOURCE = "manual"                    # "manual" | "reactome"+  def pick_stage(manifest):     """Latest official input; fall back to the latest stage of any source."""@@ -64,18 +76,90 @@ def weighted_sample(w: np.ndarray, n: int, rng: np.random.Generator) -> np.ndarr     return np.sort(idx)  +def load_pathway_genes(view_dir: str, pathway: str) -> set[str]:+    """Genes of a named Reactome pathway from prior/.++    Generic pathway-annotation knowledge (not stage-specific measurements).+    Returns an empty set if the prior is unavailable.+    """+    gmt = os.path.join(view_dir, "prior", "reactome", "gene_sets.gmt")+    if not os.path.exists(gmt):+        return set()+    try:+        with open(gmt) as fh:+            for line in fh:+                parts = line.rstrip("\n").split("\t")+                if len(parts) >= 3 and parts[1] == pathway:+                    return set(parts[2:])+    except OSError:+        return set()+    return set()+++def type_ols_slopes(X, idx_by_type, score, n_types, n_genes, slope_min,+                    t_min=0.0):+    """Per type OLS slope of each gene on `score`, thresholded by |slope|+    and (optionally) by the regression |t-statistic|."""+    S = np.zeros((n_types, n_genes), dtype=np.float32)+    tmean = np.zeros(n_types, dtype=np.float64)+    for t, idx in enumerate(idx_by_type):+        n = len(idx)+        if n < MIN_TYPE_CELLS:+            continue+        c = score[idx]+        tmean[t] = c.mean()+        pc = c - tmean[t]+        denom = float(pc @ pc)+        if denom <= 1e-8:+            continue+        Xt = X[idx]+        slope = np.asarray(Xt.T @ pc, dtype=np.float64).ravel() / denom+        keep = np.abs(slope) > slope_min+        if t_min > 0.0 and np.any(keep):+            sumsq = np.asarray(Xt.multiply(Xt).sum(axis=0),+                               dtype=np.float64).ravel()+            gmean = np.asarray(Xt.mean(axis=0), dtype=np.float64).ravel()+            rss = np.maximum(sumsq - n * gmean**2 - slope**2 * denom, 1e-12)+            tstat = np.abs(slope) * np.sqrt(denom * (n - 2) / rss)+            keep &= tstat > t_min+        slope[~keep] = 0.0+        S[t] = slope.astype(np.float32)+    return S, tmean+++def apply_axis(Xsel, S, trow, c, chunk=1024):+    """x += c_i * S[t,g], clipped to >= 0, blockwise; returns csr."""+    parts = []+    for a in range(0, Xsel.shape[0], chunk):+        b = min(a + chunk, Xsel.shape[0])+        D = np.asarray(Xsel[a:b].todense(), dtype=np.float32)+        D += (c[a:b, None] * S[trow[a:b]]).astype(np.float32)+        np.clip(D, 0.0, None, out=D)+        parts.append(sparse.csr_matrix(D))+    return sparse.vstack(parts).tocsr()++ 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("--gamma", type=float, default=None)+    parser.add_argument("--gamma-diff", type=float, default=None)     parser.add_argument("--beta-cell", type=float, default=None)     parser.add_argument("--beta-type", type=float, default=None)+    parser.add_argument("--slope-min", type=float, default=None)+    parser.add_argument("--t-min", type=float, default=0.0)+    parser.add_argument("--prolif-source", default=PROLIF_SOURCE,+                        choices=["manual", "reactome"])+    parser.add_argument("--debug-cos", action="store_true",+                        help="print cosine similarity of the two slope axes")     args = parser.parse_args()     gamma = GAMMA if args.gamma is None else args.gamma+    gamma_diff = GAMMA_DIFF if args.gamma_diff is None else args.gamma_diff     beta_cell = BETA_CELL if args.beta_cell is None else args.beta_cell     beta_type = BETA_TYPE if args.beta_type is None else args.beta_type+    slope_min = SLOPE_MIN if args.slope_min is None else args.slope_min      manifest = load_manifest(args.data)     genes = panel_genes(args.data, manifest)@@ -84,14 +168,31 @@ def main() -> None:     rng = np.random.default_rng(args.seed)     n_target = target_n_cells(manifest, last.n_obs) +    gene_set = set(genes)     score = None-    have = [g for g in PROLIF if g in set(genes)]+    if args.prolif_source == "reactome":+        cc = load_pathway_genes(args.data, CELLCYCLE_PATHWAY)+        have = [g for g in genes if g in cc]+    else:+        have = [g for g in PROLIF if g in gene_set]     if len(have) >= 3:         cols = np.array([genes.index(g) for g in have], dtype=np.int64)         Xc = last.X[:, cols]         Xc = Xc.toarray() if sparse.issparse(Xc) else np.asarray(Xc)         score = Xc.mean(axis=1).astype(np.float64) +    # Differentiation score: mean log-expression of Reactome Developmental+    # Biology genes (proliferation markers removed), snapshot-internal.+    diff_score = None+    if score is not None and gamma_diff != 0.0:+        dev = load_pathway_genes(args.data, REACTOME_PATHWAY)+        dev_cols = [i for i, g in enumerate(genes)+                    if g in dev and g not in set(PROLIF)]+        if len(dev_cols) >= MIN_DIFF_GENES:+            Xd = last.X[:, np.asarray(dev_cols, dtype=np.int64)]+            Xd = Xd.toarray() if sparse.issparse(Xd) else np.asarray(Xd)+            diff_score = Xd.mean(axis=1).astype(np.float64)+     inv = None     if score is not None:         lab = labels_of(last)@@ -105,44 +206,45 @@ def main() -> None:         ts = type_score[inv]         mean_s = np.average(type_score, weights=counts)         w_type = np.maximum(1.0 + beta_type * (ts - mean_s), 1e-6)-        w_cell = np.maximum(1.0 + beta_cell * (score - ts), 1e-6) if beta_cell != 0 else np.ones_like(w_type)+        w_cell = (np.maximum(1.0 + beta_cell * (score - ts), 1e-6)+                  if beta_cell != 0 else np.ones_like(w_type))         rows = weighted_sample(w_type * w_cell, n_target, rng)     else:         rows = sample_rows(last.n_obs, n_target, rng)      X = last.X     Xsel = X[rows]-    # Guard: only adjust expression on official stages. In external-only views-    # (e.g. test question X3: different platform/technology, author cell-type-    # labels) the within-type proliferation regression proved harmful-    # (X3 A-half 50.0 -> 43.2), so we copy the cells untouched there.-    degenerate = n_target >= last.n_obs  # no real subsampling (e.g. test X3)-    if gamma != 0.0 and score is not None and not is_external(entry) and not degenerate:+    # Guard: only adjust expression on official stages with real subsampling.+    # On external platforms (test X3) the regression proved harmful+    # (50.0 -> 43.2), so cells are copied untouched there.+    degenerate = n_target >= last.n_obs+    adjust = (score is not None and not is_external(entry) and not degenerate)+    if adjust and (gamma != 0.0 or gamma_diff != 0.0):         n_types = int(inv.max()) + 1-        S = np.zeros((n_types, X.shape[1]), dtype=np.float32)-        for t in range(n_types):-            idx = np.flatnonzero(inv == t)-            if len(idx) < MIN_TYPE_CELLS:-                continue-            pc = score[idx] - score[idx].mean()-            denom = float(pc @ pc)-            if denom <= 1e-8:-                continue-            slope = np.asarray(X[idx].T @ pc, dtype=np.float64).ravel() / denom-            slope[np.abs(slope) <= SLOPE_MIN] = 0.0-            S[t] = slope.astype(np.float32)-        if np.any(S):-            c = (-gamma) * (score[rows] - type_score[inv[rows]])-            trow = inv[rows]-            parts = []-            chunk = 1024-            for a in range(0, len(rows), chunk):-                b = min(a + chunk, len(rows))-                D = np.asarray(Xsel[a:b].todense(), dtype=np.float32)-                D += (c[a:b, None] * S[trow[a:b]]).astype(np.float32)-                np.clip(D, 0.0, None, out=D)-                parts.append(sparse.csr_matrix(D))-            Xsel = sparse.vstack(parts).tocsr()+        idx_by_type = [np.flatnonzero(inv == t) for t in range(n_types)]+        trow = inv[rows]+        if gamma != 0.0:+            S, tmean = type_ols_slopes(X, idx_by_type, score, n_types,+                                       X.shape[1], slope_min, args.t_min)+            if args.debug_cos and gamma_diff != 0.0 and diff_score is not None:+                Sd, _ = type_ols_slopes(X, idx_by_type, diff_score, n_types,+                                        X.shape[1], slope_min)+                big = [(t, S[t], Sd[t]) for t in range(n_types)+                       if np.any(S[t]) and np.any(Sd[t])]+                for t, sp, sd in big[:5]:+                    cos = float(sp @ sd / (np.linalg.norm(sp)+                                           * np.linalg.norm(sd)))+                    print(f"[debug] type {t} cos(prolif_slope, diff_slope)="+                          f"{cos:.3f}")+            if np.any(S):+                c = (-gamma) * (score[rows] - tmean[inv[rows]])+                Xsel = apply_axis(Xsel, S, trow, c)+        if gamma_diff != 0.0 and diff_score is not None:+            Sd, dmean = type_ols_slopes(X, idx_by_type, diff_score, n_types,+                                        X.shape[1], slope_min, args.t_min)+            if np.any(Sd):+                cd = (-gamma_diff) * (diff_score[rows] - dmean[inv[rows]])+                Xsel = apply_axis(Xsel, Sd, trow, cd)      write_prediction(Xsel, genes, args.out, seed=args.seed) 

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

Final state verified. Summary:

**做了什么**：按 PLAN 在节点 7/18 基底（β_type=-4、β_cell=-1、γ=-1.6、SLOPE_MIN=0.1）上实现了 Reactome「Developmental Biology」分化得分的第二表达回归轴（γ_diff 可调、与增殖轴独立叠加、同样的外部/退化保护），并额外实现了斜率 t 统计量过滤与 Reactome Cell Cycle 增殖得分两个开关。

**查分结果**（proxy A 半，用 8/20 次）：γ_diff=-0.4 → 57.21（covariation 崩 4.9 分）、+0.4 → 55.29，两符号均低于对照 57.59，触发 PLAN 中止规则，γ_diff=0；Reactome Cell Cycle 得分灾难性失败（48.28，手工 10 基因列表保留）；t_min=2（57.61）、γ=-2.4（57.31）、slope_min=0.15（57.60）均为噪声内无效。最终提交 = 对照配置，3 seeds 57.59/57.31/57.37；proxy/proxy2/X3 三视图 vec-check 全部 ok，proxy2 与 proxy 输出 md5 相同，X3 保护路径生效。预期节点分 ≈ 节点 18 水平（~54.6-54.7），非超越。

**下一步最值得试**：de_recovery 对所有表达级扰动完全不敏感（52.48 恒定），提升它只能在组成/抽样层面动手（如类型权重的非线性形式、按类型分化速率差异化 β_type）；covariation 与 cell_state 在 γ 轴上存在明确 trade-off，二维联合扫描（β_cell × γ）比单轴细扫更可能找到超噪声的点。

调研员

{
  "name": "双轴表达调整：增殖斜率 + Reactome分化轴回归斜率（γ_diff扫描）",
  "motivation": "父节点7最弱组为de_recovery（51.69），Analyst指出其对抽样和表达扰动'完全不敏感'（节点18确认51.69→51.69）。当前表达调整仅沿增殖-表达回归斜率（单一潜轴），而E8.5→E9.5的DE基因大量来自分化程序而非增殖程序。节点19曾将Reactome Developmental Biology基因集用作抽样权重轴（β_diff），实测与基线同分——但从未将其用作表达回归调整的潜轴。机制假设：分化相关基因在类型内与分化得分的回归斜率携带增殖轴无法捕获的DE方向信息，沿该斜率放大可提升de_recovery而不伤及已被增殖轴覆盖的cell_state。",
  "approach": "在节点7/18基底（β_type=-4、β_cell=-1、γ_prolif=-1.6、SLOPE_MIN=0.1）上新增第二表达调整轴：\n1. 从prior/加载Reactome 'Developmental Biology'基因列表，去除10个PROLIF基因，与面板取交集（需≥10个基因才启用，否则跳过第二轴）。\n2. 计算每细胞分化得分diff_score = 交集基因log表达均值（与prolif_score计算方式完全对称）。\n3. 对每个细胞类型（≥20细胞），用全部源细胞做OLS：slope_diff_g = cov(x_g, diff_score)/var(diff_score)；只保留|slope_diff_g|>SLOPE_MIN（0.1，与增殖轴一致）的基因。\n4. 表达调整叠加：x_adj[i,g] += γ_diff · slope_diff_g · (type_mean_diff − diff_i)，clip≥0。与增殖轴调整独立相加（先做增殖轴，再做分化轴，各自clip）。\n5. γ_diff扫描（proxy A半，3 seeds）：{0, -0.2, -0.4, -0.8}。判定：γ_diff最优值需3种子均值比γ_diff=0高≥1分（proxy A半），否则保留γ_diff=0（输出等价父节点）。\n6. 保护：外部来源/退化视图（X3）跳过两轴调整（现有逻辑不变）。\n7. 单输入退路：两轴得分均在单快照内计算，无需第二时间点，proxy/final逻辑相同。\n8. vec-score快速筛选：先只跑γ_diff∈{0,-0.4}两个点×1 seed确认方向，再全扫。\n9. 若γ_diff最优≠0，额外检查covariation是否下降>1.5分；若是则减半γ_diff或提高SLOPE_MIN到0.15。",
  "expected_groups": ["de_recovery", "covariation"],
  "risks": "1) Reactome Developmental Biology基因集过大/过杂，回归斜率信噪比低，γ_diff扫描全部在噪声内→Engineer应在第一轮1-seed快扫后若γ_diff=-0.4与0差距<0.5分即中止，节省查分次数。2) 分化轴与增殖轴高度相关（分化慢的细胞增殖高），两轴调整近似重复→检查两轴斜率向量余弦相似度，若>0.8则第二轴退化为噪声，应放弃。3) covariation因双重调整而下降（节点7教训：γ=-1.6时cov掉2分）→用SLOPE_MIN=0.1保护，且γ_diff上限-0.8。4) prior/中Reactome基因列表格式/命名与面板不匹配→Engineer先打印交集数量，<10个则跳过第二轴并报告。",
  "sources": []
}

审查员

REVIEW.json written