总览 · ← 返回运行 20261001-024931-search-t1-spark-eng-default

节点 n4

父节点重加权锚（4000 真实细胞）保持不变，新增可按 β 开关的族内单快照成熟方向平移；β=0（默认）逐位复现父节点。

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 4da4a5ed26:solution/METHOD.md

python solution/run.py --data <view> --out pred.h5ad --seed 0                 # β=0，= 父节点，1.5 s
python solution/run.py --data <view> --out pred.h5ad --seed 0 --beta 0.15 --verbose   # 28 s，打印每族 |corr| 与 top 基因

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 2be2358fd9。改动的文件：solution/METHOD.md +56 −0、solution/dpt.py +269 −0、solution/run.py +137 −11

diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..77f2eac--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,56 @@+父节点重加权锚（4000 真实细胞）保持不变，新增可按 β 开关的族内单快照成熟方向平移；β=0（默认）逐位复现父节点。++# 方法++## 组成（父节点已验证部件，未动）+- 复制 `reweight.py` 的规则：丢 `Neural Tube`，心脏侧类型 ×1.6，`Surface Ectoderm / EXEM / Paraxial Mesoderm` ×0.25，其余 ×1，`largest_remainder` 分配 4000 个细胞。+- `run.py:resample_with_labels` 用与 `heart_reweight` 完全相同的类型遍历顺序和 `rng.choice` 调用序列，因此 β=0 时输出与父节点**逐位相同**（已用 `np.array_equal` 核对：True）。+- 抽样保真检查（PLAN 步骤 1）：E8.5 每个类型的分配数都小于该类型细胞数（最大 Foregut 518/2405、pSHF 756/2192），`replace=pool.size<n` 全为 False，**重复行 = 0**，因此不需要换成加权不放回抽样，省下一次查分。++## 表达（新增，默认关闭）+`solution/dpt.py`：对最新输入阶段的每个细胞类型（≥200 个细胞才做）+1. 族内 log1p(CP10k) 面板，子采样 ≤3000 细胞；只保留族内平均表达 ≥0.05 的基因；+2. 族内中心化并投掉「每细胞 log 总读数」方向（`_detotal`，去技术深度轴）；+3. `randomized_svd(n_components=5, random_state=0)` 取前 5 个 PC；+4. 增殖打分 p = 细胞周期基因集在族内的 z 分数均值。基因集来自视图 `prior/go/gene_sets_bp.gmt` 里 GO:0044770 / GO:0044772 / GO:0051301 / GO:0140014（面板命中 470 个基因）；prior 缺失或命中 <10 时退回内置通用细胞周期列表，不会崩；+5. 取 |corr(score_k, p)| 最大的 PC，要求 ≥0.20，否则该族 δ=0；方向指向增殖下降的一侧（时间前进）；+6. δ = mean(x | 顶三分位) − mean(x | 底三分位)（对数空间）；+7. 稳定性收缩：族内随机对半划分 4 次，每半重算 (5)(6)，逐基因统计两半符号一致次数 → w∈{0,.25,.5,.75,1}，δ̃=w·δ，|δ̃| 按族内 99 分位截断；+8. 生成：x'=clip(x+β·δ̃,0)，可选 `--renorm` 把每细胞重归一到输入总读数中位数（数据是 CP10k，线性总量恒为 1e4，所以 renorm on/off 几乎等价）。++两个及以上输入阶段（final）时，额外算同名族的伪批量差 `pseudobulk_delta`（同样的表达阈值 + 4 次细胞 bootstrap 符号收缩），合成 δ=β·δ̃_dpt+α·Δ̃_pb，α=0.35 固定，且有符号门：cos(δ_dpt,Δ_pb)<0 的族丢掉 DPT 项只保留 Δ_pb。族归属沿用父节点的精确名集合 + `type_weights`；该分支在 proxy 上不执行。++# 验证过什么++- **护栏**：`--beta 0` 输出与父节点逐位相同（已核对），`vec-check` ok，1.5 s / <1.3 GB。+- **方向合理性**（不查分，先看数）：8/18 个族 |corr|≥0.20 拿到 δ≠0（AVC-CM .42、Foregut .38、IFT-CM .41、LV-CM .46、OFT/RV-CM .51、RV-CM .41、SV-CM .42、aSHF .21）；其余（Blood/EXEM/Endothelium/JCF/NCC/Neural Tube/Paraxial/Pericardium/Surface Ectoderm/pSHF）|corr|<0.20 → δ=0，退回真实细胞。δ 的 top 基因与「增殖下降 + 心肌成熟」一致：IFT-CM 晚期侧升 Nkx2-5、Tnnt2、Tnni1、Myl4、Des、Actn2、Nexn、Crip2；OFT/RV-CM 晚期侧降 Mki67、Top2a、Cenpe、Cks1b、升 Cdkn1c。没有发现技术轴主导的迹象。+- **替代评测（seed 0，1 次/配置）**：++| 配置 | 榜分 | cell_state | covariation | de_recovery | direction |+|---|---|---|---|---|---|+| 父节点（β=0，复现） | 55.9694 | 56.90 | 54.97 | 53.06 | 58.56 |+| β=0.15 + renorm | 55.9437 | 57.41 | 53.63 | 53.06 | 58.92 |+| β=0.15 − renorm | 55.9335 | 57.43 | 53.57 | 53.06 | 58.90 |+| β=0.30 + renorm | 55.8012 | 57.84 | 52.19 | 53.06 | 58.98 |++  趋势单调且清楚：平移让 `mmd_u` 变小（cell_state +0.5/+0.9）、`de_direction` 微升（direction +0.36/+0.42），但 `variogram` 变大（covariation −1.3/−2.8），`de_score` 一位都没动（0.1111 三次完全相同）。净效果 ≤0，且 β 越大越差。+- 因此**默认 β=0**，把这套机制作为可开关的代码保留。按 PLAN 步骤 5 的接受判据（4-seed 均值 +1.5 且 ≥3/4 同向），β>0 在 seed 0 上就没有正收益，不值得再花 12 次查分做 4-seed 复核（时间预算也不允许）。++# 没验证 / 零结果++- **单快照伪时间平移在 T1 proxy 上对 de_recovery 完全无效**（de_score 三位小数不变），对 covariation 有害。原因推测：δ 是族内第三分位对比，量级 |δ̃| 中位远小于真实 E8.5→E9.5 的族间 DE，且它把细胞推离真实流形（variogram 变差正是共变结构被破坏的信号）。+- 没验证多 seed（时间不够）；没验证 β<0.15；没验证 final 的 Δ_pb 分支（proxy 只有一个输入阶段，该分支不执行，只做了「不崩」级别的代码检查）。+- 没验证只平移 δ≠0 的 8 个族之外的其它组合、也没试「只对增殖最强的族平移」。++# 下一步最值得试++1. covariation 是父节点第二弱的组，而且它对任何表达扰动都很敏感（β=0.15 就掉 1.34）。先做**不改变每个细胞流形位置**的表达改动：例如按族做「细胞间的重排/混合」（把同族细胞的基因表达按 rank 做轻度混合）而不是加常数向量。+2. de_recovery 对常数位移完全不响应，说明它比较的是族间 DE 结构；试试**族特异的少量高置信基因**（只在 δ 的 top-50 基因、且 w=1 的基因上平移，β 放大到 0.5–1），其余基因不动，避免全基因噪声推高 variogram。+3. 组成侧还有空间：父节点的心脏 ×1.6 是在 2 分噪声上选的，但 `cell_state` 56.9 明显受 mmd_u 驱动；用 E8.5 自身的 marker 结构（视图内数据）做一次「组织占比 → 期望细胞数」的重新标定，比手调 1.6/0.25 更可能迁移到 final。++# 运行++```bash+python solution/run.py --data <view> --out pred.h5ad --seed 0                 # β=0，= 父节点，1.5 s+python solution/run.py --data <view> --out pred.h5ad --seed 0 --beta 0.15 --verbose   # 28 s，打印每族 |corr| 与 top 基因+```diff --git a/solution/dpt.py b/solution/dpt.pynew file mode 100644index 0000000..b5db2bd--- /dev/null+++ b/solution/dpt.py@@ -0,0 +1,269 @@+"""Single-snapshot maturation direction (DPT-like) per cell group.++Everything is computed at run time from the input stage(s) given by the+manifest: no stage names, proportions or measured statistics are baked in.++For one cell group we look for the axis inside the group that best tracks a+decline in proliferation (canonical proxy for embryonic time passing within a+snapshot), take the high-minus-low expression contrast along that axis, and+shrink it gene-wise by a split-half sign-agreement weight so noise genes do+not dominate the shift.+"""++from __future__ import annotations++from pathlib import Path++import numpy as np+from scipy import sparse++GO_CC_TERMS = {+    "GO:0044770",  # cell cycle phase transition+    "GO:0044772",  # mitotic cell cycle phase transition+    "GO:0051301",  # cell division+    "GO:0140014",  # mitotic nuclear division+}+GO_CC_NAMES = {+    "cell cycle phase transition",+    "mitotic cell cycle phase transition",+    "cell division",+    "mitotic nuclear division",+}+# Fallback only used when the prior gene sets are unavailable/too sparse.+FALLBACK_CC = [+    "Mki67", "Top2a", "Cdk1", "Ccna2", "Ccnb1", "Ccnb2", "Ccne1", "Ccnd1",+    "Pcna", "Mcm2", "Mcm3", "Mcm4", "Mcm5", "Mcm6", "Mcm7", "Plk1", "Aurka",+    "Aurkb", "Birc5", "Bub1", "Bub1b", "Cdc20", "Cdc6", "Cdt1", "E2f1",+    "Foxm1", "Kif11", "Kif23", "Nusap1", "Rrm2", "Tyms", "Ube2c", "Cenpa",+    "Cenpe", "Cenpf", "Chek1", "Gins1", "Orc1", "Rad51", "Aspm",+]+MIN_EXP = 0.05+MIN_CELLS = 200+MAX_SUB = 3000+N_PC = 5+MIN_CORR = 0.20+N_SPLIT = 4+Q_LO, Q_HI = 1.0 / 3.0, 2.0 / 3.0+++def cell_cycle_genes(view, genes: list[str]) -> np.ndarray:+    """Boolean mask over the panel marking proliferation genes (prior first)."""+    mask = np.zeros(len(genes), dtype=bool)+    idx = {g: i for i, g in enumerate(genes)}+    path = Path(view) / "prior" / "go" / "gene_sets_bp.gmt"+    picked: set[str] = set()+    if path.exists():+        try:+            with path.open("r", encoding="utf-8") as fh:+                for ln in fh:+                    parts = ln.rstrip("\n").split("\t")+                    if len(parts) < 3:+                        continue+                    if parts[0].strip() in GO_CC_TERMS or parts[1].strip().lower() in GO_CC_NAMES:+                        picked.update(g for g in parts[2:] if g)+        except OSError:+            picked = set()+    if len(picked & set(idx)) < 10:+        picked = set(FALLBACK_CC)+    for g in picked:+        j = idx.get(g)+        if j is not None:+            mask[j] = True+    return mask+++def _detotal(Dc: np.ndarray) -> np.ndarray:+    """Remove the per-cell log-total-reads direction (technical depth axis)."""+    tot = Dc.sum(axis=1, dtype=np.float64)+    t = tot - tot.mean()+    nrm = np.linalg.norm(t)+    if nrm < 1e-8:+        return Dc+    u = (t / nrm).astype(np.float32)+    return Dc - np.outer(u, Dc.T @ u).astype(np.float32)+++def _dense(X, rows: np.ndarray) -> np.ndarray:+    if sparse.issparse(X):+        return np.asarray(X[rows].todense(), dtype=np.float32)+    return np.asarray(X)[rows].astype(np.float32)+++class _Axis:+    """Fitted centring + PC directions of one cell group."""++    __slots__ = ("keep", "mu", "u", "Vt", "D", "cc", "active", "prol", "sc", "k", "corr", "lo", "hi", "late_first")++    def __init__(self, D: np.ndarray, cc_mask: np.ndarray, n_pc: int = N_PC):+        self.D = D+        self.cc = cc_mask+        keep = D.std(axis=0) > 0+        self.keep = keep+        Dk = np.ascontiguousarray(D[:, keep], dtype=np.float32)+        self.mu = Dk.mean(axis=0)+        self.u = None+        Dc = _detotal(Dk - self.mu)+        npc = max(1, min(n_pc, Dc.shape[0] - 1, Dc.shape[1] - 1))+        try:+            from sklearn.utils.extmath import randomized_svd++            _, _, Vt = randomized_svd(Dc, n_components=npc, random_state=0)+            self.Vt = Vt.astype(np.float32)+        except Exception:+            v = np.linalg.eigh((Dc.T @ Dc).astype(np.float64))[1]+            self.Vt = np.ascontiguousarray(v[:, np.argsort(-(Dc.T @ Dc).diagonal())][:, :npc].T.astype(np.float32))+        self.active = np.ones(D.shape[1], dtype=bool)+        self.sc = self.project(None)+        self.prol = self._proliferation(None)+        self.k, self.corr, self.lo, self.hi, self.late_first = self._orient(self.sc, self.prol)++    def project(self, rows):+        D = self.D if rows is None else self.D[rows]+        Dk = np.ascontiguousarray(D[:, self.keep], dtype=np.float32) - self.mu+        return (_detotal(Dk) @ self.Vt.T).astype(np.float64)++    def _proliferation(self, rows: np.ndarray) -> np.ndarray:+        ccm = self.cc+        if ccm.sum() < 5:+            return np.zeros(len(rows), dtype=np.float64)+        Z = (self.D[:, ccm] if rows is None else self.D[np.ix_(rows, ccm)]).astype(np.float64)+        sd = Z.std(axis=0)+        sd[sd < 1e-8] = 1.0+        return ((Z - Z.mean(axis=0)) / sd).mean(axis=1)++    @staticmethod+    def _orient(sc: np.ndarray, prol: np.ndarray):+        if np.allclose(prol, prol[0]):+            return 0, 0.0, -np.inf, np.inf, False+        cs = np.zeros(sc.shape[1])+        for j in range(sc.shape[1]):+            s = sc[:, j]+            if s.std() < 1e-12:+                continue+            cs[j] = np.nan_to_num(np.corrcoef(s, prol)[0, 1])+        k = int(np.argmax(np.abs(cs)))+        s = sc[:, k]+        lo, hi = np.quantile(s, [Q_LO, Q_HI])+        # forward in time = proliferation goes down+        late_first = cs[k] < 0+        return k, float(abs(cs[k])), float(lo), float(hi), bool(late_first)++    def contrast(self, rows: np.ndarray):+        """delta over the *active* columns given by ``self.active``, or None."""+        if self.corr < MIN_CORR:+            return None+        n = self.D.shape[0] if rows is None else rows.size+        if n < 30:+            return None+        sc = self.project(rows)[:, self.k]+        prol = self._proliferation(rows)+        _, c, lo, hi, late_first = self._orient(sc[:, None], prol)+        if c < MIN_CORR:+            return None+        late = sc >= hi if late_first else sc <= lo+        early = sc <= lo if late_first else sc >= hi+        if late.sum() < 10 or early.sum() < 10:+            return None+        Da = self.D[:, self.active] if rows is None else self.D[np.ix_(rows, self.active)]+        return (Da[late].mean(axis=0) - Da[early].mean(axis=0)).astype(np.float64)+++def group_delta(+    X,+    labels: np.ndarray,+    genes: list[str],+    cc_mask: np.ndarray,+    seed: int = 0,+    min_cells: int = MIN_CELLS,+    n_split: int = N_SPLIT,+    verbose: bool = False,+) -> dict[str, np.ndarray]:+    """delta per cell group, on the panel index, shrunk by split-half agreement."""+    out: dict[str, np.ndarray] = {}+    rng = np.random.default_rng(seed + 104729)+    all_rows = np.arange(labels.size)+    for g in [str(x) for x in np.unique(labels)]:+        zero = np.zeros(len(genes), dtype=np.float32)+        rows = all_rows[labels == g]+        if rows.size < min_cells:+            out[g] = zero+            continue+        if rows.size > MAX_SUB:+            rows = np.sort(rng.choice(rows, size=MAX_SUB, replace=False))+        D = _dense(X, rows)+        active = D.mean(axis=0) >= MIN_EXP+        cc_local = cc_mask & active+        try:+            ax = _Axis(D, cc_local)+        except Exception as exc:  # pragma: no cover - never fail the run+            if verbose:+                print(f"[dpt] {g}: axis fit failed ({exc})")+            out[g] = zero+            continue+        ax.active = active+        delta = ax.contrast(None)+        if delta is None:+            out[g] = zero+            if verbose:+                print(f"[dpt] {g:24s} n={rows.size:5d} |corr|={ax.corr:.3f} -> delta=0")+            continue+        agree = np.zeros(delta.size, dtype=np.float64)+        m = rows.size+        for _ in range(n_split):+            perm = rng.permutation(m)+            h1, h2 = perm[: m // 2], perm[m // 2:]+            d1 = ax.contrast(h1)+            d2 = ax.contrast(h2)+            if d1 is None or d2 is None:+                continue+            agree += (np.sign(d1) == np.sign(d2)) & (np.abs(d1) > 1e-9) & (np.abs(d2) > 1e-9)+        d = (agree / max(1, n_split)) * delta+        cap = float(np.quantile(np.abs(d), 0.99))+        if cap > 0:+            d = np.clip(d, -cap, cap)+        full = np.zeros(len(genes), dtype=np.float32)+        full[active] = d.astype(np.float32)+        out[g] = full+        if verbose:+            top = np.argsort(-np.abs(full))[:12]+            print(f"[dpt] {g:24s} n={rows.size:5d} |corr|={ax.corr:.3f} late_high={ax.late_first} "+                  f"nz={(np.abs(d)>0).sum()} cap={cap:.3f} top=" ++                  ",".join(f"{genes[i]}{'+' if full[i]>0 else '-'}" for i in top))+    return out+++def pseudobulk_delta(+    X_prev, lab_prev, X_last, lab_last, genes: list[str], seed: int = 0+) -> dict[str, np.ndarray]:+    """Per-group pseudobulk difference between the two latest input stages."""+    rng = np.random.default_rng(seed + 7919)+    out: dict[str, np.ndarray] = {}+    shared = sorted(set(str(x) for x in np.unique(lab_prev)) & set(str(x) for x in np.unique(lab_last)))+    for g in shared:+        zero = np.zeros(len(genes), dtype=np.float32)+        a = np.flatnonzero(lab_prev == g)+        b = np.flatnonzero(lab_last == g)+        if a.size < 50 or b.size < 50:+            out[g] = zero+            continue+        if a.size > MAX_SUB:+            a = np.sort(rng.choice(a, size=MAX_SUB, replace=False))+        if b.size > MAX_SUB:+            b = np.sort(rng.choice(b, size=MAX_SUB, replace=False))+        Da, Db = _dense(X_prev, a), _dense(X_last, b)+        ma, mb = Da.mean(axis=0).astype(np.float64), Db.mean(axis=0).astype(np.float64)+        active = (ma >= MIN_EXP) | (mb >= MIN_EXP)+        d = np.zeros(len(genes), dtype=np.float64)+        d[active] = (mb - ma)[active]+        boot = np.zeros(len(genes), dtype=np.float64)+        for _ in range(N_SPLIT):+            ia = rng.integers(0, a.size, a.size)+            ib = rng.integers(0, b.size, b.size)+            bb = Db[ib].mean(axis=0).astype(np.float64) - Da[ia].mean(axis=0).astype(np.float64)+            boot += (np.sign(bb) == np.sign(d)) & (np.abs(bb) > 1e-9)+        d = (boot / N_SPLIT) * d+        cap = float(np.quantile(np.abs(d), 0.99))+        if cap > 0:+            d = np.clip(d, -cap, cap)+        out[g] = d.astype(np.float32)+    return outdiff --git a/solution/run.py b/solution/run.pyindex a752b21..b059965 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,18 +1,44 @@ #!/usr/bin/env python3-"""heart_jcf_peri: run2's T1 winner. Reweight the latest input stage by cell type.+"""Reweighted anchors + a stability-shrunk single-snapshot maturation shift. -Neural Tube dropped; heart types (CM, endothelium/endocardium, SHF/PHM, JCF,-pericardium, proepicardium) x1.6; surface ectoderm, EXEM, paraxial mesoderm-x0.25; everything else x1. 4000 real cells, no expression shift. Proxy reads-E8.5, final reads E9.5; the type set already names both stages' labels.+Composition side (verified in the parent, untouched): drop the groups absent+from the later dissection, scale heart-side groups up and the peripheral+edge groups down, then draw ``N_CELLS`` real cells from the latest input+stage without replacement.++Expression side (new): every output cell stays anchored on a real cell, and+its group's maturation contrast is added in log space with a small step+``beta``.  The contrast is estimated inside the latest snapshot from the axis+that best tracks declining proliferation, oriented towards "later", and shrunk+gene-wise by split-half sign agreement.  With two or more input stages the+group pseudobulk difference between the two latest stages is added as well+(step ``alpha``) behind a sign gate, so nothing degenerates to a plain copy.++``--beta 0`` reproduces the parent prediction bit for bit. """  from __future__ import annotations  import argparse+import os+import sys++for _v in ("OMP_NUM_THREADS", "OPENBLAS_NUM_THREADS", "MKL_NUM_THREADS", "NUMEXPR_NUM_THREADS"):+    os.environ.setdefault(_v, "4")  # bounded, reproducible BLAS++from pathlib import Path++import numpy as np+from scipy import sparse -from src.task1_temporal.reweight import heart_reweight-from src.task1_temporal.view_io import (+sys.path.insert(0, str(Path(__file__).resolve().parent))++from src.task1_temporal.reweight import (  # noqa: E402+    DROP_TYPES,+    largest_remainder,+    type_weights,+)+from src.task1_temporal.view_io import (  # noqa: E402     inputs_by_time,     labels_of,     load_manifest,@@ -22,7 +48,71 @@ from src.task1_temporal.view_io import (     write_prediction, ) +import dpt  # noqa: E402+ N_CELLS = 4000+ALPHA = 0.35  # conservative: a full-strength constant shift scores below copy_last on T1+++def resample_with_labels(X, labels, n_cells, heart_weight, edge_weight, seed):+    """Same allocation / RNG sequence as ``heart_reweight``, plus output labels."""+    rng = np.random.default_rng(seed)+    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)+    w = type_weights(types, heart_weight, edge_weight)+    alloc = largest_remainder(counts * w, n_cells)+    blocks, out_lab, dup = [], [], 0+    for t, n in zip(types, alloc):+        if n <= 0:+            continue+        pool = np.flatnonzero(labels == t)+        replace = pool.size < n+        choice = rng.choice(pool, size=int(n), replace=replace)+        if replace:+            dup += int(n) - int(np.unique(choice).size)+        blocks.append(X[choice])+        out_lab.append(np.full(int(n), t, dtype=object))+    out = sparse.vstack(blocks, format="csr").astype(np.float32)+    out.eliminate_zeros()+    return out, np.concatenate(out_lab) if out_lab else np.array([], dtype=object), dup+++def apply_shift(X, out_lab, deltas, beta, renorm, alpha_deltas=None, alpha=ALPHA):+    """x' = clip(x + beta*delta, 0), then per-cell renormalisation in linear space."""+    groups = {}+    for g, d in deltas.items():+        groups[g] = d+    step = {}+    for g, d in groups.items():+        v = beta * d.astype(np.float32)+        if alpha_deltas is not None and g in alpha_deltas and alpha != 0.0:+            ap = alpha * alpha_deltas[g].astype(np.float32)+            both = np.abs(d) > 0+            cos = float(np.dot(d[both], alpha_deltas[g][both])) if both.any() else 0.0+            if cos < 0:  # snapshot axis disagrees with the observed trend: drop it+                v = ap+            else:+                v = v + ap+        step[g] = v+    if not step or all(not np.any(v) for v in step.values()):+        return X+    Xd = np.asarray(X.todense(), dtype=np.float32)+    tot_before = np.expm1(np.clip(Xd, 0, None)).sum(axis=1)+    uniq = [g for g in np.unique(out_lab) if g in step]+    for g in uniq:+        m = out_lab == g+        if not m.any() or not np.any(step[g]):+            continue+        Xd[m] = np.clip(Xd[m] + step[g][None, :], 0.0, None)+    if renorm:+        lin = np.expm1(Xd)+        tot = lin.sum(axis=1)+        scale = np.ones_like(tot)+        ok = tot > 1e-6+        target = np.median(tot_before[tot_before > 1e-6]) if (tot_before > 1e-6).any() else 1.0+        scale[ok] = target / tot[ok]+        Xd = np.log1p(lin * scale[:, None]).astype(np.float32)+    return sparse.csr_matrix(Xd, dtype=np.float32)   def main() -> None:@@ -30,14 +120,50 @@ def main() -> None:     parser.add_argument("--data", required=True)     parser.add_argument("--out", required=True)     parser.add_argument("--seed", type=int, default=0)+    parser.add_argument("--beta", type=float, default=0.0)+    parser.add_argument("--alpha", type=float, default=ALPHA)+    parser.add_argument("--renorm", type=int, default=1)+    parser.add_argument("--n-cells", type=int, default=N_CELLS)+    parser.add_argument("--heart-weight", type=float, default=1.6)+    parser.add_argument("--edge-weight", type=float, default=0.25)+    parser.add_argument("--verbose", action="store_true")     args = parser.parse_args()      manifest = load_manifest(args.data)     genes = panel_genes(args.data, manifest)-    last = read_stage(args.data, inputs_by_time(manifest)[-1], genes)-    n = target_n_cells(manifest, N_CELLS)-    X = heart_reweight(last.X, labels_of(last), n_cells=n, seed=args.seed)-    write_prediction(X, genes, args.out, seed=args.seed)+    entries = inputs_by_time(manifest)+    last = read_stage(args.data, entries[-1], genes)+    labels = labels_of(last)+    n = target_n_cells(manifest, args.n_cells)++    X, out_lab, dup = resample_with_labels(+        last.X, labels, n, args.heart_weight, args.edge_weight, args.seed+    )+    if args.verbose:+        print(f"[sample] {X.shape} duplicated_rows={dup}")++    if args.beta == 0.0 and (len(entries) < 2 or args.alpha == 0.0):+        write_prediction(X, genes, args.out, seed=args.seed)+        return++    cc_mask = dpt.cell_cycle_genes(args.data, genes)+    if args.verbose:+        print(f"[prior] cell-cycle genes on panel: {int(cc_mask.sum())}")+    deltas = dpt.group_delta(+        last.X, labels, genes, cc_mask, seed=args.seed, verbose=args.verbose+    )++    alpha_deltas = None+    if len(entries) >= 2 and args.alpha != 0.0:+        prev = read_stage(args.data, entries[-2], genes)+        alpha_deltas = dpt.pseudobulk_delta(+            prev.X, labels_of(prev), last.X, labels, genes, seed=args.seed+        )+        if args.verbose:+            print(f"[pb] groups with pseudobulk delta: {len(alpha_deltas)}")++    Xs = apply_shift(X, out_lab, deltas, args.beta, bool(args.renorm), alpha_deltas, args.alpha)+    write_prediction(Xs, 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

做了什么：保留父节点全部组成部件（family 权重、4000 抽样、write_prediction），新增可开关的族内单快照成熟方向平移（GO 细胞周期集 → PC 定向 → 分位对比 → 4 次对半符号收缩），并实现 final 的伪批量分支（α=0.35 + 符号门）。抽样保真检查显示重复行=0，PLAN 步骤 1 不需要。

查分（4 次，剩 16）：β=0 逐位复现父节点 55.9694；β=0.15 → 55.94（cell_state +0.5、direction +0.36、covariation −1.34）；β=0.30 → 55.80（covariation −2.78）；renorm on/off 无差。**de_score 三个 β 完全不变（0.1111）**，即常数式表达对 de_recovery 零作用，且 β 越大 variogram 越差 → 零结果，默认锁回 β=0（1.5 s，vec-check ok）。

下一步最值得试：① 只在 w=1 的族内 top-50 高置信基因上做大步长平移（de_recovery 可能要稀疏、族特异才响应）；② 用「族内 rank 混合」这类不离开真实流形的表达扰动救 covariation；③ 用视图内 marker 结构重新标定族占比，替代手调 1.6/0.25。

调研员

PLAN.json written

审查员

REVIEW.json written