总览 · ← 返回运行 20261002-202908-search-t1-scr-C

节点 n45

改了什么

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 9c30f07cd7: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 882a3542e2。改动的文件：solution/METHOD.md +8 −6、solution/run.py +9 −29

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex d910c3d..3abaf58 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,10 +1,12 @@ ## 改了什么-相对父节点（ot_moscot seed）的全部改动：-1. 加入谱系分支感知步长调制（lineage_branching 最小实现）：利用耦合矩阵 P 和前一阶段的 celltype 标签，计算每个输出细胞的祖先标签分布熵；高熵（祖先分散于多个类型）的细胞位移步长被收缩。BRANCH_SHRINK 从第 1 轮的 0.5 提高到 0.8（第 1 轮 0.5 导致 de_recovery 从 51.38 降到 50.0，过于激进）。-2. 新增逐基因经验贝叶斯收缩（k018）：对每个基因计算步长的信噪比 SNR = |mean_step| / std_step，收缩因子 = SNR / (SNR + prior)。方向一致的基因保持步长，噪声基因被压缩。目标是改善 direction 分数（当前最弱 49.07）。-关闭对照：BRANCH_SHRINK=1.0 且 GENE_SHRINK_PRIOR=0 时退化为父节点行为。+相对父节点（node 43）的全部改动：+1. 移除分支熵步长收缩（branch_shrink_factors）：ANALYSIS 明确指出"按细胞熵收缩步长这件事本身在压低 de_recovery"，且 BRANCH_SHRINK 从 0.5 调到 0.8 并未救回，说明机制本身有害而非参数问题。+2. 移除逐基因 SNR 收缩（gene_factor = snr/(snr+prior)）：ANALYSIS 指出它近似全局重加权，cell_state +3.29 与 covariation -3.48 互相抵消。+3. 新增步长 PC 截断（truncate_step_pc）：将 step 在 HVG 子空间投影到 PCA 分量，只保留前 STEP_PCS=10 个分量再投影回基因空间。原理：生物学协调变化（方向信号）集中在前几个 PC，高频噪声分散在所有 PC；截断去噪改善 direction，同时保留低秩共变结构不损害 covariation。不做逐基因缩放，避免 covariation 损失。+关闭对照：STEP_PCS = N_PCS（30）时退化为父节点的父节点（node 38）行为（无截断、无分支收缩、无基因收缩）。 ## 用到的知识与出处-- 方向库 lineage_branching 条目：分支概率依赖亲本细胞自身的表达程序，变化来自局部邻域，保留单细胞残差。-- k018（方法卡）：「Optionally shrink per gene (empirical Bayes on the delta's variance) so noisy genes move less」；本实现用 SNR 做逐基因收缩。+- ANALYSIS next_suggestions 第 4 条："直接对 step 做 PC 截断（保留前 10 个 PC）后再解码，再测 direction 是否出噪声"。+- ANALYSIS lessons："在 OT 位移步长上做逐基因线性缩放近似全局重加权"——故移除。+- ANALYSIS lessons："按细胞熵收缩步长这件事本身在压低 de_recovery"——故移除。 - Waddington-OT / moscot：Schiebinger et al. Cell 2019 (doi:10.1016/j.cell.2019.01.006)；Klein et al. Nature 2025 (doi:10.1038/s41586-024-08453-2)。 - 标签来自视图输入阶段的 obs.celltype（官方词汇），未使用禁窗内数据。diff --git a/solution/run.py b/solution/run.pyindex bccc6fc..e10b4b7 100644--- a/solution/run.py+++ b/solution/run.py@@ -49,8 +49,7 @@ K_SMOOTH = 30 LAMBDA = 1.0     # 1 = continue the observed displacement at full rate GROWTH = True    # resample output cells by g^dt_out (WOT birth-death model) N_THREADS = 8-BRANCH_SHRINK = 0.8  # shrink factor for branch-point cells (0=no step, 1=full step)-GENE_SHRINK_PRIOR = 1.0  # prior SNR for per-gene step shrinkage (k018)+STEP_PCS = 10  # number of PCs to keep when truncating step for denoising   def weighted_rows(w: np.ndarray, n: int, rng: np.random.Generator) -> np.ndarray:@@ -58,22 +57,12 @@ def weighted_rows(w: np.ndarray, n: int, rng: np.random.Generator) -> np.ndarray     return np.sort(rng.choice(len(w), size=n, replace=n > len(w), p=p))  -def branch_shrink_factors(P, prev_labels, rows):-    Pc = P[:, rows]-    col_sums = np.maximum(Pc.sum(axis=0, keepdims=True), 1e-30)-    Pc_norm = Pc / col_sums-    unique_labels = np.unique(prev_labels)-    n_labels = len(unique_labels)-    label_idx = np.searchsorted(unique_labels, prev_labels)-    label_mass = np.zeros((n_labels, Pc_norm.shape[1]), dtype=np.float32)-    for li in range(n_labels):-        label_mass[li] = Pc_norm[label_idx == li].sum(axis=0)-    p = np.maximum(label_mass, 1e-30)-    entropy = -(p * np.log(p)).sum(axis=0)-    max_entropy = np.log(max(n_labels, 2))-    norm_entropy = np.clip(entropy / max_entropy, 0, 1)-    factors = 1.0 - norm_entropy * (1.0 - BRANCH_SHRINK)-    return factors.astype(np.float32)+def truncate_step_pc(step, hvg, components, n_keep):+    hvg_step = step[:, hvg]+    pc_proj = hvg_step @ components.T+    pc_proj[:, n_keep:] = 0.0+    step[:, hvg] = pc_proj @ components+    return step   def main() -> None:@@ -101,7 +90,7 @@ def main() -> None:     import torch     from moscot.problems.time import TemporalProblem -    (Zp, Zl), _ = embed([prev.X, last.X], mask, N_HVG, N_PCS, args.seed)+    (Zp, Zl), embed_info = embed([prev.X, last.X], mask, N_HVG, N_PCS, args.seed)     g_prev, g_last = growth_rates([prev, last], genes, mask, args.seed)      obs = pd.DataFrame({"time": np.r_[np.zeros(prev.n_obs), np.ones(last.n_obs)],@@ -121,10 +110,6 @@ def main() -> None:     rows = weighted_rows(g_last ** dt_out if GROWTH else np.ones(last.n_obs), n, rng)     print(f"diag: growth last min/median/max {g_last.min():.3f}/{np.median(g_last):.3f}/{g_last.max():.3f}", file=sys.stderr) -    prev_labels = prev.obs["celltype"].values if "celltype" in prev.obs.columns else np.zeros(prev.n_obs, dtype=str)-    b_factors = branch_shrink_factors(P, prev_labels, rows)-    print(f"diag: branch_shrink min/median/max {b_factors.min():.3f}/{np.median(b_factors):.3f}/{b_factors.max():.3f}", file=sys.stderr)-     Pc = P[:, rows]     del P     Pc /= np.maximum(Pc.sum(axis=0, keepdims=True), 1e-30)@@ -137,12 +122,7 @@ def main() -> None:     step = (smooth - anc) * factor     del smooth, anc     step[:, ~mask] = 0.0-    step *= b_factors[:, None]-    mu_step = step.mean(axis=0)-    sd_step = step.std(axis=0)-    snr = np.abs(mu_step) / np.maximum(sd_step, 1e-10)-    gene_factor = snr / (snr + GENE_SHRINK_PRIOR)-    step *= gene_factor[None, :]+    step = truncate_step_pc(step, embed_info["hvg"], embed_info["components"], STEP_PCS)     X = last.X[rows].toarray()     step *= X > 0  # "addnz": move only measured (non-zero) entries, keeps each cell's zero pattern     X += step

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）。

没有记录调研来源。

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

native

stream

Looking at the results:

| Metric | Grandparent (38) | Round 0 (PC trunc) | Round 1 (kNN smooth) |
|---|---|---|---|
| Score | 50.65 | 50.30 | 49.22 |
| cell_state | 50.07 | 53.88 | 50.77 |
| covariation | 52.52 | 49.85 | 47.0 |
| de_recovery | 51.38 | 46.49 | 48.18 |
| direction | 49.14 | 50.18 | 50.17 |

kNN step smoothing was catastrophic (covariation 47.0, score 49.22). Every step modification has failed vs the grandparent. Strategy: **revert to grandparent behavior** (no step modification) and reduce K_SMOOTH from 30→15 to give more cell-specific displacement directions, targeting the weakest dimension (direction=49.14 in grandparent).

<<<<<<< SEARCH
K_SMOOTH = 30
LAMBDA = 1.0     # 1 = continue the observed displacement at full rate
GROWTH = True    # resample output cells by g^dt_out (WOT birth-death model)
N_THREADS = 8
K_STEP_SMOOTH = 15
=======
K_SMOOTH = 15
LAMBDA = 1.0     # 1 = continue the observed displacement at full rate
GROWTH = True    # resample output cells by g^dt_out (WOT birth-death model)
N_THREADS = 8
>>>>>>> REPLACE

<<<<<<< SEARCH
def knn_smooth_step(Z_rows, step, k):
    if k <= 1:
        return step
    from sklearn.neighbors import NearestNeighbors
    nn = NearestNeighbors(n_neighbors=min(k, len(Z_rows))).fit(Z_rows)
    idx = nn.kneighbors(Z_rows, return_distance=False)
    return step[idx].mean(axis=1).astype(np.float32)


=======
>>>>>>> REPLACE

<<<<<<< SEARCH
    step[:, ~mask] = 0.0