总览 · ← 返回运行 20261003-093415-search-t1-r2-D-s0

节点 n6

composition_trend — type-level composition reweighting + within-type maturity selection

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 911056ceda: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 6bb7e1fdf1。改动的文件：solution/run.py +83 −5

diff --git a/solution/run.py b/solution/run.pyindex b7ef52b..2d139e2 100644--- a/solution/run.py+++ b/solution/run.py@@ -20,6 +20,7 @@ Deterministic for a given --seed. CPU only. from __future__ import annotations  import argparse+import os  import numpy as np @@ -32,12 +33,17 @@ from src.task1_temporal.view_io import (     write_prediction, ) -A_TP = -0.55      # type layer: lower mean proliferation -> larger share-A_CM = 1.2        # cell layer: metabolic maturity (OXPHOS - glycolysis)-A_FATE = 0.6      # cell layer: within-type diffusion pseudotime (more committed cells)+A_TP = float(os.environ.get("VEC_A_TP", "-0.55"))   # type layer: lower mean proliferation -> larger share+A_CM = float(os.environ.get("VEC_A_CM", "1.2"))     # cell layer: metabolic maturity (OXPHOS - glycolysis)+A_FATE = float(os.environ.get("VEC_A_FATE", "0.6")) # cell layer: within-type diffusion pseudotime K_STRAT = 5       # z_met strata per type MIN_TYPE_CELLS = 5 +# --- node 6 mechanisms (defaults ON; VEC_K_SHRINK=0 VEC_GAMMA=0 recovers node 3 exactly) ---+K_SHRINK = float(os.environ.get("VEC_K_SHRINK", "30"))   # empirical-Bayes shrink of type-mean proliferation+GAMMA = float(os.environ.get("VEC_GAMMA", "2.0"))        # direction-consistency guard strength (0 = off)+DET_THRESH = 0.10                                        # guard uses genes detected in >10% of cells+ CYCLE = [     "Mki67", "Top2a", "Pcna", "Ccna2", "Ccnb1", "Ccnb2", "Ccnd1", "Ccneg", "Ccne1",     "Cdk1", "Cdk2", "Cdk4", "Cdk6", "Mcm2", "Mcm3", "Mcm4", "Mcm5", "Mcm6", "Mcm7",@@ -154,6 +160,52 @@ def stratified_sample(w, z_strat, inv, n_types, n_out, K, rng):     return np.sort(np.concatenate(out)) if out else np.empty(0, dtype=np.int64)  +def gene_trend_spearman(X, z_fate, det_thresh, chunk=2000):+    """Per-gene Spearman correlation of expression with z_fate, only for genes detected+    in > det_thresh of cells (0 elsewhere). Ordinal ranks; chunked dense columns."""+    from scipy.stats import rankdata++    n, g = X.shape+    det = np.asarray((X != 0).sum(axis=0), dtype=np.float64).ravel() / max(n, 1)+    cols = np.where(det > det_thresh)[0]+    e = np.zeros(g, dtype=np.float64)+    if len(cols) == 0:+        return e+    rz = rankdata(z_fate).astype(np.float64)+    rz -= rz.mean()+    rzn = float(np.linalg.norm(rz))+    Xcsc = X.tocsc()+    ar = np.arange(n)+    for s in range(0, len(cols), chunk):+        cc = cols[s:s + chunk]+        D = np.asarray(Xcsc[:, cc].todense(), dtype=np.float32)+        order = np.argsort(D, axis=0, kind="stable")+        R = np.empty(D.shape, dtype=np.int32)+        np.put_along_axis(R, order, np.broadcast_to(ar[:, None], D.shape).astype(np.int32), axis=0)+        R = R.astype(np.float64)+        R -= R.mean(axis=0)+        num = R.T @ rz+        den = np.linalg.norm(R, axis=0) * rzn+        e[cc] = np.where(den > 1e-12, num / np.maximum(den, 1e-12), 0.0)+    return e+++def direction_cosine(X, idx, e):+    """Cosine between the pseudobulk shift dp = mean(sampled) - mean(all cells) and the+    within-stage maturity trend e, restricted to genes where e is defined."""+    mask = e != 0.0+    if not mask.any():+        return 0.0+    pb_all = np.asarray(X.mean(axis=0), dtype=np.float64).ravel()+    pb_sel = np.asarray(X[idx].mean(axis=0), dtype=np.float64).ravel()+    dp = (pb_sel - pb_all)[mask]+    ee = e[mask]+    na, ne = np.linalg.norm(dp), np.linalg.norm(ee)+    if na < 1e-12 or ne < 1e-12:+        return 0.0+    return float(dp @ ee / (na * ne))++ def main() -> None:     ap = argparse.ArgumentParser()     ap.add_argument("--data", required=True)@@ -181,13 +233,39 @@ def main() -> None:     raw = raw - tmean(raw)[inv]          # only the within-type order matters     z_fate = zscore(raw) -    w_type = np.exp(A_TP * z_prolif_t)+    # mechanism 1: empirical-Bayes size shrinkage of the type-level proliferation z-score+    shrink = counts.astype(np.float64) / (counts.astype(np.float64) + K_SHRINK)+    z_tp = z_prolif_t * shrink++    w_type = np.exp(A_TP * z_tp)     w_type[counts < MIN_TYPE_CELLS] = 1.0-    w = np.clip(w_type[inv] * np.exp(A_CM * z_met + A_FATE * z_fate), 1e-6, 1e6)+    w_cell = np.exp(A_CM * z_met + A_FATE * z_fate)+    w = np.clip(w_type[inv] * w_cell, 1e-6, 1e6)      n_out = target_n_cells(manifest, adata.n_obs)     rng = np.random.default_rng(args.seed)     idx = stratified_sample(w, z_met, inv, len(uniq), n_out, K_STRAT, rng)++    # mechanism 2: direction-consistency guard on the type layer+    if GAMMA > 0.0:+        e = gene_trend_spearman(X, z_fate, DET_THRESH)+        c = direction_cosine(X, idx, e)+        print(f"[guard] cosine(dp, trend) = {c:.4f}, genes_used = {int((e != 0).sum())}", flush=True)+        if c < 0.0:+            damp = float(np.exp(c * GAMMA))     # in (0, 1): shrinks w_type toward 1+            w_type = np.exp(A_TP * z_tp * damp)+            w_type[counts < MIN_TYPE_CELLS] = 1.0+            w = np.clip(w_type[inv] * w_cell, 1e-6, 1e6)+            idx = stratified_sample(w, z_met, inv, len(uniq), n_out, K_STRAT, rng)+            print(f"[guard] fired: damping = {damp:.4f}", flush=True)++    print(f"[shrink] k = {K_SHRINK}; type weights (raw -> shrunk):", flush=True)+    for t in range(len(uniq)):+        wr = float(np.exp(A_TP * z_prolif_t[t]))+        ws = float(w_type[t])+        if abs(wr - ws) > 1e-6:+            print(f"  {uniq[t]}: n={int(counts[t])} w_raw={wr:.3f} w_final={ws:.3f}", flush=True)+     write_prediction(X[idx], 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

A_FATE=0 gives X3 ≈ 52.2 (+3.5 over parent). Verifying on proxy10 now while preparing METHOD.md:

调研员

 is near 0 on all views, the guard never fires and the node equals the parent (zero progress). Engineer checks the printed cosine value. 4. Runtime: Spearman over ~20k genes × ~17k cells adds ~5 s; acceptable within 30 min.",
  "family_id": "composition_program",
  "mechanism": "Data-adaptive shrinkage of type-level composition reweighting: (a) empirical Bayes shrinkage of type-mean proliferation toward zero based on type size, preventing small/noisy types from driving composition shifts; (b) a direction-consistency guard that dampens the composition shift when the resulting pseudobulk change is anti-correlated with the within-stage gene-level maturity trend.",
  "vs_constant_shift": "Constant displacement shifts every cell's expression by the same vector per type, changing expression values. This mechanism changes which cells are sampled (composition), never modifies expression. The shrinkage and guard are data-driven (depend on type sizes and gene-level correlations in the input), not fixed offsets. The direction guard is a feedback mechanism absent from any constant-shift or fixed-reweighting scheme.",
  "mechanism_evidence": "Engineer should report: (1) the cosine c value per view (X3 vs proxy10) — if c<0 on X3 but c>0 on proxy10, the guard is selectively protecting X3; (2) the effective type weights before and after shrinkage — show which types had weights changed; (3) per-metric breakdown: X3 de_score skill should rise from 0.445, X3 mmd_u skill from 0.472; proxy10 metrics should stay within noise; (4) the number of cells whose sampling probability changed by >20%.",
  "mechanism_off_control": "Set k=0 (no size shrinkage, recover raw type-mean z_prolif) and γ=0 (direction guard disabled, damping factor always 1). This recovers node 3's exact weights and sampling. The output should be byte-identical to node 3. If it is not, there is a code bug. If it is, any score difference in the mechanism-on run is attributable to the shrinkage+guard.",
  "sources": []
}
```

审查员

REVIEW.json written