总览 · ← 返回运行 20261002-034201-search-t1-abc-r1-A-era

节点 n3

官方最新输入阶段 copy_last：修复 proxy2 误复制外部心脏细胞的 bug；乘法位移实测不胜过复制，默认关闭。

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 57bd129c62: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 f87cd7b03b。改动的文件：solution/METHOD.md +42 −0、solution/README.md +0 −4、solution/run.py +61 −11

diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..76c3be0--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,42 @@+官方最新输入阶段 copy_last：修复 proxy2 误复制外部心脏细胞的 bug；乘法位移实测不胜过复制，默认关闭。++# 方法++对每个视图：读 `manifest["inputs"]` 里的**官方**阶段（`inputs_by_time(manifest, include_external=False)`），取时间最晚的一个作为基底，按 `target_n_cells` 无放回抽样（rng(seed)，确定性），直接输出。输出基因 = 该视图 `genes.txt`（X3 的面板与官方不同，代码不写死）。++保留了一个可选的位移分支（`SHRINK > 0` 才生效，默认 0）：两个官方阶段时，per-type pseudobulk delta = mean(last|t) − mean(prev|t)，乘时间比例 `(t_target − t_last)/(t_last − t_prev)`（cap 2.0），以**乘法**方式施加：`x' = log1p(expm1(x) · exp(scale·delta))`。乘法形式非负、逐基因单调、零点不动，避免父节点 `clip(x+delta, 0)` 在 0 处堆积质量破坏基因共变。++# 与父节点（node 1, pseudobulk_shift, 39.34）的差异++1. **proxy2 基底修复**（主要收益）：父节点用默认 `inputs_by_time`，在 proxy2 上"最新阶段"是外部 Qiu E9.0（2174 个心脏细胞、27,883/32,285 基因、另一技术），把全胚 E9.5 目标预测成了纯心脏细胞群。改为始终用官方最新阶段作基底。+2. **位移默认关闭**：三把尺子上实测没有任何位移配置胜过 copy_last（见下）。+3. 加法位移的 clip-at-0 换成乘法 fold-change（保留在代码里，SHRINK>0 时可用）。++# 实测（vec-score，A 半，seed 0）++| 配置 | proxy | proxy2 | X3 |+|---|---|---|---|+| 父节点 | 50.04 | 27.43 | 40.53 |+| 官方基底 copy_last（本提交） | 50.40 | 50.40 | 50.00 |+| X3 时间比例×2 加法+clip | - | - | 37.65 |+| X3 时间比例×2 乘法 | - | - | 47.99 |+| X3 时间比例×1 加法+clip（=父行为） | - | - | (40.53) |+| proxy2 Qiu 心脏 delta（谱系映射 LV/AVC-CM←FHF，aSHF/pSHF/OFT-RV/RV/IFT/SV←SHF，Endothelium←Endocardial，乘法 ×1） | - | 37.03 | - |++评分器把 copy_last 定为 50（X3 copy_last 全组恰为 50.0），且拒绝负值 X（"metrics expect log-normalized nonnegative data"）。预期节点分 ≈ (50.40+50.40+50.00)/3 ≈ 50.3。++# 生物学知识来源++- Qiu 心脏 delta 实验（已否决，未进提交）用的谱系映射来自通用心脏发育知识：FHF→LV/AVC 心肌，SHF→RV/OFT/流入道及 aSHF/pSHF，心内膜来自心区内皮（Pijuan-Sala et al. 2019 Nature 小鼠原肠胚图谱的心脏谱系注释）。该实验 direction 51.14（方向略对）但 cell_state 20.04（跨数据集批次效应破坏分布），整体 37.03，故不用。+- 提交代码本身不含任何保留阶段/基因型来源的信息：只用 manifest 给出的输入阶段现场抽样，不写死类型名、比例或表达值。++# 验证过 / 没验证++- 验证过：三个视图 `vec-check` ok；seed 0 下 proxy/proxy2/X3 查分如上；run.py 在三个视图完整跑通（runtime ~5s，内存远低于 28GB）。+- 没验证：final 视图（两官方阶段，本代码退化为 copy_last E9.5——官方报告 copy_last > 常数位移 48.6，方向一致但未实测）；SHRINK>0 的乘法位移在 final 上是否更好；多种子稳定性（抽样 rng 固定 seed，确定性成立）。++# 下一步建议++1. copy_last 恰为 50 = 分数中点，要往上必须产生真实的发育位移：final 视图有 E8.5→E9.5 两个官方阶段，试乘法位移 + 类型级收缩（每类型按 delta 与残差的信噪比缩放），以及"新类型出现"的处理（E9.5 独有类型在 final 是输入，proxy 上没有）。+2. X3 上 E8.75→E9.0 delta 与 E9.0→E9.5 真值方向为负相关（de_direction −0.069），提示 0.25 天短窗 delta 噪声大或发育非线性；可试只取 |delta| 大且在两数据集一致的基因。+3. proxy2 跨数据集 delta 的批次效应可用基因级校正（如按覆盖基因的分位数对齐 Qiu 与官方 E8.5 后再取 delta）重试，direction 信号是正的。diff --git a/solution/README.md b/solution/README.mddeleted file mode 100644index ba29577..0000000--- a/solution/README.md+++ /dev/null@@ -1,4 +0,0 @@-# pseudobulk_shift--最新阶段抽样后，每个细胞加上所属类型在最后一步的伪批量差值 mean(last|type) − mean(prev|type)，夹到 ≥0；前一阶段没有的类型原样复制。-T1 proxy 只有一个输入阶段，没有差值可取，退化成 copy_last（同样的抽样），所以 proxy 分 = copy_last（seed 0 实测 49.77）。final 才真正平移；官方在真实 T1 上报的常数位移是 48.6，低于地板。diff --git a/solution/run.py b/solution/run.pyindex f3a0f25..bca9694 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,13 +1,25 @@ #!/usr/bin/env python3-"""pseudobulk_shift: latest stage + per-cell-type pseudobulk delta of the last step.+"""copy_last_official: sample the latest OFFICIAL input stage; optional+time-scaled multiplicative pseudobulk shift (disabled by default, SHRINK=0). -The delta is mean(last|type) - mean(prev|type) over the two latest inputs,-computed on the full stages and added once to a subsample of the latest stage-(clipped at 0). Types missing from the earlier stage are copied unchanged.+Changes vs parent (pseudobulk_shift seed, node 1): -With a single input stage (T1 proxy: E8.5 only) there is no step to take a-delta from, so this falls back to copy_last with the same sampling. The proxy-therefore cannot tell this seed from copy_last; that gap is expected.+1. BASE = latest OFFICIAL input stage on every view. The parent used+   ``inputs_by_time`` default, which on proxy2 makes the external Qiu E9.0+   heart-only file the "latest" stage, so the prediction became 2174 cardiac+   cells for a whole-embryo target (proxy2 27.43). Official base lifts proxy2+   to ~50.4 (measured, A-half).++2. Shift mechanism kept for two-official-stage views (X3/final) but SHRINK=0:+   measured on X3, copy_last=50.0, time-scaled additive shift with clip=37.6,+   multiplicative (fold-change) shift=48.0 -- no shift beat copying the last+   stage there, so the shift is off. "mult" mode x'=log1p(expm1(x)*exp(d*scale))+   (scale=(t_target-t_last)/(t_last-t_prev), capped) avoids the clip-at-0 mass+   that destroyed the parent's covariation; the scorer rejects negative X, so+   an unclipped additive shift is not usable.++Single official input (proxy, proxy2): pure copy_last with the same sampling+as the parent, output identical. """  from __future__ import annotations@@ -15,8 +27,9 @@ from __future__ import annotations import argparse  import numpy as np+from scipy import sparse -from src.task1_temporal.baselines import shift_rows, type_deltas+from src.task1_temporal.baselines import as_csr, type_deltas from src.task1_temporal.view_io import (     inputs_by_time,     labels_of,@@ -28,6 +41,37 @@ from src.task1_temporal.view_io import (     write_prediction, ) +SCALE_CAP = 2.0+SHRINK = 0.0         # delta multiplier; 0 = copy_last (empirically best, see docstring)+SHIFT_MODE = "mult"  # "mult": log1p(expm1(x)*exp(delta)); "add_floor": clip(x+delta, 0)+++def shift_rows_nc(X, labels, deltas: dict[str, np.ndarray]) -> sparse.csr_matrix:+    """Per-type pseudobulk shift; types without a delta are copied."""+    X = as_csr(X)+    blocks = []+    order = []+    for t in np.unique(labels):+        idx = np.flatnonzero(labels == t)+        order.append(idx)+        if str(t) in deltas:+            d = deltas[str(t)].astype(np.float32)+            if SHIFT_MODE == "mult":+                r = np.exp(np.clip(d, -20.0, 20.0))+                dense = np.expm1(np.asarray(X[idx].toarray(), dtype=np.float32))+                dense *= r+                np.log1p(dense, out=dense)+            else:+                dense = np.asarray(X[idx].toarray(), dtype=np.float32) + d+                np.maximum(dense, 0.0, out=dense)+            blocks.append(sparse.csr_matrix(dense))+        else:+            blocks.append(X[idx])+    out = sparse.vstack(blocks, format="csr")+    inv = np.empty(X.shape[0], dtype=np.int64)+    inv[np.concatenate(order)] = np.arange(X.shape[0])+    return out[inv]+  def main() -> None:     parser = argparse.ArgumentParser()@@ -38,16 +82,22 @@ def main() -> None:      manifest = load_manifest(args.data)     genes = panel_genes(args.data, manifest)-    stages = inputs_by_time(manifest)+    stages = inputs_by_time(manifest, include_external=False)     last = read_stage(args.data, stages[-1], genes)     rng = np.random.default_rng(args.seed)     rows = sample_rows(last.n_obs, target_n_cells(manifest, last.n_obs), rng)     X = last.X[rows]-    if len(stages) >= 2:+    labels = labels_of(last)[rows]+    if SHRINK > 0 and len(stages) >= 2:         prev = read_stage(args.data, stages[-2], genes)         deltas = type_deltas(prev.X, labels_of(prev), last.X, labels_of(last))         del prev-        X = shift_rows(X, labels_of(last)[rows], deltas)+        dt_step = float(stages[-1]["time"]) - float(stages[-2]["time"])+        dt_pred = float(manifest["target"]["time"]) - float(stages[-1]["time"])+        scale = float(np.clip(dt_pred / dt_step, 0.0, SCALE_CAP)) * SHRINK if dt_step > 0 else 0.0+        if scale > 0:+            deltas = {t: d * np.float32(scale) for t, d in deltas.items()}+            X = shift_rows_nc(X, labels, deltas)     write_prediction(X, 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

All views pass. Final state verified: `solution/run.py` + `METHOD.md`, deterministic, ~5s runtime.

**做了什么**：修复父节点在 proxy2 上把外部 Qiu E9.0 心脏细胞当"最新阶段"整体复制的 bug（改为始终用最新**官方**阶段作基底）；把 clip-at-0 加法位移换成乘法 fold-change 位移并加时间比例缩放，但三把尺子实测无一胜过 copy_last，故位移默认关闭（SHRINK=0）。另实测并否决了 proxy2 上跨数据集心脏 delta（谱系映射 Qiu→官方类型）：direction 略升但批次效应把 cell_state 打到 20。

**查分**（6/20 次，A 半）：proxy 50.40、proxy2 50.40（父 27.43）、X3 50.00（父 40.53）→ 预期节点分 ≈ 50.3（父 39.34）。发现评分器以 copy_last=50 为中点且拒绝负值 X。

**下一步最值得试**：① final 视图有两官方阶段，试类型级信噪比收缩的乘法位移 + E9.5 独有类型处理；② proxy2 上先对 Qiu 做基因级分位数对齐再取 delta（方向信号为正，批次是主要损伤）；③ X3 的 0.25 天窗 delta 与真值方向负相关，试只保留跨数据集一致的大 |delta| 基因。

调研员

s smooth and covariance-preserving). This alone should recover several covariation points.\n2) WITHIN-TYPE CORRELATED NOISE: after shifting, add ε_i ~ N(0, σ²·Σ_c) where Σ_c is the within-type correlation matrix estimated from the last input stage (shrinkage estimator: Σ_c = (1-λ)·sample_cov + λ·I, λ=0.3). σ is small (0.05–0.15 of within-type std). This injects realistic gene-gene covariance into the prediction. Compute Σ_c once per type on ≤2000 cells (fast). For types with <30 cells, use global covariance.\n3) COMPOSITION TREND (only when ≥2 input stages, i.e. proxy2 and final): compute per-type proportions at each stage, linearly extrapolate one step forward, clip proportions to [0.001, 0.6], renormalize, then sample cells per type according to extrapolated proportions. SINGLE-INPUT FALLBACK (proxy): skip step 3 entirely (copy_last sampling), skip step 1–2 as well since there is no delta — identical to parent's proxy behaviour.\nKey params initial values: σ=0.1, λ=0.3, softplus threshold=-2. Search: σ∈{0.05,0.1,0.2}, λ∈{0.2,0.3,0.5}.\nvec-score strategy: run once on proxy (expect ~50, same as parent since single-input fallback), once on proxy2 (expect improvement from unclipped + noise), once on X3. If proxy2 improves >3 points, keep; else try σ=0.05. Total ≤4 queries.",
  "expected_groups": ["covariation", "cell_state"],
  "risks": "1) Removing clip may produce negative log-values that the scorer doesn't expect — Engineer should check write_prediction handles floats and verify output shape after first local run. 2) Correlated noise with σ too large adds variance and could hurt cell_state; monitor cell_state group after first vec-score. 3) Composition extrapolation on proxy2 (E8.5→Qiu E9.0 cardiac) may give nonsense proportions since Qiu is cardiac-only; safeguard: if extrapolated proportion of any type exceeds 2× observed, fall back to observed proportions. 4) Improvement may be <2-point noise floor; confirm with 2 vec-score seeds on proxy2 before committing."
}