总览 · ← 返回运行 20261003-094241-search-t2-heart-interp-g24-D-s2

节点 n4

括号真实细胞按 t 混抽（坐标原样）+ 共有细胞类型的类型特异加性表达位移（无全局回退），f=0 关闭机制退化为纯 mix。

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 184446411c: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 6b309b4ac0。改动的文件：solution/EXECUTION.json +1 −0、solution/METHOD.md +44 −0、solution/README.md +0 −4、solution/run.py +111 −15

diff --git a/solution/EXECUTION.json b/solution/EXECUTION.jsonnew file mode 100644index 0000000..6d8012e--- /dev/null+++ b/solution/EXECUTION.json@@ -0,0 +1 @@+{"gpu": false}\ No newline at end of filediff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..91a6e1a--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,44 @@+括号真实细胞按 t 混抽（坐标原样）+ 共有细胞类型的类型特异加性表达位移（无全局回退），f=0 关闭机制退化为纯 mix。++# 方法（family T2HI-01：type-specific expression interpolation）++1. `interp_bracket` 从 manifest 取目标两侧的输入 (a, b, t)，t = (target − t_a)/(t_b − t_a)（只用时间差，视图无关）。+2. 细胞数 n = log 线性插值夹到 [min_cells, max_cells]（proxy 上 = 17,616）。+3. `mix_indices` 按 (1−t, t) 从两侧分层（按 celltype）无放回混抽真实细胞，坐标原样保留。+4. **机制**：对两括号中都有 ≥30 个细胞的类型 k（proxy 上为 5 个共有类型），+   delta_k = mean_b[k] − mean_a[k]（log 空间，按抽样行计算）；a 侧 k 型细胞 +t·delta_k，+   b 侧 k 型细胞 −(1−t)·delta_k，使每个共有类型的伪批量落到线性插值点。+   非共有类型（a 侧 28 个、b 侧 17 个）**不位移**（FALLBACK=none，见下）。+5. 单输入 / 目标不被括号夹住：退化为 copy_last（分层抽 anchor 阶段）。+6. 开关：环境变量 T2HI_SHIFT（默认 1.0；0 = 关闭机制）、T2HI_MODE（add/mult）、+   T2HI_FALLBACK（none/global）、T2HI_MASK（all/nonzero）。提交默认 add + none + all + 1.0。++# 对照与查分（proxy A 半，seed 0，6 次查询）++| 配置 | 总分 | 表达 | 状态 | 形状 | 邻域 | de_score | variogram |+|---|---:|---:|---:|---:|---:|---:|---:|+| 父节点 copy_last | 48.68 | 44.3 | 51.3 | 49.4 | 49.8 | −0.404 | 0.054 |+| 机制关（SHIFT=0，纯 mix） | 58.90 | 63.8 | 65.9 | 53.3 | 52.6 | 0.326 | 0.0275 |+| **机制开（提交默认，add+fb-none）** | **58.59** | **65.0** | 63.8 | 53.3 | 52.3 | **0.359** | 0.0329 |+| add + 全局回退（28 类型也位移） | 53.55 | 65.5 | 49.8 | 51.0 | 47.9 | 0.380 | 0.0773 |+| mult（线性空间乘性，保零） | 58.53 | 63.5 | 65.8 | 53.3 | 51.6 | 0.315 | 0.0274 |+| add + 只移非零元 | 58.75 | 63.8 | 66.0 | 53.3 | 52.0 | 0.326 | 0.0274 |++## 机制生效的证据++- 位移只作用于 5 个共有类型（约 3.5k/17.6k 个细胞），实际改变了这些细胞的 500 个基因值；+  四组分变化：表达 63.8→65.0（de_score 0.326→0.359，de_direction 0.519→0.541，方向正确），+  状态 65.9→63.8（variogram 0.0275→0.0329，零位抬高破坏共变），形状不变（坐标未动），+  邻域 52.6→52.3（表达-位置配对轻微失配）。+- 对照（SHIFT=0）恢复纯 mix 的全部指标，符合预期；与 PLAN 不同点是基座为 mix 而非 copy_last+  （copy_last 基座 + 加性位移即 add 全局回退一行的近似，已验证大幅劣化）。++## 结论与已知弱点++- 机制在 DE 上真实有效（+1.2 表达组分），但等量地损害 variogram/邻域，总分 58.59 与纯 mix+  58.90 的差在噪声（~1 分）内；相对父节点 copy_last 48.68 是明确改进（+9.9）。+- 全局回退位移（把 28 个 a-独有类型也推向全局差）确定有害（−5.4），不要再用。+- mult 保零但不改 DE 排序（mix 的伪批量已在插值点附近），无收益。+- 未验证：不同 t / 括号（final 视图 3 输入时 bracket 由时间差决定，代码路径相同）；+  SHIFT∈(0,1) 的细扫（预期在 58.6–58.9 之间，噪声内）。+- 生物学知识来源：仅使用视图内两个已发布括号阶段的实测差值，无外部/文献先验。diff --git a/solution/README.md b/solution/README.mddeleted file mode 100644index 49dbaae..0000000--- a/solution/README.md+++ /dev/null@@ -1,4 +0,0 @@-# copy_last（T2:heart:val_interp）--输出目标之前最近的输入阶段（proxy 与 final 都是 E8.25_late），表达和坐标原样。E8.25_late 有 58,716 个细胞，超过榜上限 17,616，按细胞类型分层无放回抽到上限。-proxy 预期 ≈ 48.6（seed 0 实测 48.64；表达 44.1 / 状态 51.4 / 形状 49.4 / 邻域 49.6）。低于地板 50 的原因：地板是整份参考阶段本身，de_score 恰为 0；抽样后的子集相对参考阶段有抽样噪声造成的假差异（`_de_raw` 0.196），de_score 变负。合规上限迫使抽样，所以这是合法 copy_last 的实际地板。diff --git a/solution/run.py b/solution/run.pyindex 0595dc9..76e3c3b 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,20 +1,51 @@ #!/usr/bin/env python3-"""copy_last (T2): the latest input at or before the target, with its own coordinates.+"""Type-specific expression interpolation on a real-cell bracket mix (T2 heart interp). -Interpolation boards take the lower bracket stage, extrapolation boards the-last input; this is the floor's reference stage on the proxy. Cells are drawn-stratified by cell type only when the stage exceeds the board's max_cells.-Expression and coordinates are unchanged.+Base: stratified mix of real cells from the two bracket stages at fraction+t = (target - t_a) / (t_b - t_a), coordinates kept as measured (family of the+`mix` seed). Mechanism (PLAN T2HI-01): every sampled cell additionally gets a+per-cell-type, per-gene shift derived from the observed bracket difference+delta_k = mean_b[k] - mean_a[k]; cells from the lower bracket move +t*delta_k,+cells from the upper bracket move -(1-t)*delta_k, so each type's pseudobulk+lands at the interpolated state. Types with too few cells in either bracket+fall back to the global (all-cell) bracket difference; types missing from one+bracket are not shifted.++Mechanism off: env T2HI_SHIFT=0 disables the shift and yields the plain mix.+Single input stage / unbracketed target: degenerates to copy_last. """  from __future__ import annotations  import argparse+import os  import numpy as np -from src.task2_spatial.sample import take-from src.task2_spatial.view_io import anchor_entry, load_manifest, panel_genes, read_stage, write_t2+from src.task2_spatial.sample import interp_count, mix_indices, take+from src.task2_spatial.view_io import (+    anchor_entry,+    interp_bracket,+    load_manifest,+    panel_genes,+    read_stage,+    write_t2,+)++SHIFT = float(os.environ.get("T2HI_SHIFT", "1.0"))+MODE = os.environ.get("T2HI_MODE", "add")+FALLBACK = os.environ.get("T2HI_FALLBACK", "none")  # "global" or "none"+MIN_TYPE_CELLS = int(os.environ.get("T2HI_MIN_TYPE_CELLS", "30"))+++def _copy_anchor(view, manifest, genes, rng, out, seed):+    stage = read_stage(view, anchor_entry(manifest), genes)+    n = int(np.clip(stage.n, manifest["min_cells"], manifest["max_cells"]))+    if n <= stage.n:+        rows = np.sort(take(stage.labels, n, rng))+    else:+        rows = np.sort(rng.choice(stage.n, size=n, replace=True))+    write_t2(out, stage.X[rows].toarray(), stage.coords[rows], genes, seed=seed)   def main() -> None:@@ -24,16 +55,81 @@ def main() -> None:     parser.add_argument("--seed", type=int, default=0)     args = parser.parse_args() +    rng = np.random.default_rng(args.seed)     manifest = load_manifest(args.data)     genes = panel_genes(args.data, manifest)-    stage = read_stage(args.data, anchor_entry(manifest), genes)-    n = int(np.clip(stage.n, manifest["min_cells"], manifest["max_cells"]))-    rng = np.random.default_rng(args.seed)-    if n <= stage.n:-        rows = np.sort(take(stage.labels, n, rng))-    else:  # fewer cells than min_cells: resample with replacement-        rows = np.sort(rng.choice(stage.n, size=n, replace=True))-    write_t2(args.out, stage.X[rows].toarray(), stage.coords[rows], genes, seed=args.seed)+    a_entry, b_entry, t = interp_bracket(manifest)+    if b_entry is None or t is None:+        _copy_anchor(args.data, manifest, genes, rng, args.out, args.seed)+        return++    A = read_stage(args.data, a_entry, genes)+    B = read_stage(args.data, b_entry, genes)+    lo, hi = int(manifest["min_cells"]), int(manifest["max_cells"])+    n = interp_count(A.n, B.n, t, lo, hi)+    rows_a, rows_b = mix_indices(A.labels, B.labels, t, n, rng)++    Xa = np.asarray(A.X[rows_a].todense(), dtype=np.float32)+    Xb = np.asarray(B.X[rows_b].todense(), dtype=np.float32)+    coords = np.vstack([A.coords[rows_a], B.coords[rows_b]])+    lab_a = np.asarray(A.labels)[rows_a].astype(str)+    lab_b = np.asarray(B.labels)[rows_b].astype(str)++    if SHIFT > 0.0 and Xa.shape[0] > 0 and Xb.shape[0] > 0:+        idx_a = {k: np.flatnonzero(lab_a == k) for k in set(lab_a)}+        idx_b = {k: np.flatnonzero(lab_b == k) for k in set(lab_b)}+        if MODE == "add":+            global_delta = Xb.mean(0) - Xa.mean(0) if FALLBACK == "global" else np.zeros(Xa.shape[1], np.float32)+            deltas: dict[str, np.ndarray] = {}+            for k in set(lab_a) | set(lab_b):+                ia, ib = idx_a.get(k), idx_b.get(k)+                if ia is not None and ib is not None and len(ia) >= MIN_TYPE_CELLS and len(ib) >= MIN_TYPE_CELLS:+                    deltas[k] = Xb[ib].mean(0) - Xa[ia].mean(0)+                else:+                    deltas[k] = global_delta+            mask = os.environ.get("T2HI_MASK", "all")  # "all" or "nonzero"+            for k, ia in idx_a.items():+                d = np.float32(t * SHIFT) * deltas[k]+                if mask == "nonzero":+                    Xa[ia] += d * (Xa[ia] > 0)+                else:+                    Xa[ia] += d+            for k, ib in idx_b.items():+                d = np.float32((1.0 - t) * SHIFT) * deltas[k]+                if mask == "nonzero":+                    Xb[ib] -= d * (Xb[ib] > 0)+                else:+                    Xb[ib] -= d+        else:  # multiplicative in linear space: preserves the zero pattern+            Ea = np.expm1(Xa, dtype=np.float64)+            Eb = np.expm1(Xb, dtype=np.float64)+            eps = 1e-6+            gm_a, gm_b = Ea.mean(0), Eb.mean(0)+            gm_t = (1.0 - t) * gm_a + t * gm_b+            gfac_a = np.clip((gm_t + eps) / (gm_a + eps), 0.05, 20.0)+            gfac_b = np.clip((gm_t + eps) / (gm_b + eps), 0.05, 20.0)+            for k in set(lab_a) | set(lab_b):+                ia, ib = idx_a.get(k), idx_b.get(k)+                if ia is not None and ib is not None and len(ia) >= MIN_TYPE_CELLS and len(ib) >= MIN_TYPE_CELLS:+                    m_a = Ea[ia].mean(0)+                    m_b = Eb[ib].mean(0)+                    m_t = (1.0 - t) * m_a + t * m_b+                    fac_a = np.clip((m_t + eps) / (m_a + eps), 0.05, 20.0)+                    fac_b = np.clip((m_t + eps) / (m_b + eps), 0.05, 20.0)+                else:+                    fac_a, fac_b = gfac_a, gfac_b+                # blend factor toward 1.0 by SHIFT (shift strength damping)+                fac_a = 1.0 + SHIFT * (fac_a - 1.0)+                fac_b = 1.0 + SHIFT * (fac_b - 1.0)+                if ia is not None:+                    Ea[ia] *= np.clip(fac_a, 0.0, None)+                if ib is not None:+                    Eb[ib] *= np.clip(fac_b, 0.0, None)+            Xa = np.log1p(Ea).astype(np.float32)+            Xb = np.log1p(Eb).astype(np.float32)++    X = np.clip(np.vstack([Xa, Xb]), 0.0, None)+    write_t2(args.out, X, coords, genes, 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

完成。

**做了什么**：在父节点 copy_last（48.68）基础上实现 PLAN 家族 T2HI-01：按 t=0.4 混抽括号两侧真实细胞（坐标原样），再对 5 个共有细胞类型施加类型特异加性位移（a 侧 +t·delta_k，b 侧 −(1−t)·delta_k），非共有类型不位移、无全局回退。

**查分**（6/20 次，proxy A 半，seed 0）：提交配置 58.59；机制关（纯 mix）58.90；全局回退位移 53.55（确定有害）；mult 58.53；非零掩码 58.75。机制使 de_score 0.326→0.359、表达组 63.8→65.0，但 variogram 等量受损，总分与纯 mix 差在噪声内；相对父节点 +9.9。确定性、vec-check、默认输出与查分文件逐字节一致均已验证。

**下一步最值得试**：位移的收益被零位抬高的 variogram 损失抵消——值得试"稀疏感知"位移（按目标均值重采样非零模式而非加常数），或在 mix 基座上只对 DE 强基因（|delta| 大）位移；另外邻域组（52）仍低于 mix 家族上限，坐标不动的表达位移天然受此制约。

调研员

l per-gene mean shift. 2) If the interpolation fraction is wrong (e.g., target is not E8.5), the shift magnitude is off; however DE metrics are rank-based so direction matters more than magnitude. 3) Shifting expression without moving coordinates could slightly degrade neighborhood_mmd if the shift creates expression-position mismatch; monitor local_spatial group. 4) If expression_change improves but cell_state drops (mmd_u worsens because shifted cells leave the lower-stage distribution), reduce f. Engineer should check all four groups after first vec-score query.",
  "family_id": "T2HI-01",
  "mechanism": "Per-cell-type, per-gene expression shift derived from the observed temporal difference between the two bracket stages, scaled by the interpolation fraction. This injects the correct direction of gene expression change into the prediction.",
  "vs_constant_shift": "The shift varies by both gene AND cell type (a different vector per type), derived from actual observed inter-stage differences. A constant shift applies one vector to all cells regardless of type. Here, e.g., cardiomyocytes and endothelial cells receive different shift vectors reflecting their distinct temporal trajectories.",
  "mechanism_evidence": "1) de_score and de_direction raw values should flip from negative to positive (currently -0.40 and -0.12). 2) Per-type shift vectors should have non-trivial dispersion (std across types > 0.1 for at least 50 genes), confirming type-specificity. 3) expression_change group should rise above 50 while local_spatial stays >= 49. 4) Compare per-gene dp correlation with the bracket-difference direction: Spearman rho(dp, delta) should be significantly positive.",
  "mechanism_off_control": "Set interpolation fraction f = 0 in the script (one flag or constant). With f = 0, no shift is applied and the output is identical to copy_last (parent node 1). Expected: all eight metrics return to parent values (score ~48.7, expression_change ~44.3).",
  "sources": []
}
```