| """Summarize the prepared labels and partitions without evaluating a model.""" |
| import json |
| from pathlib import Path |
|
|
| import pandas as pd |
|
|
| ROOT = Path(__file__).resolve().parents[1] |
|
|
|
|
| def main(): |
| manifest = json.loads((ROOT / "data/processed/manifest.json").read_text()) |
| task_counts = {platform: stats["unique_VH_VL_count"] for platform, stats in manifest["datasets"].items()} |
| total_samples = sum(task_counts.values()) |
| folder = ROOT / "reports" |
| folder.mkdir(exist_ok=True) |
| headline = [] |
| folds = [] |
| warnings = [] |
| duplicates = [] |
| quantiles = [] |
| for platform in ("OmniFlic", "OmniClic"): |
| frame = pd.read_csv(ROOT / f"data/processed/{platform.lower()}_nkp46.csv") |
| stats = manifest["datasets"][platform] |
| kd = frame.KD_M * 1e9 |
| headline.append(f"| {platform} | {stats['raw_clone_count']} | {manifest['deduplicated_counts_before_completeness_filter'][platform]} | {len(frame)} | {frame.family_id.nunique()} | {kd.min():.4g} | {kd.median():.4g} | {kd.max():.4g} |") |
| for fold, group in frame.groupby("fold_id"): |
| folds.append({"platform": platform, "fold_id": fold, "samples": len(group), |
| "families": group.family_id.nunique(), "pKD_min": group.pKD.min(), |
| "pKD_median": group.pKD.median(), "pKD_max": group.pKD.max()}) |
| for q in (0, 0.25, 0.5, 0.75, 1): |
| quantiles.append({"platform": platform, "quantile": q, |
| "KD_nM": float(kd.quantile(q)), "pKD": float(frame.pKD.quantile(q))}) |
| warnings.extend(frame.loc[frame.annotation_warning.notna(), ["sample_id", "clone_id", "vh_seq_id", "annotation_warning"]].assign(platform=platform).to_dict("records")) |
| duplicates.extend(frame.loc[frame.measurement_count > 1, ["sample_id", "clone_id", "measurement_count", "pKD_min", "pKD_max", "pKD"]].assign(platform=platform).to_dict("records")) |
| pd.DataFrame(folds).to_csv(folder / "fold_summary.csv", index=False) |
| pd.DataFrame(quantiles).to_csv(folder / "label_quantiles.csv", index=False) |
| pd.DataFrame(warnings).to_csv(folder / "annotation_warnings.csv", index=False) |
| pd.DataFrame(duplicates, columns=["sample_id", "clone_id", "measurement_count", "pKD_min", "pKD_max", "pKD", "platform"]).to_csv(folder / "duplicate_measurements.csv", index=False) |
| content = """# 数据构建结果 |
| |
| | 平台 | 原始克隆 | VH＋VL 去重后 | 最终样本 | 家族数 | 最小 KD（nM） | 中位 KD（nM） | 最大 KD（nM） | |
| |---|---:|---:|---:|---:|---:|---:|---:| |
| """ + "\n".join(headline) + f""" |
| |
| 最终两个任务共 {total_samples} 个样本。原始 415 条克隆记录保留在来源表中。OmniClic 排除全部 8 条 `_v2`，并完整排除 `50 / 0050_v2` 对应的序列组；最终 {task_counts['OmniClic']} 个样本各自使用非 `_v2` 记录的原始 KD。OmniFlic 的 30 组相同 VH＋VL 统一使用 `Rat_` 的原始 KD；31 条数字编号测量由对应 `Rat_` 替代（其中一组有两条数字编号），最终保留 138 条 `Rat_` 和 75 条数字编号，共 {task_counts['OmniFlic']} 个样本。两个子集的 measurement_count 均为 1。构建前排除的 39 条测量列于 `data/processed/excluded_measurements.csv`，在 `clone_to_sample.csv` 中标记为 included=false。 |
| |
| 序列排除记录列于 `data/processed/excluded.csv`：OmniFlic 的 `Gen_1`（完整 VH 表中称 `Generated_1`，SEQ ID 1364）原始序列末端止于 IMGT 121，缺少 122–128；OmniClic 的 `50` 所属序列组因两条原始 KD 相差约 12.75 倍而排除。`0050_v2` 已包含在 `_v2` 测量排除记录中。序列组排除在家族和外层 fold 分配后应用，其余样本的 family_id、fold_id 和标签沿用 0.5.0。 |
| |
| OmniClic 的 1446、1742 有保守位点氨基酸替换的注释提示。两条序列均直接来自原始 XML，保留原序列及提示，不作修复。 |
| |
| {manifest['datasets']['OmniFlic']['family_count']} 个 OmniFlic 家族和 {manifest['datasets']['OmniClic']['family_count']} 个 OmniClic 家族分别划入五折；最大家族分别有 {manifest['datasets']['OmniFlic']['largest_family']}、{manifest['datasets']['OmniClic']['largest_family']} 个样本。每折的样本数、家族数及标签分布见 `fold_summary.csv`。 |
| |
| 每轮外层一折做 test，剩余四折直接轮流做内层 val，其余三折训练预测头；以四轮平均验证 MAE 选择 Ridge alpha，再在完整外层训练池上重新拟合并评测 test。 |
| |
| 模型评测未运行。数据构建和评测代码的验证使用原始文件一致性检查及合成数据测试。 |
| """ |
| (folder / "data_summary.md").write_text(content) |
| (folder / "preparation_status.json").write_text(json.dumps({ |
| "date": "2026-11-04", "timezone": "Asia/Hong_Kong", "datasets_ready": True, |
| "raw_clone_records": 415, "benchmark_samples": total_samples, |
| "task_samples": task_counts, |
| "dataset_version": manifest["protocol"]["version"], |
| "excluded_clone_measurements": {platform: sum(row["platform"] == platform for row in manifest["excluded_measurements"]) for platform in manifest["protocol"]["platforms"]}, |
| "excluded_sequence_samples": {platform: sum(row["platform"] == platform for row in manifest["excluded_sequences"]) for platform in manifest["protocol"]["platforms"]}, |
| "excluded_original_records": {platform: stats["excluded_measurement_count"] for platform, stats in manifest["datasets"].items()}, |
| "model_evaluations_run": False, |
| }, indent=2) + "\n") |
| print(content) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
