Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261002T124820Z-crosscohort-liu-default/diffusion/dataset/server_1/model8/plans/paired_crosscohort_20261002/PREPARE_EXTRA.json

@@ -0,0 +1,76 @@
+{
+ "schema": "pbisc_sdd.crosscohort.extra.v1",
+ "written_utc": "2026-10-02T12:44:24.416802+00:00",
+ "extra_training_cohorts": [
+  "liu2021"
+ ],
+ "split_prefix": "hoL",
+ "script_sha256": "1887a7b3d1b8a459a420924623f3f3349b53b5fe874f8a7384c8607a31ab6c7c",
+ "prepare_loco_sha256": "94a628303b2326c109341a53730efc5a23c973c7ff905b86e1f42671f059a89c",
+ "inputs": {
+  "liu2021/features.npz": "124ca7198c529f8fd0856731ab04b2bfefba3691c3ed960056ff10f8ffba29e3",
+  "liu2021/folds.csv": "1d8aa9a8af520e73d14386b2573655286e1f5282d0209e0a6c83d4ce33e3dc5f",
+  "liu2021/cells_panel6k.npz": "481bddfaaf2980a10cbe4d0f57a1c6f02e6146141dc00736bc898ac915390c8c"
+ },
+ "files": {
+  "data/liu2021/conditions_raw_K1000.npz": "ad3131eabcfb11b166ca37c7585c50abb11efb60e2a736fe62de604794535e0c",
+  "data/liu2021/conditions_raw_K2000.npz": "9e2d84da00e8d729e57b934b6fefc8123d7e1581db68751092af383de89925b2",
+  "splits/hoL_immicro/test_xcal_K1000.npz": "c347c9c871de798fbb17b83a7804192a84b66ef4e7fa55e2774342b5d161c399",
+  "splits/hoL_immicro/test_xcal_K2000.npz": "62516a6ad3fa5f1fa944595e6e8ccfcd8f5424cadc7f8032cb8610a17e896e72",
+  "splits/hoL_immicro.json": "d5646bee4a5f735e64f490824ca865a2cf8191234a8de40c0963cd590ebe6743",
+  "splits/hoL_ye2026_sepsis/test_xcal_K1000.npz": "b4c2b51f0669e2f0236e7981a6b5b5ddabcd22cbf9e3949384144fb55f5564ed",
+  "splits/hoL_ye2026_sepsis/test_xcal_K2000.npz": "03d9fa67aac63bf8b4a33b220dc4eed8a04b7fc72dbf4c0f6169ebfa24e6a5dd",
+  "splits/hoL_ye2026_sepsis.json": "39cad7b25faf6c15cd65f0512977b9e526353cfb085dc2c45d18f8edc56db426",
+  "splits/hoL_imperial_eskd/test_xcal_K1000.npz": "a6ad2125a8ea0bf0c4e28a6b85b4ef9204a5514f919befbfcb6cbf75e2d2ce08",
+  "splits/hoL_imperial_eskd/test_xcal_K2000.npz": "b576896df92bb45fa60e3be46712f3c7e453a5e93f34e8ec6d606858e1d9b2fc",
+  "splits/hoL_imperial_eskd.json": "dc497bb9a090f5762cf897e1e1b297cb63e33e459eedfcc3c0a181d4e66b62c4",
+  "splits/hoL_giroux2022/test_xcal_K1000.npz": "b791302e74c54f0117048083448e10447112b0ca9ded8af001a107792532ba53",
+  "splits/hoL_giroux2022/test_xcal_K2000.npz": "79d1176f1e9b9382ce0b351143818876001a7ef18731f71e9f85e2d416076340",
+  "splits/hoL_giroux2022.json": "3c26e00bd5691416d6ad89c8a5335dcdc42a252198ba0966b80d2a0d2b68495a",
+  "splits/hoL_ji2025_chili/test_xcal_K1000.npz": "c6dc46b8d5eb7b5507e67d77a000322b483d90963b9df82291af8f4d520b877d",
+  "splits/hoL_ji2025_chili/test_xcal_K2000.npz": "d5733ec713c666feb2c66f21cefe9cd1a4a12ff68c75610854b59dd8ebe0c0a8",
+  "splits/hoL_ji2025_chili.json": "1efe2e8a540137380f689ea29d64f99189981b8fcc636824c7ae858db989f787"
+ },
+ "cohorts": {
+  "liu2021": {
+   "role": "training only",
+   "training_exclusion": {
+    "file": "/home/sj_server_1/gblee/datalake/paired-pbmc-bulk-sc/processing/liu2021_v1_20261002/identity/hto_negative_sensitivity.csv",
+    "column": "negative_fraction",
+    "max": "0.3",
+    "reason": "HTO-negative cells > 30 % of the row's unsorted cells: retained-cell composition is biased against the all-cell bulk",
+    "file_sha256": "336827fdb731d10c081b02fd49d041aea2690b915c9585891ad25c7fef13aeda",
+    "sample_ids": [
+     "HGR0000042_T0",
+     "HGR0000051_T2",
+     "HGR0000069_T2",
+     "HGR0000092_T1",
+     "HGR0000094_T1",
+     "HGR0000101_T0",
+     "HGR0000124_T1",
+     "HGR0000134_T1",
+     "HGR0000142_T0",
+     "HGR0000144_T0",
+     "HGR0000392_T3",
+     "HGR0000430_T1",
+     "SHD2_HC",
+     "SHD6_HC"
+    ]
+   },
+   "usable_rows": 58,
+   "training_pool_rows": 44,
+   "training_pool_donors": 37,
+   "inner_val_donors": [
+    "HGR0000098",
+    "HGR0000115",
+    "HGR0000143",
+    "SHD1"
+   ],
+   "inner_val_rows": 4,
+   "groups": {
+    "COVID": 28,
+    "HC": 9
+   }
+  }
+ }
+}

paper-releases/pbisc-diffusion/updates/20261002T124820Z-crosscohort-liu-default/diffusion/dataset/server_1/model8/plans/paired_crosscohort_20261002/splits/hoL_giroux2022.json

@@ -0,0 +1,497 @@
+{
+ "schema": "pbisc_sdd.crosscohort.split.v1",
+ "split": "hoL_giroux2022",
+ "held_out": "giroux2022",
+ "held_out_bulk": "PBMC",
+ "training_cohorts": [
+  {
+   "cohort": "immicro",
+   "dataset_dir": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/immicro",
+   "condition_npz": {
+    "1000": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/data/immicro/conditions_raw_K1000.npz",
+    "2000": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/data/immicro/conditions_raw_K2000.npz"
+   },
+   "train_sample_ids": [
+    "HS1",
+    "HS2",
+    "HS3",
+    "HS6",
+    "HS7",
+    "HS8",
+    "HS9",
+    "HS10",
+    "HS12",
+    "HS13",
+    "HS14",
+    "HS15",
+    "HS16",
+    "HS17",
+    "HS18",
+    "HS19",
+    "HS20",
+    "HS21",
+    "HS23",
+    "HS24",
+    "HS25",
+    "HS28",
+    "HS29",
+    "HS30",
+    "HS32",
+    "HS34",
+    "HS35",
+    "HS36",
+    "HS37",
+    "HS38",
+    "HS39",
+    "HS40",
+    "HS41",
+    "HS42",
+    "HS43",
+    "HS44",
+    "HS45",
+    "HS47",
+    "HS48",
+    "HS49",
+    "HS50",
+    "HS51",
+    "HS52",
+    "HS53",
+    "HS54",
+    "HS56",
+    "HS57",
+    "HS58",
+    "HS59",
+    "HS61",
+    "HS63",
+    "HS64",
+    "HS65",
+    "HS66",
+    "HS67",
+    "HS68",
+    "HS69",
+    "HS70",
+    "HS71",
+    "HS72",
+    "HS73",
+    "HS74",
+    "HS75",
+    "HS76",
+    "HS77",
+    "HS78",
+    "HS79",
+    "HS80",
+    "HS81",
+    "HS82",
+    "HS83",
+    "HS84",
+    "HS85",
+    "HS86",
+    "HS88",
+    "HS89",
+    "HS90",
+    "HS91",
+    "HS92",
+    "HS93",
+    "HS94",
+    "HS95",
+    "HS96",
+    "HS97",
+    "HS98",
+    "HS100",
+    "HS101",
+    "HS102",
+    "HS103",
+    "HS104",
+    "HS105",
+    "HS106",
+    "HS107",
+    "HS110",
+    "HS111"
+   ],
+   "inner_val_sample_ids": [
+    "HS4",
+    "HS11",
+    "HS26",
+    "HS27",
+    "HS31",
+    "HS46",
+    "HS55",
+    "HS60",
+    "HS62",
+    "HS87",
+    "HS108",
+    "HS109"
+   ]
+  },
+  {
+   "cohort": "ye2026_sepsis",
+   "dataset_dir": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/ye2026_sepsis",
+   "condition_npz": {
+    "1000": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/data/ye2026_sepsis/conditions_raw_K1000.npz",
+    "2000": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/data/ye2026_sepsis/conditions_raw_K2000.npz"
+   },
+   "train_sample_ids": [
+    "Abd-PS04",
+    "Abd-PS05",
+    "Abd-PS07",
+    "Abd-PS19",
+    "Abd-PS20",
+    "Abd-PS27",
+    "Abd-S103",
+    "Abd-S120",
+    "Abd-S123",
+    "Abd-S156",
+    "Abd-S39",
+    "Abd-S42",
+    "Abd-S47",
+    "Abd-S51",
+    "Abd-S57",
+    "Abd-S66",
+    "Abd-S79",
+    "Abd-S83",
+    "HC05",
+    "HC07",
+    "HC09",
+    "HC12",
+    "HC13",
+    "HC16",
+    "IC10",
+    "IC15",
+    "IC16",
+    "IC22",
+    "IC23",
+    "PHC01",
+    "PHC02",
+    "PHC03",
+    "PHC05",
+    "PIC01",
+    "PIC02",
+    "PIC04",
+    "PIC05",
+    "Res-PS02",
+    "Res-PS03",
+    "Res-PS12",
+    "Res-PS21",
+    "Res-PS22",
+    "Res-PS25",
+    "Res-PS26",
+    "Res-S10",
+    "Res-S125",
+    "Res-S127",
+    "Res-S136",
+    "Res-S146",
+    "Res-S149",
+    "Res-S150",
+    "Res-S164",
+    "Res-S166",
+    "Res-S177",
+    "Res-S184",
+    "Res-S19",
+    "Res-S26",
+    "Res-S40",
+    "Res-S77",
+    "Res-S87",
+    "Skin-S140",
+    "Skin-S142",
+    "Skin-S152",
+    "Skin-S38",
+    "Skin-S43",
+    "Skin-S46",
+    "Skin-S52",
+    "Skin-S74",
+    "Uri-S134",
+    "Uri-S147",
+    "Uri-S157",
+    "Uri-S20",
+    "Uri-S93"
+   ],
+   "inner_val_sample_ids": [
+    "Abd-PS23",
+    "Abd-S102",
+    "Abd-S54",
+    "Abd-S98",
+    "HC01",
+    "IC19",
+    "Res-PS16",
+    "Res-S139",
+    "Res-S174"
+   ]
+  },
+  {
+   "cohort": "imperial_eskd",
+   "dataset_dir": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/imperial_eskd",
+   "condition_npz": {
+    "1000": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/data/imperial_eskd/conditions_raw_K1000.npz",
+    "2000": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/data/imperial_eskd/conditions_raw_K2000.npz"
+   },
+   "train_sample_ids": [
+    "C20_1004",
+    "C20_1404",
+    "C21_1004",
+    "C21_1704",
+    "C21_2504",
+    "C23_2204",
+    "C31_1304",
+    "C31_2004",
+    "C31_2704",
+    "C33_1401",
+    "C33_1601",
+    "C33_1901",
+    "C34_1404",
+    "C34_2104",
+    "C34_2504",
+    "C36_1504",
+    "C36_1704",
+    "C36_2104",
+    "C40_1504",
+    "C40_2204",
+    "C40_0105",
+    "C42_1504",
+    "C42_2204",
+    "C42_2704",
+    "C58_1604",
+    "C58_2501",
+    "C58_2901",
+    "C58_0102",
+    "C58_0402",
+    "C58_1003",
+    "C60_2504",
+    "C65_1704",
+    "C65_2104",
+    "C65_2504",
+    "C69_1804",
+    "C82_2504",
+    "C85_2004",
+    "C93_2104",
+    "C93_2704",
+    "C93_0405",
+    "C101_0601",
+    "C101_1101",
+    "C101_1301",
+    "C101_1801",
+    "C101_2001",
+    "C101_2502",
+    "C103_2204",
+    "C104_2204",
+    "C106_2204",
+    "C108_2204",
+    "C108_1501",
+    "C108_1801",
+    "C108_2201",
+    "C113_0605",
+    "C123_2404",
+    "C123_2904",
+    "C123_0505",
+    "C124_2404",
+    "C124_2704",
+    "C124_0105",
+    "C126_2504",
+    "C126_3004",
+    "C126_0605",
+    "C127_2504",
+    "C127_0405",
+    "C127_1105",
+    "C132_2804",
+    "C132_0205",
+    "C132_1105",
+    "C137_3004",
+    "C137_2502",
+    "C138_3004",
+    "C138_2301",
+    "C138_3001",
+    "C138_0202",
+    "C138_0402",
+    "C140_3004",
+    "C140_2801",
+    "C140_3001",
+    "C140_0202",
+    "C140_0402",
+    "C140_0602",
+    "C140_0902",
+    "C140_1103",
+    "C141_2001",
+    "C141_2201",
+    "C141_3001",
+    "C141_0202",
+    "C145_3004",
+    "C145_0701",
+    "C145_1201",
+    "C145_1401",
+    "C145_1901",
+    "C145_2301",
+    "C146_3004",
+    "C146_1201",
+    "C146_1401",
+    "C146_1601",
+    "C146_1901",
+    "C146_2301",
+    "C146_2701",
+    "C146_2502",
+    "C152_0405",
+    "C161_1105",
+    "C164_1105",
+    "C167_1205",
+    "C168_1205",
+    "C168_2001",
+    "C168_2301",
+    "C168_3001",
+    "C168_0202",
+    "C168_0403",
+    "C169_1205",
+    "C169_2001",
+    "C169_2301",
+    "C169_0202",
+    "C169_0403",
+    "C170_1205",
+    "C170_2301",
+    "C170_0202",
+    "C170_0402",
+    "C170_0403",
+    "C187_2605",
+    "C187_1201",
+    "C187_1401",
+    "C187_1901",
+    "C187_2301",
+    "C187_2502",
+    "C190_2605",
+    "C190_2901",
+    "C190_0102",
+    "C190_0302",
+    "C190_0502",
+    "C190_0802",
+    "C190_1002",
+    "C190_0903",
+    "C193_2605",
+    "C195_2605",
+    "C197_2605",
+    "C212_2705",
+    "C214_2705",
+    "C237_2805",
+    "C238_2905",
+    "C240_2905",
+    "C246_2905",
+    "C251_2905"
+   ],
+   "inner_val_sample_ids": [
+    "C63_1704",
+    "C63_2404",
+    "C63_2904",
+    "C128_2504",
+    "C128_0405",
+    "C128_1105",
+    "C142_3004",
+    "C147_3004",
+    "C147_1401",
+    "C147_1901",
+    "C234_2805"
+   ]
+  },
+  {
+   "cohort": "liu2021",
+   "dataset_dir": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/liu2021",
+   "condition_npz": {
+    "1000": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/data/liu2021/conditions_raw_K1000.npz",
+    "2000": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/data/liu2021/conditions_raw_K2000.npz"
+   },
+   "train_sample_ids": [
+    "AA220014_HC",
+    "AA220534_HC",
+    "AA220907_HC",
+    "HGR0000071_T0",
+    "HGR0000074_T0",
+    "HGR0000083_T1",
+    "HGR0000089_T1",
+    "HGR0000091_T0",
+    "HGR0000093_T0",
+    "HGR0000136_T0",
+    "HGR0000137_T0",
+    "HGR0000141_T0",
+    "IM000040_HC",
+    "HGR0000042_T1",
+    "HGR0000069_T0",
+    "HGR0000069_T1",
+    "HGR0000092_T0",
+    "HGR0000094_T0",
+    "HGR0000124_T0",
+    "HGR0000144_T2",
+    "HGR0000391_T0",
+    "HGR0000395_T0",
+    "HGR0000418_T1",
+    "SHD4_HC",
+    "SHD7_HC",
+    "HGR0000051_T0",
+    "HGR0000051_T1",
+    "HGR0000051_T3",
+    "HGR0000101_T1",
+    "HGR0000101_T3",
+    "HGR0000102_T0",
+    "HGR0000102_T1",
+    "HGR0000134_T0",
+    "HGR0000135_T0",
+    "HGR0000135_T1",
+    "HGR0000392_T0",
+    "HGR0000429_T1",
+    "HGR0000429_T2",
+    "SHD3_HC",
+    "SHD5_HC"
+   ],
+   "inner_val_sample_ids": [
+    "HGR0000098_T0",
+    "HGR0000143_T0",
+    "HGR0000115_T1",
+    "SHD1_HC"
+   ]
+  }
+ ],
+ "test": {
+  "cohort": "giroux2022",
+  "dataset_dir": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/giroux2022",
+  "sample_ids": [
+   "450905_d0",
+   "450905_d14",
+   "0B943B_d0",
+   "0B943B_d14",
+   "7085CA_d0",
+   "7085CA_d7",
+   "7085CA_d14",
+   "2AD75E_d0",
+   "2AD75E_d7",
+   "2AD75E_d14",
+   "B85D75_d0",
+   "B85D75_d7",
+   "180E1A_d0",
+   "180E1A_d7",
+   "180E1A_d14",
+   "82CCF5_d0",
+   "82CCF5_d7",
+   "82CCF5_d14",
+   "BAAF62_d0",
+   "BAAF62_d7",
+   "BAAF62_d14",
+   "0BF51C_d0",
+   "0BF51C_d7",
+   "0BF51C_d14"
+  ],
+  "condition_npz": {
+   "1000": {
+    "raw": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/data/giroux2022/test_raw_K1000.npz",
+    "xcal": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/splits/hoL_giroux2022/test_xcal_K1000.npz"
+   },
+   "2000": {
+    "raw": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/data/giroux2022/test_raw_K2000.npz",
+    "xcal": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/splits/hoL_giroux2022/test_xcal_K2000.npz"
+   }
+  }
+ },
+ "xcal": {
+  "alpha": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/splits/hoL_giroux2022/alpha_xcal.npy",
+  "genes_estimated": 11275,
+  "rule": "cohort-balanced mean CPM of the training cohorts' usable pairs"
+ },
+ "n_train_rows": 354,
+ "n_inner_val_rows": 36,
+ "extra_training_cohorts": [
+  "liu2021"
+ ],
+ "base_split": "ho_giroux2022"
+}

paper-releases/pbisc-diffusion/updates/20261002T124820Z-crosscohort-liu-default/diffusion/dataset/server_1/model8/plans/paired_crosscohort_20261002/splits/hoL_immicro.json

@@ -0,0 +1,496 @@
+{
+ "schema": "pbisc_sdd.crosscohort.split.v1",
+ "split": "hoL_immicro",
+ "held_out": "immicro",
+ "held_out_bulk": "PBMC",
+ "training_cohorts": [
+  {
+   "cohort": "ye2026_sepsis",
+   "dataset_dir": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/ye2026_sepsis",
+   "condition_npz": {
+    "1000": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/data/ye2026_sepsis/conditions_raw_K1000.npz",
+    "2000": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/data/ye2026_sepsis/conditions_raw_K2000.npz"
+   },
+   "train_sample_ids": [
+    "Abd-PS04",
+    "Abd-PS05",
+    "Abd-PS07",
+    "Abd-PS19",
+    "Abd-PS20",
+    "Abd-PS27",
+    "Abd-S103",
+    "Abd-S120",
+    "Abd-S123",
+    "Abd-S156",
+    "Abd-S39",
+    "Abd-S42",
+    "Abd-S47",
+    "Abd-S51",
+    "Abd-S57",
+    "Abd-S66",
+    "Abd-S79",
+    "Abd-S83",
+    "HC05",
+    "HC07",
+    "HC09",
+    "HC12",
+    "HC13",
+    "HC16",
+    "IC10",
+    "IC15",
+    "IC16",
+    "IC22",
+    "IC23",
+    "PHC01",
+    "PHC02",
+    "PHC03",
+    "PHC05",
+    "PIC01",
+    "PIC02",
+    "PIC04",
+    "PIC05",
+    "Res-PS02",
+    "Res-PS03",
+    "Res-PS12",
+    "Res-PS21",
+    "Res-PS22",
+    "Res-PS25",
+    "Res-PS26",
+    "Res-S10",
+    "Res-S125",
+    "Res-S127",
+    "Res-S136",
+    "Res-S146",
+    "Res-S149",
+    "Res-S150",
+    "Res-S164",
+    "Res-S166",
+    "Res-S177",
+    "Res-S184",
+    "Res-S19",
+    "Res-S26",
+    "Res-S40",
+    "Res-S77",
+    "Res-S87",
+    "Skin-S140",
+    "Skin-S142",
+    "Skin-S152",
+    "Skin-S38",
+    "Skin-S43",
+    "Skin-S46",
+    "Skin-S52",
+    "Skin-S74",
+    "Uri-S134",
+    "Uri-S147",
+    "Uri-S157",
+    "Uri-S20",
+    "Uri-S93"
+   ],
+   "inner_val_sample_ids": [
+    "Abd-PS23",
+    "Abd-S102",
+    "Abd-S54",
+    "Abd-S98",
+    "HC01",
+    "IC19",
+    "Res-PS16",
+    "Res-S139",
+    "Res-S174"
+   ]
+  },
+  {
+   "cohort": "imperial_eskd",
+   "dataset_dir": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/imperial_eskd",
+   "condition_npz": {
+    "1000": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/data/imperial_eskd/conditions_raw_K1000.npz",
+    "2000": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/data/imperial_eskd/conditions_raw_K2000.npz"
+   },
+   "train_sample_ids": [
+    "C20_1004",
+    "C20_1404",
+    "C21_1004",
+    "C21_1704",
+    "C21_2504",
+    "C23_2204",
+    "C31_1304",
+    "C31_2004",
+    "C31_2704",
+    "C33_1401",
+    "C33_1601",
+    "C33_1901",
+    "C34_1404",
+    "C34_2104",
+    "C34_2504",
+    "C36_1504",
+    "C36_1704",
+    "C36_2104",
+    "C40_1504",
+    "C40_2204",
+    "C40_0105",
+    "C42_1504",
+    "C42_2204",
+    "C42_2704",
+    "C58_1604",
+    "C58_2501",
+    "C58_2901",
+    "C58_0102",
+    "C58_0402",
+    "C58_1003",
+    "C60_2504",
+    "C65_1704",
+    "C65_2104",
+    "C65_2504",
+    "C69_1804",
+    "C82_2504",
+    "C85_2004",
+    "C93_2104",
+    "C93_2704",
+    "C93_0405",
+    "C101_0601",
+    "C101_1101",
+    "C101_1301",
+    "C101_1801",
+    "C101_2001",
+    "C101_2502",
+    "C103_2204",
+    "C104_2204",
+    "C106_2204",
+    "C108_2204",
+    "C108_1501",
+    "C108_1801",
+    "C108_2201",
+    "C113_0605",
+    "C123_2404",
+    "C123_2904",
+    "C123_0505",
+    "C124_2404",
+    "C124_2704",
+    "C124_0105",
+    "C126_2504",
+    "C126_3004",
+    "C126_0605",
+    "C127_2504",
+    "C127_0405",
+    "C127_1105",
+    "C132_2804",
+    "C132_0205",
+    "C132_1105",
+    "C137_3004",
+    "C137_2502",
+    "C138_3004",
+    "C138_2301",
+    "C138_3001",
+    "C138_0202",
+    "C138_0402",
+    "C140_3004",
+    "C140_2801",
+    "C140_3001",
+    "C140_0202",
+    "C140_0402",
+    "C140_0602",
+    "C140_0902",
+    "C140_1103",
+    "C141_2001",
+    "C141_2201",
+    "C141_3001",
+    "C141_0202",
+    "C145_3004",
+    "C145_0701",
+    "C145_1201",
+    "C145_1401",
+    "C145_1901",
+    "C145_2301",
+    "C146_3004",
+    "C146_1201",
+    "C146_1401",
+    "C146_1601",
+    "C146_1901",
+    "C146_2301",
+    "C146_2701",
+    "C146_2502",
+    "C152_0405",
+    "C161_1105",
+    "C164_1105",
+    "C167_1205",
+    "C168_1205",
+    "C168_2001",
+    "C168_2301",
+    "C168_3001",
+    "C168_0202",
+    "C168_0403",
+    "C169_1205",
+    "C169_2001",
+    "C169_2301",
+    "C169_0202",
+    "C169_0403",
+    "C170_1205",
+    "C170_2301",
+    "C170_0202",
+    "C170_0402",
+    "C170_0403",
+    "C187_2605",
+    "C187_1201",
+    "C187_1401",
+    "C187_1901",
+    "C187_2301",
+    "C187_2502",
+    "C190_2605",
+    "C190_2901",
+    "C190_0102",
+    "C190_0302",
+    "C190_0502",
+    "C190_0802",
+    "C190_1002",
+    "C190_0903",
+    "C193_2605",
+    "C195_2605",
+    "C197_2605",
+    "C212_2705",
+    "C214_2705",
+    "C237_2805",
+    "C238_2905",
+    "C240_2905",
+    "C246_2905",
+    "C251_2905"
+   ],
+   "inner_val_sample_ids": [
+    "C63_1704",
+    "C63_2404",
+    "C63_2904",
+    "C128_2504",
+    "C128_0405",
+    "C128_1105",
+    "C142_3004",
+    "C147_3004",
+    "C147_1401",
+    "C147_1901",
+    "C234_2805"
+   ]
+  },
+  {
+   "cohort": "giroux2022",
+   "dataset_dir": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/giroux2022",
+   "condition_npz": {
+    "1000": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/data/giroux2022/conditions_raw_K1000.npz",
+    "2000": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/data/giroux2022/conditions_raw_K2000.npz"
+   },
+   "train_sample_ids": [
+    "450905_d0",
+    "450905_d14",
+    "0B943B_d0",
+    "0B943B_d14",
+    "7085CA_d0",
+    "7085CA_d7",
+    "7085CA_d14",
+    "2AD75E_d0",
+    "2AD75E_d7",
+    "2AD75E_d14",
+    "B85D75_d0",
+    "B85D75_d7",
+    "180E1A_d0",
+    "180E1A_d7",
+    "180E1A_d14",
+    "82CCF5_d0",
+    "82CCF5_d7",
+    "82CCF5_d14",
+    "BAAF62_d0",
+    "BAAF62_d7",
+    "BAAF62_d14",
+    "0BF51C_d0",
+    "0BF51C_d7",
+    "0BF51C_d14"
+   ],
+   "inner_val_sample_ids": []
+  },
+  {
+   "cohort": "liu2021",
+   "dataset_dir": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/liu2021",
+   "condition_npz": {
+    "1000": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/data/liu2021/conditions_raw_K1000.npz",
+    "2000": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/data/liu2021/conditions_raw_K2000.npz"
+   },
+   "train_sample_ids": [
+    "AA220014_HC",
+    "AA220534_HC",
+    "AA220907_HC",
+    "HGR0000071_T0",
+    "HGR0000074_T0",
+    "HGR0000083_T1",
+    "HGR0000089_T1",
+    "HGR0000091_T0",
+    "HGR0000093_T0",
+    "HGR0000136_T0",
+    "HGR0000137_T0",
+    "HGR0000141_T0",
+    "IM000040_HC",
+    "HGR0000042_T1",
+    "HGR0000069_T0",
+    "HGR0000069_T1",
+    "HGR0000092_T0",
+    "HGR0000094_T0",
+    "HGR0000124_T0",
+    "HGR0000144_T2",
+    "HGR0000391_T0",
+    "HGR0000395_T0",
+    "HGR0000418_T1",
+    "SHD4_HC",
+    "SHD7_HC",
+    "HGR0000051_T0",
+    "HGR0000051_T1",
+    "HGR0000051_T3",
+    "HGR0000101_T1",
+    "HGR0000101_T3",
+    "HGR0000102_T0",
+    "HGR0000102_T1",
+    "HGR0000134_T0",
+    "HGR0000135_T0",
+    "HGR0000135_T1",
+    "HGR0000392_T0",
+    "HGR0000429_T1",
+    "HGR0000429_T2",
+    "SHD3_HC",
+    "SHD5_HC"
+   ],
+   "inner_val_sample_ids": [
+    "HGR0000098_T0",
+    "HGR0000143_T0",
+    "HGR0000115_T1",
+    "SHD1_HC"
+   ]
+  }
+ ],
+ "test": {
+  "cohort": "immicro",
+  "dataset_dir": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/immicro",
+  "sample_ids": [
+   "HS1",
+   "HS2",
+   "HS3",
+   "HS4",
+   "HS6",
+   "HS7",
+   "HS8",
+   "HS9",
+   "HS10",
+   "HS11",
+   "HS12",
+   "HS13",
+   "HS14",
+   "HS15",
+   "HS16",
+   "HS17",
+   "HS18",
+   "HS19",
+   "HS20",
+   "HS21",
+   "HS23",
+   "HS24",
+   "HS25",
+   "HS26",
+   "HS27",
+   "HS28",
+   "HS29",
+   "HS30",
+   "HS31",
+   "HS32",
+   "HS34",
+   "HS35",
+   "HS36",
+   "HS37",
+   "HS38",
+   "HS39",
+   "HS40",
+   "HS41",
+   "HS42",
+   "HS43",
+   "HS44",
+   "HS45",
+   "HS46",
+   "HS47",
+   "HS48",
+   "HS49",
+   "HS50",
+   "HS51",
+   "HS52",
+   "HS53",
+   "HS54",
+   "HS55",
+   "HS56",
+   "HS57",
+   "HS58",
+   "HS59",
+   "HS60",
+   "HS61",
+   "HS62",
+   "HS63",
+   "HS64",
+   "HS65",
+   "HS66",
+   "HS67",
+   "HS68",
+   "HS69",
+   "HS70",
+   "HS71",
+   "HS72",
+   "HS73",
+   "HS74",
+   "HS75",
+   "HS76",
+   "HS77",
+   "HS78",
+   "HS79",
+   "HS80",
+   "HS81",
+   "HS82",
+   "HS83",
+   "HS84",
+   "HS85",
+   "HS86",
+   "HS87",
+   "HS88",
+   "HS89",
+   "HS90",
+   "HS91",
+   "HS92",
+   "HS93",
+   "HS94",
+   "HS95",
+   "HS96",
+   "HS97",
+   "HS98",
+   "HS100",
+   "HS101",
+   "HS102",
+   "HS103",
+   "HS104",
+   "HS105",
+   "HS106",
+   "HS107",
+   "HS108",
+   "HS109",
+   "HS110",
+   "HS111"
+  ],
+  "condition_npz": {
+   "1000": {
+    "raw": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/data/immicro/test_raw_K1000.npz",
+    "xcal": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/splits/hoL_immicro/test_xcal_K1000.npz"
+   },
+   "2000": {
+    "raw": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/data/immicro/test_raw_K2000.npz",
+    "xcal": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/splits/hoL_immicro/test_xcal_K2000.npz"
+   }
+  }
+ },
+ "xcal": {
+  "alpha": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002/splits/hoL_immicro/alpha_xcal.npy",
+  "genes_estimated": 11102,
+  "rule": "cohort-balanced mean CPM of the training cohorts' usable pairs"
+ },
+ "n_train_rows": 283,
+ "n_inner_val_rows": 24,
+ "extra_training_cohorts": [
+  "liu2021"
+ ],
+ "base_split": "ho_immicro"
+}

paper-releases/pbisc-diffusion/updates/20261002T124820Z-crosscohort-liu-default/diffusion/dataset/server_1/model8/plans/paired_finetune_intradataset_20260927/data/liu2021/QC_REPORT.md

@@ -0,0 +1,178 @@
+# Liu 2021 COVID-19 (NIH): paired PBMC bulk and CITE-seq (QC report)
+
+Study: Liu C, Martins AJ, Lau WW et al., Cell 2021 ("Time-resolved systems immunology reveals a late juncture linked to fatal COVID-19"). GEO GSE161918 holds both the bulk RNA-seq and the CITE-seq.
+Built 2026-10-02 on server_1, CPU only. Processing directory: `/home/sj_server_1/gblee/datalake/paired-pbmc-bulk-sc/processing/liu2021_v1_20261002/`.
+
+**Use: extra training data only.** The single cells of this study are in the model's pretraining corpus, so Liu must never be a test cohort.
+
+**The bulk is PBMC.** All 73 bulk GSMs have source "Peripheral Blood Mononuclear Cells" and "Unsorted". GEO protocol: 100,000 PBMCs per sample in Trizol, miRNeasy micro kit, Illumina TruSeq library, STAR, then htseq-count 0.9.1 with the UCSC hg19 annotation. The same bulk RNA-seq gave the SNP genotypes that demuxlet used to assign single cells to subjects.
+
+## Result
+
+**65 pairs from 44 subjects. 58 rows are usable** (identity pass and at least 200 retained cells). One row is one subject at one timepoint.
+
+| group | subjects | pairs | timepoints | usable rows (subjects) | retained cells per sample, median (min to max) |
+|---|---|---|---|---|---|
+| COVID | 33 | 54 | T0 28, T1 19, T2 4, T3 3 | 47 (30) | 2,610 (114 to 8,731) |
+| HC | 11 | 11 | HC 11 | 11 (11) | 3,484 (353 to 7,126) |
+| all | 44 | 65 | | 58 (41) | 2,705 (114 to 8,731); 191,544 cells in total |
+
+- Usable rows by timepoint: T0 24, T1 16, T2 4, T3 3, HC 11.
+- COVID subjects by author severity: Critical 24, Severe 6, Moderate 3. Four are deceased (author `outcome`).
+- A COVID subject has 1 to 4 rows (27 subjects have 1 row). Donor = subject, so `donor_ids` repeat.
+
+**Cell numbers depend on the batch.**
+
+| batch | pairs | usable | retained cells per sample, median (min to max) |
+|---|---|---|---|
+| B1 | 22 | 15 | 315 (114 to 451) |
+| B2 | 20 | 20 | 4,418 (908 to 8,731) |
+| B3 | 23 | 23 | 4,829 (920 to 7,249) |
+
+In batch 1 the authors pooled FACS-sorted B, T, CD14+ monocyte and CD14− innate cells with unsorted PBMC at 1:1:1:1:1 in the same lanes. Only the unsorted fifth matches an unsorted PBMC bulk (see Cells), so batch-1 samples keep only 114 to 451 cells.
+
+**Seven rows go to fold −1** in `folds.csv`. All seven are batch-1 rows with fewer than 200 retained cells:
+
+| row | cells | identity |
+|---|---|---|
+| HGR0000078_T0 | 184 | borderline: own subject rank 4 (bulk to sc) and 1 (sc to bulk) |
+| HGR0000078_T1 | 199 | pass |
+| HGR0000079_T0 | 114 | fail: sex (see below) |
+| HGR0000079_T1 | 141 | fail: sex (see below) |
+| HGR0000083_T0 | 158 | pass |
+| HGR0000093_T1 | 152 | pass |
+| HGR0000110_T0 | 128 | fail: own subject rank 14 and 13 |
+
+The 2026-09-27 metadata pair map listed 45 subjects and 66 pairs; the earlier survey said 46 and about 69. The difference is SHD8 and the controls:
+- SHD8 has a bulk column but no cells in the author object. The authors excluded this donor (lymphoma-like B cells).
+- The technical controls (CHI014 in all three batches, and the HD_ML / HD_VO buffy-coat test donors) are not paired, by the task rule.
+
+## Pairing rule
+
+- **Bulk column to GSM.** The two author count tables (batch 1; batches 2+3) have 23 + 50 = 73 columns, one per bulk GSM. A column matches the GSM whose GEO title names it, within the same table. The match is 1:1 for all 73.
+- **Column to subject and timepoint.** The column name is `<subject>_T<k>`, or `<subject>` for healthy controls (timepoint HC from GEO). Two columns need a note:
+  - `PHGR0000074_T0` has a stray leading "P". GEO titles it HGR0000074 (Timepoint 0), and the HTO table has HGR0000074 at T0. It is paired as HGR0000074_T0 (`id_match` records the correction).
+  - `HGR0000089_T1`: the column name says T1, but the GEO characteristic of its GSM says Timepoint 0. The GEO HTO table, the author sc sample and the author clinical sample ID (`BS-EQ-0027:T1:PBMC`) all say T1. It is paired as T1 and flagged in `id_match`.
+- **sc sample.** This is the author `sample_id` (batch + subject + timepoint). demuxlet gives the subject from SNPs, and the hashtag gives the timepoint.
+  - Check: every cell's (batch, donor, hashtag) was looked up in the GEO HTO/timepoint table. The timepoint agrees for 411,902 of 411,902 cells. In batch 1, the sorted flag agrees for 41,677 of 41,677 cells.
+- **A pair** is a bulk column and an sc sample with the same subject and the same timepoint. All 65 also agree on batch, and on group (GEO label vs author `Class`).
+- `sample_id` = `<subject>_<timepoint>` (e.g. HGR0000083_T0, SHD1_HC). `donor_ids` = subject. `groups` = author `Class` (COVID or HC). `timepoints` = the author label (T0 to T3, or HC), stored as a string. Days since symptom onset is a separate key.
+
+**Not paired:**
+- **Bulk (8 columns).**
+  - CHI-014, CHI014_B2 and CHI014_B3 are technical controls.
+  - HD_ML_bc and HD_VO_bc are the buffy-coat freezing test.
+  - HD_ML_PBMC and HD_VO_PBMC are the same two donors' PBMC. They are excluded by the task rule, although their unsorted sc samples exist (7,350 and 6,684 cells).
+  - SHD8 was excluded by the authors and has no cells.
+- **sc (5 samples).** CHI014 (3 batches), HDML_PBMC and HDVO_PBMC.
+
+These unpaired samples stay in the identity null pool.
+
+## Identity checks
+
+Subject identity is matched by construction here, because demuxlet assigned cells using the bulk-derived genotypes. The checks therefore mainly guard against timepoint (hashtag) and processing errors.
+
+**Sex.** This uses XIST against RPS4Y1 + DDX3Y + UTY + KDM5D, in CPM on the source axis (the Ye 2026 rule). All four Y genes are rows of the bulk table, so bulk and sc use the same score.
+- **Fallback for pooled lanes.** In female sc samples, XIST is only 44 to 120 CPM (5′ capture) and Y is 9 to 33 CPM (ambient RNA from male pool-mates). So 5 female samples came out "ambiguous" at pseudobulk level: AA220907_HC, HGR0000078_T0, HGR0000078_T1, HGR0000144_T0 and HGR0000144_T2.
+- For those 5, the call uses per-cell detection instead (the Imperial ESKD rule, unchanged). Every metadata-female sample (HGR0000079 aside, see below) has a Y UMI in at most 15.1 % of cells; every male sample in at least 41 %.
+- **Result: 63 of 65 pairs agree** across bulk, sc and metadata (52 male, 13 female rows).
+- **HGR0000079 T0 and T1 fail.** Bulk and sc are both clearly male (Y 951 to 980 CPM in bulk and 613 to 972 CPM in sc; XIST about 0), but the author metadata says Female. Because the sc cells were assigned with this subject's bulk SNPs, the metadata sex is the likely error. Both rows have fewer than 200 cells anyway.
+
+**Donor-sensitive expression.** This is the Giroux test.
+- Within each modality, remove the top **k = 5** sample PCs. Then use the 2,000 genes with the highest remaining variance in both modalities, and compute Pearson over all 73 bulk columns × 70 sc samples.
+- chrY genes are left out. Ranks are over subjects, and the controls get one label per person.
+- Own subject top match: **62 of 65 (bulk to sc) and 63 of 65 (sc to bulk)** at k = 5. At k = 3 it is 61 and 64; at k = 10, 60 and 62. The own sample is the top match for 61 of 65.
+- Own-pair r for passing rows: median 0.68, minimum 0.30.
+- The simple Spearman test on the 2,000 highest-mean genes does not separate donors (25 of 65 top-1), as in Ye and Giroux.
+
+**Same timepoint.** There are 38 rows from subjects with more than one paired timepoint.
+- The own timepoint ranks first in 37 of 38 in both directions (residualised), and in 30 and 34 of 38 (centred). No row has a timepoint failure.
+- The interferon score peaks at the same timepoint in both assays for 16 of 17 such subjects. The exception is HGR0000078 (bulk T0, sc T1), whose rows have 184 and 199 cells.
+
+**Verdict:** 61 pass, 3 fail and 1 is borderline (the table above). All four non-passing rows have fewer than 200 cells. `identity_notes` records the per-cell sex calls, the two pairing notes and weaker within-subject timepoint disagreements (10 rows, none a failure; 6 of them are usable rows).
+
+## Cells
+
+**Source.** The cells come from the authors' Seurat object `GSE161918_AllBatches_SeuratObj.rds.gz` (Seurat 3.1.0). It was exported in R without changes: the RNA counts slot plus all 94 metadata columns.
+- The object has 411,902 cells and 32,738 RNA genes. All cells are demuxlet singlets. ADT, HTO and the log-normalised data slot are not used.
+- **Counts are raw UMIs.** Every stored value is a positive integer. Each cell's sum equals `nCount_RNA`, and its stored-entry count equals `nFeature_RNA` (checked in R and again per cell in Python).
+- A spot check of 8 random cells against the CELLxGENE raw counts gives equal values on every shared Ensembl ID. Two cells also have 1 UMI on a gene that CELLxGENE files under a newer Ensembl ID (SOD2; PAXX = C9orf142 in hg19); our cells hold the same UMI under the hg19 ID.
+
+**Why not the corpus copy.** The CELLxGENE files behind the corpus hold 372,081 of the object's cells on a 30,000-gene axis. They include 136,988 sorted-fraction cells and 42,315 HTO doublets or negatives, but no `Sorted` or HTO column, so the unsorted HTO-singlet subset cannot be selected there. The corpus parquet also has no barcodes.
+
+**Cells kept for a sample:**
+- unsorted fraction (`Sorted` = N);
+- HTO singlet (`HTO_classification.global` = Singlet; Doublet and Negative are dropped, as instructed);
+- demuxlet singlet.
+
+Then the ImMicro rule (`flag_conservative_review_pass`: at least 200 genes, at least 500 UMIs, at most 20 % mito) is applied.
+- In the 65 paired samples: 365,732 cells in the object; 135,870 sorted cells dropped; 1,748 HTO doublets and 31,147 HTO negatives dropped. That leaves 196,967 cells, and **191,544 are retained**.
+- The author object was already filtered: at least 201 genes and at most 30 % mito. The author `percent.mito` × 100 equals our mito % to 5e-14.
+- The author cluster labels "Dblt" (3,935 retained cells) and "gated_out" (2,494) are not HTO calls; these cells are kept.
+
+**HTO negatives.** Negatives come from weakly stained pools (B2 HTO3, B3 HTO5). 14 rows lose more than 30 % of their unsorted cells, up to 69 % (HGR0000392_T3).
+- Their author cell-type mix differs from the singlets' (total variation 0.23). Negatives have about 20 percentage points fewer classical monocytes and more CD4 T cells, so the retained cells of those rows over-represent monocytes.
+- Sensitivity check over the 20 rows with at least 50 negatives (`processing/.../identity/HTO_NEGATIVE_SENSITIVITY.json`):
+  - Adding the negatives back leaves the bulk-vs-pseudobulk r unchanged (median 0.918 both ways; it rises in 7 of 20 rows).
+  - Negatives alone match their own timepoint's bulk better than the subject's other timepoints in 10 of 14 rows; singlets do so in 12 of 14.
+- The dataset keeps the task rule (singlets only).
+
+**Other cell facts.**
+- UMIs per retained cell: median 4,613 (source axis) and 4,384 (canonical axis).
+- Chemistry (GEO): 10x 5′ Chromium Next GEM v1.1 CITE-seq; Cell Ranger 3.1.0, hg19-1.2.0 reference.
+- `author_cell_type` = `WCTcoursecelltype` (the label CELLxGENE published). `author_cell_type_fine` = `WCTmergedcelltype`, an extra column.
+
+## Genes
+
+**scRNA**
+- The Seurat rownames are Seurat's version of the Cell Ranger hg19-1.2.0 symbols: make.unique, with "_" replaced by "-". They equal that genes table at all 32,738 positions, so Ensembl IDs were restored by position. The proof is in `processing/.../reference_crosswalk/CROSSWALK.json`; 114 positions are renamed.
+- The table is a GEO copy, already extracted and receipt-checked by the Giroux build.
+- Mapped by ID with the frozen GeneMapper: 18,976 canonical genes are measured. 90 genes whose symbol disagrees with the ID are kept under the ID, as in Giroux and ImMicro. 1,824 hg19 IDs are unknown to the axis.
+
+**Bulk**
+- 55,765 UCSC hg19 symbols, all integer. There are no htseq "__" rows and no duplicate symbols. Mapped by symbol: **19,060 canonical genes are measured**.
+- 10 many-to-one groups were summed (the bundle rule for verified counts). 59 ambiguous symbols were dropped.
+- Unlike Giroux, ribosomal and mitochondrial genes are present.
+- Assigned counts per paired column: median 18.3 M (11.3 to 28.2 M). The batch-1 table has median 13.3 M; batches 2+3 have 19.7 M.
+
+**Overlap.** `pair_valid` has 18,553 genes. On the 6k panel, 5,806 genes are measured in the sc, 5,753 in the bulk and 5,681 in both.
+
+## Bulk vs pseudobulk (sanity)
+
+For each usable row: Pearson r of log1p CPM, bulk against the sc pseudobulk, over the row's `pair_valid` genes (CPM over the same genes). The other datasets were computed the same way (`processing/.../SANITY.json`).
+
+| dataset | usable rows | r, median (min to max) | retained cells per row, median | UMIs per cell, median |
+|---|---|---|---|---|
+| liu2021 | 58 | 0.898 (0.869 to 0.912) | 2,932 | 4,617 |
+| giroux2022 | 24 | 0.899 (0.858 to 0.912) | 10,958 | 3,318 |
+| ye2026_sepsis | 82 | 0.914 (0.827 to 0.935) | 7,154 | 4,044 |
+| imperial_eskd | 157 | 0.875 (0.815 to 0.904) | 2,483 | 2,946 |
+
+By batch, the Liu median r is 0.887 (B1), 0.898 (B2) and 0.902 (B3).
+
+## Processing
+
+This build uses the frozen strict-QC code (bundle v5, copied unchanged and sha-checked) and the same 20,097-gene GENCODE v49 axis.
+- Each author sample was cut from the exported matrix into its own source H5AD and run through the bundle as one library, as for Imperial.
+- Every step ran as a bounded systemd user unit (MemoryMax 40G, no swap, at most 8 CPUs, CUDA hidden), after checking that at least 80 GiB was available.
+
+## Files
+
+| file | contents |
+|---|---|
+| `features.npz` | One row per pair, in bulk GSM order. Core keys follow Giroux. `bulk_raw` is **NaN** where not measured (the task contract and ImMicro; Giroux, Ye and Imperial store 0; `prepare_dataset.py` accepts both). `pseudo_raw_sum` is 0 where not measured. `timepoints` is a string label. Extra keys: `bulk_gsm`, `bulk_count_column`, `bulk_title`, `sc_sample`, `id_match`, `batch`, `geo_bulk_timepoint`, `severity`, `outcome`, `status_meta`, `days_since_onset`, `days_since_hospitalized`, `ever_admitted_to_icu`, `ward`, `sex_meta`, `age_meta`, `below_200_cells`. |
+| `cells_panel6k.npz` | CSR int32: 191,544 × 6,000, 82.8 M nonzero. The panel equals the giroux2022, imperial_eskd, ye2026_sepsis and ji2025_chili panels and `gene_positions.npy` (sha checked). |
+| `cells_full.npz` | The same cells on the full axis: CSR int32, 288.8 M nonzero, 2.3 GB (`cells_full.BUILD.json`). |
+| `cells_obs.parquet` | One row per retained cell. The columns are Giroux's, plus `author_cell_type_fine`. |
+| `samples.csv`, `pairs.csv`, `identity_*.csv` | Row index, pairing fields, cells dropped per step, sex and expression checks, and the verdict for each pair. |
+| `folds.csv`, `conditions_*_fixed.npz`, `fold*/`, `PREPARE.json` | From `tools/prepare_dataset.py --dataset liu2021`: 5 subject-level folds over 41 usable subjects. No subject is split across folds. |
+| `MANIFEST.json` | Sources and code with sha256, all checks, the excluded lists, the sex-fallback details and `post_assembly` (cells_full, prepare run, sanity, HTO sensitivity). |
+
+**Checks passed:**
+- For every row, the summed panel counts equal `pseudo_raw_sum[:, positions]` exactly. The full-axis cells sum to `pseudo_raw_sum`, and their panel columns equal `cells_panel6k`; this was re-checked on the written file.
+- Cell counts equal `qc_cell_counts`. There are no counts outside the masks, and `bulk_raw` is NaN exactly where `bulk_valid` is False.
+- The `gene_positions` sha256 matches.
+
+**Notes for downstream use**
+- `tools/ft_common.py load_features` refuses datasets with more than one sample per donor. As for Giroux and Imperial, Liu needs the multi-sample adapter.
+- The intermediate `processing/.../rds_export/` (4.8 GB of binary counts) can be regenerated from the RDS in about 2 minutes with `code/run_step.sh rds`.