Archive 4 checkpoints (1-4)

paper-releases/20260928/diffusion/data/annotator/aifi_scanvi_f01common_v1/metrics.json

@@ -0,0 +1,503 @@
+{
+  "contract": {
+    "layer": "counts",
+    "labels_key": "aifi_l2_train",
+    "unlabeled_category": "Unknown",
+    "batch_key": "batch_key",
+    "query_train_epochs": 0
+  },
+  "in_sample_l2": {
+    "n": 20000,
+    "accuracy": 0.87645,
+    "macro_f1_present_classes": 0.8696383455603359,
+    "macro_f1_all_labels": 0.8696383455603359,
+    "n_classes_present_in_truth": 29,
+    "per_class": {
+      "ASDC": {
+        "precision": 0.9565217391304348,
+        "recall": 1.0,
+        "f1": 0.9777777777777777,
+        "support": 22
+      },
+      "CD14 monocyte": {
+        "precision": 0.9969135802469136,
+        "recall": 0.8532075471698113,
+        "f1": 0.9194794631964213,
+        "support": 2650
+      },
+      "CD16 monocyte": {
+        "precision": 0.951063829787234,
+        "recall": 0.9738562091503268,
+        "f1": 0.9623250807319699,
+        "support": 459
+      },
+      "CD56bright NK cell": {
+        "precision": 0.8673469387755102,
+        "recall": 0.9883720930232558,
+        "f1": 0.9239130434782609,
+        "support": 258
+      },
+      "CD56dim NK cell": {
+        "precision": 0.9910130718954249,
+        "recall": 0.8919117647058824,
+        "f1": 0.9388544891640866,
+        "support": 1360
+      },
+      "CD8aa": {
+        "precision": 0.6733167082294265,
+        "recall": 0.989010989010989,
+        "f1": 0.8011869436201781,
+        "support": 273
+      },
+      "DN T cell": {
+        "precision": 0.6111111111111112,
+        "recall": 0.9734513274336283,
+        "f1": 0.7508532423208191,
+        "support": 113
+      },
+      "Effector B cell": {
+        "precision": 0.760932944606414,
+        "recall": 0.9190140845070423,
+        "f1": 0.8325358851674641,
+        "support": 284
+      },
+      "Erythrocyte": {
+        "precision": 0.9102564102564102,
+        "recall": 0.9726027397260274,
+        "f1": 0.9403973509933775,
+        "support": 73
+      },
+      "ILC": {
+        "precision": 0.7307692307692307,
+        "recall": 1.0,
+        "f1": 0.8444444444444444,
+        "support": 38
+      },
+      "Intermediate monocyte": {
+        "precision": 0.4312393887945671,
+        "recall": 0.9694656488549618,
+        "f1": 0.5969447708578144,
+        "support": 262
+      },
+      "MAIT": {
+        "precision": 0.7555938037865749,
+        "recall": 0.9690949227373068,
+        "f1": 0.8491295938104448,
+        "support": 453
+      },
+      "Memory B cell": {
+        "precision": 0.9072164948453608,
+        "recall": 0.8349146110056926,
+        "f1": 0.8695652173913043,
+        "support": 527
+      },
+      "Memory CD4 T cell": {
+        "precision": 0.9453662484823958,
+        "recall": 0.7397086763774541,
+        "f1": 0.8299875643986498,
+        "support": 3158
+      },
+      "Memory CD8 T cell": {
+        "precision": 0.9319338422391857,
+        "recall": 0.8102876106194691,
+        "f1": 0.8668639053254438,
+        "support": 1808
+      },
+      "Naive B cell": {
+        "precision": 0.9506666666666667,
+        "recall": 0.7783842794759825,
+        "f1": 0.8559423769507803,
+        "support": 916
+      },
+      "Naive CD4 T cell": {
+        "precision": 0.917295188556567,
+        "recall": 0.9382814578345304,
+        "f1": 0.927669647553919,
+        "support": 3759
+      },
+      "Naive CD8 T cell": {
+        "precision": 0.9371165644171779,
+        "recall": 0.9721559268098647,
+        "f1": 0.9543147208121827,
+        "support": 1257
+      },
+      "Plasma cell": {
+        "precision": 1.0,
+        "recall": 0.989247311827957,
+        "f1": 0.9945945945945946,
+        "support": 93
+      },
+      "Platelet": {
+        "precision": 0.9926739926739927,
+        "recall": 1.0,
+        "f1": 0.9963235294117647,
+        "support": 271
+      },
+      "Progenitor cell": {
+        "precision": 0.9848484848484849,
+        "recall": 0.9848484848484849,
+        "f1": 0.9848484848484849,
+        "support": 66
+      },
+      "Proliferating NK cell": {
+        "precision": 0.5787037037037037,
+        "recall": 0.9398496240601504,
+        "f1": 0.7163323782234957,
+        "support": 133
+      },
+      "Proliferating T cell": {
+        "precision": 0.7704918032786885,
+        "recall": 0.9494949494949495,
+        "f1": 0.8506787330316742,
+        "support": 99
+      },
+      "Transitional B cell": {
+        "precision": 0.5714285714285714,
+        "recall": 0.9156626506024096,
+        "f1": 0.7037037037037037,
+        "support": 249
+      },
+      "Treg": {
+        "precision": 0.6443298969072165,
+        "recall": 0.9715025906735751,
+        "f1": 0.7747933884297521,
+        "support": 386
+      },
+      "cDC1": {
+        "precision": 1.0,
+        "recall": 1.0,
+        "f1": 1.0,
+        "support": 48
+      },
+      "cDC2": {
+        "precision": 0.8079710144927537,
+        "recall": 0.9695652173913043,
+        "f1": 0.8814229249011858,
+        "support": 230
+      },
+      "gdT": {
+        "precision": 0.5710594315245479,
+        "recall": 0.8308270676691729,
+        "f1": 0.6768759571209801,
+        "support": 532
+      },
+      "pDC": {
+        "precision": 1.0,
+        "recall": 0.9955156950672646,
+        "f1": 0.9977528089887641,
+        "support": 223
+      }
+    }
+  },
+  "held_out_donors": {
+    "n_cells": 39690,
+    "donors": [
+      "BR1007",
+      "BR1016",
+      "BR1033",
+      "BR1042",
+      "BR1059",
+      "BR2013",
+      "BR2024",
+      "BR2037",
+      "BR2053",
+      "UP1022",
+      "UP1025"
+    ],
+    "predict_seconds": 0.1,
+    "atlas_L1_vs_map_collapse_agreement": 1.0,
+    "l2": {
+      "n": 39690,
+      "accuracy": 0.8692869740488788,
+      "macro_f1_present_classes": 0.8500925599737216,
+      "macro_f1_all_labels": 0.8500925599737216,
+      "n_classes_present_in_truth": 29,
+      "per_class": {
+        "ASDC": {
+          "precision": 0.9482758620689655,
+          "recall": 0.9482758620689655,
+          "f1": 0.9482758620689655,
+          "support": 58
+        },
+        "CD14 monocyte": {
+          "precision": 0.9984364529819075,
+          "recall": 0.8068592057761733,
+          "f1": 0.8924827792752321,
+          "support": 5540
+        },
+        "CD16 monocyte": {
+          "precision": 0.9467570183930301,
+          "recall": 0.9839034205231388,
+          "f1": 0.9649728663048841,
+          "support": 994
+        },
+        "CD56bright NK cell": {
+          "precision": 0.800718132854578,
+          "recall": 0.975929978118162,
+          "f1": 0.8796844181459567,
+          "support": 457
+        },
+        "CD56dim NK cell": {
+          "precision": 0.9756852343059239,
+          "recall": 0.8567546583850931,
+          "f1": 0.9123604795369987,
+          "support": 2576
+        },
+        "CD8aa": {
+          "precision": 0.4610951008645533,
+          "recall": 0.9580838323353293,
+          "f1": 0.622568093385214,
+          "support": 167
+        },
+        "DN T cell": {
+          "precision": 0.6060606060606061,
+          "recall": 0.9486166007905138,
+          "f1": 0.7395993836671803,
+          "support": 253
+        },
+        "Effector B cell": {
+          "precision": 0.663093415007657,
+          "recall": 0.9232409381663113,
+          "f1": 0.7718360071301248,
+          "support": 469
+        },
+        "Erythrocyte": {
+          "precision": 0.9230769230769231,
+          "recall": 0.9795918367346939,
+          "f1": 0.9504950495049505,
+          "support": 343
+        },
+        "ILC": {
+          "precision": 0.8536585365853658,
+          "recall": 0.9459459459459459,
+          "f1": 0.8974358974358975,
+          "support": 111
+        },
+        "Intermediate monocyte": {
+          "precision": 0.3270758122743682,
+          "recall": 0.9476987447698745,
+          "f1": 0.4863123993558776,
+          "support": 478
+        },
+        "MAIT": {
+          "precision": 0.7421746293245469,
+          "recall": 0.9688172043010753,
+          "f1": 0.8404850746268657,
+          "support": 930
+        },
+        "Memory B cell": {
+          "precision": 0.892485549132948,
+          "recall": 0.7790110998990918,
+          "f1": 0.8318965517241379,
+          "support": 991
+        },
+        "Memory CD4 T cell": {
+          "precision": 0.9367541766109785,
+          "recall": 0.7681017612524462,
+          "f1": 0.8440860215053764,
+          "support": 6132
+        },
+        "Memory CD8 T cell": {
+          "precision": 0.9470023186485591,
+          "recall": 0.8019635343618513,
+          "f1": 0.8684690157958688,
+          "support": 3565
+        },
+        "Naive B cell": {
+          "precision": 0.9472789115646258,
+          "recall": 0.7262059973924381,
+          "f1": 0.822140221402214,
+          "support": 1534
+        },
+        "Naive CD4 T cell": {
+          "precision": 0.9268385621745205,
+          "recall": 0.9384002057613169,
+          "f1": 0.9325835516646431,
+          "support": 7776
+        },
+        "Naive CD8 T cell": {
+          "precision": 0.9565373563218391,
+          "recall": 0.9761730205278593,
+          "f1": 0.966255442670537,
+          "support": 2728
+        },
+        "Plasma cell": {
+          "precision": 1.0,
+          "recall": 1.0,
+          "f1": 1.0,
+          "support": 170
+        },
+        "Platelet": {
+          "precision": 0.9846547314578005,
+          "recall": 0.9897172236503856,
+          "f1": 0.9871794871794872,
+          "support": 389
+        },
+        "Progenitor cell": {
+          "precision": 1.0,
+          "recall": 0.9939759036144579,
+          "f1": 0.9969788519637462,
+          "support": 166
+        },
+        "Proliferating NK cell": {
+          "precision": 0.536036036036036,
+          "recall": 0.918918918918919,
+          "f1": 0.6770981507823614,
+          "support": 259
+        },
+        "Proliferating T cell": {
+          "precision": 0.6543624161073825,
+          "recall": 0.9154929577464789,
+          "f1": 0.7632093933463796,
+          "support": 213
+        },
+        "Transitional B cell": {
+          "precision": 0.5990566037735849,
+          "recall": 0.9304029304029304,
+          "f1": 0.7288378766140603,
+          "support": 546
+        },
+        "Treg": {
+          "precision": 0.6267409470752089,
+          "recall": 0.9401114206128134,
+          "f1": 0.7520891364902507,
+          "support": 718
+        },
+        "cDC1": {
+          "precision": 1.0,
+          "recall": 0.9850746268656716,
+          "f1": 0.9924812030075187,
+          "support": 67
+        },
+        "cDC2": {
+          "precision": 0.8149253731343283,
+          "recall": 0.9891304347826086,
+          "f1": 0.8936170212765957,
+          "support": 552
+        },
+        "gdT": {
+          "precision": 0.5901183020180932,
+          "recall": 0.8514056224899599,
+          "f1": 0.6970817920263049,
+          "support": 996
+        },
+        "pDC": {
+          "precision": 0.9941176470588236,
+          "recall": 0.990234375,
+          "f1": 0.9921722113502935,
+          "support": 512
+        }
+      }
+    },
+    "l1": {
+      "n": 39690,
+      "accuracy": 0.9918367346938776,
+      "macro_f1_present_classes": 0.9715232379009036,
+      "macro_f1_all_labels": 0.9715232379009036,
+      "n_classes_present_in_truth": 9,
+      "per_class": {
+        "B cell": {
+          "precision": 0.9994612068965517,
+          "recall": 1.0,
+          "f1": 0.9997305308542171,
+          "support": 3710
+        },
+        "DC": {
+          "precision": 0.9056748466257669,
+          "recall": 0.9932716568544996,
+          "f1": 0.9474528680304853,
+          "support": 1189
+        },
+        "Erythrocyte": {
+          "precision": 0.9230769230769231,
+          "recall": 0.9795918367346939,
+          "f1": 0.9504950495049505,
+          "support": 343
+        },
+        "ILC": {
+          "precision": 0.8536585365853658,
+          "recall": 0.9459459459459459,
+          "f1": 0.8974358974358975,
+          "support": 111
+        },
+        "Monocyte": {
+          "precision": 0.9991298042059463,
+          "recall": 0.9824586423274386,
+          "f1": 0.990724095779104,
+          "support": 7012
+        },
+        "NK cell": {
+          "precision": 0.9816120134845234,
+          "recall": 0.9729647630619684,
+          "f1": 0.9772692601067887,
+          "support": 3292
+        },
+        "Platelet": {
+          "precision": 0.9846547314578005,
+          "recall": 0.9897172236503856,
+          "f1": 0.9871794871794872,
+          "support": 389
+        },
+        "Progenitor cell": {
+          "precision": 1.0,
+          "recall": 0.9939759036144579,
+          "f1": 0.9969788519637462,
+          "support": 166
+        },
+        "T cell": {
+          "precision": 0.996549226771184,
+          "recall": 0.9963369963369964,
+          "f1": 0.9964431002534557,
+          "support": 23478
+        }
+      }
+    },
+    "pbmc5_plus_other": {
+      "n": 39690,
+      "accuracy": 0.9920886873267826,
+      "macro_f1_present_classes": 0.9812088972148597,
+      "macro_f1_all_labels": 0.9812088972148597,
+      "n_classes_present_in_truth": 6,
+      "per_class": {
+        "T cell": {
+          "precision": 0.996549226771184,
+          "recall": 0.9963369963369964,
+          "f1": 0.9964431002534557,
+          "support": 23478
+        },
+        "B cell": {
+          "precision": 0.9994612068965517,
+          "recall": 1.0,
+          "f1": 0.9997305308542171,
+          "support": 3710
+        },
+        "Monocyte": {
+          "precision": 0.9991298042059463,
+          "recall": 0.9824586423274386,
+          "f1": 0.990724095779104,
+          "support": 7012
+        },
+        "NK cell": {
+          "precision": 0.9816120134845234,
+          "recall": 0.9729647630619684,
+          "f1": 0.9772692601067887,
+          "support": 3292
+        },
+        "DC": {
+          "precision": 0.9056748466257669,
+          "recall": 0.9932716568544996,
+          "f1": 0.9474528680304853,
+          "support": 1189
+        },
+        "Other": {
+          "precision": 0.959731543624161,
+          "recall": 0.9920713577799801,
+          "f1": 0.9756335282651072,
+          "support": 1009
+        }
+      }
+    },
+    "mean_max_prob": 0.9731612739454262,
+    "frac_max_prob_ge_0.5": 0.9962711010330058
+  }
+}

paper-releases/20260928/diffusion/data/annotator/aifi_scanvi_f01common_v1/predict_scanvi.py

@@ -0,0 +1,307 @@
+#!/usr/bin/env python
+"""Annotate query cells with a frozen AIFI scANVI model (zero query-training epochs).
+
+Rebuilt 2026-09-16 to the surviving query-side contract of ``aifi_scanvi_new_blocks_v1``
+(``aifi_recon_annotation_grid.py`` lines 103-128):
+
+    obs["aifi_l2_train"] = "Unknown"   obs["batch_key"] = <one batch per sample>
+    layers["counts"]      = integer counts on the model's gene space
+    SCANVI.prepare_query_anndata -> SCANVI.load_query_data -> predict(soft=True)   (no training)
+
+Input
+-----
+``--query`` is either
+* an ``.npz`` of generated cells (``generated_counts``, ``gene_ids``, ``bulk_index``,
+  ``sample_ids``) -> Poisson-integerised by ``poisson_readout.py`` (seeded), or
+* an ``.h5ad`` (real cells) carrying integer counts in ``layers["counts"]`` or ``X``.
+
+Gene space
+----------
+The model's genes must all be present in the query.  ``prepare_query_anndata`` would pad
+missing genes with zeros; that fabricates measurements, so it is refused unless
+``--allow-missing-genes`` is passed (and then the count is recorded in the manifest).
+Query genes outside the model panel are dropped (recorded).
+
+Output (``--out-dir``)
+----------------------
+``labels.parquet``  -- same schema as the benchmark's representative centroid transfer
+  (``label_transfer_representative_centroid.py`` scheme ``pbmc5``), consumed by
+  ``evaluate_generated_vs_truth.py``::
+
+      sample_id, group, source, cell_index, celltype_l1_pred, label_confidence,
+      aifi_l1, aifi_l2, aifi_l2_confidence            (+ aifi_l1_confidence, npz_row)
+
+  ``celltype_l1_pred`` is the AIFI L1 collapsed to PBMC5; L1 classes outside PBMC5
+  (Platelet, Erythrocyte, Progenitor cell, ILC) become ``--other-label`` (default
+  ``Other``) -- kept, never dropped.  ``aifi_l1`` keeps the full L1 name.
+  ``label_confidence`` = summed L2 probability mass of the predicted L1;
+  ``aifi_l2_confidence`` = max L2 probability.
+``labels.manifest.json`` -- provenance for the evaluator (``target_truth_used_for_fitting``
+  = false, ``training_source`` = reference h5ad, ``labels`` = PBMC5 + Other so
+  ``--labels-from-manifest`` yields the 6-way scheme).
+``soft_probabilities_l2.parquet`` / ``probabilities_l1.parquet`` -- full soft outputs.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import sys
+import time
+from datetime import datetime, timezone
+from importlib.metadata import version as _pkg_version
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+import scipy.sparse as sp
+
+PLAN_DIR = Path(__file__).resolve().parent
+sys.path.insert(0, str(PLAN_DIR))
+import scanvi_common as C  # noqa: E402
+
+UNSPEC = "__unspecified__"
+
+
+def parse_args(argv=None) -> argparse.Namespace:
+    p = argparse.ArgumentParser(description="Frozen scANVI annotation of query cells.", formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+    p.add_argument("--model-dir", required=True, type=Path, help=".../scanvi_model directory (contains model.pt)")
+    p.add_argument("--query", required=True, type=Path, help=".npz (generated) or .h5ad (real) query cells")
+    p.add_argument("--out-dir", required=True, type=Path)
+    p.add_argument("--device", choices=["cpu", "cuda"], default="cpu")
+    p.add_argument("--seed", type=int, default=20260916, help="Poisson readout seed (npz input)")
+    p.add_argument("--npz-matrix-key", default="generated_counts")
+    p.add_argument("--library-scale", type=float, default=1.0, help="Poisson rate multiplier; 1.0 = no rescaling")
+    p.add_argument("--max-cells", type=int, default=None, help="SMOKE: seeded random subset of query cells (output is then NOT evaluator-aligned)")
+    p.add_argument("--subset-seed", type=int, default=0)
+    p.add_argument("--method-name", default=None, help="obs['source'] / manifest method (default: query stem)")
+    p.add_argument("--dataset-id", default="")
+    p.add_argument("--split", default="")
+    p.add_argument("--batch-prefix", default="", help="batch_key = prefix + sample_id (e.g. '004_A_GSE135779_SLE_PBMC_')")
+    p.add_argument("--sample-id-col", default=None, help="h5ad: obs column with sample ids (default: sample_id, sample_ids, donor_id)")
+    p.add_argument("--batch-key-col", default=None, help="h5ad: obs column to use as batch_key (default: prefix + sample id)")
+    p.add_argument("--other-label", default=C.OTHER_LABEL)
+    p.add_argument("--confidence-threshold", type=float, default=0.0, help="cells with max L2 prob below this get aifi_l2='__unspecified__' and celltype_l1_pred=other-label")
+    p.add_argument("--allow-missing-genes", action="store_true", help="permit zero-padding of model genes absent from the query (recorded)")
+    p.add_argument("--eval-batch-size", type=int, default=2048)
+    p.add_argument("--num-threads", type=int, default=8)
+    p.add_argument("--model-manifest", type=Path, default=None, help="MODEL_MANIFEST.json (default: <model-dir>/../MODEL_MANIFEST.json)")
+    return p.parse_args(argv)
+
+
+def _load_h5ad_query(path: Path, *, sample_id_col: str | None, batch_key_col: str | None, batch_prefix: str, method: str):
+    import anndata as ad
+
+    adata = ad.read_h5ad(path)
+    if "counts" in adata.layers:
+        counts = adata.layers["counts"]
+    else:
+        counts = adata.X
+    counts = sp.csr_matrix(counts) if not sp.issparse(counts) else counts.tocsr()
+    if counts.nnz and not np.all(counts.data == np.round(counts.data)):
+        raise SystemExit(f"{path}: counts are not integer-valued; scANVI needs integer counts (Poisson-integerise generated cells first)")
+    counts = counts.astype(np.float32)
+    adata.layers["counts"] = counts
+    adata.X = counts.copy()
+    sid_col = sample_id_col or next((c for c in ("sample_id", "sample_ids", "donor_id") if c in adata.obs.columns), None)
+    if sid_col is None:
+        raise SystemExit(f"{path}: obs lacks a sample id column (pass --sample-id-col)")
+    adata.obs["sample_id"] = adata.obs[sid_col].astype(str).to_numpy()
+    if batch_key_col is not None:
+        adata.obs["batch_key"] = pd.Categorical(adata.obs[batch_key_col].astype(str))
+    else:
+        adata.obs["batch_key"] = pd.Categorical([f"{batch_prefix}{s}" for s in adata.obs["sample_id"]])
+    adata.obs["aifi_l2_train"] = pd.Categorical([C.UNLABELED_CATEGORY] * adata.n_obs)
+    adata.obs["npz_row"] = np.arange(adata.n_obs, dtype=np.int64)
+    adata.obs["cell_index"] = np.arange(adata.n_obs, dtype=np.int64)
+    if "source" not in adata.obs.columns:
+        adata.obs["source"] = method
+    query_input = {"kind": "h5ad", "path": str(path), "counts_from": "layers['counts']" if "counts" in adata.layers else "X", "n_input_cells": int(adata.n_obs)}
+    # Carry provenance written by build_benchmark_queries.py (Poisson readout seed, source, join rule).
+    for key in ("poisson_readout", "query_build"):
+        if key in adata.uns:
+            query_input[key] = _plain(adata.uns[key])
+    adata.uns["query_input"] = query_input
+    return adata
+
+
+def _plain(obj):
+    """h5ad-loaded uns -> plain JSON-able Python."""
+    if isinstance(obj, dict):
+        return {str(k): _plain(v) for k, v in obj.items()}
+    if isinstance(obj, (list, tuple)):
+        return [_plain(v) for v in obj]
+    if isinstance(obj, np.ndarray):
+        return obj.tolist()
+    if isinstance(obj, np.generic):
+        return obj.item()
+    return obj
+
+
+def main(argv=None) -> int:
+    args = parse_args(argv)
+    if args.device == "cpu":
+        os.environ["CUDA_VISIBLE_DEVICES"] = ""
+    import anndata as ad  # noqa: E402
+    import scvi  # noqa: E402
+    import torch  # noqa: E402
+    from scvi.model import SCANVI  # noqa: E402
+
+    import poisson_readout as PR  # noqa: E402
+
+    torch.set_num_threads(int(args.num_threads))
+    if args.device == "cuda" and not torch.cuda.is_available():
+        raise SystemExit("cuda requested but not available")
+    accelerator = "gpu" if args.device == "cuda" else "cpu"
+    scvi.settings.seed = int(args.seed)
+    t_start = time.monotonic()
+    out_dir: Path = args.out_dir
+    out_dir.mkdir(parents=True, exist_ok=True)
+    model_dir: Path = args.model_dir
+    if not (model_dir / "model.pt").exists():
+        raise SystemExit(f"{model_dir} has no model.pt")
+    method = args.method_name or args.query.stem
+    l2_to_l1 = C.load_l2_to_l1()
+
+    # ---- query -------------------------------------------------------------
+    suffix = args.query.suffix.lower()
+    readout_info = None
+    if suffix == ".npz":
+        adata = PR.readout_npz(
+            args.query,
+            seed=args.seed,
+            matrix_key=args.npz_matrix_key,
+            max_cells=args.max_cells,
+            subset_seed=args.subset_seed,
+            batch_prefix=args.batch_prefix,
+            method_name=method,
+            library_scale=args.library_scale,
+        )
+        readout_info = dict(adata.uns["poisson_readout"])
+        adata.obs["group"] = "unknown"
+    elif suffix == ".h5ad":
+        adata = _load_h5ad_query(args.query, sample_id_col=args.sample_id_col, batch_key_col=args.batch_key_col, batch_prefix=args.batch_prefix, method=method)
+        if args.max_cells is not None and args.max_cells < adata.n_obs:
+            rows = np.sort(np.random.default_rng(args.subset_seed).choice(adata.n_obs, size=args.max_cells, replace=False))
+            adata = adata[rows].copy()
+            adata.obs["npz_row"] = rows
+            adata.obs["cell_index"] = np.arange(adata.n_obs, dtype=np.int64)
+        adata.obs["group"] = adata.obs["condition"].astype(str).to_numpy() if "condition" in adata.obs.columns else "unknown"
+        readout_info = adata.uns["query_input"].get("poisson_readout")
+    else:
+        raise SystemExit(f"unsupported query format {suffix!r}")
+    query_build_info = adata.uns.get("query_input", {}).get("query_build") if suffix == ".h5ad" else None
+    n_query_genes = adata.n_vars
+    n_input_cells = int(readout_info["input_n_cells"]) if readout_info else int(adata.uns.get("query_input", {}).get("n_input_cells", adata.n_obs))
+    subset = adata.n_obs < n_input_cells
+
+    # ---- gene-space guard (no fabricated zeros) -------------------------------
+    ref_vars = SCANVI.prepare_query_anndata(None, str(model_dir), return_reference_var_names=True)
+    ref_set, q_set = set(map(str, ref_vars)), set(map(str, adata.var_names))
+    missing = sorted(ref_set - q_set)
+    dropped = len(q_set - ref_set)
+    if missing and not args.allow_missing_genes:
+        raise SystemExit(
+            f"{len(missing)} of {len(ref_set)} model genes are absent from the query "
+            f"(e.g. {missing[:5]}); prepare_query_anndata would zero-pad them. Use a model "
+            "trained on a panel contained in the query, or pass --allow-missing-genes."
+        )
+    SCANVI.prepare_query_anndata(adata, str(model_dir))
+    assert list(map(str, adata.var_names)) == list(map(str, ref_vars))
+
+    # ---- frozen prediction ----------------------------------------------------
+    query = SCANVI.load_query_data(adata, str(model_dir), accelerator=accelerator)
+    t0 = time.monotonic()
+    soft = query.predict(soft=True, batch_size=int(args.eval_batch_size))
+    predict_seconds = time.monotonic() - t0
+    pred_l2, conf_l2, probs_l2 = C.soft_to_predictions(soft)
+    l1_mass = C.l1_probability_mass(probs_l2, l2_to_l1)
+    pred_l1 = C.collapse_l1(pred_l2, l2_to_l1)
+    conf_l1 = l1_mass.to_numpy()[np.arange(len(pred_l1)), [list(l1_mass.columns).index(x) for x in pred_l1]]
+    pred_p5 = C.collapse_pbmc5(pred_l1, other=args.other_label)
+    if args.confidence_threshold > 0:
+        low = conf_l2 < args.confidence_threshold
+        pred_l2 = np.where(low, UNSPEC, pred_l2).astype(object)
+        pred_l1 = np.where(low, UNSPEC, pred_l1).astype(object)
+        pred_p5 = np.where(low, args.other_label, pred_p5).astype(object)
+
+    # ---- outputs --------------------------------------------------------------
+    sample_ids = adata.obs["sample_id"].astype(str).to_numpy()
+    frame = pd.DataFrame(
+        {
+            "sample_id": sample_ids.astype(object),
+            "group": adata.obs["group"].astype(str).to_numpy().astype(object),
+            "source": adata.obs["source"].astype(str).to_numpy().astype(object),
+            "cell_index": adata.obs["cell_index"].to_numpy().astype(np.int64),
+            "celltype_l1_pred": pred_p5,
+            "label_confidence": conf_l1.astype(np.float32),
+            "aifi_l1": pred_l1,
+            "aifi_l2": pred_l2,
+            "aifi_l2_confidence": conf_l2.astype(np.float32),
+            "aifi_l1_confidence": conf_l1.astype(np.float32),
+            "npz_row": adata.obs["npz_row"].to_numpy().astype(np.int64),
+        }
+    )
+    frame.to_parquet(out_dir / "labels.parquet", index=False)
+    probs_l2.astype(np.float32).reset_index(names="obs_name").to_parquet(out_dir / "soft_probabilities_l2.parquet", index=False)
+    l1_mass.astype(np.float32).reset_index(names="obs_name").to_parquet(out_dir / "probabilities_l1.parquet", index=False)
+
+    model_manifest_path = args.model_manifest or (model_dir.parent / "MODEL_MANIFEST.json")
+    model_manifest = json.loads(model_manifest_path.read_text()) if model_manifest_path.exists() else {}
+    training_source = model_manifest.get("reference", {}).get("h5ad")
+    labels_6 = list(C.PBMC5_LABELS) + [args.other_label]
+    payload = {
+        "dataset_id": args.dataset_id,
+        "split": args.split,
+        "input_h5ad": str(args.query),
+        "method": method,
+        "out_parquet": str(out_dir / "labels.parquet"),
+        "training_source": training_source,
+        "target_truth_used_for_fitting": False,
+        "classifier": f"aifi_scanvi_rebuild_20260916_frozen_predict[{model_manifest.get('run_name', model_dir.parent.name)}]",
+        "labels": labels_6,
+        "l2_labels": [str(c) for c in probs_l2.columns],
+        "n_cells": int(adata.n_obs),
+        "n_genes": int(adata.n_vars),
+        "note": "Evaluation-only annotation with a frozen AIFI scANVI (zero query-training epochs); L1 outside PBMC5 -> Other, never dropped.",
+        "label_source_column": "AIFI_L2 (reference) -> AIFI_L2_TO_L1 -> PBMC5 + Other",
+        "query_train_epochs": 0,
+        "confidence_threshold": float(args.confidence_threshold),
+        "subset": bool(subset),
+        "evaluator_aligned": not bool(subset),
+        "gene_space": {
+            "join_key": "ensembl_id",
+            "model_genes": int(len(ref_set)),
+            "query_genes": int(n_query_genes),
+            "model_genes_missing_from_query_zero_padded": int(len(missing)),
+            "query_genes_outside_model_dropped": int(dropped),
+            "rule": "query is subset/reordered to the model's gene list; zero-padding of absent model genes is refused unless --allow-missing-genes",
+        },
+        "readout": readout_info,
+        "query_build": query_build_info,
+        "model": {
+            "scanvi_model_dir": str(model_dir),
+            "model_pt_sha256": C.sha256_file(model_dir / "model.pt"),
+            "model_manifest": str(model_manifest_path) if model_manifest_path.exists() else None,
+            "held_out_summary": model_manifest.get("held_out_summary"),
+        },
+        "label_counts": {
+            "celltype_l1_pred": {k: int(v) for k, v in pd.Series(pred_p5).value_counts().items()},
+            "aifi_l1": {k: int(v) for k, v in pd.Series(pred_l1).value_counts().items()},
+            "aifi_l2": {k: int(v) for k, v in pd.Series(pred_l2).value_counts().items()},
+        },
+        "confidence": {"mean_l2_max_prob": float(conf_l2.mean()), "mean_l1_mass": float(conf_l1.mean()), "frac_l2_max_prob_ge_0.5": float(np.mean(conf_l2 >= 0.5))},
+        "timings_s": {"predict": round(predict_seconds, 1), "total": round(time.monotonic() - t_start, 1)},
+        "versions": {"scvi-tools": scvi.__version__, "torch": torch.__version__, "anndata": _pkg_version("anndata")},
+        "device": args.device,
+        "created_at_utc": datetime.now(timezone.utc).isoformat(),
+    }
+    C.write_json(out_dir / "labels.manifest.json", payload)
+    print(json.dumps({k: payload[k] for k in ("method", "n_cells", "n_genes", "label_counts", "confidence", "timings_s", "subset")}, indent=2))
+    print(f"-> {out_dir / 'labels.parquet'}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

paper-releases/20260928/diffusion/data/annotator/aifi_scanvi_f01common_v1/reference/REFERENCE_MANIFEST.json

@@ -0,0 +1,572 @@
+{
+  "schema_version": "aifi_scanvi_rebuild.reference_manifest.v1",
+  "built_at_utc": "2026-09-15T16:05:57.806936+00:00",
+  "atlas": {
+    "path": "/home/sj_server_1/gblee/datalake/single-cell-RNA-seq-datasets/PBMC-cells/009_A_Human_Immune_Health_Atlas_Human_Immune_Health_Atlas_AIFI_immune_dynamics_across_healthy_age/39832a61-3df1-4f1f-971d-3c2a74539047.h5ad",
+    "size_bytes": 10162450774,
+    "mtime": "2025-10-30T00:12:32+00:00",
+    "n_cells": 1821725,
+    "n_genes": 32357
+  },
+  "panel": {
+    "path": "/home/sj_server_1/PBISC/PBISC-data/single-cell-data-block-blood/metadata/var.parquet",
+    "n_genes": 20097
+  },
+  "gene_match": {
+    "canonical_20097": {
+      "matched": 19642,
+      "unmatched": 455
+    },
+    "tier_6k": {
+      "n": 6000,
+      "matched": 5944,
+      "unmatched": 56
+    },
+    "tier_2k": {
+      "n": 2000,
+      "matched": 1984
+    },
+    "tier_broad": {
+      "n": 11589,
+      "matched": 11517
+    },
+    "rule": "join on Ensembl ID without version; panel genes absent from the atlas are absent from the reference (no zero columns)",
+    "intersection_2980": {
+      "n": 2980,
+      "matched": 2977,
+      "unmatched": 3,
+      "absent_gene_ids": [
+        "ENSG00000233276",
+        "ENSG00000168255",
+        "ENSG00000158427"
+      ],
+      "list_source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/analysis/scsemi_benchmark_diffusion_20260913/metrics/panel_6k_e4_2980genes/pbmc5/evaluation_manifest.json",
+      "list_sha256": "f40ee3fd60a0b75ba35cd79d76837eecc1c7698a245b810e8cf28ab1dd3b447a",
+      "rule": "verbatim 2,980-gene intersection (6k tier x benchmark working panel); join on Ensembl ID; absent genes are absent"
+    },
+    "f01_common": {
+      "n": 1006,
+      "matched": 1005,
+      "unmatched": 1,
+      "absent_gene_ids": [
+        "ENSG00000233276"
+      ],
+      "list_source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/floor01_campaign_20260923/evaluation_manifest_common.json",
+      "list_sha256": "0749a7af9eb20bf301f63e9dc5dc6735782248823e6f732c643584c89c5af5c1",
+      "rule": "verbatim 2,980-gene intersection (6k tier x benchmark working panel); join on Ensembl ID; absent genes are absent"
+    }
+  },
+  "sampling": {
+    "seed": 20260916,
+    "target_cells": 400000,
+    "min_per_class": 5000,
+    "proportional_rate": 0.20016128631272764,
+    "strata": "AIFI_L2 x donor_id, largest-remainder allocation within class, uniform without replacement within stratum",
+    "n_selected": 400000,
+    "dev_row_limit": null
+  },
+  "donor_split": {
+    "val_donor_frac": 0.1,
+    "stratified_by": "cohort",
+    "n_train_donors": 97,
+    "n_val_donors": 11,
+    "train_donors": [
+      "BR1001",
+      "BR1002",
+      "BR1003",
+      "BR1004",
+      "BR1005",
+      "BR1006",
+      "BR1008",
+      "BR1009",
+      "BR1010",
+      "BR1011",
+      "BR1012",
+      "BR1013",
+      "BR1014",
+      "BR1015",
+      "BR1017",
+      "BR1018",
+      "BR1019",
+      "BR1021",
+      "BR1023",
+      "BR1024",
+      "BR1025",
+      "BR1026",
+      "BR1028",
+      "BR1030",
+      "BR1031",
+      "BR1032",
+      "BR1037",
+      "BR1040",
+      "BR1041",
+      "BR1043",
+      "BR1044",
+      "BR1045",
+      "BR1047",
+      "BR1048",
+      "BR1049",
+      "BR1050",
+      "BR1051",
+      "BR1052",
+      "BR1054",
+      "BR1056",
+      "BR1057",
+      "BR1058",
+      "BR2002",
+      "BR2003",
+      "BR2005",
+      "BR2008",
+      "BR2009",
+      "BR2010",
+      "BR2012",
+      "BR2014",
+      "BR2015",
+      "BR2016",
+      "BR2018",
+      "BR2019",
+      "BR2020",
+      "BR2021",
+      "BR2022",
+      "BR2023",
+      "BR2025",
+      "BR2026",
+      "BR2028",
+      "BR2029",
+      "BR2030",
+      "BR2031",
+      "BR2032",
+      "BR2033",
+      "BR2034",
+      "BR2035",
+      "BR2036",
+      "BR2038",
+      "BR2039",
+      "BR2040",
+      "BR2041",
+      "BR2042",
+      "BR2043",
+      "BR2044",
+      "BR2045",
+      "BR2046",
+      "BR2047",
+      "BR2048",
+      "BR2050",
+      "BR2051",
+      "BR2052",
+      "UP1008",
+      "UP1011",
+      "UP1012",
+      "UP1014",
+      "UP1015",
+      "UP1017",
+      "UP1018",
+      "UP1021",
+      "UP1023",
+      "UP1024",
+      "UP1026",
+      "UP1031",
+      "UP1032",
+      "UP1034"
+    ],
+    "val_donors": [
+      "BR1007",
+      "BR1016",
+      "BR1033",
+      "BR1042",
+      "BR1059",
+      "BR2013",
+      "BR2024",
+      "BR2037",
+      "BR2053",
+      "UP1022",
+      "UP1025"
+    ]
+  },
+  "cells_by_split": {
+    "train": 360310,
+    "val": 39690
+  },
+  "classes_AIFI_L2": {
+    "Naive CD4 T cell": {
+      "train": 67899,
+      "val": 7776,
+      "total": 75675
+    },
+    "Memory CD4 T cell": {
+      "train": 58278,
+      "val": 6132,
+      "total": 64410
+    },
+    "CD14 monocyte": {
+      "train": 48369,
+      "val": 5540,
+      "total": 53909
+    },
+    "Memory CD8 T cell": {
+      "train": 33084,
+      "val": 3565,
+      "total": 36649
+    },
+    "CD56dim NK cell": {
+      "train": 24222,
+      "val": 2576,
+      "total": 26798
+    },
+    "Naive CD8 T cell": {
+      "train": 21525,
+      "val": 2728,
+      "total": 24253
+    },
+    "Naive B cell": {
+      "train": 15822,
+      "val": 1534,
+      "total": 17356
+    },
+    "gdT": {
+      "train": 9130,
+      "val": 996,
+      "total": 10126
+    },
+    "Memory B cell": {
+      "train": 8594,
+      "val": 991,
+      "total": 9585
+    },
+    "MAIT": {
+      "train": 8306,
+      "val": 930,
+      "total": 9236
+    },
+    "CD16 monocyte": {
+      "train": 8197,
+      "val": 994,
+      "total": 9191
+    },
+    "Treg": {
+      "train": 7106,
+      "val": 718,
+      "total": 7824
+    },
+    "cDC2": {
+      "train": 4448,
+      "val": 552,
+      "total": 5000
+    },
+    "Transitional B cell": {
+      "train": 4454,
+      "val": 546,
+      "total": 5000
+    },
+    "Platelet": {
+      "train": 4611,
+      "val": 389,
+      "total": 5000
+    },
+    "pDC": {
+      "train": 4488,
+      "val": 512,
+      "total": 5000
+    },
+    "Intermediate monocyte": {
+      "train": 4522,
+      "val": 478,
+      "total": 5000
+    },
+    "Effector B cell": {
+      "train": 4531,
+      "val": 469,
+      "total": 5000
+    },
+    "CD8aa": {
+      "train": 4833,
+      "val": 167,
+      "total": 5000
+    },
+    "CD56bright NK cell": {
+      "train": 4543,
+      "val": 457,
+      "total": 5000
+    },
+    "Proliferating NK cell": {
+      "train": 2566,
+      "val": 259,
+      "total": 2825
+    },
+    "DN T cell": {
+      "train": 2096,
+      "val": 253,
+      "total": 2349
+    },
+    "Proliferating T cell": {
+      "train": 2107,
+      "val": 213,
+      "total": 2320
+    },
+    "Plasma cell": {
+      "train": 1981,
+      "val": 170,
+      "total": 2151
+    },
+    "Progenitor cell": {
+      "train": 1360,
+      "val": 166,
+      "total": 1526
+    },
+    "Erythrocyte": {
+      "train": 1165,
+      "val": 343,
+      "total": 1508
+    },
+    "cDC1": {
+      "train": 876,
+      "val": 67,
+      "total": 943
+    },
+    "ILC": {
+      "train": 733,
+      "val": 111,
+      "total": 844
+    },
+    "ASDC": {
+      "train": 464,
+      "val": 58,
+      "total": 522
+    }
+  },
+  "classes_AIFI_L1": {
+    "T cell": {
+      "train": 214364,
+      "val": 23478,
+      "total": 237842
+    },
+    "Monocyte": {
+      "train": 61088,
+      "val": 7012,
+      "total": 68100
+    },
+    "B cell": {
+      "train": 35382,
+      "val": 3710,
+      "total": 39092
+    },
+    "NK cell": {
+      "train": 31331,
+      "val": 3292,
+      "total": 34623
+    },
+    "DC": {
+      "train": 10276,
+      "val": 1189,
+      "total": 11465
+    },
+    "Platelet": {
+      "train": 4611,
+      "val": 389,
+      "total": 5000
+    },
+    "Progenitor cell": {
+      "train": 1360,
+      "val": 166,
+      "total": 1526
+    },
+    "Erythrocyte": {
+      "train": 1165,
+      "val": 343,
+      "total": 1508
+    },
+    "ILC": {
+      "train": 733,
+      "val": 111,
+      "total": 844
+    }
+  },
+  "obs_columns": [
+    "AIFI_L1",
+    "AIFI_L2",
+    "AIFI_L3",
+    "donor_id",
+    "batch_id",
+    "cohort",
+    "age_group",
+    "sex",
+    "cmv",
+    "n_umis_atlas",
+    "n_genes_atlas",
+    "atlas_row",
+    "split",
+    "batch_key",
+    "aifi_l2_train"
+  ],
+  "outputs": {
+    "full": {
+      "path": "reference/aifi_reference_20097.h5ad",
+      "n_cells": 400000,
+      "n_genes": 19642,
+      "nnz": 658152192,
+      "size_bytes": 5307926173,
+      "sha256": "e80b6047a0983a47fa8fca3c6d44c66d425ea647fe519fd1c7b404869e5459cb",
+      "library_size_median": 4560.0,
+      "library_size_mean": 5204.40234375,
+      "integer_valued_fraction": 1.0
+    },
+    "tier": {
+      "path": "reference/aifi_reference_6k.h5ad",
+      "n_cells": 400000,
+      "n_genes": 5944,
+      "nnz": 240804030,
+      "size_bytes": 1967088729,
+      "sha256": "60683789c961e7e3b8241d004f6d881fc6b42b3393aab8e1fd623e58884e082f",
+      "library_size_median": 2885.0,
+      "library_size_mean": 3206.773681640625,
+      "integer_valued_fraction": 1.0
+    },
+    "intersection_2980": {
+      "path": "reference/aifi_reference_2980.h5ad",
+      "n_cells": 400000,
+      "n_genes": 2977,
+      "nnz": 133667869,
+      "size_bytes": 1107879357,
+      "sha256": "e7c8a45724e978a5f32112a4bdb0db1f76821d883d9f1b6ff444c55ddf00b39d",
+      "library_size_median": 1361.0,
+      "library_size_mean": 1614.7176513671875,
+      "integer_valued_fraction": 1.0
+    },
+    "f01_common": {
+      "path": "reference/aifi_reference_f01common.h5ad",
+      "n_cells": 400000,
+      "n_genes": 1005,
+      "nnz": 40618563,
+      "size_bytes": 362998569,
+      "sha256": "e90de26e48ec905215e6f33057fb4f57d0cecc7f4c820b6bbf26497b0e45e1e1",
+      "library_size_median": 537.0,
+      "library_size_mean": 677.409423828125,
+      "integer_valued_fraction": 1.0
+    }
+  },
+  "side_files": {
+    "panel_presence_20097": "reference/panel_presence_20097.csv",
+    "panel_presence_6k": "reference/panel_presence_6k.csv",
+    "donor_split": "reference/donor_split.csv",
+    "stratum_allocation": "reference/stratum_allocation.csv",
+    "cell_selection": "reference/cell_selection.parquet",
+    "log": "reference/build_reference.log",
+    "panel_presence_2980": "reference/panel_presence_2980.csv",
+    "gene_list_2980_json": "queries_2980/gene_list_2980.json",
+    "build_reference_2980_log": "reference/build_reference_2980.log"
+  },
+  "environment": {
+    "python": "3.12.14",
+    "argv": [
+      "build_reference.py",
+      "--out-dir",
+      "reference"
+    ]
+  },
+  "derived_builds": [
+    {
+      "name": "intersection_2980",
+      "built_at_utc": "2026-09-15T16:53:20.289987+00:00",
+      "script": "build_reference_2980.py",
+      "derived_from": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/aifi_scanvi_rebuild_20260916/reference/aifi_reference_6k.h5ad",
+      "derived_from_sha256": "60683789c961e7e3b8241d004f6d881fc6b42b3393aab8e1fd623e58884e082f",
+      "cells": "identical rows, order, obs and counts to the source (verified against cell_selection.parquet / donor_split.csv)",
+      "cell_check": {
+        "rows": 400000,
+        "obs_columns_equal": [
+          "atlas_row",
+          "split",
+          "donor_id",
+          "batch_id",
+          "batch_key",
+          "aifi_l2_train",
+          "AIFI_L1",
+          "AIFI_L2",
+          "AIFI_L3",
+          "cohort"
+        ],
+        "cells_by_split": {
+          "train": 360310,
+          "val": 39690
+        },
+        "val_donors": [
+          "BR1007",
+          "BR1016",
+          "BR1033",
+          "BR1042",
+          "BR1059",
+          "BR2013",
+          "BR2024",
+          "BR2037",
+          "BR2053",
+          "UP1022",
+          "UP1025"
+        ]
+      },
+      "gene_list_source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/analysis/scsemi_benchmark_diffusion_20260913/metrics/panel_6k_e4_2980genes/pbmc5/evaluation_manifest.json",
+      "n_genes": 2977,
+      "output": "reference/aifi_reference_2980.h5ad",
+      "argv": [
+        "build_reference_2980.py"
+      ]
+    },
+    {
+      "name": "f01_common",
+      "built_at_utc": "2026-09-23T17:24:11.636292+00:00",
+      "script": "build_reference_2980.py",
+      "derived_from": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/aifi_scanvi_rebuild_20260916/reference/aifi_reference_6k.h5ad",
+      "derived_from_sha256": "60683789c961e7e3b8241d004f6d881fc6b42b3393aab8e1fd623e58884e082f",
+      "cells": "identical rows, order, obs and counts to the source (verified against cell_selection.parquet / donor_split.csv)",
+      "cell_check": {
+        "rows": 400000,
+        "obs_columns_equal": [
+          "atlas_row",
+          "split",
+          "donor_id",
+          "batch_id",
+          "batch_key",
+          "aifi_l2_train",
+          "AIFI_L1",
+          "AIFI_L2",
+          "AIFI_L3",
+          "cohort"
+        ],
+        "cells_by_split": {
+          "train": 360310,
+          "val": 39690
+        },
+        "val_donors": [
+          "BR1007",
+          "BR1016",
+          "BR1033",
+          "BR1042",
+          "BR1059",
+          "BR2013",
+          "BR2024",
+          "BR2037",
+          "BR2053",
+          "UP1022",
+          "UP1025"
+        ]
+      },
+      "gene_list_source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/floor01_campaign_20260923/evaluation_manifest_common.json",
+      "n_genes": 1005,
+      "output": "reference/aifi_reference_f01common.h5ad",
+      "argv": [
+        "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/aifi_scanvi_rebuild_20260916/build_reference_2980.py",
+        "--gene-list-source",
+        "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/floor01_campaign_20260923/evaluation_manifest_common.json",
+        "--expected-n",
+        "1006",
+        "--gene-list-json",
+        "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/floor01_campaign_20260923/queries/gene_list_common.json",
+        "--gene-list-txt",
+        "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/floor01_campaign_20260923/queries/gene_list_common.txt",
+        "--name",
+        "aifi_reference_f01common.h5ad",
+        "--manifest-key",
+        "f01_common"
+      ]
+    }
+  ]
+}

paper-releases/20260928/diffusion/data/annotator/aifi_scanvi_f01common_v1/reference/donor_split.csv

@@ -0,0 +1,109 @@
+donor_id,cohort,batch_id,n_cells,split
+BR1001,BR1,B001,18527,train
+BR1002,BR1,B001,18957,train
+BR1003,BR1,B001,19002,train
+BR1004,BR1,B002,17350,train
+BR1005,BR1,B002,17688,train
+BR1006,BR1,B078,17318,train
+BR1007,BR1,B002,17096,val
+BR1008,BR1,B002,15905,train
+BR1009,BR1,B078,17723,train
+BR1010,BR1,B002,15060,train
+BR1011,BR1,B078,16031,train
+BR1012,BR1,B078,15153,train
+BR1013,BR1,B078,20880,train
+BR1014,BR1,B078,18561,train
+BR1015,BR1,B041,15071,train
+BR1016,BR1,B040,14992,val
+BR1017,BR1,B041,15909,train
+BR1018,BR1,B007,16299,train
+BR1019,BR1,B039,19495,train
+BR1021,BR1,B036,13097,train
+BR1023,BR1,B040,15076,train
+BR1024,BR1,B040,16748,train
+BR1025,BR1,B040,16667,train
+BR1026,BR1,B047,19307,train
+BR1028,BR1,B039,17840,train
+BR1030,BR1,B045,19469,train
+BR1031,BR1,B041,15670,train
+BR1032,BR1,B040,16626,train
+BR1033,BR1,B010,17164,val
+BR1037,BR1,B036,17890,train
+BR1040,BR1,B040,15892,train
+BR1041,BR1,B056,17489,train
+BR1042,BR1,B040,18473,val
+BR1043,BR1,B041,13953,train
+BR1044,BR1,B041,14594,train
+BR1045,BR1,B041,14985,train
+BR1047,BR1,B046,17666,train
+BR1048,BR1,B046,17836,train
+BR1049,BR1,B046,17964,train
+BR1050,BR1,B072,15463,train
+BR1051,BR1,B072,22931,train
+BR1052,BR1,B072,20144,train
+BR1054,BR1,B073,14198,train
+BR1056,BR1,B072,15475,train
+BR1057,BR1,B072,16962,train
+BR1058,BR1,B073,20720,train
+BR1059,BR1,B085,14974,val
+BR2002,BR2,B002,16670,train
+BR2003,BR2,B002,15590,train
+BR2005,BR2,B142,14993,train
+BR2008,BR2,B078,17620,train
+BR2009,BR2,B039,11724,train
+BR2010,BR2,B078,23108,train
+BR2012,BR2,B045,14953,train
+BR2013,BR2,B046,15529,val
+BR2014,BR2,B046,16482,train
+BR2015,BR2,B040,13527,train
+BR2016,BR2,B040,17686,train
+BR2018,BR2,B047,18471,train
+BR2019,BR2,B048,17234,train
+BR2020,BR2,B048,15328,train
+BR2021,BR2,B048,16802,train
+BR2022,BR2,B043,18494,train
+BR2023,BR2,B046,16175,train
+BR2024,BR2,B045,18458,val
+BR2025,BR2,B050,24539,train
+BR2026,BR2,B053,18955,train
+BR2028,BR2,B054,20762,train
+BR2029,BR2,B132,12658,train
+BR2030,BR2,B054,22728,train
+BR2031,BR2,B056,16208,train
+BR2032,BR2,B073,14161,train
+BR2033,BR2,B067,12746,train
+BR2034,BR2,B053,16734,train
+BR2035,BR2,B055,10217,train
+BR2036,BR2,B053,14677,train
+BR2037,BR2,B056,17483,val
+BR2038,BR2,B056,15915,train
+BR2039,BR2,B077,18063,train
+BR2040,BR2,B082,13397,train
+BR2041,BR2,B064,18764,train
+BR2042,BR2,B060,16943,train
+BR2043,BR2,B080,16064,train
+BR2044,BR2,B080,17604,train
+BR2045,BR2,B073,13951,train
+BR2046,BR2,B145,14290,train
+BR2047,BR2,B077,18200,train
+BR2048,BR2,B077,26131,train
+BR2050,BR2,B084,18111,train
+BR2051,BR2,B084,18490,train
+BR2052,BR2,B060,16312,train
+BR2053,BR2,B145,12571,val
+UP1008,UP1,B053,20902,train
+UP1011,UP1,B060,14645,train
+UP1012,UP1,B053,16176,train
+UP1014,UP1,B063,20613,train
+UP1015,UP1,B063,18620,train
+UP1017,UP1,B094,16871,train
+UP1018,UP1,B094,17486,train
+UP1021,UP1,B094,14124,train
+UP1022,UP1,B094,21689,val
+UP1023,UP1,B096,13291,train
+UP1024,UP1,B096,6632,train
+UP1025,UP1,B096,13282,val
+UP1026,UP1,B096,17517,train
+UP1031,UP1,B138,17706,train
+UP1032,UP1,B138,16345,train
+UP1034,UP1,B138,18018,train