File size: 54,009 Bytes
6d1bbc7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
"""ML dataset export pipeline for NegBioDB."""

from __future__ import annotations

import logging
import sqlite3
from collections import defaultdict
from itertools import groupby
from pathlib import Path

import numpy as np
import pandas as pd
import pyarrow as pa
import pyarrow.parquet as pq
from rdkit import Chem
from rdkit.Chem.Scaffolds.MurckoScaffold import GetScaffoldForMol

logger = logging.getLogger(__name__)


# ------------------------------------------------------------------
# Split helpers
# ------------------------------------------------------------------

def _register_split(
    conn: sqlite3.Connection,
    name: str,
    strategy: str,
    seed: int | None,
    ratios: dict[str, float],
) -> int:
    """Insert or retrieve a split definition and return its split_id.

    If a split with this name already exists, deletes its old assignments
    so the split can be regenerated cleanly.
    """
    row = conn.execute(
        "SELECT split_id FROM split_definitions WHERE split_name = ?",
        (name,),
    ).fetchone()

    if row is not None:
        # Clear old assignments for re-entrancy
        split_id = int(row[0])
        conn.execute(
            "DELETE FROM split_assignments WHERE split_id = ?",
            (split_id,),
        )
        return split_id

    conn.execute(
        """INSERT INTO split_definitions
        (split_name, split_strategy, random_seed,
         train_ratio, val_ratio, test_ratio)
        VALUES (?, ?, ?, ?, ?, ?)""",
        (name, strategy, seed,
         ratios["train"], ratios["val"], ratios["test"]),
    )
    row = conn.execute(
        "SELECT split_id FROM split_definitions WHERE split_name = ?",
        (name,),
    ).fetchone()
    return int(row[0])


def _assign_folds_by_group(
    conn: sqlite3.Connection,
    split_id: int,
    group_col: str,
    seed: int,
    ratios: dict[str, float],
) -> dict[str, int]:
    """Assign folds by grouping on a column (cold-compound or cold-target).

    Uses pair-count-aware greedy assignment so that each fold receives
    approximately the target ratio of *pairs* (not groups). This prevents
    high-degree groups from causing extreme imbalance (e.g., val = 0.6%).

    All pairs sharing the same group_col value get the same fold.
    Returns dict with fold counts.
    """
    _VALID_GROUP_COLS = {"compound_id", "target_id"}
    if group_col not in _VALID_GROUP_COLS:
        raise ValueError(f"Invalid group_col: {group_col!r}. Must be one of {_VALID_GROUP_COLS}")

    # Count pairs per group
    rows = conn.execute(
        f"SELECT {group_col}, COUNT(*) FROM compound_target_pairs"
        f" GROUP BY {group_col}"
    ).fetchall()

    rng = np.random.RandomState(seed)
    group_pairs = [(gid, cnt) for gid, cnt in rows]
    rng.shuffle(group_pairs)

    total = sum(cnt for _, cnt in group_pairs)
    target_train = int(total * ratios["train"])
    target_val = int(total * ratios["val"])

    # Greedy bin-packing by pair count
    fold_counts = {"train": 0, "val": 0, "test": 0}
    group_to_fold: dict[int, str] = {}
    for gid, cnt in group_pairs:
        if fold_counts["train"] < target_train:
            fold = "train"
        elif fold_counts["val"] < target_val:
            fold = "val"
        else:
            fold = "test"
        group_to_fold[gid] = fold
        fold_counts[fold] += cnt

    # Write via temp table + JOIN for performance
    conn.execute("DROP TABLE IF EXISTS _group_folds")
    conn.execute(
        f"CREATE TEMP TABLE _group_folds ({group_col} INTEGER PRIMARY KEY, fold TEXT)"
    )
    conn.executemany(
        f"INSERT INTO _group_folds ({group_col}, fold) VALUES (?, ?)",
        group_to_fold.items(),
    )
    conn.execute(
        f"""INSERT INTO split_assignments (pair_id, split_id, fold)
        SELECT ctp.pair_id, ?, gf.fold
        FROM compound_target_pairs ctp
        JOIN _group_folds gf ON ctp.{group_col} = gf.{group_col}""",
        (split_id,),
    )
    conn.execute("DROP TABLE _group_folds")

    counts: dict[str, int] = {}
    for fold, cnt in conn.execute(
        "SELECT fold, COUNT(*) FROM split_assignments WHERE split_id = ? GROUP BY fold",
        (split_id,),
    ).fetchall():
        counts[fold] = cnt

    return counts


# ------------------------------------------------------------------
# Must-have splits
# ------------------------------------------------------------------

BATCH_SIZE = 500_000


def generate_random_split(
    conn: sqlite3.Connection,
    seed: int = 42,
    ratios: dict[str, float] | None = None,
) -> dict:
    """Generate a random 70/10/20 split across all pairs."""
    if ratios is None:
        ratios = {"train": 0.7, "val": 0.1, "test": 0.2}

    split_id = _register_split(conn, "random_v1", "random", seed, ratios)

    pair_ids = np.array(
        [r[0] for r in conn.execute(
            "SELECT pair_id FROM compound_target_pairs ORDER BY pair_id"
        ).fetchall()],
        dtype=np.int64,
    )
    n = len(pair_ids)
    logger.info("Random split: %d pairs", n)

    rng = np.random.RandomState(seed)
    indices = rng.permutation(n)

    n_train = int(n * ratios["train"])
    n_val = int(n * ratios["val"])

    fold_labels = np.empty(n, dtype="U5")
    fold_labels[indices[:n_train]] = "train"
    fold_labels[indices[n_train:n_train + n_val]] = "val"
    fold_labels[indices[n_train + n_val:]] = "test"

    # Batch insert
    for start in range(0, n, BATCH_SIZE):
        end = min(start + BATCH_SIZE, n)
        batch = [
            (int(pair_ids[i]), split_id, fold_labels[i])
            for i in range(start, end)
        ]
        conn.executemany(
            "INSERT INTO split_assignments (pair_id, split_id, fold) VALUES (?, ?, ?)",
            batch,
        )
    conn.commit()

    counts = {}
    for fold, cnt in conn.execute(
        "SELECT fold, COUNT(*) FROM split_assignments WHERE split_id = ? GROUP BY fold",
        (split_id,),
    ).fetchall():
        counts[fold] = cnt

    logger.info("Random split done: %s", counts)
    return {"split_id": split_id, "counts": counts}


def generate_cold_compound_split(
    conn: sqlite3.Connection,
    seed: int = 42,
    ratios: dict[str, float] | None = None,
) -> dict:
    """Generate cold-compound split: test compounds unseen in train."""
    if ratios is None:
        ratios = {"train": 0.7, "val": 0.1, "test": 0.2}

    split_id = _register_split(
        conn, "cold_compound_v1", "cold_compound", seed, ratios
    )
    counts = _assign_folds_by_group(conn, split_id, "compound_id", seed, ratios)
    conn.commit()
    logger.info("Cold-compound split done: %s", counts)
    return {"split_id": split_id, "counts": counts}


def generate_cold_target_split(
    conn: sqlite3.Connection,
    seed: int = 42,
    ratios: dict[str, float] | None = None,
) -> dict:
    """Generate cold-target split: test targets unseen in train."""
    if ratios is None:
        ratios = {"train": 0.7, "val": 0.1, "test": 0.2}

    split_id = _register_split(
        conn, "cold_target_v1", "cold_target", seed, ratios
    )
    counts = _assign_folds_by_group(conn, split_id, "target_id", seed, ratios)
    conn.commit()
    logger.info("Cold-target split done: %s", counts)
    return {"split_id": split_id, "counts": counts}


# ------------------------------------------------------------------
# Should-have splits
# ------------------------------------------------------------------

def generate_temporal_split(
    conn: sqlite3.Connection,
    train_cutoff: int = 2020,
    val_cutoff: int = 2023,
) -> dict:
    """Generate temporal split based on earliest_year.

    Pairs with earliest_year < train_cutoff → train,
    train_cutoff <= earliest_year < val_cutoff → val,
    earliest_year >= val_cutoff → test.
    Pairs with NULL earliest_year → train (conservative).
    """
    ratios = {"train": 0.0, "val": 0.0, "test": 0.0}  # not ratio-based
    split_id = _register_split(
        conn, "temporal_v1", "temporal", None, ratios
    )

    conn.execute(
        """INSERT INTO split_assignments (pair_id, split_id, fold)
        SELECT pair_id, ?,
            CASE
                WHEN earliest_year IS NULL OR earliest_year < ? THEN 'train'
                WHEN earliest_year < ? THEN 'val'
                ELSE 'test'
            END
        FROM compound_target_pairs""",
        (split_id, train_cutoff, val_cutoff),
    )
    conn.commit()

    counts: dict[str, int] = {}
    for fold, cnt in conn.execute(
        "SELECT fold, COUNT(*) FROM split_assignments WHERE split_id = ? GROUP BY fold",
        (split_id,),
    ).fetchall():
        counts[fold] = cnt

    total = sum(counts.values())
    for fold in ("train", "val", "test"):
        pct = counts.get(fold, 0) / total * 100 if total else 0
        logger.info("Temporal %s: %d (%.1f%%)", fold, counts.get(fold, 0), pct)
    if total > 0 and counts.get("test", 0) / total < 0.05:
        logger.warning(
            "Temporal test set is very small (%.1f%%). "
            "Consider adjusting cutoff years.",
            counts.get("test", 0) / total * 100,
        )

    return {"split_id": split_id, "counts": counts}


def _compute_scaffolds(
    conn: sqlite3.Connection,
) -> dict[str, list[int]]:
    """Compute Murcko scaffolds for all compounds, return scaffold→[compound_ids].

    Uses Murcko frameworks (ring systems + linkers, heteroatoms preserved).
    Compounds that fail RDKit parsing get scaffold='NONE'.
    """
    scaffold_to_compounds: dict[str, list[int]] = defaultdict(list)

    rows = conn.execute(
        "SELECT compound_id, canonical_smiles FROM compounds ORDER BY compound_id"
    ).fetchall()

    for compound_id, smiles in rows:
        mol = Chem.MolFromSmiles(smiles)
        if mol is None:
            scaffold_to_compounds["NONE"].append(compound_id)
            continue
        try:
            core = GetScaffoldForMol(mol)
            scaffold_smi = Chem.MolToSmiles(core)
            if not scaffold_smi:
                scaffold_smi = "NONE"
        except Exception:
            scaffold_smi = "NONE"
        scaffold_to_compounds[scaffold_smi].append(compound_id)

    logger.info(
        "Scaffold computation: %d compounds → %d unique scaffolds",
        len(rows),
        len(scaffold_to_compounds),
    )
    return dict(scaffold_to_compounds)


def generate_scaffold_split(
    conn: sqlite3.Connection,
    seed: int = 42,
    ratios: dict[str, float] | None = None,
) -> dict:
    """Generate scaffold split using Murcko scaffolds.

    Groups compounds by scaffold, then assigns groups to folds using
    a greedy size-based approach (largest scaffolds first → train).
    All pairs for a compound inherit its scaffold's fold.
    """
    if ratios is None:
        ratios = {"train": 0.7, "val": 0.1, "test": 0.2}

    split_id = _register_split(
        conn, "scaffold_v1", "scaffold", seed, ratios
    )

    # Get scaffold → compound_ids mapping
    scaffold_to_compounds = _compute_scaffolds(conn)

    # Sort scaffold groups by size (largest first) for greedy assignment
    # Tie-break by scaffold SMILES for determinism
    sorted_scaffolds = sorted(
        scaffold_to_compounds.items(),
        key=lambda x: (-len(x[1]), x[0]),
    )

    # Count total pairs per compound for size-aware assignment
    compound_pair_counts: dict[int, int] = {}
    for cid, cnt in conn.execute(
        "SELECT compound_id, COUNT(*) FROM compound_target_pairs GROUP BY compound_id"
    ).fetchall():
        compound_pair_counts[cid] = cnt

    total_pairs = sum(compound_pair_counts.values())
    target_train = int(total_pairs * ratios["train"])
    target_val = int(total_pairs * ratios["val"])

    # Greedy assignment: fill train first, then val, then test
    compound_to_fold: dict[int, str] = {}
    current_train = 0
    current_val = 0

    # Shuffle scaffolds with same size for randomness
    rng = np.random.RandomState(seed)

    # Group scaffolds by size, shuffle within each size group
    size_groups = []
    for size, group in groupby(sorted_scaffolds, key=lambda x: len(x[1])):
        group_list = list(group)
        rng.shuffle(group_list)
        size_groups.extend(group_list)

    for scaffold_smi, compound_ids in size_groups:
        group_pairs = sum(compound_pair_counts.get(c, 0) for c in compound_ids)

        if current_train + group_pairs <= target_train:
            fold = "train"
            current_train += group_pairs
        elif current_val + group_pairs <= target_val:
            fold = "val"
            current_val += group_pairs
        else:
            fold = "test"

        for cid in compound_ids:
            compound_to_fold[cid] = fold

    # Write via temp table
    conn.execute("DROP TABLE IF EXISTS _scaffold_folds")
    conn.execute(
        "CREATE TEMP TABLE _scaffold_folds (compound_id INTEGER PRIMARY KEY, fold TEXT)"
    )
    conn.executemany(
        "INSERT INTO _scaffold_folds (compound_id, fold) VALUES (?, ?)",
        compound_to_fold.items(),
    )
    conn.execute(
        """INSERT INTO split_assignments (pair_id, split_id, fold)
        SELECT ctp.pair_id, ?, sf.fold
        FROM compound_target_pairs ctp
        JOIN _scaffold_folds sf ON ctp.compound_id = sf.compound_id""",
        (split_id,),
    )
    conn.execute("DROP TABLE _scaffold_folds")
    conn.commit()

    counts: dict[str, int] = {}
    for fold, cnt in conn.execute(
        "SELECT fold, COUNT(*) FROM split_assignments WHERE split_id = ? GROUP BY fold",
        (split_id,),
    ).fetchall():
        counts[fold] = cnt

    logger.info("Scaffold split done: %s", counts)
    return {"split_id": split_id, "counts": counts}


def generate_degree_balanced_split(
    conn: sqlite3.Connection,
    seed: int = 42,
    ratios: dict[str, float] | None = None,
    n_bins: int = 10,
) -> dict:
    """Generate degree-distribution-balanced (DDB) split.

    Bins pairs by (log compound_degree, log target_degree) and performs
    stratified sampling so each fold preserves the degree distribution.
    Essential for Experiment 4 (degree bias evaluation).
    """
    if ratios is None:
        ratios = {"train": 0.7, "val": 0.1, "test": 0.2}

    split_id = _register_split(
        conn, "degree_balanced_v1", "degree_balanced", seed, ratios
    )

    # Fetch pair_id, compound_degree, target_degree
    rows = conn.execute(
        """SELECT pair_id, COALESCE(compound_degree, 1), COALESCE(target_degree, 1)
        FROM compound_target_pairs ORDER BY pair_id"""
    ).fetchall()

    pair_ids = np.array([r[0] for r in rows], dtype=np.int64)
    c_deg = np.array([r[1] for r in rows], dtype=np.float64)
    t_deg = np.array([r[2] for r in rows], dtype=np.float64)

    # Log-scale binning
    c_log = np.log1p(c_deg)
    t_log = np.log1p(t_deg)

    c_bins = np.minimum(
        (c_log / (c_log.max() + 1e-9) * n_bins).astype(int), n_bins - 1
    )
    t_bins = np.minimum(
        (t_log / (t_log.max() + 1e-9) * n_bins).astype(int), n_bins - 1
    )

    # Combined bin label
    bin_labels = c_bins * n_bins + t_bins

    # Stratified split within each bin
    rng = np.random.RandomState(seed)
    fold_labels = np.empty(len(pair_ids), dtype="U5")

    for bin_id in np.unique(bin_labels):
        mask = bin_labels == bin_id
        idx = np.where(mask)[0]
        rng.shuffle(idx)
        n = len(idx)
        n_train = int(n * ratios["train"])
        n_val = int(n * ratios["val"])
        fold_labels[idx[:n_train]] = "train"
        fold_labels[idx[n_train:n_train + n_val]] = "val"
        fold_labels[idx[n_train + n_val:]] = "test"

    # Batch insert
    for start in range(0, len(pair_ids), BATCH_SIZE):
        end = min(start + BATCH_SIZE, len(pair_ids))
        batch = [
            (int(pair_ids[i]), split_id, fold_labels[i])
            for i in range(start, end)
        ]
        conn.executemany(
            "INSERT INTO split_assignments (pair_id, split_id, fold) VALUES (?, ?, ?)",
            batch,
        )
    conn.commit()

    counts: dict[str, int] = {}
    for fold, cnt in conn.execute(
        "SELECT fold, COUNT(*) FROM split_assignments WHERE split_id = ? GROUP BY fold",
        (split_id,),
    ).fetchall():
        counts[fold] = cnt

    logger.info("Degree-balanced split done: %s", counts)
    return {"split_id": split_id, "counts": counts}


# ------------------------------------------------------------------
# Dataset export
# ------------------------------------------------------------------

EXPORT_CHUNKSIZE = 500_000

# Split strategies in the order they appear as inline columns
SPLIT_STRATEGIES = [
    "random", "cold_compound", "cold_target",
    "temporal", "scaffold", "degree_balanced",
]


def _resolve_split_id(conn: sqlite3.Connection, strategy: str) -> int | None:
    """Resolve the preferred split_id for a strategy using version-aware ordering."""
    rows = conn.execute(
        """SELECT split_id, split_name, version
        FROM split_definitions
        WHERE split_strategy = ?""",
        (strategy,),
    ).fetchall()
    if not rows:
        return None

    def sort_key(row: tuple[int, str, str | None]) -> tuple[int, str, int]:
        split_id, split_name, version = row
        version_num = -1
        if version:
            try:
                version_num = int(str(version).split(".")[0])
            except ValueError:
                version_num = -1
        if "_v" in split_name:
            suffix = split_name.rsplit("_v", 1)[1]
            if suffix.isdigit():
                version_num = max(version_num, int(suffix))
        return (version_num, split_name, split_id)

    return sorted(rows, key=sort_key)[-1][0]


def _build_export_query(split_ids: dict[str, int | None]) -> str:
    """Build the pivot SQL for exporting pairs with inline split columns."""
    split_cols = []
    join_clauses = []
    for i, strategy in enumerate(split_ids):
        alias_sa = f"sa{i}"
        col_name = f"split_{strategy}"
        split_cols.append(f"{alias_sa}.fold AS {col_name}")
        split_id = split_ids[strategy]
        split_id_sql = "NULL" if split_id is None else str(int(split_id))
        join_clauses.append(
            f"LEFT JOIN split_assignments {alias_sa} "
            f"ON ctp.pair_id = {alias_sa}.pair_id "
            f"AND {alias_sa}.split_id = {split_id_sql}"
        )

    join_sql = "\n".join(join_clauses)

    base_cols = """ctp.pair_id,
       c.canonical_smiles AS smiles,
       c.inchikey,
       t.uniprot_accession AS uniprot_id,
       t.amino_acid_sequence AS target_sequence,
       t.gene_symbol,
       0 AS Y,
       ctp.best_confidence AS confidence_tier,
       ctp.best_result_type,
       ctp.num_assays,
       ctp.num_sources,
       ctp.earliest_year,
       ctp.compound_degree,
       ctp.target_degree"""

    if split_cols:
        select_clause = base_cols + ",\n       " + ",\n       ".join(split_cols)
    else:
        select_clause = base_cols

    query = f"""SELECT
       {select_clause}
FROM compound_target_pairs ctp
JOIN compounds c ON ctp.compound_id = c.compound_id
JOIN targets t ON ctp.target_id = t.target_id
{join_sql}
ORDER BY ctp.pair_id"""

    return query


def export_negative_dataset(
    db_path: str | Path,
    output_dir: str | Path,
    split_strategies: list[str] | None = None,
    chunksize: int = EXPORT_CHUNKSIZE,
) -> dict:
    """Export negative DTI pairs as Parquet and lightweight CSV.

    Produces:
      - negbiodb_dti_pairs.parquet (full dataset with sequences)
      - negbiodb_splits.csv (pair_id + smiles + uniprot + split columns only)

    Returns dict with file paths and row count.
    """
    output_dir = Path(output_dir)
    output_dir.mkdir(parents=True, exist_ok=True)

    parquet_path = output_dir / "negbiodb_dti_pairs.parquet"
    splits_csv_path = output_dir / "negbiodb_splits.csv"

    if split_strategies is None:
        split_strategies = SPLIT_STRATEGIES

    conn = sqlite3.connect(str(db_path))
    conn.execute("PRAGMA journal_mode = WAL")
    split_ids = {
        strategy: _resolve_split_id(conn, strategy)
        for strategy in split_strategies
    }
    query = _build_export_query(split_ids)
    split_cols = [f"split_{s}" for s in split_strategies]

    total_rows = 0
    pq_writer = None

    try:
        for chunk in pd.read_sql_query(query, conn, chunksize=chunksize):
            total_rows += len(chunk)

            table = pa.Table.from_pandas(chunk, preserve_index=False)
            if pq_writer is None:
                pq_writer = pq.ParquetWriter(
                    str(parquet_path), table.schema, compression="zstd"
                )
            pq_writer.write_table(table)

            logger.info("Exported %d rows so far...", total_rows)

        if pq_writer is not None:
            pq_writer.close()

        # Lightweight splits CSV (no target_sequence)
        if total_rows > 0:
            splits_columns = ["pair_id", "smiles", "inchikey", "uniprot_id"] + split_cols
            # Read back from parquet for splits CSV (avoids re-querying)
            pf = pq.ParquetFile(str(parquet_path))
            first = True
            for batch in pf.iter_batches(
                batch_size=chunksize, columns=splits_columns
            ):
                df = batch.to_pandas()
                df.to_csv(
                    str(splits_csv_path),
                    mode="w" if first else "a",
                    header=first,
                    index=False,
                )
                first = False

    finally:
        conn.close()

    logger.info(
        "Export complete: %d rows → %s (%.1f MB), %s",
        total_rows,
        parquet_path.name,
        parquet_path.stat().st_size / 1e6 if parquet_path.exists() else 0,
        splits_csv_path.name,
    )

    return {
        "total_rows": total_rows,
        "parquet_path": str(parquet_path),
        "splits_csv_path": str(splits_csv_path),
    }


# ------------------------------------------------------------------
# ChEMBL positive extraction + M1 merge
# ------------------------------------------------------------------

# ------------------------------------------------------------------
# DataFrame-level split helpers (for M1 / random negative datasets)
# ------------------------------------------------------------------

_DEFAULT_RATIOS = {"train": 0.7, "val": 0.1, "test": 0.2}


def add_random_split(
    df: pd.DataFrame,
    seed: int = 42,
    ratios: dict[str, float] | None = None,
) -> pd.DataFrame:
    """Add split_random column with deterministic 70/10/20 assignment."""
    if ratios is None:
        ratios = _DEFAULT_RATIOS
    rng = np.random.RandomState(seed)
    n = len(df)
    if n == 0:
        df = df.copy()
        df["split_random"] = pd.Series(dtype=str)
        return df
    indices = np.arange(n)
    rng.shuffle(indices)
    folds = np.empty(n, dtype=object)
    n_train = int(n * ratios["train"])
    n_val = int(n * ratios["val"])
    folds[indices[:n_train]] = "train"
    folds[indices[n_train:n_train + n_val]] = "val"
    folds[indices[n_train + n_val:]] = "test"
    df = df.copy()
    df["split_random"] = folds
    return df


def add_cold_compound_split(
    df: pd.DataFrame,
    seed: int = 42,
    ratios: dict[str, float] | None = None,
) -> pd.DataFrame:
    """Add split_cold_compound column — group by InChIKey connectivity.

    All rows sharing the same InChIKey[:14] get the same fold,
    ensuring no compound leaks between train and test.
    """
    if ratios is None:
        ratios = _DEFAULT_RATIOS
    if len(df) == 0:
        df = df.copy()
        df["split_cold_compound"] = pd.Series(dtype=str)
        return df
    compounds = np.array(df["inchikey"].str[:14].unique())
    rng = np.random.RandomState(seed)
    rng.shuffle(compounds)
    n = len(compounds)
    n_train = int(n * ratios["train"])
    n_val = int(n * ratios["val"])
    comp_fold = {}
    for i, c in enumerate(compounds):
        if i < n_train:
            comp_fold[c] = "train"
        elif i < n_train + n_val:
            comp_fold[c] = "val"
        else:
            comp_fold[c] = "test"
    df = df.copy()
    df["split_cold_compound"] = df["inchikey"].str[:14].map(comp_fold)
    return df


def add_cold_target_split(
    df: pd.DataFrame,
    seed: int = 42,
    ratios: dict[str, float] | None = None,
) -> pd.DataFrame:
    """Add split_cold_target column — group by UniProt accession.

    All rows sharing the same uniprot_id get the same fold,
    ensuring no target leaks between train and test.
    """
    if ratios is None:
        ratios = _DEFAULT_RATIOS
    if len(df) == 0:
        df = df.copy()
        df["split_cold_target"] = pd.Series(dtype=str)
        return df
    targets = np.array(df["uniprot_id"].unique())
    rng = np.random.RandomState(seed)
    rng.shuffle(targets)
    n = len(targets)
    n_train = int(n * ratios["train"])
    n_val = int(n * ratios["val"])
    tgt_fold = {}
    for i, t in enumerate(targets):
        if i < n_train:
            tgt_fold[t] = "train"
        elif i < n_train + n_val:
            tgt_fold[t] = "val"
        else:
            tgt_fold[t] = "test"
    df = df.copy()
    df["split_cold_target"] = df["uniprot_id"].map(tgt_fold)
    return df


def add_degree_balanced_split(
    df: pd.DataFrame,
    seed: int = 42,
    ratios: dict[str, float] | None = None,
    n_bins: int = 10,
) -> pd.DataFrame:
    """Add split_degree_balanced using the full benchmark graph degrees.

    This is the DataFrame analogue of the DB-level DDB split, but operates on a
    merged benchmark table so both positives and negatives are assigned using the
    same degree distribution.
    """
    if ratios is None:
        ratios = _DEFAULT_RATIOS
    if len(df) == 0:
        df = df.copy()
        df["split_degree_balanced"] = pd.Series(dtype=str)
        return df

    df = df.copy()
    compound_keys = df["inchikey"].str[:14]
    target_keys = df["uniprot_id"]

    compound_degree = compound_keys.map(compound_keys.value_counts()).to_numpy(dtype=np.float64)
    target_degree = target_keys.map(target_keys.value_counts()).to_numpy(dtype=np.float64)

    c_log = np.log1p(compound_degree)
    t_log = np.log1p(target_degree)

    c_bins = np.minimum(
        (c_log / (c_log.max() + 1e-9) * n_bins).astype(int), n_bins - 1
    )
    t_bins = np.minimum(
        (t_log / (t_log.max() + 1e-9) * n_bins).astype(int), n_bins - 1
    )
    bin_labels = c_bins * n_bins + t_bins

    rng = np.random.RandomState(seed)
    fold_labels = np.empty(len(df), dtype=object)
    for bin_id in np.unique(bin_labels):
        idx = np.where(bin_labels == bin_id)[0]
        rng.shuffle(idx)
        n = len(idx)
        n_train = int(n * ratios["train"])
        n_val = int(n * ratios["val"])
        fold_labels[idx[:n_train]] = "train"
        fold_labels[idx[n_train:n_train + n_val]] = "val"
        fold_labels[idx[n_train + n_val:]] = "test"

    df["split_degree_balanced"] = fold_labels
    return df


def apply_m1_splits(
    df: pd.DataFrame,
    seed: int = 42,
) -> pd.DataFrame:
    """Apply all three M1 split strategies to a DataFrame.

    Adds columns: split_random, split_cold_compound, split_cold_target.
    Uses the same seed for reproducibility across conditions.
    """
    df = add_random_split(df, seed=seed)
    df = add_cold_compound_split(df, seed=seed)
    df = add_cold_target_split(df, seed=seed)
    return df


_CHEMBL_POSITIVE_SQL = """
SELECT
    cs.canonical_smiles,
    cs.standard_inchi_key AS inchikey,
    cp.accession AS uniprot_id,
    cp.sequence AS target_sequence,
    a.pchembl_value,
    a.standard_type AS activity_type,
    a.standard_value AS activity_value_nm,
    docs.year AS publication_year
FROM activities a
JOIN assays ass ON a.assay_id = ass.assay_id
JOIN target_dictionary td ON ass.tid = td.tid
JOIN target_components tc ON td.tid = tc.tid
JOIN component_sequences cp ON tc.component_id = cp.component_id
JOIN molecule_dictionary md ON a.molregno = md.molregno
JOIN compound_structures cs ON md.molregno = cs.molregno
LEFT JOIN docs ON a.doc_id = docs.doc_id
WHERE a.pchembl_value >= ?
  AND a.standard_type IN ('IC50', 'Ki', 'Kd', 'EC50')
  AND td.target_type = 'SINGLE PROTEIN'
  AND td.organism = 'Homo sapiens'
  AND a.data_validity_comment IS NULL
  AND cs.canonical_smiles IS NOT NULL
  AND cp.accession IS NOT NULL
"""


def extract_chembl_positives(
    chembl_db_path: str | Path,
    negbiodb_path: str | Path,
    pchembl_min: float = 6.0,
    chunksize: int = 100_000,
) -> pd.DataFrame:
    """Extract active compounds from ChEMBL for M1 binary task.

    Filters:
    - pChEMBL >= pchembl_min (default 6.0 = IC50 <= 1 uM)
    - Single protein, human, valid data
    - Target must exist in NegBioDB target pool
    - Deduplicates by (inchikey_connectivity, uniprot_id), keeping max pChEMBL

    Returns DataFrame with columns matching negative export format.
    """
    from negbiodb.standardize import standardize_smiles

    # Load NegBioDB target pool
    neg_conn = sqlite3.connect(str(negbiodb_path))
    neg_targets = {
        r[0] for r in neg_conn.execute(
            "SELECT uniprot_accession FROM targets"
        ).fetchall()
    }
    neg_conn.close()
    logger.info("NegBioDB target pool: %d targets", len(neg_targets))

    # Query ChEMBL
    chembl_conn = sqlite3.connect(str(chembl_db_path))
    rows = []

    for chunk in pd.read_sql_query(
        _CHEMBL_POSITIVE_SQL, chembl_conn,
        params=(pchembl_min,), chunksize=chunksize,
    ):
        # Filter to shared target pool
        chunk = chunk[chunk["uniprot_id"].isin(neg_targets)]
        if chunk.empty:
            continue

        # Standardize SMILES → canonical + InChIKey
        std_results = []
        for _, row in chunk.iterrows():
            result = standardize_smiles(row["canonical_smiles"])
            if result is None:
                continue
            std_results.append({
                "smiles": result["canonical_smiles"],
                "inchikey": result["inchikey"],
                "inchikey_connectivity": result["inchikey"][:14],
                "uniprot_id": row["uniprot_id"],
                "target_sequence": row["target_sequence"],
                "pchembl_value": row["pchembl_value"],
                "activity_type": row["activity_type"],
                "activity_value_nm": row["activity_value_nm"],
                "publication_year": row["publication_year"],
            })
        if std_results:
            rows.extend(std_results)

        logger.info("Processed %d positive candidates so far...", len(rows))

    chembl_conn.close()

    if not rows:
        logger.warning("No positive records extracted from ChEMBL")
        return pd.DataFrame(columns=[
            "smiles", "inchikey", "uniprot_id", "target_sequence",
            "pchembl_value", "activity_type", "activity_value_nm",
            "publication_year",
        ])

    df = pd.DataFrame(rows)

    # Deduplicate: keep highest pChEMBL per (inchikey_connectivity, uniprot_id)
    df = df.sort_values("pchembl_value", ascending=False)
    df = df.drop_duplicates(subset=["inchikey_connectivity", "uniprot_id"], keep="first")
    df = df.drop(columns=["inchikey_connectivity"])

    logger.info("ChEMBL positives after dedup: %d unique pairs", len(df))
    return df


def merge_positive_negative(
    positives: pd.DataFrame,
    negbiodb_path: str | Path,
    output_dir: str | Path,
    seed: int = 42,
) -> dict:
    """Merge positives (Y=1) and negatives (Y=0) for M1 binary DTI task.

    Validates zero overlap between positives and negatives by InChIKey
    connectivity × UniProt, then creates:
    - Balanced (1:1) dataset
    - Realistic (1:10 pos:neg) dataset

    Returns dict with file paths and statistics.
    """
    output_dir = Path(output_dir)
    output_dir.mkdir(parents=True, exist_ok=True)

    # Load negatives from NegBioDB
    neg_conn = sqlite3.connect(str(negbiodb_path))
    negatives = pd.read_sql_query(
        """SELECT c.canonical_smiles AS smiles, c.inchikey,
                  t.uniprot_accession AS uniprot_id,
                  t.amino_acid_sequence AS target_sequence,
                  ctp.best_confidence AS confidence_tier,
                  ctp.num_assays, ctp.num_sources, ctp.earliest_year
        FROM compound_target_pairs ctp
        JOIN compounds c ON ctp.compound_id = c.compound_id
        JOIN targets t ON ctp.target_id = t.target_id""",
        neg_conn,
    )
    neg_conn.close()

    # Overlap check: InChIKey connectivity × UniProt
    pos_keys = set(
        zip(
            positives["inchikey"].str[:14],
            positives["uniprot_id"],
        )
    )
    neg_keys = set(
        zip(
            negatives["inchikey"].str[:14],
            negatives["uniprot_id"],
        )
    )
    overlap = pos_keys & neg_keys
    if overlap:
        logger.warning(
            "Found %d overlapping (inchikey_conn, uniprot) pairs! "
            "Removing from BOTH positives and negatives.",
            len(overlap),
        )
        pos_mask = ~pd.Series(
            [k in overlap for k in zip(positives["inchikey"].str[:14], positives["uniprot_id"])],
            index=positives.index,
        )
        neg_mask = ~pd.Series(
            [k in overlap for k in zip(negatives["inchikey"].str[:14], negatives["uniprot_id"])],
            index=negatives.index,
        )
        positives = positives[pos_mask]
        negatives = negatives[neg_mask]
        logger.info("After overlap removal: %d pos, %d neg", len(positives), len(negatives))

    # Prepare label columns
    positives = positives.copy()
    positives["Y"] = 1
    negatives = negatives.copy()
    negatives["Y"] = 0

    # Common columns for merge
    common_cols = ["smiles", "inchikey", "uniprot_id", "target_sequence", "Y"]
    pos_export = positives[common_cols]
    neg_export = negatives[common_cols]

    rng = np.random.RandomState(seed)
    results = {}

    # Balanced (1:1)
    n_pos = len(pos_export)
    n_neg = len(neg_export)
    n_balanced = min(n_pos, n_neg)

    pos_balanced = pos_export.sample(n=n_balanced, random_state=rng)
    neg_balanced = neg_export.sample(n=n_balanced, random_state=rng)
    balanced = pd.concat([pos_balanced, neg_balanced], ignore_index=True)
    balanced = balanced.sample(frac=1, random_state=rng).reset_index(drop=True)

    # Apply M1 splits (random, cold-compound, cold-target)
    balanced = apply_m1_splits(balanced, seed=seed)

    balanced_path = output_dir / "negbiodb_m1_balanced.parquet"
    balanced.to_parquet(str(balanced_path), index=False, compression="zstd")
    results["balanced"] = {
        "path": str(balanced_path),
        "n_pos": n_balanced,
        "n_neg": n_balanced,
        "total": len(balanced),
    }

    # Realistic (1:10 pos:neg)
    n_realistic_neg = min(n_pos * 10, n_neg)
    n_realistic_pos = min(n_pos, n_realistic_neg // 10) if n_realistic_neg >= 10 else n_pos

    rng2 = np.random.RandomState(seed)
    pos_realistic = pos_export.sample(n=n_realistic_pos, random_state=rng2)
    neg_realistic = neg_export.sample(n=n_realistic_neg, random_state=rng2)
    realistic = pd.concat([pos_realistic, neg_realistic], ignore_index=True)
    realistic = realistic.sample(frac=1, random_state=rng2).reset_index(drop=True)

    # Apply M1 splits (random, cold-compound, cold-target)
    realistic = apply_m1_splits(realistic, seed=seed)

    realistic_path = output_dir / "negbiodb_m1_realistic.parquet"
    realistic.to_parquet(str(realistic_path), index=False, compression="zstd")
    results["realistic"] = {
        "path": str(realistic_path),
        "n_pos": n_realistic_pos,
        "n_neg": n_realistic_neg,
        "total": len(realistic),
    }

    logger.info(
        "M1 merge done: balanced=%d (1:1), realistic=%d (1:%d)",
        len(balanced),
        len(realistic),
        n_realistic_neg // n_realistic_pos if n_realistic_pos > 0 else 0,
    )

    return results


# ------------------------------------------------------------------
# Random negative generation (Exp 1)
# ------------------------------------------------------------------

def _load_tested_pairs(
    negbiodb_path: str | Path,
    positives: pd.DataFrame | None = None,
) -> set[tuple[str, str]]:
    """Load all tested (compound, target) pairs as (inchikey_conn, uniprot_id).

    Includes NegBioDB negative pairs and optionally positive pairs.
    """
    conn = sqlite3.connect(str(negbiodb_path))
    try:
        tested = set()
        for row in conn.execute(
            """SELECT c.inchikey_connectivity, t.uniprot_accession
            FROM compound_target_pairs ctp
            JOIN compounds c ON ctp.compound_id = c.compound_id
            JOIN targets t ON ctp.target_id = t.target_id"""
        ):
            tested.add((row[0], row[1]))
    finally:
        conn.close()

    if positives is not None:
        for ik, uid in zip(positives["inchikey"].str[:14], positives["uniprot_id"]):
            tested.add((ik, uid))

    return tested


def _load_compound_target_pools(
    negbiodb_path: str | Path,
    positives: pd.DataFrame | None = None,
) -> tuple[list[dict], list[dict]]:
    """Load compound and target pools with SMILES/sequence for output.

    Returns (compounds_list, targets_list) where each element is a dict
    with the fields needed for M1 output.
    """
    conn = sqlite3.connect(str(negbiodb_path))
    try:
        # Compounds: inchikey_connectivity → {smiles, inchikey}
        compound_map: dict[str, dict] = {}
        for row in conn.execute(
            "SELECT inchikey_connectivity, canonical_smiles, inchikey FROM compounds"
        ):
            compound_map[row[0]] = {"smiles": row[1], "inchikey": row[2]}

        # Targets: uniprot_accession → {target_sequence}
        target_map: dict[str, dict] = {}
        for row in conn.execute(
            "SELECT uniprot_accession, amino_acid_sequence FROM targets"
        ):
            target_map[row[0]] = {"target_sequence": row[1]}
    finally:
        conn.close()

    # Add compounds from positives that might not be in NegBioDB
    if positives is not None:
        for _, r in positives[["smiles", "inchikey"]].drop_duplicates(
            subset=["inchikey"]
        ).iterrows():
            ik_conn = r["inchikey"][:14]
            if ik_conn not in compound_map:
                compound_map[ik_conn] = {
                    "smiles": r["smiles"],
                    "inchikey": r["inchikey"],
                }

    compounds = [
        {"inchikey_conn": k, **v} for k, v in compound_map.items()
    ]
    targets = [
        {"uniprot_id": k, **v} for k, v in target_map.items()
    ]
    return compounds, targets


def generate_uniform_random_negatives(
    negbiodb_path: str | Path,
    positives: pd.DataFrame,
    n_samples: int,
    output_dir: str | Path,
    seed: int = 42,
) -> dict:
    """Generate uniform random negative pairs for Exp 1 control.

    Samples untested compound-target pairs uniformly from the cross-product
    of all compounds × all targets, excluding any tested pairs (both
    NegBioDB negatives and ChEMBL positives).

    Merges with the same positive set and applies M1 splits.

    Returns dict with file path and statistics.
    """
    output_dir = Path(output_dir)
    output_dir.mkdir(parents=True, exist_ok=True)

    logger.info("Loading tested pairs for exclusion...")
    tested = _load_tested_pairs(negbiodb_path, positives)
    logger.info("Tested pairs: %d", len(tested))

    logger.info("Loading compound/target pools...")
    compounds, targets = _load_compound_target_pools(negbiodb_path, positives)
    logger.info("Compound pool: %d, Target pool: %d", len(compounds), len(targets))

    # Rejection sampling
    rng = np.random.RandomState(seed)
    neg_rows = []
    attempts = 0
    max_attempts = n_samples * 100  # safety limit

    while len(neg_rows) < n_samples and attempts < max_attempts:
        batch = min((n_samples - len(neg_rows)) * 2, 1_000_000)
        c_idx = rng.randint(0, len(compounds), batch)
        t_idx = rng.randint(0, len(targets), batch)

        for ci, ti in zip(c_idx, t_idx):
            comp = compounds[ci]
            tgt = targets[ti]
            key = (comp["inchikey_conn"], tgt["uniprot_id"])
            if key not in tested:
                neg_rows.append({
                    "smiles": comp["smiles"],
                    "inchikey": comp["inchikey"],
                    "uniprot_id": tgt["uniprot_id"],
                    "target_sequence": tgt["target_sequence"],
                    "Y": 0,
                })
                tested.add(key)  # prevent duplicate negatives
                if len(neg_rows) >= n_samples:
                    break
            attempts += 1

    if len(neg_rows) == 0:
        logger.warning("Uniform random: 0 negatives generated (pool exhausted)")
    else:
        logger.info(
            "Uniform random: generated %d negatives (rejection rate: %.2f%%)",
            len(neg_rows),
            (1 - len(neg_rows) / max(attempts, 1)) * 100,
        )

    neg_df = pd.DataFrame(neg_rows)

    # Merge with positives (same as M1 balanced)
    pos_export = positives[["smiles", "inchikey", "uniprot_id", "target_sequence"]].copy()
    pos_export["Y"] = 1

    n_balanced = min(len(pos_export), len(neg_df))
    # Use independent RNG streams for pos/neg sampling to avoid correlated draws.
    pos_sample = pos_export.sample(n=n_balanced, random_state=np.random.RandomState(seed))
    neg_sample = neg_df.sample(n=n_balanced, random_state=np.random.RandomState(seed + 1))

    merged = pd.concat([pos_sample, neg_sample], ignore_index=True)
    merged = merged.sample(frac=1, random_state=np.random.RandomState(seed + 2)).reset_index(drop=True)
    merged = apply_m1_splits(merged, seed=seed)

    out_path = output_dir / "negbiodb_m1_uniform_random.parquet"
    merged.to_parquet(str(out_path), index=False, compression="zstd")

    result = {
        "path": str(out_path),
        "n_pos": n_balanced,
        "n_neg": n_balanced,
        "total": len(merged),
    }
    logger.info("Uniform random M1: %d total → %s", len(merged), out_path.name)
    return result


def generate_degree_matched_negatives(
    negbiodb_path: str | Path,
    positives: pd.DataFrame,
    n_samples: int,
    output_dir: str | Path,
    seed: int = 42,
) -> dict:
    """Generate degree-matched random negatives for Exp 1 control.

    Samples untested pairs whose compounds and targets have degree
    distributions matching NegBioDB's, isolating the effect of
    experimental confirmation vs. degree bias.

    Uses log-scale binning of compound_degree × target_degree.
    """
    output_dir = Path(output_dir)
    output_dir.mkdir(parents=True, exist_ok=True)

    logger.info("Loading tested pairs and degree info...")
    tested = _load_tested_pairs(negbiodb_path, positives)

    conn = sqlite3.connect(str(negbiodb_path))
    try:
        # Query 1: degree distribution only (no strings — fast, ~200 MB RAM)
        logger.info("Loading degree distribution from pairs table...")
        deg_dist = pd.read_sql_query(
            "SELECT compound_degree, target_degree FROM compound_target_pairs",
            conn,
        )
        # Query 2: unique compounds with smiles (919K rows × small strings)
        logger.info("Loading unique compound pool...")
        comp_df = pd.read_sql_query(
            """SELECT inchikey_connectivity, canonical_smiles, inchikey,
                      MAX(ctp.compound_degree) AS compound_degree
               FROM compounds c
               JOIN compound_target_pairs ctp ON c.compound_id = ctp.compound_id
               GROUP BY c.compound_id""",
            conn,
        )
        # Query 3: unique targets with sequences (3711 rows × long strings, ~3 MB)
        logger.info("Loading unique target pool...")
        tgt_df = pd.read_sql_query(
            """SELECT t.uniprot_accession AS uniprot_id, t.amino_acid_sequence AS target_sequence,
                      MAX(ctp.target_degree) AS target_degree
               FROM targets t
               JOIN compound_target_pairs ctp ON t.target_id = ctp.target_id
               GROUP BY t.target_id""",
            conn,
        )
    finally:
        conn.close()

    logger.info(
        "Loaded: %d pair-degree rows, %d unique compounds, %d unique targets",
        len(deg_dist), len(comp_df), len(tgt_df),
    )

    # Log-scale binning of NegBioDB degree distribution
    deg_dist["cdeg_bin"] = np.floor(
        np.log2(deg_dist["compound_degree"].clip(lower=1))
    ).astype(int)
    deg_dist["tdeg_bin"] = np.floor(
        np.log2(deg_dist["target_degree"].clip(lower=1))
    ).astype(int)

    bin_counts = deg_dist.groupby(["cdeg_bin", "tdeg_bin"]).size()
    total_pairs = bin_counts.sum()
    del deg_dist  # free memory

    # Compute target samples per bin
    bin_targets = {}
    remaining = n_samples
    for (cb, tb), count in bin_counts.items():
        n = int(count / total_pairs * n_samples)
        bin_targets[(cb, tb)] = n
        remaining -= n
    # Distribute remainder to largest bins
    if remaining > 0:
        for key in bin_counts.sort_values(ascending=False).index:
            if remaining <= 0:
                break
            bin_targets[key] = bin_targets.get(key, 0) + 1
            remaining -= 1

    # Build per-bin compound and target pools
    compounds_by_bin: dict[int, list[dict]] = defaultdict(list)
    for _, row in comp_df.iterrows():
        cb = int(np.floor(np.log2(max(row["compound_degree"], 1))))
        compounds_by_bin[cb].append({
            "inchikey_conn": row["inchikey_connectivity"],
            "smiles": row["canonical_smiles"],
            "inchikey": row["inchikey"],
        })

    targets_by_bin: dict[int, list[dict]] = defaultdict(list)
    for _, row in tgt_df.iterrows():
        tb = int(np.floor(np.log2(max(row["target_degree"], 1))))
        targets_by_bin[tb].append({
            "uniprot_id": row["uniprot_id"],
            "target_sequence": row["target_sequence"],
        })

    # Rejection sampling per bin
    rng = np.random.RandomState(seed)
    neg_rows = []

    for (cb, tb), target_n in bin_targets.items():
        if target_n <= 0:
            continue
        c_pool = compounds_by_bin.get(cb, [])
        t_pool = targets_by_bin.get(tb, [])
        if not c_pool or not t_pool:
            logger.warning("Empty pool for bin (%d, %d), skipping %d samples", cb, tb, target_n)
            continue

        sampled = 0
        attempts = 0
        max_attempts = target_n * 200

        while sampled < target_n and attempts < max_attempts:
            ci = rng.randint(0, len(c_pool))
            ti = rng.randint(0, len(t_pool))
            comp = c_pool[ci]
            tgt = t_pool[ti]
            key = (comp["inchikey_conn"], tgt["uniprot_id"])
            if key not in tested:
                neg_rows.append({
                    "smiles": comp["smiles"],
                    "inchikey": comp["inchikey"],
                    "uniprot_id": tgt["uniprot_id"],
                    "target_sequence": tgt["target_sequence"],
                    "Y": 0,
                })
                tested.add(key)  # prevent duplicate negatives
                sampled += 1
            attempts += 1

        if sampled < target_n:
            logger.warning(
                "Bin (%d, %d): sampled %d/%d (max_attempts reached)",
                cb, tb, sampled, target_n,
            )

    if len(neg_rows) == 0:
        logger.warning("Degree-matched: 0 negatives generated (pool exhausted)")
    else:
        logger.info("Degree-matched: generated %d negatives", len(neg_rows))

    neg_df = pd.DataFrame(neg_rows)

    # Merge with positives (same as M1 balanced)
    pos_export = positives[["smiles", "inchikey", "uniprot_id", "target_sequence"]].copy()
    pos_export["Y"] = 1

    n_balanced = min(len(pos_export), len(neg_df))
    # Use independent RNG streams for pos/neg sampling to avoid correlated draws.
    pos_sample = pos_export.sample(n=n_balanced, random_state=np.random.RandomState(seed))
    neg_sample = neg_df.sample(n=n_balanced, random_state=np.random.RandomState(seed + 1)) if len(neg_df) > n_balanced else neg_df

    merged = pd.concat([pos_sample, neg_sample], ignore_index=True)
    merged = merged.sample(frac=1, random_state=np.random.RandomState(seed + 2)).reset_index(drop=True)
    merged = apply_m1_splits(merged, seed=seed)

    out_path = output_dir / "negbiodb_m1_degree_matched.parquet"
    merged.to_parquet(str(out_path), index=False, compression="zstd")

    result = {
        "path": str(out_path),
        "n_pos": n_balanced,
        "n_neg": n_balanced,
        "total": len(merged),
    }
    logger.info("Degree-matched M1: %d total → %s", len(merged), out_path.name)
    return result


# ------------------------------------------------------------------
# Data leakage check
# ------------------------------------------------------------------

def check_cold_split_integrity(
    conn: sqlite3.Connection,
) -> dict:
    """Verify cold splits have zero entity leakage between train and test.

    Returns dict with per-strategy leak counts (should all be 0).
    """
    results = {}

    for strategy, entity_col in [
        ("cold_compound", "compound_id"),
        ("cold_target", "target_id"),
    ]:
        sid = _resolve_split_id(conn, strategy)
        if sid is None:
            results[strategy] = {"status": "not_found"}
            continue

        leaks = conn.execute(
            f"""SELECT COUNT(DISTINCT ctp1.{entity_col})
            FROM split_assignments sa1
            JOIN compound_target_pairs ctp1 ON sa1.pair_id = ctp1.pair_id
            WHERE sa1.split_id = ? AND sa1.fold = 'train'
            AND ctp1.{entity_col} IN (
                SELECT ctp2.{entity_col}
                FROM split_assignments sa2
                JOIN compound_target_pairs ctp2 ON sa2.pair_id = ctp2.pair_id
                WHERE sa2.split_id = ? AND sa2.fold = 'test'
            )""",
            (sid, sid),
        ).fetchone()[0]

        results[strategy] = {"split_id": sid, "leaks": leaks}

    return results


def check_cross_db_overlap(
    conn: sqlite3.Connection,
) -> dict:
    """Check overlap between sources at compound×target pair level.

    Returns overlap statistics between each pair of sources.
    """
    sources = [r[0] for r in conn.execute(
        "SELECT DISTINCT source_db FROM negative_results ORDER BY source_db"
    ).fetchall()]

    overlaps = {}
    for i, s1 in enumerate(sources):
        for s2 in sources[i + 1:]:
            count = conn.execute(
                """SELECT COUNT(DISTINCT nr1.compound_id || ':' || nr1.target_id)
                FROM negative_results nr1
                WHERE nr1.source_db = ?
                AND EXISTS (
                    SELECT 1 FROM negative_results nr2
                    WHERE nr2.compound_id = nr1.compound_id
                    AND nr2.target_id = nr1.target_id
                    AND nr2.source_db = ?
                )""",
                (s1, s2),
            ).fetchone()[0]
            overlaps[f"{s1}_vs_{s2}"] = count

    return overlaps


def generate_leakage_report(
    db_path: str | Path,
    output_path: str | Path | None = None,
) -> dict:
    """Generate comprehensive data leakage and integrity report.

    Checks:
    1. Cold split integrity (zero entity leakage)
    2. Split fold counts and ratios
    3. Database summary statistics

    Returns report dict, optionally writes to JSON file.
    """
    import json

    conn = sqlite3.connect(str(db_path))
    conn.execute("PRAGMA journal_mode = WAL")

    report: dict = {}

    # 1. DB summary
    report["db_summary"] = {
        "compounds": conn.execute("SELECT COUNT(*) FROM compounds").fetchone()[0],
        "targets": conn.execute("SELECT COUNT(*) FROM targets").fetchone()[0],
        "negative_results": conn.execute("SELECT COUNT(*) FROM negative_results").fetchone()[0],
        "pairs": conn.execute("SELECT COUNT(*) FROM compound_target_pairs").fetchone()[0],
    }

    # 2. Source breakdown
    source_counts = {}
    for source, cnt in conn.execute(
        "SELECT source_db, COUNT(*) FROM negative_results GROUP BY source_db"
    ).fetchall():
        source_counts[source] = cnt
    report["source_counts"] = source_counts

    # 3. Split summary
    split_summary = {}
    for sid, name, strategy in conn.execute(
        "SELECT split_id, split_name, split_strategy FROM split_definitions"
    ).fetchall():
        fold_counts = {}
        for fold, cnt in conn.execute(
            "SELECT fold, COUNT(*) FROM split_assignments WHERE split_id = ? GROUP BY fold",
            (sid,),
        ).fetchall():
            fold_counts[fold] = cnt
        total = sum(fold_counts.values())
        split_summary[name] = {
            "strategy": strategy,
            "fold_counts": fold_counts,
            "total": total,
            "ratios": {
                f: round(c / total, 4) if total > 0 else 0
                for f, c in fold_counts.items()
            },
        }
    report["splits"] = split_summary

    # 4. Cold split integrity
    report["cold_split_integrity"] = check_cold_split_integrity(conn)

    # 5. Cross-source overlap
    report["cross_source_overlap"] = check_cross_db_overlap(conn)

    # 6. Pairs by num_sources
    multi_source = {}
    for ns, cnt in conn.execute(
        "SELECT num_sources, COUNT(*) FROM compound_target_pairs GROUP BY num_sources ORDER BY num_sources"
    ).fetchall():
        multi_source[str(ns)] = cnt
    report["pairs_by_num_sources"] = multi_source

    conn.close()

    if output_path is not None:
        output_path = Path(output_path)
        output_path.parent.mkdir(parents=True, exist_ok=True)
        with open(output_path, "w") as f:
            json.dump(report, f, indent=2)
        logger.info("Leakage report written to %s", output_path)

    return report