Joblib
File size: 33,017 Bytes
baf3373
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
import os
import math
from pathlib import Path
import sys
from contextlib import contextmanager

import numpy as np
import pandas as pd
import torch

# tqdm is optional; we’ll disable it by default in notebooks
from tqdm import tqdm

sys.path.append("/vast/projects/pranam/lab/yz927/projects/Classifier_Weight")
from tokenizer.my_tokenizers import SMILES_SPE_Tokenizer

from datasets import Dataset, DatasetDict, Features, Value, Sequence as HFSequence
from transformers import AutoTokenizer, EsmModel, AutoModelForMaskedLM

# -------------------------
# Config
# -------------------------
CSV_PATH = Path("/vast/projects/pranam/lab/yz927/projects/Classifier_Weight/c-binding_with_openfold_scores.csv")

OUT_ROOT = Path(
    "/vast/projects/pranam/lab/yz927/projects/Classifier_Weight/training_data_cleaned/binding_affinity"
)

# WT (seq) embedding model
WT_MODEL_NAME = "facebook/esm2_t33_650M_UR50D"
WT_MAX_LEN = 1022
WT_BATCH = 32

# SMILES embedding model + tokenizer
SMI_MODEL_NAME = "aaronfeller/PeptideCLM-23M-all"
TOKENIZER_VOCAB = "/vast/projects/pranam/lab/yz927/projects/Classifier_Weight/tokenizer/new_vocab.txt"
TOKENIZER_SPLITS = "/vast/projects/pranam/lab/yz927/projects/Classifier_Weight/tokenizer/new_splits.txt"
SMI_MAX_LEN = 768
SMI_BATCH = 128

# Split config
TRAIN_FRAC = 0.80
RANDOM_SEED = 1986
AFFINITY_Q_BINS = 30

# Columns expected in CSV
COL_SEQ1 = "seq1"
COL_SEQ2 = "seq2"
COL_AFF = "affinity"
COL_F2S = "Fasta2SMILES"
COL_REACT = "REACT_SMILES"
COL_WT_IPTM = "wt_iptm_score"
COL_SMI_IPTM = "smiles_iptm_score"

# Device
DEVICE = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

# -------------------------
# Quiet / notebook-safe output controls
# -------------------------
QUIET = True       # suppress most prints
USE_TQDM = False   # disable tqdm bars (recommended in Jupyter to avoid crashing)
LOG_FILE = None    # optionally: OUT_ROOT / "build.log"

def log(msg: str):
    if LOG_FILE is not None:
        Path(LOG_FILE).parent.mkdir(parents=True, exist_ok=True)
        with open(LOG_FILE, "a") as f:
            f.write(msg.rstrip() + "\n")
    if not QUIET:
        print(msg)

def pbar(it, **kwargs):
    return tqdm(it, **kwargs) if USE_TQDM else it

@contextmanager
def section(title: str):
    log(f"\n=== {title} ===")
    yield
    log(f"=== done: {title} ===")


# -------------------------
# Helpers
# -------------------------
def has_uaa(seq: str) -> bool:
    return "X" in str(seq).upper()

def affinity_to_class(a: float) -> str:
    # High: >= 9 ; Moderate: [7, 9) ; Low: < 7
    if a >= 9.0:
        return "High"
    elif a >= 7.0:
        return "Moderate"
    else:
        return "Low"

def make_distribution_matched_split(df: pd.DataFrame) -> pd.DataFrame:
    df = df.copy()

    df[COL_AFF] = pd.to_numeric(df[COL_AFF], errors="coerce")
    df = df.dropna(subset=[COL_AFF]).reset_index(drop=True)

    df["affinity_class"] = df[COL_AFF].apply(affinity_to_class)

    try:
        df["aff_bin"] = pd.qcut(df[COL_AFF], q=AFFINITY_Q_BINS, duplicates="drop")
        strat_col = "aff_bin"
    except Exception:
        df["aff_bin"] = df["affinity_class"]
        strat_col = "aff_bin"

    rng = np.random.RandomState(RANDOM_SEED)

    df["split"] = None
    for _, g in df.groupby(strat_col, observed=True):
        idx = g.index.to_numpy()
        rng.shuffle(idx)
        n_train = int(math.floor(len(idx) * TRAIN_FRAC))
        df.loc[idx[:n_train], "split"] = "train"
        df.loc[idx[n_train:], "split"] = "val"

    df["split"] = df["split"].fillna("train")
    return df

def _summ(x):
    x = np.asarray(x, dtype=float)
    x = x[~np.isnan(x)]
    if len(x) == 0:
        return {"n": 0, "mean": np.nan, "std": np.nan, "p50": np.nan, "p95": np.nan}
    return {
        "n": int(len(x)),
        "mean": float(np.mean(x)),
        "std": float(np.std(x)),
        "p50": float(np.quantile(x, 0.50)),
        "p95": float(np.quantile(x, 0.95)),
    }

def _len_stats(seqs):
    lens = np.asarray([len(str(s)) for s in seqs], dtype=float)
    if len(lens) == 0:
        return {"n": 0, "mean": np.nan, "std": np.nan, "p50": np.nan, "p95": np.nan}
    return {
        "n": int(len(lens)),
        "mean": float(lens.mean()),
        "std": float(lens.std()),
        "p50": float(np.quantile(lens, 0.50)),
        "p95": float(np.quantile(lens, 0.95)),
    }

def verify_split_before_embedding(
    df2: pd.DataFrame,
    affinity_col: str,
    split_col: str,
    seq_col: str,
    iptm_col: str,
    aff_class_col: str = "affinity_class",
    aff_bins: int = 30,
    save_report_prefix: str | None = None,
    verbose: bool = False,
):
    """
    Notebook-safe: by default prints only ONE line via `log()`.
    Optionally writes CSV reports (stats + class proportions).
    """
    df2 = df2.copy()
    df2[affinity_col] = pd.to_numeric(df2[affinity_col], errors="coerce")
    df2[iptm_col] = pd.to_numeric(df2[iptm_col], errors="coerce")

    assert split_col in df2.columns, f"Missing split col: {split_col}"
    assert set(df2[split_col].dropna().unique()).issubset({"train", "val"}), f"Unexpected split values: {df2[split_col].unique()}"
    assert df2[affinity_col].notna().any(), "No valid affinity values after coercion."

    try:
        df2["_aff_bin_dbg"] = pd.qcut(df2[affinity_col], q=aff_bins, duplicates="drop")
    except Exception:
        df2["_aff_bin_dbg"] = df2[aff_class_col].astype(str)

    tr = df2[df2[split_col] == "train"].reset_index(drop=True)
    va = df2[df2[split_col] == "val"].reset_index(drop=True)

    tr_aff = _summ(tr[affinity_col].to_numpy())
    va_aff = _summ(va[affinity_col].to_numpy())
    tr_len = _len_stats(tr[seq_col].tolist())
    va_len = _len_stats(va[seq_col].tolist())

    # bin drift
    bin_ct = (
        df2.groupby([split_col, "_aff_bin_dbg"])
           .size()
           .groupby(level=0)
           .apply(lambda s: s / s.sum())
    )
    tr_bins = bin_ct.loc["train"]
    va_bins = bin_ct.loc["val"]
    all_bins = tr_bins.index.union(va_bins.index)
    tr_bins = tr_bins.reindex(all_bins, fill_value=0.0)
    va_bins = va_bins.reindex(all_bins, fill_value=0.0)
    max_bin_diff = float(np.max(np.abs(tr_bins.values - va_bins.values)))

    msg = (
        f"[split-check] rows={len(df2)} train={len(tr)} val={len(va)} | "
        f"aff(mean±std) train={tr_aff['mean']:.3f}±{tr_aff['std']:.3f} val={va_aff['mean']:.3f}±{va_aff['std']:.3f} | "
        f"len(p50/p95) train={tr_len['p50']:.1f}/{tr_len['p95']:.1f} val={va_len['p50']:.1f}/{va_len['p95']:.1f} | "
        f"max_bin_diff={max_bin_diff:.4f}"
    )
    log(msg)

    if verbose and (not QUIET):
        class_ct = df2.groupby([split_col, aff_class_col]).size().unstack(fill_value=0)
        class_prop = class_ct.div(class_ct.sum(axis=1), axis=0)
        print("\n[verbose] affinity_class counts:\n", class_ct)
        print("\n[verbose] affinity_class proportions:\n", class_prop.round(4))

    if save_report_prefix is not None:
        out = Path(save_report_prefix)
        out.parent.mkdir(parents=True, exist_ok=True)

        stats_df = pd.DataFrame([
            {"split": "train", **{f"aff_{k}": v for k, v in tr_aff.items()}, **{f"len_{k}": v for k, v in tr_len.items()}},
            {"split": "val",   **{f"aff_{k}": v for k, v in va_aff.items()}, **{f"len_{k}": v for k, v in va_len.items()}},
        ])
        class_ct = df2.groupby([split_col, aff_class_col]).size().unstack(fill_value=0)
        class_prop = class_ct.div(class_ct.sum(axis=1), axis=0).reset_index()

        stats_df.to_csv(out.with_suffix(".stats.csv"), index=False)
        class_prop.to_csv(out.with_suffix(".class_prop.csv"), index=False)


# -------------------------
# WT pooled (ESM2)
# -------------------------
@torch.no_grad()
def wt_pooled_embeddings(seqs, tokenizer, model, batch_size=32, max_length=1022):
    embs = []
    for i in pbar(range(0, len(seqs), batch_size)):
        batch = seqs[i:i + batch_size]
        inputs = tokenizer(
            batch,
            padding=True,
            truncation=True,
            max_length=max_length,
            return_tensors="pt",
        )
        inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
        out = model(**inputs)
        h = out.last_hidden_state  # (B, L, H)

        attn = inputs["attention_mask"].unsqueeze(-1)  # (B, L, 1)
        summed = (h * attn).sum(dim=1)                 # (B, H)
        denom = attn.sum(dim=1).clamp(min=1e-9)        # (B, 1)
        pooled = (summed / denom).detach().cpu().numpy()
        embs.append(pooled)

    return np.vstack(embs)


# -------------------------
# WT unpooled (ESM2)
# -------------------------
@torch.no_grad()
def wt_unpooled_one(seq, tokenizer, model, cls_id, eos_id, max_length=1022):
    tok = tokenizer(seq, padding=False, truncation=True, max_length=max_length, return_tensors="pt")
    tok = {k: v.to(DEVICE) for k, v in tok.items()}
    out = model(**tok)
    h = out.last_hidden_state[0]           # (L, H)
    attn = tok["attention_mask"][0].bool() # (L,)
    ids = tok["input_ids"][0]

    keep = attn.clone()
    if cls_id is not None:
        keep &= (ids != cls_id)
    if eos_id is not None:
        keep &= (ids != eos_id)

    return h[keep].detach().cpu().to(torch.float16).numpy()

def build_wt_unpooled_dataset(df_split: pd.DataFrame, out_dir: Path, tokenizer, model):
    """
    Expects df_split to have:
      - target_sequence  (seq1)
      - sequence         (binder seq2; WT binder)
      - label, affinity_class, COL_AFF, COL_WT_IPTM
    Saves a dataset where each row contains BOTH:
      - target_embedding (Lt,H), target_attention_mask, target_length
      - binder_embedding (Lb,H), binder_attention_mask, binder_length
    """
    cls_id = tokenizer.cls_token_id
    eos_id = tokenizer.eos_token_id
    H = model.config.hidden_size

    features = Features({
        "target_sequence": Value("string"),
        "sequence": Value("string"),
        "label": Value("float32"),
        "affinity": Value("float32"),
        "affinity_class": Value("string"),

        "target_embedding": HFSequence(HFSequence(Value("float16"), length=H)),
        "target_attention_mask": HFSequence(Value("int8")),
        "target_length": Value("int64"),

        "binder_embedding": HFSequence(HFSequence(Value("float16"), length=H)),
        "binder_attention_mask": HFSequence(Value("int8")),
        "binder_length": Value("int64"),

        COL_WT_IPTM: Value("float32"),
        COL_AFF: Value("float32"),
    })

    def gen_rows(df: pd.DataFrame):
        for r in pbar(df.itertuples(index=False), total=len(df)):
            tgt = str(getattr(r, "target_sequence")).strip()
            bnd = str(getattr(r, "sequence")).strip()

            y = float(getattr(r, "label"))
            aff = float(getattr(r, COL_AFF))
            acls = str(getattr(r, "affinity_class"))

            iptm = getattr(r, COL_WT_IPTM)
            iptm = float(iptm) if pd.notna(iptm) else np.nan

            # token embeddings for target + binder (both ESM)
            t_emb = wt_unpooled_one(tgt, tokenizer, model, cls_id, eos_id, max_length=WT_MAX_LEN)  # (Lt,H)
            b_emb = wt_unpooled_one(bnd, tokenizer, model, cls_id, eos_id, max_length=WT_MAX_LEN)  # (Lb,H)

            t_list = t_emb.tolist()
            b_list = b_emb.tolist()
            Lt = len(t_list)
            Lb = len(b_list)

            yield {
                "target_sequence": tgt,
                "sequence": bnd,
                "label": np.float32(y),
                "affinity": np.float32(aff),
                "affinity_class": acls,

                "target_embedding": t_list,
                "target_attention_mask": [1] * Lt,
                "target_length": int(Lt),

                "binder_embedding": b_list,
                "binder_attention_mask": [1] * Lb,
                "binder_length": int(Lb),

                COL_WT_IPTM: np.float32(iptm) if not np.isnan(iptm) else np.float32(np.nan),
                COL_AFF: np.float32(aff),
            }

    out_dir.mkdir(parents=True, exist_ok=True)
    ds = Dataset.from_generator(lambda: gen_rows(df_split), features=features)
    ds.save_to_disk(str(out_dir), max_shard_size="1GB")
    return ds

def build_smiles_unpooled_paired_dataset(df_split: pd.DataFrame, out_dir: Path, wt_tokenizer, wt_model_unpooled,
                                        smi_tok, smi_roformer):
    """
    df_split must have:
      - target_sequence (seq1)
      - sequence        (binder smiles string)
      - label, affinity_class, COL_AFF, COL_SMI_IPTM
    Saves rows with:
      target_embedding (Lt,Ht) from ESM
      binder_embedding (Lb,Hb) from PeptideCLM
    """
    cls_id = wt_tokenizer.cls_token_id
    eos_id = wt_tokenizer.eos_token_id
    Ht = wt_model_unpooled.config.hidden_size

    # Infer Hb from one forward pass? easiest: run one mini batch outside in main if you want.
    # Here: we’ll infer from model config if available.
    Hb = getattr(smi_roformer.config, "hidden_size", None)
    if Hb is None:
        Hb = getattr(smi_roformer.config, "dim", None)
    if Hb is None:
        raise ValueError("Cannot infer Hb from smi_roformer config; print(smi_roformer.config) and set Hb manually.")

    features = Features({
        "target_sequence": Value("string"),
        "sequence": Value("string"),
        "label": Value("float32"),
        "affinity": Value("float32"),
        "affinity_class": Value("string"),

        "target_embedding": HFSequence(HFSequence(Value("float16"), length=Ht)),
        "target_attention_mask": HFSequence(Value("int8")),
        "target_length": Value("int64"),

        "binder_embedding": HFSequence(HFSequence(Value("float16"), length=Hb)),
        "binder_attention_mask": HFSequence(Value("int8")),
        "binder_length": Value("int64"),

        COL_SMI_IPTM: Value("float32"),
        COL_AFF: Value("float32"),
    })

    def gen_rows(df: pd.DataFrame):
        for r in pbar(df.itertuples(index=False), total=len(df)):
            tgt = str(getattr(r, "target_sequence")).strip()
            bnd = str(getattr(r, "sequence")).strip()

            y = float(getattr(r, "label"))
            aff = float(getattr(r, COL_AFF))
            acls = str(getattr(r, "affinity_class"))

            iptm = getattr(r, COL_SMI_IPTM)
            iptm = float(iptm) if pd.notna(iptm) else np.nan

            # target token embeddings (ESM)
            t_emb = wt_unpooled_one(tgt, wt_tokenizer, wt_model_unpooled, cls_id, eos_id, max_length=WT_MAX_LEN)
            t_list = t_emb.tolist()
            Lt = len(t_list)

            # binder token embeddings (PeptideCLM) — single-item batch
            _, tok_list, mask_list, lengths = smiles_embed_batch_return_both(
                [bnd], smi_tok, smi_roformer, max_length=SMI_MAX_LEN
            )
            b_emb = tok_list[0]  # np.float16 (Lb, Hb)
            b_list = b_emb.tolist()
            Lb = int(lengths[0])
            b_mask = mask_list[0].astype(np.int8).tolist()

            yield {
                "target_sequence": tgt,
                "sequence": bnd,
                "label": np.float32(y),
                "affinity": np.float32(aff),
                "affinity_class": acls,

                "target_embedding": t_list,
                "target_attention_mask": [1] * Lt,
                "target_length": int(Lt),

                "binder_embedding": b_list,
                "binder_attention_mask": [int(x) for x in b_mask],
                "binder_length": int(Lb),

                COL_SMI_IPTM: np.float32(iptm) if not np.isnan(iptm) else np.float32(np.nan),
                COL_AFF: np.float32(aff),
            }

    out_dir.mkdir(parents=True, exist_ok=True)
    ds = Dataset.from_generator(lambda: gen_rows(df_split), features=features)
    ds.save_to_disk(str(out_dir), max_shard_size="1GB")
    return ds


# -------------------------
# SMILES pooled + unpooled (PeptideCLM)
# -------------------------
def get_special_ids(tokenizer_obj):
    cand = [
        getattr(tokenizer_obj, "pad_token_id", None),
        getattr(tokenizer_obj, "cls_token_id", None),
        getattr(tokenizer_obj, "sep_token_id", None),
        getattr(tokenizer_obj, "bos_token_id", None),
        getattr(tokenizer_obj, "eos_token_id", None),
        getattr(tokenizer_obj, "mask_token_id", None),
    ]
    return sorted({x for x in cand if x is not None})

@torch.no_grad()
def smiles_embed_batch_return_both(batch_sequences, tokenizer_obj, model_roformer, max_length):
    tok = tokenizer_obj(
        batch_sequences,
        return_tensors="pt",
        padding=True,
        truncation=True,
        max_length=max_length,
    )
    input_ids = tok["input_ids"].to(DEVICE)
    attention_mask = tok["attention_mask"].to(DEVICE)

    outputs = model_roformer(input_ids=input_ids, attention_mask=attention_mask)
    last_hidden = outputs.last_hidden_state  # (B, L, H)

    special_ids = get_special_ids(tokenizer_obj)
    valid = attention_mask.bool()
    if len(special_ids) > 0:
        sid = torch.tensor(special_ids, device=DEVICE, dtype=torch.long)
        if hasattr(torch, "isin"):
            valid = valid & (~torch.isin(input_ids, sid))
        else:
            m = torch.zeros_like(valid)
            for s in special_ids:
                m |= (input_ids == s)
            valid = valid & (~m)

    valid_f = valid.unsqueeze(-1).float()
    summed = torch.sum(last_hidden * valid_f, dim=1)
    denom = torch.clamp(valid_f.sum(dim=1), min=1e-9)
    pooled = (summed / denom).detach().cpu().numpy()

    token_emb_list, mask_list, lengths = [], [], []
    for b in range(last_hidden.shape[0]):
        emb = last_hidden[b, valid[b]]  # (Li, H)
        token_emb_list.append(emb.detach().cpu().to(torch.float16).numpy())
        li = emb.shape[0]
        lengths.append(int(li))
        mask_list.append(np.ones((li,), dtype=np.int8))

    return pooled, token_emb_list, mask_list, lengths

def smiles_generate_embeddings_batched_both(seqs, tokenizer_obj, model_roformer, batch_size, max_length):
    pooled_all = []
    token_emb_all = []
    mask_all = []
    lengths_all = []

    for i in pbar(range(0, len(seqs), batch_size)):
        batch = seqs[i:i + batch_size]
        pooled, tok_list, m_list, lens = smiles_embed_batch_return_both(
            batch, tokenizer_obj, model_roformer, max_length
        )
        pooled_all.append(pooled)
        token_emb_all.extend(tok_list)
        mask_all.extend(m_list)
        lengths_all.extend(lens)

    return np.vstack(pooled_all), token_emb_all, mask_all, lengths_all

# -------------------------
# Target embedding cache (NO extra ESM runs)
# We will compute target pooled embeddings ONCE from WT view, then reuse for SMILES.
# -------------------------
def build_target_cache_from_wt_view(wt_view_train: pd.DataFrame, wt_view_val: pd.DataFrame):
    wt_tok = AutoTokenizer.from_pretrained(WT_MODEL_NAME)
    wt_model = EsmModel.from_pretrained(WT_MODEL_NAME).to(DEVICE).eval()

    # compute target pooled embeddings once
    tgt_wt_train = wt_view_train["target_sequence"].astype(str).tolist()
    tgt_wt_val   = wt_view_val["target_sequence"].astype(str).tolist()

    wt_train_tgt_emb = wt_pooled_embeddings(
        tgt_wt_train, wt_tok, wt_model, batch_size=WT_BATCH, max_length=WT_MAX_LEN
    )
    wt_val_tgt_emb = wt_pooled_embeddings(
        tgt_wt_val, wt_tok, wt_model, batch_size=WT_BATCH, max_length=WT_MAX_LEN
    )

    # build dict: target_sequence -> embedding (float32 array)
    # if duplicates exist, last wins; you can add checks if needed
    train_map = {s: e for s, e in zip(tgt_wt_train, wt_train_tgt_emb)}
    val_map   = {s: e for s, e in zip(tgt_wt_val,   wt_val_tgt_emb)}
    return wt_tok, wt_model, wt_train_tgt_emb, wt_val_tgt_emb, train_map, val_map
# -------------------------
# Main
# -------------------------
def main():
    log(f"[INFO] DEVICE: {DEVICE}")
    OUT_ROOT.mkdir(parents=True, exist_ok=True)

    # 1) Load
    with section("load csv + dedup"):
        df = pd.read_csv(CSV_PATH)
        for c in [COL_SEQ1, COL_SEQ2, COL_F2S, COL_REACT]:
            if c in df.columns:
                df[c] = df[c].apply(lambda x: x.strip() if isinstance(x, str) else x)
        
        # Dedup on the full identity tuple you want
        DEDUP_COLS = [COL_SEQ1, COL_SEQ2, COL_F2S, COL_REACT]
        df = df.drop_duplicates(subset=DEDUP_COLS).reset_index(drop=True)
        
        print("Rows after dedup on", DEDUP_COLS, ":", len(df))

        need = [COL_SEQ1, COL_SEQ2, COL_AFF, COL_F2S, COL_REACT, COL_WT_IPTM, COL_SMI_IPTM]
        missing = [c for c in need if c not in df.columns]
        if missing:
            raise ValueError(f"Missing required columns: {missing}")

        # numeric affinity for both branches
        df[COL_AFF] = pd.to_numeric(df[COL_AFF], errors="coerce")

    # 2) Build WT subset + SMILES subset separately (NO global dropping)
    with section("prepare wt/smiles subsets"):
        # WT: requires a canonical peptide sequence (no X) + affinity
        df_wt = df.copy()
        df_wt["wt_sequence"] = df_wt[COL_SEQ2].astype(str).str.strip()
        df_wt = df_wt.dropna(subset=[COL_AFF]).reset_index(drop=True)
        df_wt = df_wt[df_wt["wt_sequence"].notna() & (df_wt["wt_sequence"] != "")]
        df_wt = df_wt[~df_wt["wt_sequence"].str.contains("X", case=False, na=False)].reset_index(drop=True)

        # SMILES: requires affinity + a usable picked SMILES (UAA->REACT, else->Fasta2SMILES)
        df_smi = df.copy()
        df_smi = df_smi.dropna(subset=[COL_AFF]).reset_index(drop=True)
        df_smi = df_smi[
            pd.to_numeric(df_smi[COL_SMI_IPTM], errors="coerce").notna()
        ].reset_index(drop=True) # empty iptm means sth wrong with their smiles sequenc

        is_uaa = df_smi[COL_SEQ2].astype(str).str.contains("X", case=False, na=False)
        df_smi["smiles_sequence"] = np.where(is_uaa, df_smi[COL_REACT], df_smi[COL_F2S])
        df_smi["smiles_sequence"] = df_smi["smiles_sequence"].astype(str).str.strip()
        df_smi = df_smi[df_smi["smiles_sequence"].notna() & (df_smi["smiles_sequence"] != "")]
        df_smi = df_smi[~df_smi["smiles_sequence"].isin(["nan", "None"])].reset_index(drop=True)

        log(f"[counts] WT rows={len(df_wt)} | SMILES rows={len(df_smi)} (after per-branch filtering)")

    # 3) Split separately (different sizes and memberships are expected)
    with section("split wt and smiles separately"):
        df_wt2 = make_distribution_matched_split(df_wt)
        df_smi2 = make_distribution_matched_split(df_smi)

        # save split tables
        wt_split_csv = OUT_ROOT / "binding_affinity_wt_meta_with_split.csv"
        smi_split_csv = OUT_ROOT / "binding_affinity_smiles_meta_with_split.csv"
        df_wt2.to_csv(wt_split_csv, index=False)
        df_smi2.to_csv(smi_split_csv, index=False)
        log(f"Saved WT split meta: {wt_split_csv}")
        log(f"Saved SMILES split meta: {smi_split_csv}")

        # lightweight double-check (one-line)
        verify_split_before_embedding(
            df2=df_wt2,
            affinity_col=COL_AFF,
            split_col="split",
            seq_col="wt_sequence",
            iptm_col=COL_WT_IPTM,
            aff_class_col="affinity_class",
            aff_bins=AFFINITY_Q_BINS,
            save_report_prefix=str(OUT_ROOT / "wt_split_doublecheck_report"),
            verbose=False,
        )
        verify_split_before_embedding(
            df2=df_smi2,
            affinity_col=COL_AFF,
            split_col="split",
            seq_col="smiles_sequence",
            iptm_col=COL_SMI_IPTM,
            aff_class_col="affinity_class",
            aff_bins=AFFINITY_Q_BINS,
            save_report_prefix=str(OUT_ROOT / "smiles_split_doublecheck_report"),
            verbose=False,
        )

    # Prepare split views
    def prep_view(df_in: pd.DataFrame, binder_seq_col: str, iptm_col: str) -> pd.DataFrame:
        out = df_in.copy()
        out["target_sequence"] = out[COL_SEQ1].astype(str).str.strip()   # <-- NEW
        out["sequence"] = out[binder_seq_col].astype(str).str.strip()   # binder
        out["label"] = pd.to_numeric(out[COL_AFF], errors="coerce")
        out[iptm_col] = pd.to_numeric(out[iptm_col], errors="coerce")
        out[COL_AFF] = pd.to_numeric(out[COL_AFF], errors="coerce")
        out = out.dropna(subset=["target_sequence", "sequence", "label"]).reset_index(drop=True)
        return out[["target_sequence", "sequence", "label", "split", iptm_col, COL_AFF, "affinity_class"]]

    wt_view = prep_view(df_wt2, "wt_sequence", COL_WT_IPTM)
    smi_view = prep_view(df_smi2, "smiles_sequence", COL_SMI_IPTM)

    # -------------------------
    # Split views
    # -------------------------
    wt_train = wt_view[wt_view["split"] == "train"].reset_index(drop=True)
    wt_val   = wt_view[wt_view["split"] == "val"].reset_index(drop=True)
    smi_train = smi_view[smi_view["split"] == "train"].reset_index(drop=True)
    smi_val   = smi_view[smi_view["split"] == "val"].reset_index(drop=True)
    
    
    # =========================
    # TARGET pooled embeddings (ESM) — SEPARATE per branch
    # =========================
    with section("TARGET pooled embeddings (ESM) — WT + SMILES separately"):
        wt_tok = AutoTokenizer.from_pretrained(WT_MODEL_NAME)
        wt_esm = EsmModel.from_pretrained(WT_MODEL_NAME).to(DEVICE).eval()
    
        # ---- WT targets ----
        wt_train_tgt_emb = wt_pooled_embeddings(
            wt_train["target_sequence"].astype(str).str.strip().tolist(),
            wt_tok, wt_esm,
            batch_size=WT_BATCH,
            max_length=WT_MAX_LEN,
        ).astype(np.float32)
    
        wt_val_tgt_emb = wt_pooled_embeddings(
            wt_val["target_sequence"].astype(str).str.strip().tolist(),
            wt_tok, wt_esm,
            batch_size=WT_BATCH,
            max_length=WT_MAX_LEN,
        ).astype(np.float32)
    
        # ---- SMILES targets (independent; may include UAA-only targets) ----
        smi_train_tgt_emb = wt_pooled_embeddings(
            smi_train["target_sequence"].astype(str).str.strip().tolist(),
            wt_tok, wt_esm,
            batch_size=WT_BATCH,
            max_length=WT_MAX_LEN,
        ).astype(np.float32)
    
        smi_val_tgt_emb = wt_pooled_embeddings(
            smi_val["target_sequence"].astype(str).str.strip().tolist(),
            wt_tok, wt_esm,
            batch_size=WT_BATCH,
            max_length=WT_MAX_LEN,
        ).astype(np.float32)
    
    
    # =========================
    # WT pooled binder embeddings (binder = WT peptide)
    # =========================
    with section("WT pooled binder embeddings + save"):
        wt_train_emb = wt_pooled_embeddings(
            wt_train["sequence"].astype(str).str.strip().tolist(),
            wt_tok, wt_esm,
            batch_size=WT_BATCH,
            max_length=WT_MAX_LEN,
        ).astype(np.float32)
    
        wt_val_emb = wt_pooled_embeddings(
            wt_val["sequence"].astype(str).str.strip().tolist(),
            wt_tok, wt_esm,
            batch_size=WT_BATCH,
            max_length=WT_MAX_LEN,
        ).astype(np.float32)
    
        wt_train_ds = Dataset.from_dict({
            "target_sequence": wt_train["target_sequence"].tolist(),
            "sequence": wt_train["sequence"].tolist(),
            "label": wt_train["label"].astype(float).tolist(),
            "target_embedding": wt_train_tgt_emb,
            "embedding": wt_train_emb,
            COL_WT_IPTM: wt_train[COL_WT_IPTM].astype(float).tolist(),
            COL_AFF: wt_train[COL_AFF].astype(float).tolist(),
            "affinity_class": wt_train["affinity_class"].tolist(),
        })
    
        wt_val_ds = Dataset.from_dict({
            "target_sequence": wt_val["target_sequence"].tolist(),
            "sequence": wt_val["sequence"].tolist(),
            "label": wt_val["label"].astype(float).tolist(),
            "target_embedding": wt_val_tgt_emb,
            "embedding": wt_val_emb,
            COL_WT_IPTM: wt_val[COL_WT_IPTM].astype(float).tolist(),
            COL_AFF: wt_val[COL_AFF].astype(float).tolist(),
            "affinity_class": wt_val["affinity_class"].tolist(),
        })
    
        wt_pooled_dd = DatasetDict({"train": wt_train_ds, "val": wt_val_ds})
        wt_pooled_out = OUT_ROOT / "pair_wt_wt_pooled"
        wt_pooled_dd.save_to_disk(str(wt_pooled_out))
        log(f"Saved WT pooled -> {wt_pooled_out}")
    
    
    # =========================
    # SMILES pooled binder embeddings (binder = SMILES via PeptideCLM)
    # =========================
    with section("SMILES pooled binder embeddings + save"):
        smi_tok = SMILES_SPE_Tokenizer(TOKENIZER_VOCAB, TOKENIZER_SPLITS)
        smi_roformer = (
            AutoModelForMaskedLM
            .from_pretrained(SMI_MODEL_NAME)
            .roformer
            .to(DEVICE)
            .eval()
        )
    
        smi_train_pooled, _, _, _ = smiles_generate_embeddings_batched_both(
            smi_train["sequence"].astype(str).str.strip().tolist(),
            smi_tok, smi_roformer,
            batch_size=SMI_BATCH,
            max_length=SMI_MAX_LEN,
        )
    
        smi_val_pooled, _, _, _ = smiles_generate_embeddings_batched_both(
            smi_val["sequence"].astype(str).str.strip().tolist(),
            smi_tok, smi_roformer,
            batch_size=SMI_BATCH,
            max_length=SMI_MAX_LEN,
        )
    
        smi_train_ds = Dataset.from_dict({
            "target_sequence": smi_train["target_sequence"].tolist(),
            "sequence": smi_train["sequence"].tolist(),
            "label": smi_train["label"].astype(float).tolist(),
            "target_embedding": smi_train_tgt_emb,
            "embedding": smi_train_pooled.astype(np.float32),
            COL_SMI_IPTM: smi_train[COL_SMI_IPTM].astype(float).tolist(),
            COL_AFF: smi_train[COL_AFF].astype(float).tolist(),
            "affinity_class": smi_train["affinity_class"].tolist(),
        })
    
        smi_val_ds = Dataset.from_dict({
            "target_sequence": smi_val["target_sequence"].tolist(),
            "sequence": smi_val["sequence"].tolist(),
            "label": smi_val["label"].astype(float).tolist(),
            "target_embedding": smi_val_tgt_emb,
            "embedding": smi_val_pooled.astype(np.float32),
            COL_SMI_IPTM: smi_val[COL_SMI_IPTM].astype(float).tolist(),
            COL_AFF: smi_val[COL_AFF].astype(float).tolist(),
            "affinity_class": smi_val["affinity_class"].tolist(),
        })
    
        smi_pooled_dd = DatasetDict({"train": smi_train_ds, "val": smi_val_ds})
        smi_pooled_out = OUT_ROOT / "pair_wt_smiles_pooled"
        smi_pooled_dd.save_to_disk(str(smi_pooled_out))
        log(f"Saved SMILES pooled -> {smi_pooled_out}")


        # =========================
    # WT unpooled paired (ESM target + ESM binder) + save
    # =========================
    with section("WT unpooled paired embeddings + save"):
        wt_tok_unpooled = wt_tok                       # reuse tokenizer
        wt_esm_unpooled = wt_esm                       # reuse model

        wt_unpooled_out = OUT_ROOT / "pair_wt_wt_unpooled"
        wt_unpooled_dd = DatasetDict({
            "train": build_wt_unpooled_dataset(wt_train, wt_unpooled_out / "train",
                                               wt_tok_unpooled, wt_esm_unpooled),
            "val":   build_wt_unpooled_dataset(wt_val,   wt_unpooled_out / "val",
                                               wt_tok_unpooled, wt_esm_unpooled),
        })
        # (Optional) also save as DatasetDict root if you want a single load_from_disk path:
        wt_unpooled_dd.save_to_disk(str(wt_unpooled_out))
        log(f"Saved WT unpooled -> {wt_unpooled_out}")


    # =========================
    # SMILES unpooled paired (ESM target + PeptideCLM binder) + save
    # =========================
    with section("SMILES unpooled paired embeddings + save"):
        # reuse already-loaded smi_tok/smi_roformer from pooled section if still in scope;
        # otherwise re-init here:
        # smi_tok = SMILES_SPE_Tokenizer(TOKENIZER_VOCAB, TOKENIZER_SPLITS)
        # smi_roformer = AutoModelForMaskedLM.from_pretrained(SMI_MODEL_NAME).roformer.to(DEVICE).eval()

        smi_unpooled_out = OUT_ROOT / "pair_wt_smiles_unpooled"
        smi_unpooled_dd = DatasetDict({
            "train": build_smiles_unpooled_paired_dataset(
                smi_train, smi_unpooled_out / "train",
                wt_tok, wt_esm,
                smi_tok, smi_roformer
            ),
            "val": build_smiles_unpooled_paired_dataset(
                smi_val, smi_unpooled_out / "val",
                wt_tok, wt_esm,
                smi_tok, smi_roformer
            ),
        })
        smi_unpooled_dd.save_to_disk(str(smi_unpooled_out))
        log(f"Saved SMILES unpooled -> {smi_unpooled_out}")

    log(f"\n[DONE] All datasets saved under: {OUT_ROOT}")


if __name__ == "__main__":
    main()