Skip to content

fasta_lake.inference.strategies#

fasta_lake.inference.strategies #

Four inference strategy functions — verbatim logic from e2e-tested scripts.

Each function takes an InferenceInput and returns an InferenceResult. The algorithms are preserved exactly as they passed 12/12 golden reference checks against the NM2026 dataset.

infer_species_budget #

infer_species_budget(inp: InferenceInput) -> InferenceResult

Species-budget parsimony: cross-species aware shared peptide resolution.

Algorithm
  1. Classify peptides as unique (1 protein) or shared (>1 protein).
  2. For within-species shared peptides: razor to protein with most unique peptides.
  3. For cross-species shared peptides: keep best representative from EACH species.
  4. Select all proteins with at least 1 assigned peptide.

When de novo scores are available (inp.peptide_scores), tie-breaking uses score-weighted unique peptide evidence instead of raw counts. A protein supported by 5 peptides at score 0.95 ranks higher than one supported by 5 peptides at score 0.70.

Parameters:

Name Type Description Default
inp InferenceInput

Pre-computed peptide-protein mappings and sequences.

 required

Returns:

Type Description
InferenceResult

Selected protein IDs and statistics.
Source code in fasta_lake/inference/strategies.py 
 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
def infer_species_budget(inp: InferenceInput) -> InferenceResult:
    """
    Species-budget parsimony: cross-species aware shared peptide resolution.

    Algorithm
    ---------
    1. Classify peptides as unique (1 protein) or shared (>1 protein).
    2. For within-species shared peptides: razor to protein with most unique peptides.
    3. For cross-species shared peptides: keep best representative from EACH species.
    4. Select all proteins with at least 1 assigned peptide.

    When de novo scores are available (``inp.peptide_scores``), tie-breaking
    uses score-weighted unique peptide evidence instead of raw counts.
    A protein supported by 5 peptides at score 0.95 ranks higher than
    one supported by 5 peptides at score 0.70.

    Parameters
    ----------
    inp : InferenceInput
        Pre-computed peptide-protein mappings and sequences.

    Returns
    -------
    InferenceResult
        Selected protein IDs and statistics.
    """
    pep2prot, prot2pep = inp.pep2prot, inp.prot2pep
    scores = inp.peptide_scores
    use_scores = bool(scores)
    tax_map = inp.taxonomy_map
    do_pruning = inp.taxonomy_pruning and bool(tax_map)

    # Step 1: Group proteins by species
    # Use taxonomy_map if provided (resolves assembly proteins to species),
    # otherwise fall back to get_species_id (parses from protein ID format).
    protein_species: dict[str, str] = {}
    species_proteins: dict[str, set[str]] = defaultdict(set)
    n_from_map = 0
    n_from_id = 0
    unresolved: list[str] = []
    for protein_id in prot2pep:
        if tax_map and protein_id in tax_map:
            species = tax_map[protein_id]
            n_from_map += 1
        else:
            species = get_species_id(protein_id)
            if species is None:
                unresolved.append(protein_id)
                continue
            n_from_id += 1
        protein_species[protein_id] = species
        species_proteins[species].add(protein_id)

    if unresolved:
        # A species budget is undefined for a protein whose species is unknown:
        # pooling such proteins into one bucket silently ran razor under the
        # name species_budget on every content-hash lake. Refuse instead.
        unresolved.sort()
        shown = ", ".join(unresolved[:5]) + (" …" if len(unresolved) > 5 else "")
        raise ValueError(
            f"species_budget: {len(unresolved)} of {len(prot2pep)} proteins with evidence "
            f"have no resolvable species (e.g. {shown}). Provide a taxonomy map "
            "(--taxonomy-map, protein_id<TAB>species) covering them, or use "
            "--strategy razor / uniform, which do not need species."
        )

    # Step 2: Count unique peptides per protein (and score-weighted sum)
    unique_count: Counter[str] = Counter()
    unique_score_sum: dict[str, float] = defaultdict(float)
    unique_peptides: set[str] = set()
    for pep, prots in pep2prot.items():
        if len(prots) == 1:
            prot = next(iter(prots))
            unique_count[prot] += 1
            unique_score_sum[prot] += scores.get(pep, 1.0)
            unique_peptides.add(pep)

    # Tie-breaking key: score-weighted if scores available, count-based otherwise
    def _protein_sort_key(p: str) -> tuple:
        """Rank by optional unique-score support, then evidence counts and accession."""
        if use_scores:
            return (
                -unique_score_sum.get(p, 0.0),
                -unique_count[p],
                -len(prot2pep.get(p, set())),
                p,
            )
        return (
            -unique_count[p],
            -len(prot2pep.get(p, set())),
            p,
        )

    # Step 2b: Species abundance profile (for optional taxonomy-based pruning)
    species_unique_count: Counter[str] = Counter()
    for prot, count in unique_count.items():
        species_unique_count[protein_species[prot]] += count

    # Step 3: Assign peptides using species budget logic
    protein_assigned: dict[str, set[str]] = defaultdict(set)
    n_unique = 0
    n_within_species = 0
    n_cross_species = 0
    n_cross_species_pruned = 0

    for pep, prots in pep2prot.items():
        if len(prots) == 1:
            prot = next(iter(prots))
            protein_assigned[prot].add(pep)
            n_unique += 1
        else:
            species_set = {protein_species[p] for p in prots}
            if len(species_set) == 1:
                best = sorted(prots, key=_protein_sort_key)[0]
                protein_assigned[best].add(pep)
                n_within_species += 1
            else:
                # Optional: frequentist taxonomy pruning.
                # Only active when taxonomy_map is provided AND taxonomy_pruning=True.
                # Uses unique peptide counts per species as abundance prior.
                # Species with < 1% of the dominant species' count are excluded.
                if do_pruning:
                    max_count = max(species_unique_count.get(sp, 0) for sp in species_set)
                    min_threshold = max(1, int(max_count * 0.01))
                    qualified = {
                        sp for sp in species_set if species_unique_count.get(sp, 0) >= min_threshold
                    }
                    if not qualified:
                        qualified = species_set
                    if len(qualified) < len(species_set):
                        n_cross_species_pruned += 1
                else:
                    qualified = species_set

                for species in qualified:
                    species_prots = [p for p in prots if protein_species[p] == species]
                    best = sorted(species_prots, key=_protein_sort_key)[0]
                    protein_assigned[best].add(pep)
                n_cross_species += 1

    # Step 4: Select proteins with assigned peptides
    selected = set(protein_assigned.keys())

    # Database breakdown
    db_stats: dict[str, dict[str, int]] = defaultdict(lambda: {"unique": 0, "shared_only": 0})
    for protein_id in selected:
        db_type = get_db_type(protein_id)
        if unique_count[protein_id] > 0:
            db_stats[db_type]["unique"] += 1
        else:
            db_stats[db_type]["shared_only"] += 1

    species_retained = {protein_species[p] for p in selected}

    proteins_with_unique = sum(1 for c in unique_count.values() if c > 0)

    stats = {
        "strategy": "species_budget",
        "score_weighted": use_scores,
        "taxonomy_map_provided": bool(tax_map),
        "taxonomy_pruning": do_pruning,
        "proteins_with_evidence": len(prot2pep),
        "selected_proteins": len(selected),
        "unique_peptides": len(unique_peptides),
        "shared_peptides": len(pep2prot) - len(unique_peptides),
        "within_species_shared": n_within_species,
        "cross_species_shared": n_cross_species,
        "cross_species_pruned": n_cross_species_pruned,
        "proteins_with_unique_peptides": proteins_with_unique,
        "species_detected": len(species_proteins),
        "species_retained": len(species_retained),
        "species_from_taxonomy_map": n_from_map,
        "species_from_protein_id": n_from_id,
        "db_breakdown": {db: dict(s) for db, s in db_stats.items()},
    }

    return InferenceResult(selected_proteins=selected, stats=stats)

infer_razor #

infer_razor(inp: InferenceInput) -> InferenceResult

Razor parsimony: assign shared peptides to protein with most unique peptides.

This is the classical MaxQuant approach used in standard proteomics. When de novo scores are available, uses score-weighted tie-breaking.

Parameters:

Name Type Description Default
inp InferenceInput

Pre-computed peptide-protein mappings and sequences.

 required

Returns:

Type Description
InferenceResult

Selected protein IDs and statistics.
Source code in fasta_lake/inference/strategies.py 
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
def infer_razor(inp: InferenceInput) -> InferenceResult:
    """
    Razor parsimony: assign shared peptides to protein with most unique peptides.

    This is the classical MaxQuant approach used in standard proteomics.
    When de novo scores are available, uses score-weighted tie-breaking.

    Parameters
    ----------
    inp : InferenceInput
        Pre-computed peptide-protein mappings and sequences.

    Returns
    -------
    InferenceResult
        Selected protein IDs and statistics.
    """
    pep2prot, prot2pep = inp.pep2prot, inp.prot2pep
    scores = inp.peptide_scores
    use_scores = bool(scores)

    # Count unique peptides per protein (and score-weighted sum)
    unique_count: Counter[str] = Counter()
    unique_score_sum: dict[str, float] = defaultdict(float)
    unique_peptides: set[str] = set()
    for pep, prots in pep2prot.items():
        if len(prots) == 1:
            prot = next(iter(prots))
            unique_count[prot] += 1
            unique_score_sum[prot] += scores.get(pep, 1.0)
            unique_peptides.add(pep)

    def _protein_sort_key(p: str) -> tuple:
        """Rank the Python razor candidates by support, with lexical accession ties."""
        if use_scores:
            return (
                -unique_score_sum.get(p, 0.0),
                -unique_count[p],
                -len(prot2pep.get(p, set())),
                p,
            )
        return (
            -unique_count[p],
            -len(prot2pep.get(p, set())),
            p,
        )

    # Assign peptides using razor logic
    protein_assigned: dict[str, set[str]] = defaultdict(set)
    for pep, prots in pep2prot.items():
        if len(prots) == 1:
            prot = next(iter(prots))
            protein_assigned[prot].add(pep)
        else:
            best = sorted(prots, key=_protein_sort_key)[0]
            protein_assigned[best].add(pep)

    selected = set(protein_assigned.keys())

    # Database breakdown
    db_stats: dict[str, dict[str, int]] = defaultdict(lambda: {"unique": 0, "shared_only": 0})
    for protein_id in selected:
        db_type = get_db_type(protein_id)
        if unique_count[protein_id] > 0:
            db_stats[db_type]["unique"] += 1
        else:
            db_stats[db_type]["shared_only"] += 1

    proteins_with_unique = sum(1 for c in unique_count.values() if c > 0)

    stats = {
        "strategy": "razor",
        "proteins_with_evidence": len(prot2pep),
        "selected_proteins": len(selected),
        "unique_peptides": len(unique_peptides),
        "shared_peptides": len(pep2prot) - len(unique_peptides),
        "proteins_with_unique_peptides": proteins_with_unique,
        "db_breakdown": {db: dict(s) for db, s in db_stats.items()},
    }

    return InferenceResult(selected_proteins=selected, stats=stats)

infer_uniform #

infer_uniform(inp: InferenceInput) -> InferenceResult

Uniform parsimony: keep ALL proteins with any peptide evidence.

Baseline approach with no filtering. Useful for comparison and for letting downstream tools (e.g., SAGE) handle protein inference.

Parameters:

Name Type Description Default
inp InferenceInput

Pre-computed peptide-protein mappings and sequences.

 required

Returns:

Type Description
InferenceResult

Selected protein IDs and statistics.
Source code in fasta_lake/inference/strategies.py 
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
def infer_uniform(inp: InferenceInput) -> InferenceResult:
    """
    Uniform parsimony: keep ALL proteins with any peptide evidence.

    Baseline approach with no filtering. Useful for comparison and
    for letting downstream tools (e.g., SAGE) handle protein inference.

    Parameters
    ----------
    inp : InferenceInput
        Pre-computed peptide-protein mappings and sequences.

    Returns
    -------
    InferenceResult
        Selected protein IDs and statistics.
    """
    pep2prot, prot2pep = inp.pep2prot, inp.prot2pep

    unique_peptides: set[str] = set()
    shared_peptides: set[str] = set()
    for pep, prots in pep2prot.items():
        if len(prots) == 1:
            unique_peptides.add(pep)
        else:
            shared_peptides.add(pep)

    selected = set(prot2pep.keys())

    # Database breakdown
    db_stats: dict[str, dict[str, int]] = defaultdict(lambda: {"unique": 0, "shared_only": 0})
    for protein_id in selected:
        db_type = get_db_type(protein_id)
        has_unique = any(len(pep2prot[pep]) == 1 for pep in prot2pep[protein_id])
        if has_unique:
            db_stats[db_type]["unique"] += 1
        else:
            db_stats[db_type]["shared_only"] += 1

    stats = {
        "strategy": "uniform",
        "proteins_with_evidence": len(prot2pep),
        "selected_proteins": len(selected),
        "unique_peptides": len(unique_peptides),
        "shared_peptides": len(shared_peptides),
        "db_breakdown": {db: dict(s) for db, s in db_stats.items()},
    }

    return InferenceResult(selected_proteins=selected, stats=stats)

infer_uniform_2pep #

infer_uniform_2pep(inp: InferenceInput, min_peptides: int = 2) -> InferenceResult

Uniform 2-peptide: require at least min_peptides per protein.

More conservative than uniform — reduces false positives by requiring multiple peptide evidence.

Parameters:

Name Type Description Default
inp InferenceInput

Pre-computed peptide-protein mappings and sequences.

 required
min_peptides int

Minimum number of peptides required (default: 2).

 2

Returns:

Type Description
InferenceResult

Selected protein IDs and statistics.
Source code in fasta_lake/inference/strategies.py 
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
def infer_uniform_2pep(
    inp: InferenceInput,
    min_peptides: int = 2,
) -> InferenceResult:
    """
    Uniform 2-peptide: require at least ``min_peptides`` per protein.

    More conservative than uniform — reduces false positives by
    requiring multiple peptide evidence.

    Parameters
    ----------
    inp : InferenceInput
        Pre-computed peptide-protein mappings and sequences.
    min_peptides : int
        Minimum number of peptides required (default: 2).

    Returns
    -------
    InferenceResult
        Selected protein IDs and statistics.
    """
    pep2prot, prot2pep = inp.pep2prot, inp.prot2pep

    unique_peptides: set[str] = set()
    shared_peptides: set[str] = set()
    for pep, prots in pep2prot.items():
        if len(prots) == 1:
            unique_peptides.add(pep)
        else:
            shared_peptides.add(pep)

    selected = {prot for prot, peps in prot2pep.items() if len(peps) >= min_peptides}
    excluded = len(prot2pep) - len(selected)

    # Database breakdown
    db_stats: dict[str, dict[str, int]] = defaultdict(lambda: {"selected": 0, "excluded": 0})
    for protein_id, peps in prot2pep.items():
        db_type = get_db_type(protein_id)
        if protein_id in selected:
            db_stats[db_type]["selected"] += 1
        else:
            db_stats[db_type]["excluded"] += 1

    stats = {
        "strategy": "uniform_2pep",
        "min_peptides": min_peptides,
        "proteins_with_evidence": len(prot2pep),
        "selected_proteins": len(selected),
        "excluded_proteins": excluded,
        "unique_peptides": len(unique_peptides),
        "shared_peptides": len(shared_peptides),
        "db_breakdown": {db: dict(s) for db, s in db_stats.items()},
    }

    return InferenceResult(selected_proteins=selected, stats=stats)

infer_info_score #

infer_info_score(inp: InferenceInput, min_info: float = 0.1) -> InferenceResult

Information-content parsimony: keep proteins above an info score threshold.

Each protein gets an info score that quantifies how much informative evidence supports it from the de novo predictions:

info_score(protein) = sum over matching peptides of:
    denovo_score(peptide) / n_proteins_matching(peptide)

A peptide with score 0.95 matching 1 protein contributes 0.95. The same peptide matching 500 proteins contributes 0.0019 each. High-confidence unique peptides dominate; shared peptides contribute proportionally to their ambiguity.

This avoids the problems of: - razor: arbitrarily picks one winner, loses real species - uniform_2pep: keeps millions of proteins with shared-only evidence - species_budget: requires species-level taxonomy (not always available)

The threshold min_info controls the tradeoff: - 0.01: ~200K proteins (broad, includes low-evidence) - 0.10: ~50K proteins (balanced, good for SAGE) - 0.50: ~6K proteins (very clean, may miss real low-abundance species)

When no scores are available, falls back to: 1/n_proteins per peptide.

Parameters:

Name Type Description Default
inp InferenceInput

Pre-computed peptide-protein mappings and sequences.

 required
min_info float

Minimum information score threshold (default: 0.1).

 0.1

Returns:

Type Description
InferenceResult

Selected protein IDs and statistics.
Source code in fasta_lake/inference/strategies.py 
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
def infer_info_score(
    inp: InferenceInput,
    min_info: float = 0.1,
) -> InferenceResult:
    """
    Information-content parsimony: keep proteins above an info score threshold.

    Each protein gets an info score that quantifies how much *informative*
    evidence supports it from the de novo predictions:

        info_score(protein) = sum over matching peptides of:
            denovo_score(peptide) / n_proteins_matching(peptide)

    A peptide with score 0.95 matching 1 protein contributes 0.95.
    The same peptide matching 500 proteins contributes 0.0019 each.
    High-confidence unique peptides dominate; shared peptides contribute
    proportionally to their ambiguity.

    This avoids the problems of:
    - razor: arbitrarily picks one winner, loses real species
    - uniform_2pep: keeps millions of proteins with shared-only evidence
    - species_budget: requires species-level taxonomy (not always available)

    The threshold ``min_info`` controls the tradeoff:
    - 0.01: ~200K proteins (broad, includes low-evidence)
    - 0.10: ~50K proteins  (balanced, good for SAGE)
    - 0.50: ~6K proteins   (very clean, may miss real low-abundance species)

    When no scores are available, falls back to: 1/n_proteins per peptide.

    Parameters
    ----------
    inp : InferenceInput
        Pre-computed peptide-protein mappings and sequences.
    min_info : float
        Minimum information score threshold (default: 0.1).

    Returns
    -------
    InferenceResult
        Selected protein IDs and statistics.
    """
    pep2prot, prot2pep = inp.pep2prot, inp.prot2pep
    scores = inp.peptide_scores
    use_scores = bool(scores)

    # Compute info score per protein
    info_scores: dict[str, float] = {}
    for prot_id, peps in prot2pep.items():
        info = 0.0
        # DETERMINISM FIX: `peps` is a set of peptide strings, so the summation
        # order depended on the per-process hash seed. Float addition is not
        # associative, so info_scores differed in the low bits between runs --
        # verified: reported info_score p75 read 7.4578 on 9/11 hash seeds and
        # 7.4577 on the other 2, and every seed produced a different digest of
        # the full info_scores vector. Summing in sorted peptide order makes the
        # value a function of the peptide set alone. (The `>= min_info` cut can
        # also flip for a protein sitting exactly on the threshold.)
        for pep in sorted(peps):
            n_match = len(pep2prot.get(pep, set()))
            if n_match == 0:
                continue
            denovo_score = scores.get(pep, 1.0) if use_scores else 1.0
            info += denovo_score / n_match
        info_scores[prot_id] = info

    # Select proteins above threshold
    selected = {p for p, score in info_scores.items() if score >= min_info}

    # Compute unique vs shared peptide stats
    unique_peptides: set[str] = set()
    for pep, prots in pep2prot.items():
        if len(prots) == 1:
            unique_peptides.add(pep)

    # Proteins with unique evidence in selected set
    selected_with_unique = sum(
        1 for p in selected if any(len(pep2prot[pep]) == 1 for pep in prot2pep[p])
    )

    # Database breakdown
    db_stats: dict[str, dict[str, int]] = defaultdict(lambda: {"selected": 0, "excluded": 0})
    for protein_id in prot2pep:
        db_type = get_db_type(protein_id)
        if protein_id in selected:
            db_stats[db_type]["selected"] += 1
        else:
            db_stats[db_type]["excluded"] += 1

    # Info score distribution of selected
    selected_scores = sorted([info_scores[p] for p in selected], reverse=True)
    percentiles = {}
    if selected_scores:
        import statistics

        percentiles = {
            "median": round(statistics.median(selected_scores), 4),
            "p25": round(selected_scores[int(len(selected_scores) * 0.75)], 4),
            "p75": round(selected_scores[int(len(selected_scores) * 0.25)], 4),
            "max": round(selected_scores[0], 4),
            "min": round(selected_scores[-1], 4),
        }

    stats = {
        "strategy": "info_score",
        "min_info": min_info,
        "score_weighted": use_scores,
        "proteins_with_evidence": len(prot2pep),
        "selected_proteins": len(selected),
        "excluded_proteins": len(prot2pep) - len(selected),
        "unique_peptides": len(unique_peptides),
        "shared_peptides": len(pep2prot) - len(unique_peptides),
        "selected_with_unique_peptides": selected_with_unique,
        "info_score_percentiles": percentiles,
        "db_breakdown": {db: dict(s) for db, s in db_stats.items()},
    }

    return InferenceResult(selected_proteins=selected, stats=stats)

infer_bayesian #

infer_bayesian(inp: InferenceInput, n_iterations: int = 10, min_probability: float = 0.01) -> InferenceResult

Bayesian protein probability: iterative evidence redistribution.

Models the probability that each protein is truly present in the sample, using de novo peptide evidence. Unlike razor (which forces a binary winner/loser decision for shared peptides), this distributes evidence proportionally and iteratively refines probabilities.

Algorithm (Expectation-Maximization)

Initialization: Each protein starts with a prior probability proportional to its total peptide evidence (number of matching peptides, weighted by de novo confidence scores). This gives proteins with more evidence a head start, but doesn't commit to winners yet.

E-step (expectation): For each shared peptide, distribute its evidence across matching proteins proportional to their current probabilities:

weight(protein, peptide) = P(protein) / sum(P(p) for p matching peptide)

A protein with high probability "claims" more of each shared peptide. Unique peptides always contribute fully to their single protein.

M-step (maximization): Update each protein's probability based on its total claimed evidence, normalized by protein length to avoid bias toward long proteins:

P(protein) ∝ sum(denovo_score(pep) × weight(protein, pep)) / sqrt(length)

The sqrt(length) normalization is gentler than dividing by length (which would over-penalize long proteins) but still corrects for the fact that longer proteins match more peptides by chance.

Convergence: Repeat E/M steps until probabilities stabilize (typically 5-10 iterations). Proteins below min_probability are excluded.

Why this is better than info_score

  1. Iterative refinement: info_score computes 1/n_matching once. Bayesian iterates — as weak proteins lose probability, strong proteins claim more of the shared peptides, amplifying signal.

  2. Length normalization: avoids bias toward long proteins that match more peptides by chance.

  3. Convergent evidence: if 5 peptides each match 50 proteins but all 5 share the same top protein, that protein accumulates evidence across iterations. Info_score misses this synergy.

  4. Soft decisions: doesn't force winner-take-all like razor. Two closely related strains with similar evidence both survive at moderate probability rather than one being eliminated.

Parameters:

Name Type Description Default
inp InferenceInput

Pre-computed peptide-protein mappings and sequences.

 required
n_iterations int

Number of EM iterations (default: 10).

 10
min_probability float

Minimum probability threshold for inclusion (default: 0.01).

 0.01

Returns:

Type Description
InferenceResult

Selected protein IDs and statistics.
Source code in fasta_lake/inference/strategies.py 
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
def infer_bayesian(
    inp: InferenceInput,
    n_iterations: int = 10,
    min_probability: float = 0.01,
) -> InferenceResult:
    """
    Bayesian protein probability: iterative evidence redistribution.

    Models the probability that each protein is truly present in the sample,
    using de novo peptide evidence. Unlike razor (which forces a binary
    winner/loser decision for shared peptides), this distributes evidence
    proportionally and iteratively refines probabilities.

    Algorithm (Expectation-Maximization)
    ------------------------------------

    **Initialization:**
    Each protein starts with a prior probability proportional to its
    total peptide evidence (number of matching peptides, weighted by
    de novo confidence scores). This gives proteins with more evidence
    a head start, but doesn't commit to winners yet.

    **E-step (expectation):**
    For each shared peptide, distribute its evidence across matching
    proteins proportional to their current probabilities:

        weight(protein, peptide) = P(protein) / sum(P(p) for p matching peptide)

    A protein with high probability "claims" more of each shared peptide.
    Unique peptides always contribute fully to their single protein.

    **M-step (maximization):**
    Update each protein's probability based on its total claimed evidence,
    normalized by protein length to avoid bias toward long proteins:

        P(protein) ∝ sum(denovo_score(pep) × weight(protein, pep)) / sqrt(length)

    The sqrt(length) normalization is gentler than dividing by length
    (which would over-penalize long proteins) but still corrects for
    the fact that longer proteins match more peptides by chance.

    **Convergence:**
    Repeat E/M steps until probabilities stabilize (typically 5-10
    iterations). Proteins below ``min_probability`` are excluded.

    Why this is better than info_score
    -----------------------------------
    1. **Iterative refinement:** info_score computes 1/n_matching once.
       Bayesian iterates — as weak proteins lose probability, strong
       proteins claim more of the shared peptides, amplifying signal.

    2. **Length normalization:** avoids bias toward long proteins that
       match more peptides by chance.

    3. **Convergent evidence:** if 5 peptides each match 50 proteins
       but all 5 share the same top protein, that protein accumulates
       evidence across iterations. Info_score misses this synergy.

    4. **Soft decisions:** doesn't force winner-take-all like razor.
       Two closely related strains with similar evidence both survive
       at moderate probability rather than one being eliminated.

    Parameters
    ----------
    inp : InferenceInput
        Pre-computed peptide-protein mappings and sequences.
    n_iterations : int
        Number of EM iterations (default: 10).
    min_probability : float
        Minimum probability threshold for inclusion (default: 0.01).

    Returns
    -------
    InferenceResult
        Selected protein IDs and statistics.
    """
    import math

    pep2prot, prot2pep = inp.pep2prot, inp.prot2pep
    proteins = inp.proteins
    scores = inp.peptide_scores
    use_scores = bool(scores)

    all_prots = list(prot2pep.keys())
    if not all_prots:
        return InferenceResult(selected_proteins=set(), stats={"strategy": "bayesian"})

    # Protein lengths (for normalization)
    prot_len = {}
    for pid in all_prots:
        seq = proteins.get(pid, "")
        prot_len[pid] = max(len(seq), 50)  # floor at 50 to avoid division issues

    # Initialize: prior proportional to total score-weighted peptide count
    prob: dict[str, float] = {}
    for pid in all_prots:
        # DETERMINISM GUARD: prot2pep[pid] is a set of peptide strings, so the
        # summation order depended on the per-process hash seed and float
        # addition is not associative. Sum in sorted peptide order.
        raw = (
            sum(scores.get(pep, 1.0) for pep in sorted(prot2pep[pid]))
            if use_scores
            else len(prot2pep[pid])
        )
        prob[pid] = raw / math.sqrt(prot_len[pid])

    # Normalize to sum to 1
    total = sum(prob.values())
    if total > 0:
        for pid in prob:
            prob[pid] /= total

    convergence_history = []

    # EM iterations
    for iteration in range(n_iterations):
        # E-step: compute weights for shared peptides
        new_evidence: dict[str, float] = defaultdict(float)

        for pep, matching_prots in pep2prot.items():
            denovo_score = scores.get(pep, 1.0) if use_scores else 1.0

            if len(matching_prots) == 1:
                # Unique peptide: full evidence to its protein
                pid = next(iter(matching_prots))
                new_evidence[pid] += denovo_score
            else:
                # Shared peptide: distribute proportional to current probabilities
                # DETERMINISM GUARD: matching_prots is a set of protein-ID
                # strings; the float sum below was therefore hash-seed ordered.
                # Sort once and reuse for the sum and both accumulation loops.
                ordered_prots = sorted(matching_prots)
                prob_sum = sum(prob.get(p, 0) for p in ordered_prots)
                if prob_sum == 0:
                    # Equal split if all probabilities are zero
                    share = denovo_score / len(ordered_prots)
                    for p in ordered_prots:
                        new_evidence[p] += share
                else:
                    for p in ordered_prots:
                        weight = prob.get(p, 0) / prob_sum
                        new_evidence[p] += denovo_score * weight

        # M-step: update probabilities with length normalization
        old_prob = dict(prob)
        for pid in all_prots:
            prob[pid] = new_evidence.get(pid, 0) / math.sqrt(prot_len[pid])

        # Normalize
        total = sum(prob.values())
        if total > 0:
            for pid in prob:
                prob[pid] /= total

        # Track convergence (max change)
        max_delta = max(abs(prob.get(pid, 0) - old_prob.get(pid, 0)) for pid in all_prots)
        convergence_history.append(max_delta)

        if max_delta < 1e-8:
            break

    # Select proteins above threshold
    # Scale probabilities so the maximum is 1.0 for interpretability
    max_prob = max(prob.values()) if prob else 1.0
    if max_prob > 0:
        scaled = {pid: prob[pid] / max_prob for pid in prob}
    else:
        scaled = prob

    selected = {pid for pid, p in scaled.items() if p >= min_probability}

    # Stats
    unique_peptides = sum(1 for pep, prots in pep2prot.items() if len(prots) == 1)
    selected_with_unique = sum(
        1 for p in selected if any(len(pep2prot[pep]) == 1 for pep in prot2pep[p])
    )

    # Database breakdown
    db_stats: dict[str, dict[str, int]] = defaultdict(lambda: {"selected": 0, "excluded": 0})
    for pid in prot2pep:
        db_type = get_db_type(pid)
        if pid in selected:
            db_stats[db_type]["selected"] += 1
        else:
            db_stats[db_type]["excluded"] += 1

    # Probability distribution of selected
    selected_probs = sorted([scaled[p] for p in selected], reverse=True)
    percentiles = {}
    if selected_probs:
        import statistics

        percentiles = {
            "median": round(statistics.median(selected_probs), 6),
            "p25": round(selected_probs[int(len(selected_probs) * 0.75)], 6),
            "p75": round(selected_probs[int(len(selected_probs) * 0.25)], 6),
            "max": round(selected_probs[0], 6),
            "min": round(selected_probs[-1], 6),
        }

    stats = {
        "strategy": "bayesian",
        "n_iterations": len(convergence_history),
        "min_probability": min_probability,
        "converged": convergence_history[-1] < 1e-8 if convergence_history else False,
        "final_max_delta": convergence_history[-1] if convergence_history else 0,
        "score_weighted": use_scores,
        "proteins_with_evidence": len(prot2pep),
        "selected_proteins": len(selected),
        "excluded_proteins": len(prot2pep) - len(selected),
        "unique_peptides": unique_peptides,
        "shared_peptides": len(pep2prot) - unique_peptides,
        "selected_with_unique_peptides": selected_with_unique,
        "probability_percentiles": percentiles,
        "db_breakdown": {db: dict(s) for db, s in db_stats.items()},
    }

    return InferenceResult(selected_proteins=selected, stats=stats)

score_lake_bayesian #

score_lake_bayesian(lake_proteins: dict[str, str], sample_peptides: dict[str, dict[str, float]], n_iterations: int = 10) -> dict[str, float]

Score every protein in the evidence lake using Bayesian EM across ALL samples.

This runs at the LAKE level, not per-sample. It uses de novo predictions from every sample to compute a single probability for each protein in the lake. Proteins consistently supported across multiple samples get boosted; proteins supported only by highly-shared peptides in one sample get downweighted.

Parameters:

Name Type Description Default
lake_proteins dict[str, str]

Protein ID → sequence for the entire evidence lake.

 required
sample_peptides dict[str, dict[str, float]]

sample_id → {peptide → denovo_score}. Each sample's top de novo predictions.

 required
n_iterations int

Number of EM iterations.

 10

Returns:

Type Description
dict[str, float]

Protein ID → probability (0 to 1, where 1 = most confident). Can be used to threshold the lake before per-sample extraction.
Algorithm
  1. Build a unified peptide→protein map across the entire lake.
  2. For each sample, each peptide contributes evidence weighted by:
  3. its de novo confidence score
  4. 1 / n_proteins matching (uniqueness)
  5. current protein probability (from previous iteration)
  6. Cross-sample evidence is SUMMED — a protein supported by unique peptides in 10 samples gets 10x the evidence of one in 1 sample.
  7. Normalize by sqrt(protein length).
  8. After convergence, the probability reflects:
  9. How many samples provide evidence
  10. How strong and specific that evidence is
  11. How consistently the protein appears across the experiment
Source code in fasta_lake/inference/strategies.py 
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
def score_lake_bayesian(
    lake_proteins: dict[str, str],
    sample_peptides: dict[str, dict[str, float]],
    n_iterations: int = 10,
) -> dict[str, float]:
    """
    Score every protein in the evidence lake using Bayesian EM across ALL samples.

    This runs at the LAKE level, not per-sample. It uses de novo predictions
    from every sample to compute a single probability for each protein in the
    lake. Proteins consistently supported across multiple samples get boosted;
    proteins supported only by highly-shared peptides in one sample get
    downweighted.

    Parameters
    ----------
    lake_proteins : dict[str, str]
        Protein ID → sequence for the entire evidence lake.
    sample_peptides : dict[str, dict[str, float]]
        sample_id → {peptide → denovo_score}.
        Each sample's top de novo predictions.
    n_iterations : int
        Number of EM iterations.

    Returns
    -------
    dict[str, float]
        Protein ID → probability (0 to 1, where 1 = most confident).
        Can be used to threshold the lake before per-sample extraction.

    Algorithm
    ---------
    1. Build a unified peptide→protein map across the entire lake.
    2. For each sample, each peptide contributes evidence weighted by:
       - its de novo confidence score
       - 1 / n_proteins matching (uniqueness)
       - current protein probability (from previous iteration)
    3. Cross-sample evidence is SUMMED — a protein supported by unique
       peptides in 10 samples gets 10x the evidence of one in 1 sample.
    4. Normalize by sqrt(protein length).
    5. After convergence, the probability reflects:
       - How many samples provide evidence
       - How strong and specific that evidence is
       - How consistently the protein appears across the experiment
    """
    import math

    all_prot_ids = list(lake_proteins.keys())
    if not all_prot_ids:
        return {}

    # Build peptide → protein map for entire lake
    logger.info("Building peptide-protein map for %d proteins...", len(all_prot_ids))
    from fasta_lake.helpers import map_peptides_to_proteins

    # Collect all unique peptides across all samples
    all_peptides: dict[str, float] = {}
    for sample_id, peps in sample_peptides.items():
        for pep, score in peps.items():
            if pep not in all_peptides or score > all_peptides[pep]:
                all_peptides[pep] = score

    pep2prot, prot2pep = map_peptides_to_proteins(lake_proteins, all_peptides)

    # Protein lengths
    prot_len = {pid: max(len(seq), 50) for pid, seq in lake_proteins.items()}

    # Count per-sample evidence for cross-sample boosting
    # sample_pep2prot: for each sample, which peptides map to which proteins
    sample_evidence: dict[str, dict[str, float]] = defaultdict(lambda: defaultdict(float))
    for sample_id, peps in sample_peptides.items():
        for pep, score in peps.items():
            if pep in pep2prot:
                for pid in pep2prot[pep]:
                    sample_evidence[sample_id][pid] += score / len(pep2prot[pep])

    # Count in how many samples each protein has evidence
    prot_sample_count = Counter()
    for sample_id, prot_scores in sample_evidence.items():
        for pid in prot_scores:
            prot_sample_count[pid] += 1

    # Initialize probabilities
    prob: dict[str, float] = {}
    for pid in all_prot_ids:
        if pid in prot2pep:
            # DETERMINISM GUARD: prot2pep[pid] is a set of peptide strings, so
            # the summation order was hash-seed dependent and float addition is
            # not associative. Sum in sorted peptide order.
            raw = sum(
                all_peptides.get(pep, 1.0) / len(pep2prot.get(pep, {pid}))
                for pep in sorted(prot2pep[pid])
            )
            # Boost by sample frequency
            sample_boost = math.sqrt(prot_sample_count.get(pid, 1))
            prob[pid] = raw * sample_boost / math.sqrt(prot_len[pid])
        else:
            prob[pid] = 0.0

    total = sum(prob.values())
    if total > 0:
        for pid in prob:
            prob[pid] /= total

    # EM iterations
    for iteration in range(n_iterations):
        new_evidence: dict[str, float] = defaultdict(float)

        for pep, matching_prots in pep2prot.items():
            denovo_score = all_peptides.get(pep, 1.0)

            if len(matching_prots) == 1:
                pid = next(iter(matching_prots))
                # Boost unique peptides by number of samples that see them
                n_samples_with_pep = sum(1 for sid, sp in sample_peptides.items() if pep in sp)
                new_evidence[pid] += denovo_score * math.sqrt(n_samples_with_pep)
            else:
                # DETERMINISM GUARD: matching_prots is a set of protein-ID
                # strings; sort once and reuse for the float sum and both
                # accumulation loops.
                ordered_prots = sorted(matching_prots)
                prob_sum = sum(prob.get(p, 0) for p in ordered_prots)
                if prob_sum == 0:
                    share = denovo_score / len(ordered_prots)
                    for p in ordered_prots:
                        new_evidence[p] += share
                else:
                    for p in ordered_prots:
                        weight = prob.get(p, 0) / prob_sum
                        new_evidence[p] += denovo_score * weight

        old_prob = dict(prob)
        for pid in all_prot_ids:
            sample_boost = math.sqrt(prot_sample_count.get(pid, 1))
            prob[pid] = new_evidence.get(pid, 0) * sample_boost / math.sqrt(prot_len[pid])

        total = sum(prob.values())
        if total > 0:
            for pid in prob:
                prob[pid] /= total

        max_delta = max(abs(prob.get(pid, 0) - old_prob.get(pid, 0)) for pid in all_prot_ids)
        if max_delta < 1e-8:
            logger.info("Converged at iteration %d (delta=%.2e)", iteration, max_delta)
            break

    # Scale to 0-1
    max_prob = max(prob.values()) if prob else 1.0
    if max_prob > 0:
        return {pid: prob[pid] / max_prob for pid in prob}
    return prob