mgnify-evo2-probes / code /scripts /extract_targeted.py
JG1310's picture
Probe artifacts: code, manifests, plots, scores, summaries, checkpoints, docs
eb69de4 verified
"""
Extract per-MAG JSONL of (positive, matched-negative) pairs for the targeted Evo2/SAE run.
Per the plan in targeted_pipeline_plan.md:
- Positives: AMR / STRESS / VIRULENCE hits (from AMRFinderPlus TSV)
- Negatives: same-MAG CDSs, length ±20%, excluding AMR/STRESS/VIRULENCE/BGC/CRISPR/defence
(mobilome included by default, fallback to strict-only if needed, then ±50%)
- 1:1 pairing, sampling without replacement, seed=42
- Forward strand always (matches Goodfire's reference notebook)
- Sanity checks run on every record (length, start codon by strand, boundary clamps)
Output: one JSONL per MAG at <out_dir>/<mag_id>.jsonl, interleaved positive then negative.
"""
import argparse
import json
import random
import re
from collections import Counter
from dataclasses import dataclass, asdict, field
from pathlib import Path
from typing import Optional
import pandas as pd
from pyfaidx import Fasta
CANONICAL_STARTS = {"ATG", "GTG", "TTG", "CTG"}
COMP = str.maketrans("ACGTNacgtn", "TGCANtgcan")
def revcomp(s: str) -> str:
return s.translate(COMP)[::-1]
# =============================================================
# GFF / interval helpers
# =============================================================
@dataclass
class CDS:
locus_tag: str
contig: str
start: int # 1-indexed inclusive (GFF convention)
end: int
strand: str # '+' or '-'
partial: str = "00" # Prodigal: '00'=full, '10'=no start, '01'=no stop, '11'=neither
@property
def length(self) -> int:
return self.end - self.start + 1
def parse_master_gff(path: Path) -> list[CDS]:
"""Parse Prodigal master GFF → list of CDS records."""
cds_list = []
if not path.exists():
return cds_list
for line in path.read_text().splitlines():
if not line or line.startswith("#"):
continue
cols = line.split("\t")
if len(cols) < 9 or cols[2] != "CDS":
continue
contig, _, _, start, end, _, strand, _, attrs = cols
m = re.search(r"locus_tag=([^;]+)", attrs)
if not m:
continue
locus_tag = m.group(1)
partial_m = re.search(r"partial=(\d{2})", attrs)
partial = partial_m.group(1) if partial_m else "00"
cds_list.append(CDS(locus_tag=locus_tag, contig=contig,
start=int(start), end=int(end), strand=strand, partial=partial))
return cds_list
def parse_interval_gff(path: Path) -> list[tuple[str, int, int]]:
"""Parse a GFF (antismash, crispr, defense_finder, mobilome) → list of (contig, start, end)
for the major 'region'-style features. Skips sub-features like 'gene', 'CRISPRdr' etc."""
intervals = []
if not path.exists():
return intervals
region_types = {
# antismash
"region",
# CRISPRCasFinder: full arrays span (we exclude any CDS overlap)
"CRISPR",
# defense_finder
"Defense system", "Antidefense system",
# mobilome (genomad)
"plasmid", "viral_sequence", "prophage", "integron",
"conjugative_integron", "phage_plasmid",
"insertion_sequence", "terminal_inverted_repeat_element", "attC_site",
}
for line in path.read_text().splitlines():
if not line or line.startswith("#"):
continue
cols = line.split("\t")
if len(cols) < 5:
continue
if cols[2] in region_types:
intervals.append((cols[0], int(cols[3]), int(cols[4])))
return intervals
def cds_overlaps_any_interval(cds: CDS, intervals: list[tuple[str, int, int]]) -> bool:
"""1-indexed inclusive interval overlap check."""
for contig, s, e in intervals:
if cds.contig == contig and cds.end >= s and cds.start <= e:
return True
return False
# =============================================================
# Sequence extraction with sanity checks
# =============================================================
def extract_region(
fa: Fasta,
contig: str,
gene_start: int,
gene_end: int,
strand: str,
flank: int,
) -> dict:
"""Extract gene + flank, forward-strand. Returns dict with sequence + ext coordinates.
Coordinates are 1-indexed inclusive, matching pyfaidx's slicing convention."""
contig_len = len(fa[contig])
ext_start = max(1, gene_start - flank)
ext_end = min(contig_len, gene_end + flank)
# pyfaidx get_seq is 1-indexed INCLUSIVE on both ends — matches GFF/AMRFinderPlus convention
seq = str(fa.get_seq(contig, ext_start, ext_end).seq)
# Sanity: length matches the (possibly truncated) range
expected_len = ext_end - ext_start + 1
assert len(seq) == expected_len, (
f"length mismatch: contig={contig} ext={ext_start}-{ext_end} got {len(seq)} expected {expected_len}"
)
return {
"ext_start": ext_start, "ext_end": ext_end, "contig_len": contig_len, "sequence": seq,
}
def check_start_codon(
sequence: str, gene_start: int, ext_start: int, gene_end: int, strand: str,
partial: str = "00",
) -> Optional[str]:
"""Return the gene's first 3 coding-strand bases, or None for partial-no-start genes.
Used for sanity checks: most full genes should start with ATG/GTG/TTG/CTG."""
if strand == "+":
if partial in {"10", "11"}: # missing start
return None
offset = gene_start - ext_start
return sequence[offset : offset + 3].upper()
else: # '-' strand: gene 5' end is at gene_end on the forward strand
if partial in {"01", "11"}: # for - strand, '01' = no stop on - = no start in coding direction
return None
# Wait: Prodigal partial flag is in CODING direction (5'→3' of the CDS)
# '10' = no start, '01' = no stop, regardless of strand
if partial in {"10", "11"}:
return None
offset_end = gene_end - ext_start + 1 # position just past gene_end (slice end)
offset_start = offset_end - 3
forward_chunk = sequence[offset_start : offset_end].upper()
return revcomp(forward_chunk)
# =============================================================
# Main pipeline per MAG
# =============================================================
@dataclass
class Stats:
mags_processed: int = 0
positives_found: int = 0
pairs_emitted: int = 0
no_match: int = 0
fallback_used: dict = field(default_factory=lambda: {"strict_no_mob_l20_g5": 0,
"strict_w_mob_l20_g5": 0,
"strict_w_mob_l20_g10": 0,
"strict_w_mob_l50_g10": 0})
start_codon_pass_plus: int = 0
start_codon_total_plus: int = 0
start_codon_pass_minus: int = 0
start_codon_total_minus: int = 0
length_check_failures: int = 0
def gc_content(seq: str) -> float:
"""GC fraction over A/C/G/T (Ns excluded). Returns 0.0 for empty input."""
if not seq:
return 0.0
s = seq.upper()
gc = sum(1 for c in s if c == "G" or c == "C")
acgt = sum(1 for c in s if c in "ACGT")
return gc / acgt if acgt else 0.0
def extract_for_mag(
mag_dir: Path,
mag_id: str,
out_root: Path, # if split_by_label: out_root/{LABEL}/{mag_id}.jsonl
# else: out_root/{mag_id}.jsonl
split_by_label: bool,
flank: int,
length_tol_strict: float,
length_tol_relaxed: float,
gc_tol_strict: float,
gc_tol_relaxed: float,
seed: int,
max_pairs: Optional[int],
stats: Stats,
) -> int:
"""Extract positives + matched negatives for one MAG. Writes JSONL to out_path.
Returns number of records emitted (positive + negative count combined)."""
fna = mag_dir / f"{mag_id}.fna"
gff = mag_dir / f"{mag_id}.gff"
amr_tsv = mag_dir / f"{mag_id}_amrfinderplus.tsv"
if not (fna.exists() and gff.exists() and amr_tsv.exists()):
print(f" [{mag_id}] missing required files, skipping")
return 0
fa = Fasta(str(fna)) # auto-creates .fna.fai if absent
all_cds = parse_master_gff(gff)
cds_by_locus = {c.locus_tag: c for c in all_cds}
# Precompute GC content for every CDS once (used for negative-pair matching + recorded as metadata)
cds_gc: dict[str, float] = {}
for c in all_cds:
try:
cds_seq = str(fa.get_seq(c.contig, c.start, c.end).seq)
except (KeyError, ValueError):
continue
cds_gc[c.locus_tag] = gc_content(cds_seq)
# Positives
try:
amr_df = pd.read_csv(amr_tsv, sep="\t")
except Exception:
return 0
if "Element type" not in amr_df.columns:
return 0
positive_rows = amr_df[amr_df["Element type"].isin(["AMR", "STRESS", "VIRULENCE"])]
if len(positive_rows) == 0:
return 0
# Exclusion regions (label intervals)
bgc_iv = parse_interval_gff(mag_dir / f"{mag_id}_antismash.gff")
crispr_iv = parse_interval_gff(mag_dir / f"{mag_id}_crisprcasfinder.gff")
defense_iv = parse_interval_gff(mag_dir / f"{mag_id}_defense_finder.gff")
mobilome_iv = parse_interval_gff(mag_dir / f"{mag_id}_mobilome.gff")
label_intervals_strict = bgc_iv + crispr_iv + defense_iv
# Map CDS → flags
cds_in_mobilome = {c.locus_tag for c in all_cds if cds_overlaps_any_interval(c, mobilome_iv)}
cds_in_strict_excluded = {c.locus_tag for c in all_cds
if cds_overlaps_any_interval(c, label_intervals_strict)}
positive_locus_tags = set()
for _, row in positive_rows.iterrows():
pi = row.get("Protein identifier") or row.get("Protein id")
if pi and pi in cds_by_locus:
positive_locus_tags.add(pi)
rng = random.Random(seed)
used_negs: set[str] = set()
records = []
# Shuffle for fairness; use iterrows for clean column-name access (handles spaces in headers)
shuffled = positive_rows.sample(frac=1, random_state=seed)
for _, prow in shuffled.iterrows():
if max_pairs is not None and stats.pairs_emitted >= max_pairs:
break
# Look up CDS by locus tag (== AMRFinderPlus's "Protein identifier")
pi = prow.get("Protein identifier") or prow.get("Protein id")
if not pi or pi not in cds_by_locus:
continue
pos_cds = cds_by_locus[pi]
pos_len = pos_cds.length
# Extract positive sequence
pos_ext = extract_region(fa, pos_cds.contig, pos_cds.start, pos_cds.end, pos_cds.strand, flank)
# Negative selection fallback hierarchy:
# strict_no_mob_l20_g5 : exclude mobilome, ±20% length, ±5% GC (primary)
# strict_w_mob_l20_g5 : include mobilome, ±20% length, ±5% GC
# strict_w_mob_l20_g10 : include mobilome, ±20% length, ±10% GC
# strict_w_mob_l50_g10 : include mobilome, ±50% length, ±10% GC
candidates = None
fallback_used = None
pos_gc = cds_gc.get(pos_cds.locus_tag, 0.0)
for excluded, len_tol, gc_t, tag in [
(positive_locus_tags | cds_in_strict_excluded | cds_in_mobilome, length_tol_strict, gc_tol_strict, "strict_no_mob_l20_g5"),
(positive_locus_tags | cds_in_strict_excluded, length_tol_strict, gc_tol_strict, "strict_w_mob_l20_g5"),
(positive_locus_tags | cds_in_strict_excluded, length_tol_strict, gc_tol_relaxed, "strict_w_mob_l20_g10"),
(positive_locus_tags | cds_in_strict_excluded, length_tol_relaxed, gc_tol_relaxed, "strict_w_mob_l50_g10"),
]:
pool = [
c for c in all_cds
if c.locus_tag not in excluded
and c.locus_tag not in used_negs
and c.locus_tag in cds_gc # has computed GC
and abs(c.length - pos_len) / pos_len <= len_tol
and abs(cds_gc[c.locus_tag] - pos_gc) <= gc_t
]
if pool:
candidates = pool; fallback_used = tag; break
if not candidates:
# No negative found — emit positive solo with status
pos_record = build_record(
pos_cds, prow, pos_ext, is_positive=True, paired_with=None,
in_mobilome=(pos_cds.locus_tag in cds_in_mobilome),
fallback_used=None, extract_status="no_matching_negative",
seed=seed, mag_id=mag_id,
gc_content_val=cds_gc.get(pos_cds.locus_tag),
)
records.append(pos_record)
stats.no_match += 1
stats.positives_found += 1
_track_start_codon(pos_record, stats)
continue
neg_cds = rng.choice(candidates)
used_negs.add(neg_cds.locus_tag)
neg_ext = extract_region(fa, neg_cds.contig, neg_cds.start, neg_cds.end, neg_cds.strand, flank)
pos_record = build_record(
pos_cds, prow, pos_ext, is_positive=True, paired_with=neg_cds.locus_tag,
in_mobilome=(pos_cds.locus_tag in cds_in_mobilome),
fallback_used=fallback_used, extract_status="ok",
seed=seed, mag_id=mag_id,
gc_content_val=cds_gc.get(pos_cds.locus_tag),
)
neg_record = build_record(
neg_cds, None, neg_ext, is_positive=False, paired_with=pos_cds.locus_tag,
in_mobilome=(neg_cds.locus_tag in cds_in_mobilome),
fallback_used=None, extract_status="ok",
seed=seed, mag_id=mag_id,
gc_content_val=cds_gc.get(neg_cds.locus_tag),
)
# Inherit class/subclass from paired positive — useful for grouping in downstream analysis.
# NOT inheriting gene_symbol or pct_identity_to_ref (those are intrinsic to the positive).
neg_record["label_class"] = pos_record["label_class"]
neg_record["label_subclass"] = pos_record["label_subclass"]
records.append(pos_record); records.append(neg_record)
stats.fallback_used[fallback_used] += 1
stats.pairs_emitted += 1
stats.positives_found += 1
_track_start_codon(pos_record, stats)
_track_start_codon(neg_record, stats)
# Write JSONL — either single file or split into per-label folders (negatives → MISC).
if not records:
return 0
if split_by_label:
# Group records by destination folder
buckets: dict[str, list[dict]] = {}
for r in records:
folder = "MISC" if not r["is_positive"] else r["label"] # AMR / VIRULENCE / STRESS / MISC
buckets.setdefault(folder, []).append(r)
for folder, recs in buckets.items():
out_dir = out_root / folder
out_dir.mkdir(parents=True, exist_ok=True)
(out_dir / f"{mag_id}.jsonl").write_text(
"\n".join(json.dumps(r) for r in recs) + "\n"
)
else:
out_root.mkdir(parents=True, exist_ok=True)
(out_root / f"{mag_id}.jsonl").write_text(
"\n".join(json.dumps(r) for r in records) + "\n"
)
return len(records)
def build_record(
cds: CDS, prow, ext: dict, *, is_positive: bool, paired_with: Optional[str],
in_mobilome: bool, fallback_used: Optional[str], extract_status: str,
seed: int, mag_id: str,
gc_content_val: Optional[float] = None,
) -> dict:
"""Construct a single JSONL record from CDS + AMRFinderPlus pandas Series + extracted region."""
label = "negative"
label_class = label_subclass = gene_symbol = None
pct_id = None
def _clean(v):
"""Convert pandas NaN (and similar) to None; passes other values through."""
if v is None: return None
# pandas NaN is float and != itself
if isinstance(v, float) and v != v: return None
return v
if is_positive and prow is not None:
et = _clean(prow.get("Element type"))
label = et if et in ("AMR", "STRESS", "VIRULENCE") else "AMR"
label_class = _clean(prow.get("Class"))
label_subclass = _clean(prow.get("Subclass"))
gene_symbol = _clean(prow.get("Gene symbol"))
pct_id = _clean(prow.get("% Identity to reference sequence"))
# Coerce numpy → Python types
if pct_id is not None:
try: pct_id = float(pct_id)
except Exception: pct_id = None
return {
"mag_id": mag_id,
"locus_tag": cds.locus_tag,
"region_id": f"{cds.locus_tag}_{label}",
"is_positive": is_positive,
"label": label,
"label_class": label_class,
"label_subclass": label_subclass,
"gene_symbol": gene_symbol,
"pct_identity_to_ref": pct_id,
"contig": cds.contig,
"gene_start": cds.start,
"gene_end": cds.end,
"strand": cds.strand,
"cds_length": cds.length,
"partial": cds.partial,
"ext_start": ext["ext_start"],
"ext_end": ext["ext_end"],
"contig_len": ext["contig_len"],
"paired_with": paired_with,
"cds_in_mobilome": in_mobilome,
"gc_content": round(gc_content_val, 4) if gc_content_val is not None else None,
"negative_pool_fallback": fallback_used,
"extract_status": extract_status,
"random_seed": seed,
"sequence": ext["sequence"],
}
def _track_start_codon(record: dict, stats: Stats):
"""Update start-codon pass-rate stats (used to detect strand-handling bugs)."""
codon = check_start_codon(
record["sequence"], record["gene_start"], record["ext_start"],
record["gene_end"], record["strand"], record.get("partial", "00"),
)
if codon is None:
return
if record["strand"] == "+":
stats.start_codon_total_plus += 1
if codon in CANONICAL_STARTS:
stats.start_codon_pass_plus += 1
else:
stats.start_codon_total_minus += 1
if codon in CANONICAL_STARTS:
stats.start_codon_pass_minus += 1
def cross_tool_check(out_dir: Path, mag_dirs: dict, n_samples: int = 5) -> str:
"""Compare pyfaidx output to samtools faidx output for n_samples random records.
Returns a status string. Skips silently if samtools not available."""
import shutil, subprocess
if not shutil.which("samtools"):
return "skipped (samtools not on PATH)"
# Pick records from any JSONL files present
records = []
for jsonl in out_dir.glob("*.jsonl"):
for line in jsonl.read_text().splitlines():
if line.strip():
records.append(json.loads(line))
if len(records) < 1:
return "no records to check"
rng = random.Random(0)
sample = rng.sample(records, min(n_samples, len(records)))
for r in sample:
mag_id = r["mag_id"]
fna_path = mag_dirs[mag_id] / f"{mag_id}.fna"
region = f"{r['contig']}:{r['ext_start']}-{r['ext_end']}"
result = subprocess.run(["samtools", "faidx", str(fna_path), region],
capture_output=True, text=True, check=False)
if result.returncode != 0:
return f"samtools failed on {region}: {result.stderr[:200]}"
# Strip header + newlines
lines = result.stdout.splitlines()
samtools_seq = "".join(l for l in lines if not l.startswith(">"))
if samtools_seq.upper() != r["sequence"].upper():
return (f"MISMATCH on {region}: pyfaidx len={len(r['sequence'])}, "
f"samtools len={len(samtools_seq)}")
return f"passed ({len(sample)} records cross-tool verified)"
# =============================================================
# CLI
# =============================================================
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--skin-dir", "--catalogue-dir", type=Path,
default=Path("/home/ror25cal/MGnify/data/human-skin/species_catalogue"),
dest="skin_dir",
help="catalogue's species_catalogue directory (skin or chicken-gut)")
ap.add_argument("--out-dir", type=Path,
default=Path("/home/ror25cal/MGnify/data/targeted_jsonl/skin"))
ap.add_argument("--split-by-label", action="store_true",
help="write to {out_dir}/{AMR|VIRULENCE|STRESS|MISC}/{mag_id}.jsonl "
"(MISC = negatives). Default: single {out_dir}/{mag_id}.jsonl per MAG.")
ap.add_argument("--mag-ids", nargs="*", default=None,
help="restrict to specific MAG IDs (default: all 579 skin MAGs)")
ap.add_argument("--top-csv", type=Path, default=None,
help="if set, restrict to MAGs listed in this CSV file (col 'mag_id')")
ap.add_argument("--max-pairs", type=int, default=None,
help="for testing: stop after this many positive-negative pairs total")
ap.add_argument("--flank", type=int, default=2000)
ap.add_argument("--length-tol-strict", type=float, default=0.20)
ap.add_argument("--length-tol-relaxed", type=float, default=0.50)
ap.add_argument("--gc-tol-strict", type=float, default=0.05,
help="absolute GC-fraction tolerance for paired negative (default ±0.05)")
ap.add_argument("--gc-tol-relaxed", type=float, default=0.10,
help="relaxed GC tolerance used in later fallbacks (default ±0.10)")
ap.add_argument("--seed", type=int, default=42)
args = ap.parse_args()
args.out_dir.mkdir(parents=True, exist_ok=True)
# Resolve MAG list
if args.mag_ids:
mag_ids = args.mag_ids
elif args.top_csv:
mag_ids = pd.read_csv(args.top_csv)["mag_id"].tolist()
else:
mag_ids = sorted(p.name for p in args.skin_dir.glob("*/MGYG*") if p.is_dir())
print(f"target: {len(mag_ids)} MAG(s)")
print(f"flank={args.flank}, length_tol_strict=±{args.length_tol_strict*100:.0f}%, "
f"relaxed=±{args.length_tol_relaxed*100:.0f}%, seed={args.seed}")
if args.max_pairs:
print(f"max_pairs={args.max_pairs} (testing mode)")
print()
stats = Stats()
mag_dirs_map = {}
for mag_id in mag_ids:
if args.max_pairs and stats.pairs_emitted >= args.max_pairs:
break
prefix = mag_id[:11]
mag_dir = args.skin_dir / prefix / mag_id / "genome"
if not mag_dir.exists():
print(f" {mag_id}: dir not found at {mag_dir}, skipping")
continue
mag_dirs_map[mag_id] = mag_dir
n_records = extract_for_mag(
mag_dir, mag_id, args.out_dir,
split_by_label=args.split_by_label,
flank=args.flank,
length_tol_strict=args.length_tol_strict,
length_tol_relaxed=args.length_tol_relaxed,
gc_tol_strict=args.gc_tol_strict,
gc_tol_relaxed=args.gc_tol_relaxed,
seed=args.seed,
max_pairs=(args.max_pairs - stats.pairs_emitted) if args.max_pairs else None,
stats=stats,
)
if n_records and stats.mags_processed % 50 == 0:
print(f" ({stats.mags_processed} MAGs processed, {stats.pairs_emitted} pairs so far)")
stats.mags_processed += 1
print()
print("=" * 65)
print("SUMMARY")
print("=" * 65)
print(f"MAGs processed: {stats.mags_processed}")
print(f"Positives found: {stats.positives_found}")
print(f"Pairs emitted: {stats.pairs_emitted}")
print(f"No-match positives: {stats.no_match}")
print(f"\nFallback usage (negative selection):")
for k, v in stats.fallback_used.items():
print(f" {k:25s} {v}")
print(f"\nStart-codon sanity check (canonical starts: ATG/GTG/TTG/CTG):")
if stats.start_codon_total_plus > 0:
rate_p = stats.start_codon_pass_plus / stats.start_codon_total_plus * 100
print(f" + strand: {stats.start_codon_pass_plus}/{stats.start_codon_total_plus} ({rate_p:.1f}%)")
if stats.start_codon_total_minus > 0:
rate_m = stats.start_codon_pass_minus / stats.start_codon_total_minus * 100
print(f" - strand: {stats.start_codon_pass_minus}/{stats.start_codon_total_minus} ({rate_m:.1f}%)")
if stats.start_codon_total_plus and stats.start_codon_total_minus:
diff = abs(rate_p - rate_m)
if rate_p < 80 or rate_m < 80 or diff > 15:
print(f" ⚠ WARNING: strand pass-rates suggest a strand-handling bug")
else:
print(f" ✓ strand handling looks correct")
# Cross-tool verification
print(f"\nCross-tool sanity check (pyfaidx vs samtools): "
f"{cross_tool_check(args.out_dir, mag_dirs_map, n_samples=10)}")
if __name__ == "__main__":
main()