run_canonical_parameter_sweep.py

#!/usr/bin/env python3
"""
STXBP1 Base-Editing Parameter Sweep - Canonical (603-aa / MANE Plus Clinical)

Re-build of the v1 sweep that was published against NM_001032221.6
(594 aa, MANE Select). This version uses NM_003165.6 (603 aa,
MANE Plus Clinical, NP_003156.1) — the same canonical reference
ARIA's Variant Explorer and Mouse Context Verifier use.

Why the rebuild matters:
  * Positions 1-575 are byte-identical between the two transcripts,
    so results for common pathogenic variants (K196X, R292H, R406H,
    and 168 others in the byte-identical zone) will match v1 exactly.
  * Positions 576+ diverge. v1 had 2 variants affected:
      - Q576X: mislabeled; canonical[576]=T (Thr), not Q (Gln). This
        variant's label is invalid under the canonical frame and has
        been removed from the input list. A separate `p.T576*` file
        would need to be seeded from ClinVar if that position has a
        true pathogenic variant reported on canonical numbering.
      - E603D: valid on canonical (pos 603, CDS nt 1807-1809); was
        invalid on 594-aa where the CDS ends at nt 1782. v1 would
        have errored on this variant; the canonical rebuild resolves it.
  * Provenance (S96 canonical rebuild, 2026-04-24): caught a years-old
    contamination where the original STXBP1 dataset had been folded
    against a 586-aa non-Munc18-1 protein. Documented in
    D:/STXBP1_datasets/scripts/stxbp1_canonical.py.

Run:
  python run_canonical_parameter_sweep.py \\
      --variants stxbp1_snv_variants_170_canonical.txt \\
      --output stxbp1_ULTIMATE_v2_canonical_YYYYMMDD.csv

Workers default to cpu_count(). One variant at a time to control memory.
"""
from __future__ import annotations

import argparse
import csv
import os
import re
import sys
import time
from datetime import datetime, timedelta
from multiprocessing import Pool, cpu_count
from pathlib import Path
from typing import Dict, List, Optional, Tuple

import requests

# =============================================================================
# CANONICAL CONFIG — the ONLY frame-dependent values
# =============================================================================

# MANE Plus Clinical frame (603 aa). Matches ARIA's Variant Explorer
# and ClinVar's variant coordinates.
HUMAN_ACCESSION = "NM_003165.6"
# S97 fix: mouse_accession was NM_011502 — that's mouse Syntaxin-3 (Stx3),
# NOT Stxbp1. Same wrong-accession bug that's still in the Mouse Context
# Verifier HF Space's GENE_CONFIGS (which propagated here when ARIA's
# MouseVerifierModal copied it). Correct mouse Stxbp1 accession is
# NM_009295 (NP_033321.2, isoform b, 594 aa, starts MAPIGLKAVV...).
# Verified directly via NCBI eutils efetch on 2026-04-24.
MOUSE_ACCESSION = "NM_009295"

# Banner written into every CSV header for downstream traceability.
REFERENCE_BANNER = (
    f"human={HUMAN_ACCESSION} (NP_003156.1, 603 aa, MANE Plus Clinical); "
    f"mouse={MOUSE_ACCESSION}"
)

CACHE_DIR = Path(__file__).parent / "sequence_cache_canonical"

# =============================================================================
# PARAMETER SPACE (UNCHANGED from v1 for apples-to-apples comparability)
# =============================================================================

SCAN_RADII = list(range(10, 201))        # 191 values
CONTEXT_RADII = list(range(5, 101))       # 96 values
EDIT_WINDOWS = [(s, e) for s in range(1, 21) for e in range(s, 21)]  # 210

TOTAL_PARAM_COMBOS = len(SCAN_RADII) * len(CONTEXT_RADII) * len(EDIT_WINDOWS)

# Worker globals (set by init_worker via fork semantics on Linux; on Windows
# they are passed via initargs and re-initialized per worker).
HUMAN_CDS: Optional[str] = None
MOUSE_CDS: Optional[str] = None


# =============================================================================
# SEQUENCE FETCHING
# =============================================================================

def fetch_cds_from_ncbi(accession: str) -> Optional[str]:
    url = "https://eutils.ncbi.nlm.nih.gov/entrez/eutils/efetch.fcgi"
    params = {
        "db": "nuccore",
        "id": accession,
        "rettype": "fasta_cds_na",
        "retmode": "text",
    }
    try:
        resp = requests.get(url, params=params, timeout=30)
        if resp.status_code == 200:
            lines = resp.text.strip().split("\n")
            return "".join(
                line.strip() for line in lines[1:] if not line.startswith(">")
            ).upper()
    except Exception as e:
        print(f"NCBI error: {e}", file=sys.stderr)
    return None


def get_sequence(accession: str) -> Optional[str]:
    CACHE_DIR.mkdir(exist_ok=True)
    # Match with or without version, returning first hit
    for f in CACHE_DIR.glob(f"{accession.split('.')[0]}*.fasta"):
        return f.read_text().strip()
    seq = fetch_cds_from_ncbi(accession)
    if seq:
        (CACHE_DIR / f"{accession}.fasta").write_text(seq)
        return seq
    return None


# =============================================================================
# CODON / EDIT-TYPE HELPERS
# =============================================================================

def parse_cdna(cdna: str) -> Tuple[Optional[int], Optional[str], Optional[str]]:
    m = re.match(r"c\.(\d+)([ACGT])>([ACGT])", cdna)
    return (int(m.group(1)), m.group(2), m.group(3)) if m else (None, None, None)


def classify_edit_type(ref: str, alt: str) -> str:
    change = f"{ref}>{alt}"
    if change in ("A>G", "T>C"):
        return "ABE"
    if change in ("C>T", "G>A"):
        return "CBE"
    return "PRIME"


# =============================================================================
# SINGLE-PARAM ANALYSIS
# =============================================================================

def analyze_one_param(params: Tuple) -> Optional[Dict]:
    label, cdna, cdna_pos, edit_type, scan_radius, context_radius, edit_window = params
    human_cds = HUMAN_CDS
    mouse_cds = MOUSE_CDS
    if human_cds is None or mouse_cds is None:
        return None
    if cdna_pos > len(human_cds) or cdna_pos > len(mouse_cds):
        return None

    # Context identity
    ctx_start = max(0, cdna_pos - 1 - context_radius)
    ctx_end = min(len(human_cds), cdna_pos + context_radius)
    human_ctx = human_cds[ctx_start:ctx_end]
    mouse_ctx = mouse_cds[ctx_start:ctx_end]
    min_len = min(len(human_ctx), len(mouse_ctx))
    if min_len == 0:
        return None
    matches = sum(1 for i in range(min_len) if human_ctx[i] == mouse_ctx[i])
    context_identity = matches / min_len

    # PAM search
    scan_start = max(0, cdna_pos - 1 - scan_radius)
    scan_end = min(len(human_cds), cdna_pos + scan_radius)
    human_scan = human_cds[scan_start:scan_end]
    mouse_scan = mouse_cds[scan_start:scan_end]
    target_in_scan = cdna_pos - 1 - scan_start

    best_mm = 999
    best_in_window = False
    num_candidates = 0

    for i in range(len(human_scan) - 22):
        if human_scan[i + 21:i + 23] == "GG":
            proto_start = i
            if proto_start <= target_in_scan < proto_start + 20:
                edit_pos = target_in_scan - proto_start + 1
                in_window = edit_window[0] <= edit_pos <= edit_window[1]
                if proto_start + 23 <= len(mouse_scan):
                    mm = sum(
                        1
                        for j in range(23)
                        if human_scan[proto_start + j] != mouse_scan[proto_start + j]
                    )
                    num_candidates += 1
                    if mm < best_mm or (mm == best_mm and in_window and not best_in_window):
                        best_mm = mm
                        best_in_window = in_window

    score = context_identity * 40
    if best_mm < 999:
        score += max(0, 30 - best_mm * 2.5)
        if best_in_window:
            score += 20
    compatible = context_identity >= 0.90 and best_mm <= 3 and best_in_window
    if compatible:
        score += 10

    return {
        "label": label,
        "cdna": cdna,
        "edit_type": edit_type,
        "scan": scan_radius,
        "ctx": context_radius,
        "win": f"{edit_window[0]}-{edit_window[1]}",
        "identity": round(context_identity, 4),
        "candidates": num_candidates,
        "best_mm": best_mm if best_mm < 999 else None,
        "in_window": best_in_window,
        "compat": compatible,
        "score": round(min(100, max(0, score)), 1),
    }


def init_worker(human_cds: str, mouse_cds: str) -> None:
    global HUMAN_CDS, MOUSE_CDS
    HUMAN_CDS = human_cds
    MOUSE_CDS = mouse_cds


# =============================================================================
# PER-VARIANT DRIVER
# =============================================================================

def process_single_variant(
    variant: Dict,
    human_cds: str,
    mouse_cds: str,
    writer: csv.DictWriter,
    num_workers: int,
) -> Tuple[int, Optional[Dict]]:
    label = variant.get("label", "")
    cdna = variant.get("cdna", "")
    cdna_pos, ref, alt = parse_cdna(cdna)
    if cdna_pos is None:
        return 0, None

    edit_type = classify_edit_type(ref, alt) if ref and alt else "?"

    work_items = [
        (label, cdna, cdna_pos, edit_type, scan, ctx, window)
        for scan in SCAN_RADII
        for ctx in CONTEXT_RADII
        for window in EDIT_WINDOWS
    ]

    best_result: Optional[Dict] = None
    best_score = -1.0
    count = 0

    with Pool(num_workers, initializer=init_worker, initargs=(human_cds, mouse_cds)) as pool:
        for result in pool.imap_unordered(analyze_one_param, work_items, chunksize=500):
            if result is None:
                continue
            writer.writerow(result)
            count += 1
            if result["score"] > best_score:
                best_score = result["score"]
                best_result = result

    return count, best_result


# =============================================================================
# VARIANT LOADING
# =============================================================================

def load_variants(filepath: str) -> List[Dict]:
    variants = []
    # encoding='utf-8-sig' strips an optional BOM; errors='replace' is defense
    # against Windows-1252 punctuation (em-dashes, smart quotes) that PowerShell
    # tools sometimes inject when piping or saving from Word. Comment lines are
    # skipped anyway, so any lossy replacement in the header is harmless.
    with open(filepath, "r", encoding="utf-8-sig", errors="replace") as f:
        for line in f:
            line = line.strip()
            if not line or line.startswith("#"):
                continue
            parts = line.split()
            if len(parts) >= 2:
                variants.append(
                    {"label": parts[0], "cdna": parts[1], "aa": parts[2] if len(parts) > 2 else ""}
                )
    return variants


# =============================================================================
# MAIN
# =============================================================================

def main() -> None:
    parser = argparse.ArgumentParser(
        description="STXBP1 Canonical Base-Editing Parameter Sweep (603-aa)"
    )
    parser.add_argument("--workers", "-w", type=int, default=None)
    parser.add_argument(
        "--variants", "-v", type=str,
        default=str(Path(__file__).parent / "stxbp1_snv_variants_170_canonical.txt"),
    )
    parser.add_argument("--output", "-o", type=str, default=None)
    args = parser.parse_args()

    num_workers = args.workers or cpu_count()

    print("=" * 70)
    print("STXBP1 CANONICAL PARAMETER SWEEP")
    print(f"Reference: {REFERENCE_BANNER}")
    print("=" * 70)
    print(f"Parameter space: {len(SCAN_RADII)} x {len(CONTEXT_RADII)} x {len(EDIT_WINDOWS)} = {TOTAL_PARAM_COMBOS:,}")
    print(f"Workers: {num_workers}")
    print("=" * 70)

    if not os.path.exists(args.variants):
        print(f"ERROR: Variant file not found: {args.variants}", file=sys.stderr)
        sys.exit(1)

    variants = load_variants(args.variants)
    print(f"Loaded {len(variants)} variants from {args.variants}")

    print("\nFetching canonical CDS sequences...")
    human_cds = get_sequence(HUMAN_ACCESSION)
    mouse_cds = get_sequence(MOUSE_ACCESSION)
    if not human_cds or not mouse_cds:
        print("ERROR: Could not get sequences", file=sys.stderr)
        sys.exit(1)

    # Canonical CDS sanity check — 603 aa + stop = 1812 nt
    if len(human_cds) < 1809:
        print(
            f"WARNING: human CDS length {len(human_cds)} is shorter than expected "
            f"1812 nt for NP_003156.1. Did NCBI return a different isoform?",
            file=sys.stderr,
        )

    print(f"Human: {len(human_cds)} nt  Mouse: {len(mouse_cds)} nt")

    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
    csv_path = args.output or f"stxbp1_ULTIMATE_v2_canonical_{timestamp}.csv"
    json_path = csv_path.replace(".csv", "_best.json")

    total_analyses = len(variants) * TOTAL_PARAM_COMBOS
    print(f"\nTotal analyses: {total_analyses:,}")
    print(f"Output CSV: {csv_path}")
    print(f"Best-config JSON: {json_path}\n")

    fieldnames = [
        "label", "cdna", "edit_type", "scan", "ctx", "win",
        "identity", "candidates", "best_mm", "in_window", "compat", "score",
    ]

    start_time = time.time()
    total_written = 0
    best_configs: Dict[str, Dict] = {}

    with open(csv_path, "w", newline="", encoding="utf-8") as f:
        # Banner comment rows so a reader grepping the CSV head sees the frame
        f.write(f"# reference_frame: {REFERENCE_BANNER}\n")
        f.write(f"# build_date: {timestamp}\n")
        f.write(f"# variants: {len(variants)} from {os.path.basename(args.variants)}\n")
        writer = csv.DictWriter(f, fieldnames=fieldnames)
        writer.writeheader()

        for i, variant in enumerate(variants):
            var_start = time.time()
            count, best = process_single_variant(variant, human_cds, mouse_cds, writer, num_workers)
            total_written += count
            if best:
                best_configs[variant["label"]] = best
            f.flush()

            var_time = time.time() - var_start
            elapsed = time.time() - start_time
            rate = total_written / elapsed if elapsed > 0 else 0
            eta = (total_analyses - total_written) / rate if rate > 0 else 0
            print(
                f"[{i+1:3d}/{len(variants)}] {variant['label']:<10} "
                f"{count:,} rows  {var_time:.1f}s  "
                f"Total: {total_written:,}  "
                f"Rate: {rate:,.0f}/s  "
                f"ETA: {timedelta(seconds=int(eta))}"
            )

    # Best-config JSON
    import json as _json
    output_json = {
        "_metadata": {
            "reference_frame": REFERENCE_BANNER,
            "build_date": timestamp,
            "total_variants": len(variants),
            "total_param_combos_per_variant": TOTAL_PARAM_COMBOS,
            "total_rows_written": total_written,
        },
        "variants": {},
    }
    for label, row in best_configs.items():
        output_json["variants"][label] = {
            "cdna": row["cdna"],
            "edit_type": row["edit_type"],
            "optimal_params": {
                "scan_radius": row["scan"],
                "context_radius": row["ctx"],
                "edit_window": row["win"],
            },
            "score": row["score"],
            "context_identity": row["identity"],
            "compatible": row["compat"],
        }
    with open(json_path, "w", encoding="utf-8") as f:
        _json.dump(output_json, f, indent=2)

    total_time = time.time() - start_time
    print("\n" + "=" * 70)
    print("COMPLETE")
    print("=" * 70)
    print(f"Time:   {timedelta(seconds=int(total_time))}")
    print(f"Rows:   {total_written:,}")
    print(f"Rate:   {total_written/total_time:,.0f}/s")
    print(f"CSV:    {csv_path}")
    print(f"Best:   {json_path}")


if __name__ == "__main__":
    main()