libs/docking/rdock_features.py

from __future__ import annotations

import math
import re
from pathlib import Path
from typing import Any, Dict, List, Tuple

import numpy as np
from rdkit import Chem
from rdkit.Chem import rdMolAlign


TAG_RE = re.compile(r"^>\s*<\s*([^>]+?)\s*>", flags=re.IGNORECASE)


def _safe_float(value: object) -> float | None:
    try:
        v = float(value)  # type: ignore[arg-type]
    except Exception:
        return None
    if not np.isfinite(v):
        return None
    return float(v)


def _split_sdf_blocks(text: str) -> List[str]:
    blocks = []
    for part in text.split("$$$$"):
        block = part.strip()
        if block:
            blocks.append(block + "\n$$$$\n")
    return blocks


def _parse_block_tags(block: str) -> Dict[str, str]:
    lines = block.splitlines()
    tags: Dict[str, str] = {}
    i = 0
    while i < len(lines):
        m = TAG_RE.match(lines[i].strip())
        if not m:
            i += 1
            continue
        tag = m.group(1).strip()
        value = ""
        if i + 1 < len(lines):
            value = lines[i + 1].strip()
        tags[tag] = value
        i += 2
    return tags


def _pose_molecules(blocks: List[str]) -> List[Chem.Mol | None]:
    out: List[Chem.Mol | None] = []
    for b in blocks:
        mol = None
        try:
            mol = Chem.MolFromMolBlock(b, sanitize=False, removeHs=False, strictParsing=False)
        except Exception:
            mol = None
        out.append(mol)
    return out


def _pose_rmsd_stats(poses: List[Dict[str, Any]], mols: List[Chem.Mol | None], top_k: int = 5) -> Dict[str, float | None]:
    if not poses:
        return {"top_pose_rmsd_consistency": None, "geometry_similarity_top5": None}

    top = sorted(poses, key=lambda x: x["score"])[: max(1, min(top_k, len(poses)))]
    top_idx = [int(x["pose_idx"]) for x in top]
    if not top_idx:
        return {"top_pose_rmsd_consistency": None, "geometry_similarity_top5": None}

    ref_idx = int(top_idx[0])
    ref_mol = mols[ref_idx] if 0 <= ref_idx < len(mols) else None
    if ref_mol is None:
        return {"top_pose_rmsd_consistency": None, "geometry_similarity_top5": None}

    rmsds: List[float] = []
    for idx in top_idx[1:]:
        if idx < 0 or idx >= len(mols):
            continue
        mol = mols[idx]
        if mol is None:
            continue
        try:
            if ref_mol.GetNumAtoms() != mol.GetNumAtoms():
                continue
            rmsd = float(rdMolAlign.GetBestRMS(ref_mol, mol))
            if np.isfinite(rmsd):
                rmsds.append(rmsd)
        except Exception:
            continue

    if not rmsds:
        return {"top_pose_rmsd_consistency": None, "geometry_similarity_top5": None}

    mean_rmsd = float(np.mean(np.asarray(rmsds, dtype=float)))
    geom_sim = float(1.0 / (1.0 + mean_rmsd))
    return {"top_pose_rmsd_consistency": mean_rmsd, "geometry_similarity_top5": geom_sim}


def parse_rdock_output(path: str | Path) -> Dict[str, Any]:
    """
    Parse real rDock SDF output and compute native + derived multi-pose features.

    Native features are direct tags parsed from rDock output (`SCORE*`).
    Derived features are computed from real generated poses and are explicitly marked as proxies where needed.
    """
    source = Path(path)
    if not source.exists():
        raise ValueError(f"rDock output file does not exist: {source}")
    if source.stat().st_size == 0:
        raise ValueError(f"rDock output file is empty: {source}")

    raw = source.read_text(encoding="utf-8", errors="ignore")
    blocks = _split_sdf_blocks(raw)
    if not blocks:
        raise ValueError(f"No molecule blocks found in real rDock output: {source}")

    poses: List[Dict[str, Any]] = []
    for i, block in enumerate(blocks):
        tags = _parse_block_tags(block)
        score = _safe_float(tags.get("SCORE"))
        if score is None:
            continue
        poses.append({"pose_idx": i, "pose_rank": int(tags.get("RI", i) or i) + 1, "score": score, "tags": tags})

    if not poses:
        raise ValueError(f"No SCORE fields found in real rDock output: {source}")

    poses = sorted(poses, key=lambda x: x["score"])
    scores = [float(p["score"]) for p in poses]
    best = poses[0]
    top3 = scores[:3]
    top5 = scores[:5]

    native = {
        "rdock_total_score": float(best["score"]),
        "rdock_pose_rank": int(best.get("pose_rank", 1)),
        "n_generated_poses": int(len(scores)),
        "best_pose_score": float(best["score"]),
        "mean_top3_pose_score": float(np.mean(top3)),
        "mean_top5_pose_score": float(np.mean(top5)),
        "std_top5_pose_score": float(np.std(np.asarray(top5, dtype=float), ddof=0)),
        "pose_score_gap_1_2": float(scores[1] - scores[0]) if len(scores) > 1 else np.nan,
        "rdock_restraint_term": _safe_float(best["tags"].get("SCORE.RESTR")),
        "rdock_internal_ligand_term": _safe_float(best["tags"].get("SCORE.INTRA")),
        "rdock_polar_term": _safe_float(best["tags"].get("SCORE.INTER.POLAR")),
        "rdock_vdw_term": _safe_float(best["tags"].get("SCORE.INTER.VDW")),
        "rdock_inter_term": _safe_float(best["tags"].get("SCORE.INTER")),
        "rdock_intra_vdw_term": _safe_float(best["tags"].get("SCORE.INTRA.VDW")),
        "rdock_intra_dih_term": _safe_float(best["tags"].get("SCORE.INTRA.DIHEDRAL")),
        "rdock_norm_score": _safe_float(best["tags"].get("SCORE.norm")),
        "rdock_heavy_atoms": _safe_float(best["tags"].get("SCORE.heavy")),
    }

    mols = _pose_molecules(blocks)
    rmsd_stats = _pose_rmsd_stats(poses=poses, mols=mols, top_k=5)

    std5 = native["std_top5_pose_score"]
    if std5 is None or (isinstance(std5, float) and (not np.isfinite(std5))):
        pose_stability = None
    else:
        pose_stability = float(math.exp(-float(std5)))

    inter_scores: List[float] = []
    for p in poses[:5]:
        v = _safe_float(p["tags"].get("SCORE.INTER"))
        if v is not None:
            inter_scores.append(float(v))
    inter_std = float(np.std(np.asarray(inter_scores), ddof=0)) if inter_scores else np.nan

    # Explicitly proxy-derived (not directly reported by rDock tags).
    derived = {
        "n_valid_poses": int(len(scores)),
        "top_pose_rmsd_consistency": rmsd_stats["top_pose_rmsd_consistency"],
        "contact_overlap_consistency": (float(1.0 / (1.0 + inter_std)) if np.isfinite(inter_std) else np.nan),
        "hotspot_contact_frequency": (
            float(np.mean([1.0 if (_safe_float(p["tags"].get("SCORE.INTER")) or 0.0) < 0.0 else 0.0 for p in poses[:5]]))
            if poses
            else np.nan
        ),
        "subpocket_match_score": pose_stability if pose_stability is not None else np.nan,
        "consensus_contact_score": (float(-np.mean(np.asarray(inter_scores, dtype=float))) if inter_scores else np.nan),
        "consensus_hotspot_coverage": (
            float(np.mean([1.0 if (_safe_float(p["tags"].get("SCORE.INTER.VDW")) or 0.0) < 0.0 else 0.0 for p in poses[:5]]))
            if poses
            else np.nan
        ),
        "pose_stability_proxy": pose_stability,
        "geometry_similarity_top5": rmsd_stats["geometry_similarity_top5"],
    }

    feature_provenance: List[Dict[str, Any]] = []
    for name, val in native.items():
        feature_provenance.append(
            {
                "feature_name": name,
                "feature_source": "rdock_native",
                "pose_source": "best_pose" if name.startswith("rdock_") else "top5",
                "raw_output_file": str(source),
                "parsed_from": f"{source}::SDF_TAG",
                "feature_type": "exact" if val is not None and np.isfinite(float(val)) else "unavailable",
                "available": bool(val is not None and np.isfinite(float(val))),
            }
        )
    for name, val in derived.items():
        feature_provenance.append(
            {
                "feature_name": name,
                "feature_source": "rdock_derived",
                "pose_source": "top5" if ("top" in name or "consensus" in name) else "best_pose",
                "raw_output_file": str(source),
                "parsed_from": f"{source}::derived",
                "feature_type": "proxy" if val is not None and np.isfinite(float(val)) else "unavailable",
                "available": bool(val is not None and np.isfinite(float(val))),
            }
        )

    return {
        "pose_path": str(source),
        "score": float(native["best_pose_score"]),
        "score_tag": "SCORE",
        "all_scores": scores,
        "pose_count": int(len(scores)),
        "poses": poses,
        "native_features": native,
        "derived_features": derived,
        "feature_provenance": feature_provenance,
    }