libs/docking/comparison.py

from __future__ import annotations

from typing import Dict

import numpy as np
import pandas as pd


def backend_score_semantics(backend_name: str) -> str:
    name = str(backend_name).strip().lower()
    if name == "rdock":
        return "rdock_docking_score_lower_better"
    if name == "haddock":
        return "haddock_emscoring_energy_proxy_lower_better"
    return "unknown_score_semantics"


def build_backend_comparison_table(df: pd.DataFrame) -> pd.DataFrame:
    """Build backend-comparison rows with explicit semantics and normalized scores."""
    if df.empty:
        return pd.DataFrame(
            columns=[
                "ligand_id",
                "backend_name",
                "backend_mode",
                "raw_score",
                "normalized_score",
                "zscore_within_backend",
                "score_semantics",
                "direct_comparison_justified",
                "provenance_path",
                "parsed_from",
            ]
        )

    work = df.copy()
    if "docking_score" in work.columns and "raw_score" not in work.columns:
        work["raw_score"] = pd.to_numeric(work["docking_score"], errors="coerce")
    work["backend_name"] = work["backend_name"].astype(str)
    work["backend_mode"] = work.get("backend_mode", "").astype(str)
    work["score_semantics"] = work["backend_name"].map(backend_score_semantics)
    work["provenance_path"] = work.get("raw_output_file", "").astype(str)
    work["parsed_from"] = work.get("parsed_from", "").astype(str)

    norm = np.full(work.shape[0], np.nan, dtype=float)
    z = np.full(work.shape[0], np.nan, dtype=float)
    for backend, idx in work.groupby("backend_name").groups.items():
        vals = pd.to_numeric(work.loc[idx, "raw_score"], errors="coerce").to_numpy(dtype=float)
        valid = np.isfinite(vals)
        if valid.any():
            v = vals[valid]
            ranks = pd.Series(v).rank(method="average", pct=True).to_numpy(dtype=float)
            # lower score is better -> invert percentile for "higher is better normalized"
            nv = 1.0 - ranks
            mu = float(np.mean(v))
            sd = float(np.std(v))
            zv = (v - mu) / (sd if sd > 1e-9 else 1.0)

            # map back
            norm_vals = np.full(vals.shape[0], np.nan, dtype=float)
            z_vals = np.full(vals.shape[0], np.nan, dtype=float)
            norm_vals[valid] = nv
            z_vals[valid] = zv
            norm[list(idx)] = norm_vals
            z[list(idx)] = z_vals

    work["normalized_score"] = norm
    work["zscore_within_backend"] = z

    unique_sem = set(work["score_semantics"].astype(str).tolist())
    direct = len(unique_sem) == 1
    work["direct_comparison_justified"] = bool(direct)
    if not direct:
        work["direct_comparison_justified"] = False

    cols = [
        "ligand_id",
        "backend_name",
        "backend_mode",
        "raw_score",
        "normalized_score",
        "zscore_within_backend",
        "score_semantics",
        "direct_comparison_justified",
        "provenance_path",
        "parsed_from",
    ]
    for c in cols:
        if c not in work.columns:
            work[c] = np.nan
    return work[cols].copy()


def summarize_backend_table(df: pd.DataFrame) -> pd.DataFrame:
    if df.empty:
        return pd.DataFrame(
            columns=[
                "backend_name",
                "score_semantics",
                "n_scores",
                "mean_raw_score",
                "std_raw_score",
                "mean_normalized_score",
                "direct_comparison_justified",
            ]
        )
    rows = []
    for backend, sub in df.groupby("backend_name"):
        rows.append(
            {
                "backend_name": str(backend),
                "score_semantics": str(sub["score_semantics"].iloc[0]),
                "n_scores": int(sub.shape[0]),
                "mean_raw_score": float(pd.to_numeric(sub["raw_score"], errors="coerce").mean()),
                "std_raw_score": float(pd.to_numeric(sub["raw_score"], errors="coerce").std()),
                "mean_normalized_score": float(pd.to_numeric(sub["normalized_score"], errors="coerce").mean()),
                "direct_comparison_justified": bool(sub["direct_comparison_justified"].all()),
            }
        )
    return pd.DataFrame(rows)


def semantic_compatibility_notes() -> Dict[str, str]:
    return {
        "rdock": "Pose generation + docking score; lower score is better.",
        "haddock": (
            "HADDOCK emscoring on provided complexes/poses; energy proxy; "
            "lower score is better; in this repository treated as rescoring semantics."
        ),
        "cross_backend": (
            "Direct raw-score comparison between rDock and HADDOCK is approximate; "
            "prefer within-backend normalization / rank statistics."
        ),
    }