scripts/og2stats.py

"""Gross statistics of the subset: base frequencies, GC, soft-masking, CpG, k-mer spectra, Markov entropies.

  python og2stats.py [--root .] [--splits train heldout valid] [--workers 8]

For every split, per subset (directory), per domain (D__ of the window's tag; '(untagged)' otherwise) and overall:
  tokens, windows; counts of A, C, G, T, N; GC fraction (of ACGT); soft-masked (lowercase) fraction;
  CpG observed/expected = f(CG) / (f(C) f(G)) from dinucleotides.
Overall and per subset, from 8-mer counts (case folded; k-mers containing N or crossing a window boundary skipped):
  k-mer counts for k = 1..8 (by marginalising the 8-mer counts over trailing positions), the entropy H_k of each
  k-mer distribution and the Markov conditional entropies h_k = H_{k+1} - H_k (bits per base given the previous k
  bases: the in-sample bits/base of an order-k model), distinct 8-mers, the 20 most frequent 8-mers, and
  Chargaff's second rule (mean |f(w) - f(revcomp w)| / (f(w) + f(revcomp w)) over 4-mers).
Windows are those listed in <split>/selected.json when present (`og2subset.py finalize`), else all.
Writes <root>/stats/<split>.json and <root>/stats/<split>_kmer8.npz (8-mer counts, overall and per subset).
"""

from __future__ import annotations

import argparse
import json
import math
import re
from collections import defaultdict
from concurrent.futures import ProcessPoolExecutor
from pathlib import Path

import numpy as np

K = 8
BASES = "ACGT"


def taxon(tag: str, rank: str) -> str:
    m = re.search(rf"{rank}__([^;|]*)", tag)
    return m.group(1) if m else "(untagged)"


def shard_stats(npy: str, ids: list[int] | None) -> dict:
    """Counts for one shard: per domain base/lower/CpG counts and the shard's 8-mer counts."""
    base = Path(npy).with_suffix("")
    arr = np.load(npy)
    rows = [json.loads(line) for line in base.with_suffix(".jsonl").read_text().splitlines()]
    if ids is not None:
        rows = [rows[i] for i in ids]
    kmer = np.zeros(4**K, dtype=np.int64)
    dom: dict[str, np.ndarray] = defaultdict(lambda: np.zeros(8, dtype=np.int64))  # A C G T N lower CG windows
    for r in rows:
        w = arr[r["offset"]: r["offset"] + r["len"]]
        b = (w & 7).astype(np.int64)
        d = dom[taxon(r["tag"], "D")]
        d[:5] += np.bincount(b, minlength=5)[:5]
        d[5] += int(np.count_nonzero(w & 0x80))
        if len(b) > 1:
            d[6] += int(np.count_nonzero((b[:-1] == 1) & (b[1:] == 2)))
        d[7] += 1
        n = len(b) - K + 1
        if n <= 0:
            continue
        bad = np.concatenate([[0], np.cumsum(b >= 4)])
        ok = (bad[K:] - bad[:-K]) == 0  # no N in positions i..i+K-1
        idx = np.zeros(n, dtype=np.int64)
        for j in range(K):
            idx = idx * 4 + np.minimum(b[j:j + n], 3)
        kmer += np.bincount(idx[ok], minlength=4**K)
    return {"subset": base.parent.name, "domains": {k: v.tolist() for k, v in dom.items()}, "kmer": kmer}


def entropy(counts: np.ndarray) -> float:
    p = counts[counts > 0] / counts.sum()
    return float(-(p * np.log2(p)).sum())


def kmer_summary(k8: np.ndarray) -> dict:
    if k8.sum() == 0:
        return {}
    t = k8.reshape([4] * K)
    H = {}
    for k in range(1, K + 1):
        ck = t.sum(axis=tuple(range(k, K))) if k < K else t
        H[k] = entropy(ck.ravel())
    h = {0: H[1], **{k: H[k + 1] - H[k] for k in range(1, K)}}
    c4 = t.sum(axis=tuple(range(4, K))).ravel().astype(float)
    rc = np.array([int("".join(str(3 - int(c)) for c in np.base_repr(i, 4).zfill(4)[::-1]), 4) for i in range(256)])
    ch2 = float(np.mean(np.abs(c4 - c4[rc]) / np.maximum(c4 + c4[rc], 1)))
    top = np.argsort(k8)[::-1][:20]
    word = lambda i: "".join(BASES[int(c)] for c in np.base_repr(int(i), 4).zfill(K))
    return {"kmers_counted": int(k8.sum()), "distinct_8mers": int((k8 > 0).sum()),
            "entropy_bits": {str(k): round(v, 5) for k, v in H.items()},
            "markov_bits_per_base": {str(k): round(v, 5) for k, v in h.items()},
            "chargaff2_4mer_asymmetry": round(ch2, 5),
            "top_8mers": [[word(i), round(float(k8[i] / k8.sum()), 6)] for i in top]}


def base_summary(c: np.ndarray) -> dict:
    a, cc, g, t, n, low, cg, win = (int(x) for x in c)
    acgt = a + cc + g + t
    tot = acgt + n
    fc, fg = cc / max(acgt, 1), g / max(acgt, 1)
    return {"tokens": tot, "windows": win,
            "freq": {"A": a / max(tot, 1), "C": cc / max(tot, 1), "G": g / max(tot, 1), "T": t / max(tot, 1),
                     "N": n / max(tot, 1)},
            "gc": (cc + g) / max(acgt, 1), "soft_masked": low / max(tot, 1),
            "cpg_obs_exp": (cg / max(acgt - win, 1)) / (fc * fg) if fc * fg > 0 else math.nan}


def split_stats(root: Path, split: str, workers: int) -> dict:
    sel_path = root / split / "selected.json"
    sel = json.loads(sel_path.read_text())["shards"] if sel_path.exists() else None
    jobs = []
    for npy in sorted((root / split).glob("*/*.npy")):
        key = str(npy.with_suffix("").relative_to(root))
        jobs.append((str(npy), None if sel is None else sel.get(key, [])))
    sub_k: dict[str, np.ndarray] = defaultdict(lambda: np.zeros(4**K, dtype=np.int64))
    sub_c: dict[str, np.ndarray] = defaultdict(lambda: np.zeros(8, dtype=np.int64))
    dom_c: dict[str, np.ndarray] = defaultdict(lambda: np.zeros(8, dtype=np.int64))
    with ProcessPoolExecutor(workers) as ex:
        for res in ex.map(shard_stats, *zip(*jobs), chunksize=4):
            sub_k[res["subset"]] += res["kmer"]
            for d, v in res["domains"].items():
                sub_c[res["subset"]] += np.array(v)
                dom_c[d] += np.array(v)
    allk = sum(sub_k.values()) if sub_k else np.zeros(4**K, dtype=np.int64)
    allc = sum(sub_c.values()) if sub_c else np.zeros(8, dtype=np.int64)
    out_dir = root / "stats"
    out_dir.mkdir(exist_ok=True)
    np.savez_compressed(out_dir / f"{split}_kmer8.npz", overall=allk, **{f"subset_{k}": v for k, v in sub_k.items()})
    st = {"split": split, "selected": sel is not None, "overall": {**base_summary(allc), **kmer_summary(allk)},
          "subsets": {k: {**base_summary(sub_c[k]), **kmer_summary(sub_k[k])}
                      for k in sorted(sub_c, key=lambda s: -sub_c[s][:5].sum())},
          "domains": {k: base_summary(v) for k, v in sorted(dom_c.items(), key=lambda kv: -kv[1][:5].sum())}}
    (out_dir / f"{split}.json").write_text(json.dumps(st, indent=1))
    o = st["overall"]
    print(f"{split}: {o['tokens'] / 1e9:.3f}B tokens, GC {o['gc']:.4f}, soft-masked {o['soft_masked']:.4f}, "
          f"CpG o/e {o['cpg_obs_exp']:.3f}, h_k " + " ".join(f"{v:.3f}" for v in o.get("markov_bits_per_base", {}).values()))
    return st


def main() -> None:
    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
    ap.add_argument("--root", type=Path, default=Path("."))
    ap.add_argument("--splits", nargs="+", default=["train", "heldout", "valid"])
    ap.add_argument("--workers", type=int, default=8)
    a = ap.parse_args()
    for split in a.splits:
        if (a.root / split).exists():
            split_stats(a.root, split, a.workers)


if __name__ == "__main__":
    main()