"""Replicate SCLC tissue signals and compare target dependencies across lung lineages.""" import argparse import csv import gzip import hashlib import json import re from collections import defaultdict from pathlib import Path import numpy as np from scipy import stats TARGETS = ("ANXA2", "CIP2A") MARKERS = ("BIRC2", "BIRC3") CONTROL_GENES = ("RPL4", "PSMA1") NF = "HALLMARK_TNFA_SIGNALING_VIA_NFKB" AU = "GOBP_MACROAUTOPHAGY" RECLASSIFIED = {"ACH-000292", "ACH-000530", "ACH-000890"} def write_csv(path, records): with path.open("w", encoding="utf-8", newline="") as handle: writer = csv.DictWriter(handle, fieldnames=records[0]) writer.writeheader() writer.writerows(records) def bh(rows, p_key="p", q_key="q"): order = sorted(range(len(rows)), key=lambda i: rows[i][p_key]) previous = 1.0 for rank, i in reversed(list(enumerate(order, start=1))): previous = min(previous, rows[i][p_key] * len(rows) / rank) rows[i][q_key] = previous def geo_rows(path, delimiter): with gzip.open(path, "rt", encoding="utf-8-sig") as f: yield from csv.reader(f, delimiter=delimiter) def gene_symbol(value): symbol = value.strip().upper() return "CIP2A" if symbol == "KIAA1524" else symbol def read_cohort(folder, accession): paired = accession == "GSE149507" key = accession.lower() meta = {} for row in geo_rows(folder / f"{key}.meta.gz", "\t"): if row and row[0] in {"!Sample_title", "!Sample_source_name_ch1", "!Sample_geo_accession"}: meta[row[0]] = [v.strip() for v in row[1:]] sample_meta = dict(zip(meta["!Sample_title"], zip(meta["!Sample_source_name_ch1"], meta["!Sample_geo_accession"], strict=True), strict=True)) data = geo_rows(folder / (f"{key}.csv.gz" if paired else f"{key}.tsv.gz"), "," if paired else "\t") offset = 2 if paired else 1 titles = [v.strip() for v in next(data)[offset:]] if not paired: for title in titles: if title.endswith(".normal"): assert sample_meta[title.removesuffix(".normal")][0] == "normal" titles = [title.removesuffix(".normal") for title in titles] assert len(titles) == len(set(titles)) and set(titles) == set(sample_meta) samples = [] for title in titles: tissue, gsm = sample_meta[title] assert tissue.lower() in {"tumor", "normal"} patient = "" if paired: patient, suffix = re.fullmatch(r"(SCLC_\d+)_(ca|n)", title).groups() assert (suffix == "ca") == (tissue.lower() == "tumor") samples.append(dict(accession=accession, sample=gsm, title=title, tissue=tissue.lower(), patient=patient)) seen = defaultdict(list) for row in data: if not row: continue symbol = gene_symbol(row[offset-1]) if not symbol or symbol in {"NA", "---"}: continue assert len(row) == len(titles) + offset values = np.array([float(x) if x not in ("NA", "") else np.nan for x in row[offset:]]) if np.all(np.isfinite(values)): seen[symbol].append(values) matrix = {name: np.mean(values, axis=0) for name, values in seen.items()} assert all(g in matrix for g in TARGETS + MARKERS), (key, [g for g in TARGETS + MARKERS if g not in matrix]) return samples, matrix def groups(samples): tumors = [i for i, r in enumerate(samples) if r["tissue"] == "tumor"] normals = [i for i, r in enumerate(samples) if r["tissue"] == "normal"] paired = bool(samples[0]["patient"]) if paired: indexed = {samples[i]["patient"]: i for i in normals} assert len(tumors) == len(normals) == len(indexed) == 18 normals = [indexed[samples[i]["patient"]] for i in tumors] else: assert (len(tumors), len(normals)) == (79, 7) return tumors, normals, paired def read_sets(folder): found = {} for file in ("hallmark.gmt", "macro.gmt"): for line in (folder / file).read_text(encoding="utf-8-sig").splitlines(): name, url, *genes = line.rstrip().split("\t") if name in {NF, AU}: found[name] = {gene_symbol(g) for g in genes} assert set(found) == {NF, AU} return found def tissue_effect(v, t, n, paired): a, b = v[t], v[n] delta = float(a.mean() - b.mean()) if paired: diff = a - b se = float(stats.sem(diff)) crit = stats.t.ppf(.975, len(diff) - 1) test = stats.ttest_rel(a, b) rank_p = stats.wilcoxon(diff).pvalue sensitivity = [np.delete(diff, i).mean() for i in range(len(diff))] up = int(np.sum(diff > 0)) else: av = a.var(ddof=1) / len(a) bv = b.var(ddof=1) / len(b) se = np.sqrt(av + bv) df = (av+bv)**2 / (av**2/(len(a)-1) + bv**2/(len(b)-1)) crit = stats.t.ppf(.975, df) test = stats.ttest_ind(a, b, equal_var=False) rank_p = stats.mannwhitneyu(a, b, alternative="two-sided").pvalue sensitivity = [a.mean() - np.delete(b, i).mean() for i in range(len(b))] up = "" return dict(n_tumor=len(a), n_normal=len(b), mean_difference=delta, ci_low=delta-crit*se, ci_high=delta+crit*se, p=float(test.pvalue), rank_p=float(rank_p), tumor_higher_pairs=up, leave_one_out_min=min(sensitivity), leave_one_out_max=max(sensitivity)) def tissue_analysis(folder, out): cohorts = {name: read_cohort(folder, name) for name in ("GSE149507", "GSE60052")} common = sorted(set.intersection(*(set(matrix) for _, matrix in cohorts.values()))) gene_sets = read_sets(folder) included = {name: sorted(members.intersection(common)) for name, members in gene_sets.items()} assert min(map(len, included.values())) >= 50 coverage = [dict(pathway=name, reference_genes=len(gene_sets[name]), common_genes=len(genes), common_rank_background=len(common), members=";".join(genes)) for name, genes in included.items()] write_csv(out / "pathway_coverage.csv", coverage) effects, sample_rows, associations, pathway_sensitivity = [], [], [], [] for accession, (samples, matrix) in cohorts.items(): ranks = stats.rankdata(np.stack([matrix[g] for g in common]), axis=0) ranks = (ranks - 0.5) / len(common) * 100 indices = {g: i for i, g in enumerate(common)} scores = {name: ranks[[indices[g] for g in genes]].mean(axis=0) for name, genes in included.items()} signals = {gene: matrix[gene] for gene in TARGETS + MARKERS} | scores t, n, paired = groups(samples) for name, genes in included.items(): values = np.stack([matrix[g] for g in genes]) variable = values.std(axis=1, ddof=1) > 0 score = stats.zscore(values[variable], axis=1, ddof=1).mean(axis=0) pathway_sensitivity.append(dict(accession=accession, feature=name, genes=int(variable.sum()), **tissue_effect(score, t, n, paired))) cohort_effects = [dict(accession=accession, feature=name, unit="rank percentile points" if name in scores else "published log2 expression", **tissue_effect(v, t, n, paired)) for name, v in signals.items()] bh(cohort_effects) effects.extend(cohort_effects) for i, sample in enumerate(samples): sample_rows.append(dict(sample, **{name: values[i] for name, values in signals.items()})) for gene, pathway in zip(TARGETS, (NF, AU)): test = stats.spearmanr(signals[gene][t], scores[pathway][t]) associations.append(dict(accession=accession, target=gene, pathway=pathway, n_tumor=len(t), rho=test.statistic, nominal_p=test.pvalue)) write_csv(out / "tissue_effects.csv", effects) write_csv(out / "tissue_sample_values.csv", sample_rows) write_csv(out / "tumor_only_target_pathway_correlations.csv", associations) write_csv(out / "pathway_standardized_sensitivity.csv", pathway_sensitivity) def cellosaurus_models(folder): """Join exact ACH IDs and unambiguous NCIt histology in official API exports.""" lookup = defaultdict(set) for group, code in (("SCLC", "C4917"), ("LUAD", "C3512")): stem = f"cello-exact-{group.lower()}" pages = [folder / f"{stem}.tsv", *sorted(folder.glob(f"{stem}-page*.tsv"))] rows = [] for path in pages: text = path.read_text(encoding="utf-8-sig") expected = int(re.search(r"response\.numFound: (\d+)", text)[1]) offset = int(re.search(r"query\.start: (\d+)", text)[1]) assert offset == len(rows), (path, offset, len(rows)) rows.extend(csv.DictReader((line for line in text.splitlines() if line and not line.startswith("##")), delimiter="\t")) assert len(rows) == expected, (group, len(rows), expected) assert len({r["ac"] for r in rows}) == expected for row in rows: disease = {x.split("; ")[1] for x in row["di"].split("|") if x.startswith("NCIt; ")} if disease != {code}: continue for ref in row["dr"].split("|"): if re.fullmatch(r"DepMap; ACH-\d+", ref): model = ref.split("; ")[1] lookup[model].add((group, row["id"], row["ac"])) return {model: next(iter(assignments)) for model, assignments in lookup.items() if len(assignments) == 1} def depmap_models(path): groups = {"SCLC": "SCLC", "LUAD": "LUAD"} models = {} with path.open(newline="", encoding="utf-8-sig") as handle: for row in csv.DictReader(handle): code = row["OncotreeCode"] if code in groups and row["OncotreeLineage"] == "Lung": models[row["ModelID"]] = groups[code], row["CellLineName"], row.get("RRID", "") return models def effect_records(effects, models): records = [] with effects.open(newline="", encoding="utf-8-sig") as f: reader = csv.reader(f) columns = next(reader) names = {gene_symbol(value.split(" (")[0]): i for i, value in enumerate(columns)} targets = TARGETS + CONTROL_GENES for row in reader: if row[0] not in models: continue group, name, source_id = models[row[0]] for gene in targets: x = row[names[gene]] if x and np.isfinite(float(x)): records.append(dict(model_id=row[0], model_name=name, source_id=source_id, group=group, gene=gene, chronos=float(x), reclassified_model=row[0] in RECLASSIFIED)) return records def group_comparison(a, b, rng): test = stats.mannwhitneyu(a, b, alternative="two-sided") median_diff = np.median(a) - np.median(b) boot = np.median(rng.choice(a, size=(10000, len(a)), replace=True), axis=1) boot -= np.median(rng.choice(b, size=(10000, len(b)), replace=True), axis=1) ci = np.quantile(boot, [0.025, .975]) # Leave-one-model-out effect size can be computed from each model's pairwise wins. pairwise = np.sign(a[:, None] - b[None, :]) leave_one = np.r_[(pairwise.sum() - pairwise.sum(axis=1))/((len(a)-1)*len(b)), (pairwise.sum() - pairwise.sum(axis=0))/(len(a)*(len(b)-1))] trimmed = stats.mannwhitneyu(np.sort(a)[3:], np.sort(b)[3:], alternative="two-sided") return dict(n_sclc=len(a), n_luad=len(b), median_sclc=np.median(a), median_luad=np.median(b), median_difference=median_diff, bootstrap_low=ci[0], bootstrap_high=ci[1], cliffs_delta=pairwise.mean(), p=test.pvalue, sclc_below_minus_0_5=int(np.sum(a < -.5)), luad_below_minus_0_5=int(np.sum(b < -.5)), sclc_below_minus_1=int(np.sum(a < -1)), luad_below_minus_1=int(np.sum(b < -1)), loo_delta_min=min(leave_one), loo_delta_max=max(leave_one), trimmed_delta=2*trimmed.statistic/((len(a)-3)*(len(b)-3))-1, trimmed_p=trimmed.pvalue) def dependency_analysis(effects, model_source, out, use_cellosaurus): models = cellosaurus_models(model_source) if use_cellosaurus else depmap_models(model_source) records = effect_records(effects, models) write_csv(out / "depmap_model_scores.csv", records) rng = np.random.default_rng(149507) results, sensitivity = [], [] for gene in TARGETS + CONTROL_GENES: subset = [r for r in records if r["gene"] == gene and not r["reclassified_model"]] a = np.array([r["chronos"] for r in subset if r["group"] == "SCLC"]) b = np.array([r["chronos"] for r in subset if r["group"] == "LUAD"]) assert len(a) >= 15 and len(b) >= 40, (gene, len(a), len(b)) results.append(dict(gene=gene, role="candidate" if gene in TARGETS else "essential control", **group_comparison(a, b, rng), q="")) if gene in TARGETS: excluded_names = {"HCC827 GR5", "MOR/CPR", "NCI-H2286"} stricter_luad = np.array([r["chronos"] for r in subset if r["group"] == "LUAD" and r["model_name"] not in excluded_names]) sensitivity.append(dict(gene=gene, excluded_luad_models=";".join(sorted(excluded_names)), **group_comparison(a, stricter_luad, rng))) bh([r for r in results if r["gene"] in TARGETS]) write_csv(out / "depmap_comparison.csv", results) write_csv(out / "depmap_luad_annotation_sensitivity.csv", sensitivity) def main(): parser = argparse.ArgumentParser(description=__doc__) parser.add_argument("inputs", type=Path) parser.add_argument("out", type=Path) parser.add_argument("--effects", type=Path) parser.add_argument("--model-source", type=Path) parser.add_argument("--cellosaurus", action="store_true") parser.add_argument("--only", choices=("tissue", "dependency")) args = parser.parse_args() args.out.mkdir(parents=True, exist_ok=True) if args.only != "dependency": tissue_analysis(args.inputs, args.out) if args.only != "tissue": dependency_analysis(args.effects, args.model_source, args.out, args.cellosaurus) if __name__ == "__main__": main()