File size: 10,459 Bytes
638ea38 fe11dd9 638ea38 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 | import numpy as np
import pandas as pd
import statsmodels.api as sm
from scipy import stats
from scipy.optimize import linear_sum_assignment
from sklearn.metrics import roc_auc_score, roc_curve
from common import RES, SEED, hom, axes, centre, effects, sgd_labels, phenotypes
CALIPERS = (0.05, 0.1, 0.25, 0.5)
PRIMARY_CALIPER = 0.25
N_BOOT = 2000
N_ORDERS = 200
STRATA = {"all annotations": None, "small-scale studies only": ["classical genetics"]}
Z = hom()
ag, ox, os_, _ = axes(Z.columns)
b = Z.median(axis=1)
E = dict(zip(("oxidative", "osmotic"), effects(centre(Z), ox, os_)))
N = {"oxidative": Z.loc[:, ox].mean(axis=1), "osmotic": Z.loc[:, os_].mean(axis=1)}
rng = np.random.default_rng(SEED)
src = phenotypes()
for stress in ("oxidative", "osmotic"):
terms = ("oxidative stress resistance",) if stress == "oxidative" else ("hyperosmotic stress resistance", "osmotic stress resistance")
s = src[(src.mutant == "null") & src.phenotype.str.rsplit(": ", n=1).str[0].isin(terms)]
s.groupby(["pmid", "exptype"]).size().rename("annotations").sort_values(ascending=False).to_csv(
RES / f"benchmark_sources_{stress}.csv")
def auc_ci(y, x, paired_idx=None):
y, x = np.asarray(y), np.asarray(x)
a = roc_auc_score(y, x)
boot = []
for _ in range(N_BOOT):
if paired_idx is None:
i = rng.integers(0, len(y), len(y))
if len(set(y[i])) < 2:
continue
else:
k = rng.integers(0, len(paired_idx), len(paired_idx))
i = paired_idx[k].ravel()
boot.append(roc_auc_score(y[i], x[i]))
lo, hi = np.percentile(boot, [2.5, 97.5])
return a, lo, hi
def smd(x1, x0):
return (x1.mean() - x0.mean()) / np.sqrt((x1.var() + x0.var()) / 2)
summary, terc, val, roc_rows = [], [], [], []
for stratum, ex in STRATA.items():
for stress in ("oxidative", "osmotic"):
lab = sgd_labels(stress, exptypes=ex)
lab = lab[lab.index.isin(Z.index)]
y = (lab == -1).astype(int).values
e = E[stress].loc[lab.index]
conc = (np.sign(e) == -lab).astype(int)
k, n = int(conc.sum()), len(conc)
row = {"stratum": stratum, "stress": stress, "genes": n, "decreased": int(y.sum()), "increased": int((1 - y).sum()),
"concordant": k, "concordance": k / n, "p_binomial": stats.binomtest(k, n, 0.5, alternative="greater").pvalue,
"smd_slow_growth": smd(b.loc[lab.index][y == 1], b.loc[lab.index][y == 0]),
"p_slow_growth_bias": stats.mannwhitneyu(b.loc[lab.index][y == 1], b.loc[lab.index][y == 0], alternative="greater").pvalue}
for name, sc in (("baseline_adjusted", e), ("naive", N[stress].loc[lab.index]), ("slow_growth", b.loc[lab.index])):
a, lo, hi = auc_ci(y, sc.values)
row.update({f"auroc_{name}": a, f"auroc_{name}_lo": lo, f"auroc_{name}_hi": hi})
summary.append(row)
if stratum == "all annotations":
val.append(pd.DataFrame({"stress": stress, "gene": lab.index, "effect": e.values, "sgd_dir": lab.values,
"concordant": conc.values, "slow_growth": b.loc[lab.index].values}))
fpr, tpr, _ = roc_curve(y, e.values)
roc_rows.append(pd.DataFrame({"stress": stress, "fpr": fpr, "tpr": tpr}))
order = np.argsort(-e.abs().values)
for bname, idx in zip(("large", "medium", "small"), np.array_split(order, 3)):
cb = conc.values[idx]
lo, hi = stats.binomtest(int(cb.sum()), len(cb)).proportion_ci()
terc.append((stress, bname, len(idx), cb.mean(), lo, hi, float(np.median(e.abs().values[idx]))))
S = pd.DataFrame(summary)
S.to_csv(RES / "benchmark_summary.csv", index=False)
pd.concat(val).to_csv(RES / "validation.csv", index=False)
pd.concat(roc_rows).round(4).to_csv(RES / "roc.csv", index=False)
T = pd.DataFrame(terc, columns=["stress", "tercile", "genes", "concordance", "ci_low", "ci_high", "median_abs_effect"])
T.to_csv(RES / "benchmark_terciles.csv", index=False)
x = pd.concat(val)
trend = []
for stress in ("oxidative", "osmotic"):
v = x[x.stress == stress]
m = sm.Logit(v.concordant.values, sm.add_constant(np.column_stack([np.abs(v.effect.values), v.slow_growth.values]))).fit(disp=0)
trend.append((stress, m.params[1], m.pvalues[1], m.params[2], m.pvalues[2], len(v)))
pd.DataFrame(trend, columns=["stress", "coef_abs_effect", "p_abs_effect", "coef_slow_growth", "p_slow_growth", "genes"]).to_csv(
RES / "benchmark_concordance_trend.csv", index=False)
def greedy(dec, inc, s, cal, order):
pool = set(inc)
pairs = []
for g in order:
if not pool:
break
p = sorted(pool)
d = np.abs(s.loc[p].values - s[g])
j = int(np.argmin(d))
if d[j] <= cal:
pairs.append((g, p[j]))
pool.remove(p[j])
return pairs
def optimal(dec, inc, s, cal):
C = np.abs(s.loc[dec].values[:, None] - s.loc[inc].values[None, :])
r, c = linear_sum_assignment(C)
return [(dec[i], inc[j]) for i, j in zip(r, c) if C[i, j] <= cal]
def evaluate(pairs, stress, method, cal):
g = [p for pr in pairs for p in pr]
y = np.array([1, 0] * len(pairs))
idx = np.arange(2 * len(pairs)).reshape(-1, 2)
d, i = b.loc[[p[0] for p in pairs]], b.loc[[p[1] for p in pairs]]
out = {"stress": stress, "method": method, "caliper_sd": cal, "pairs": len(pairs), "smd_after": smd(d, i),
"variance_ratio": d.var() / i.var(), "ks_after": stats.ks_2samp(d, i).statistic}
for name, sc in (("baseline_adjusted", E[stress]), ("naive", N[stress]), ("slow_growth", b)):
a, lo, hi = auc_ci(y, sc.loc[g].values, idx)
out.update({f"auroc_{name}": a, f"auroc_{name}_lo": lo, f"auroc_{name}_hi": hi})
return out
match = []
for stress in ("oxidative", "osmotic"):
lab = sgd_labels(stress)
lab = lab[lab.index.isin(Z.index)]
dec, inc = list(lab.index[lab == -1]), list(lab.index[lab == 1])
sd = b.loc[lab.index].std()
for cal in CALIPERS:
match.append(evaluate(greedy(dec, inc, b, cal * sd, dec), stress, "greedy nearest neighbour", cal))
match.append(evaluate(optimal(dec, inc, b, cal * sd), stress, "optimal 1:1", cal))
orders = []
for _ in range(N_ORDERS):
pr = greedy(dec, inc, b, PRIMARY_CALIPER * sd, list(rng.permutation(dec)))
g = [p for q in pr for p in q]
yy = np.array([1, 0] * len(pr))
orders.append((len(pr), roc_auc_score(yy, E[stress].loc[g]), roc_auc_score(yy, b.loc[g])))
O = np.array(orders)
match.append({"stress": stress, "method": f"greedy, {N_ORDERS} random orders (median; 2.5-97.5%)",
"caliper_sd": PRIMARY_CALIPER, "pairs": float(np.median(O[:, 0])),
"auroc_baseline_adjusted": float(np.median(O[:, 1])), "auroc_baseline_adjusted_lo": float(np.percentile(O[:, 1], 2.5)),
"auroc_baseline_adjusted_hi": float(np.percentile(O[:, 1], 97.5)), "auroc_slow_growth": float(np.median(O[:, 2]))})
q = pd.qcut(b.loc[lab.index], 5, labels=False)
w, aucs = [], []
for k in range(5):
ii = lab.index[q == k]
yy = (lab[ii] == -1).astype(int)
if yy.nunique() == 2:
aucs.append(roc_auc_score(yy, E[stress].loc[ii]))
w.append(len(ii))
match.append({"stress": stress, "method": "stratified (slow-growth quintiles), weighted mean", "caliper_sd": np.nan,
"pairs": np.nan, "auroc_baseline_adjusted": float(np.average(aucs, weights=w))})
y = (lab == -1).astype(int).values
Xr = sm.add_constant(np.column_stack([stats.zscore(E[stress].loc[lab.index].values), stats.zscore(b.loc[lab.index].values)]))
m = sm.Logit(y, Xr).fit(disp=0)
ci = m.conf_int()[1]
match.append({"stress": stress, "method": "logistic regression adjusting for slow growth (odds ratio per SD)",
"caliper_sd": np.nan, "pairs": np.nan, "odds_ratio": float(np.exp(m.params[1])),
"odds_ratio_lo": float(np.exp(ci[0])), "odds_ratio_hi": float(np.exp(ci[1])), "p": float(m.pvalues[1])})
M = pd.DataFrame(match)
M.to_csv(RES / "benchmark_matching.csv", index=False)
rng_d = np.random.default_rng(SEED + 1)
def auc_difference(y, x1, x2, paired_idx=None):
y, x1, x2 = np.asarray(y), np.asarray(x1), np.asarray(x2)
obs = roc_auc_score(y, x1) - roc_auc_score(y, x2)
boot = []
for _ in range(N_BOOT):
if paired_idx is None:
i = rng_d.integers(0, len(y), len(y))
if len(set(y[i])) < 2:
continue
else:
i = paired_idx[rng_d.integers(0, len(paired_idx), len(paired_idx))].ravel()
boot.append(roc_auc_score(y[i], x1[i]) - roc_auc_score(y[i], x2[i]))
boot = np.array(boot)
lo, hi = np.percentile(boot, [2.5, 97.5])
return obs, lo, hi, min(1.0, 2 * min((boot <= 0).mean(), (boot >= 0).mean()))
diff = []
for stress in ("oxidative", "osmotic"):
lab = sgd_labels(stress)
lab = lab[lab.index.isin(Z.index)]
y = (lab == -1).astype(int).values
for name, x2 in (("naive", N[stress]), ("slow_growth", b)):
d, lo, hi, p = auc_difference(y, E[stress].loc[lab.index].values, x2.loc[lab.index].values)
diff.append((stress, "all annotations", f"baseline_adjusted - {name}", len(lab), d, lo, hi, p))
dec, inc = list(lab.index[lab == -1]), list(lab.index[lab == 1])
pairs = greedy(dec, inc, b, PRIMARY_CALIPER * b.loc[lab.index].std(), dec)
g = [q for pr in pairs for q in pr]
yy = np.array([1, 0] * len(pairs))
idx = np.arange(2 * len(pairs)).reshape(-1, 2)
for name, x2 in (("naive", N[stress]), ("slow_growth", b)):
d, lo, hi, p = auc_difference(yy, E[stress].loc[g].values, x2.loc[g].values, idx)
diff.append((stress, f"growth-matched (greedy, caliper {PRIMARY_CALIPER} SD)", f"baseline_adjusted - {name}",
len(pairs), d, lo, hi, p))
pd.DataFrame(diff, columns=["stress", "benchmark", "comparison", "n", "auroc_difference", "ci_low", "ci_high",
"p_two_sided"]).to_csv(RES / "benchmark_auroc_differences.csv", index=False)
pd.set_option("display.width", 250)
pd.set_option("display.max_columns", 30)
|