| """Evaluation metrics for synthetic recovery experiments.""" |
|
|
| from __future__ import annotations |
|
|
| import numpy as np |
| from scipy import stats |
|
|
|
|
| def gamma_recovery( |
| gamma_true: np.ndarray, |
| gamma_pred: np.ndarray, |
| per_gene: bool = True, |
| ) -> float | np.ndarray: |
| """Spearman correlation between true and predicted gamma. |
| |
| Parameters |
| ---------- |
| gamma_true, gamma_pred |
| Shape ``(n_cells, n_genes)``. |
| per_gene |
| If True, compute per-gene Spearman r and return the median. |
| If False, flatten and compute a single global r. |
| |
| Returns |
| ------- |
| float or np.ndarray |
| Median per-gene r (if per_gene=True) or global r. |
| """ |
| if per_gene: |
| n_genes = gamma_true.shape[1] |
| rs = np.empty(n_genes, dtype=np.float64) |
| for g in range(n_genes): |
| gt = gamma_true[:, g] |
| gp = gamma_pred[:, g] |
| if gt.std() < 1e-10 or gp.std() < 1e-10: |
| rs[g] = 0.0 |
| else: |
| rs[g] = stats.spearmanr(gt, gp).statistic |
| return float(np.nanmedian(rs)) |
| else: |
| return float(stats.spearmanr(gamma_true.ravel(), gamma_pred.ravel()).statistic) |
|
|
|
|
| def ci_coverage( |
| gamma_true: np.ndarray, |
| gamma_mean: np.ndarray, |
| gamma_var: np.ndarray, |
| level: float = 0.95, |
| ) -> float: |
| """Fraction of true gamma values within the posterior credible interval. |
| |
| Uses a Gaussian approximation: CI = mean +/- z * std. |
| |
| Parameters |
| ---------- |
| gamma_true |
| Ground-truth gamma, shape ``(n_cells, n_genes)``. |
| gamma_mean |
| Posterior mean, same shape. |
| gamma_var |
| Posterior variance, same shape. |
| level |
| Credible interval level (e.g. 0.95). |
| |
| Returns |
| ------- |
| float |
| Coverage fraction in [0, 1]. |
| """ |
| from scipy.stats import norm |
|
|
| z = norm.ppf(0.5 + level / 2) |
| std = np.sqrt(np.clip(gamma_var, 1e-10, None)) |
| lower = gamma_mean - z * std |
| upper = gamma_mean + z * std |
| inside = (gamma_true >= lower) & (gamma_true <= upper) |
| return float(inside.mean()) |
|
|
|
|
| def latent_recovery( |
| z_true: np.ndarray, |
| z_pred: np.ndarray, |
| ) -> float: |
| """Mean canonical correlation (CCA) between true and predicted latents. |
| |
| Parameters |
| ---------- |
| z_true, z_pred |
| Shape ``(n_cells, d_latent)``. |
| |
| Returns |
| ------- |
| float |
| Mean canonical correlation across components. |
| """ |
| from sklearn.cross_decomposition import CCA |
|
|
| d = min(z_true.shape[1], z_pred.shape[1]) |
| cca = CCA(n_components=d, max_iter=500) |
| X_c, Y_c = cca.fit_transform(z_true, z_pred) |
|
|
| correlations = np.array( |
| [np.corrcoef(X_c[:, i], Y_c[:, i])[0, 1] for i in range(d)] |
| ) |
| return float(np.mean(np.abs(correlations))) |
|
|
