Release v0.1.0 · JinHongDu-Lab/causarray

[0.1.0]

Inference fix for small perturbation arms, and the GLM engine moves to crispyx
(>= 0.1.5 now required).

crispyx >= 0.1.5 is required. causarray/glm_onehot.py, the two-stage
per-perturbation fitter and the design preconditioner are removed in favour
of crispyx's StructuredGLMBatchFitter and dense batch fitter, which agree
with the statsmodels path to 2e-5 on well-conditioned genes.
fit_glm_auto routes on gene count only (_FAST_MIN_P = 10), with the
_FAST_MAX_COEF divergence check falling back to statsmodels.
Fitted means are no longer floored at min_mu = 0.5, which biased genes
below about one count per cell.
LFC uses the influence-function variance (usevar='pooled') only; the
Welch 'unequal' variance overstated standard errors.
Propensity scores default to calibrated probabilities
(class_weight=None); 'balanced' remains available.
ps_clip defaults to 'auto', a prevalence-aware bound per treatment;
resolved bounds are returned as estimation['ps_clip_bounds'].
LFC default thres_min='auto' tests a gene only when about min_counts
(default 5) counts are expected in the smaller arm.

A Poisson variance floor 1/(n1*tau1) + 1/(n0*tau0) in LFC, so chance
all-zero arms of sparse genes are not called while complete knockouts stay
significant (var_floored column).
Small-sample correction for K=1: variance scaled by n/(n-d) and a t
reference with n-d degrees of freedom.
Per-pair support columns n_treated, n_control, count_treated,
count_control and std_raw in every LFC result.
RuntimeWarnings for arms under 200 cells hitting the variance floor, and
for backend='fast' without crispyx.
tune_penalty_factor picks the smallest per-treatment L2 penalty on one
propensity covariate that meets a support target, checking the
dropped-covariate limit first to detect infeasible treatments.
gcate_lfc_batch(save_nuisances=True) saves each batch's outcome-model
predictions to <cache_path stem>.nuisances.h5 for re-estimation under a
different propensity model; budget about
8 * n_cells * n_genes * n_treatments bytes per batch.
estimate_r accepts backend and returns a time_s column.

Designs with one-hot treatment blocks are solved by a Schur complement,
flat in the number of treatments (n = p = 3,000: 5.1 s with 10 treatments,
7.7 s with 200).
GCATE's objective and gradients use parallel numba kernels with a fixed
reduction order (results independent of thread count); one update on a
3,000 x 3,000 problem drops from ~2.3 s to 0.5 s.
estimate_r fits the initial [X | A] GLM once for all candidate r.

NB dispersion falls back to the per-gene estimate_disp when no batched
estimate is available, instead of a fixed size of 1.
estimate_disp and fit_glm accept offset=False.
fit_glm with several treatments and impute=False returned zero
coefficients.
mem_limit_gb is honoured on every imputation route.
fit_glm_ondisk drops cells with no counts (with a warning) instead of
returning NaN coefficients.