pertpy 1.5.0

1.5.0 🌈

on Python PyPI

🚀 Features

• pt.md.Cytokine provides the cytokine response signatures of the Human Cytokine Dictionary for 90 cytokines in 24 PBMC cell types, as gene sets that the Enrichment tests take as targets (#1186) @Jenniliu12, @Zethson
• Statsmodels fits linear mixed models when the design contains a random intercept such as ~condition + (1 | donor), by REML with Satterthwaite t-tests like lmerTest, so single cells can be tested while accounting for the samples they come from; the kernel is shared with Dialogue and the (1 | variable) syntax with Milo (#1147) @Sizerta, @Zethson
• assign_mixture_model also writes the binary cell-by-guide assignment to adata.layers[output_layer], since the per-cell string in .obs caps and joins the assigned guides (#1173) @Zethson

⚡ Performance

• Augur scores each cross-validation fold in one pass over the forest instead of six scorer passes, skips scikit-learn's per-tree parameter validation and no longer ships the fitted trees back from the workers, with identical metrics and importances on seeded runs (#1178) @Zethson
• Dialogue runs the PMD iterations in a numba kernel (20x per fit) and solves the iterative NNLS on the normal equations, replacing the sparsecca dependency; whole runs take 3 to 13 s on 5k to 27k cells, about 2x faster than before with up to half the peak memory, where R needs 2 to 16 minutes (#1177) @Zethson
• assign_mixture_model fits each guide on its distinct count values, 20x faster on Papalexi (2.9 s to 0.14 s) and with a third of the peak memory at 100k cells x 500 guides (#1172) @Zethson

⚠️ Changed results

• Dialogue reproduces the R reference end to end: REML mixed models with Satterthwaite t-tests like lmerTest, R's PMD penalty, permutation test and cell-type pruning, Pearson pre-selection of signature genes, one-sided gene p-values, and test_phenotype_association now follows DIALOGUE.pheno (first level or True as the indicator, pooled All row, no Fisher combination) (#1177) @Zethson
• Dialogue orders the shared samples like R's ICU collation, since that order decides the signs of the PMD initialization and with them the programs found (#1177) @Zethson
• Augur is reproducible by default: unseeded estimators and folds are seeded by the subsample index like R Augur, and predict_differential_prioritization draws its permuted means without replacement from a seeded generator, as R does (#1178) @Zethson
• Dialogue's min_cells_per_sample defaults to R's 15 and only gates the ANOVA feature filter, which now raises below five retained features instead of keeping all of them; n_permutations counts the observed fit, so values below 11 cannot pass the default empirical_alpha (#1177) @Zethson
• assign_mixture_model fits the Poisson-Gaussian mixture by multi-start EM instead of 500 Adam steps from one start, which left thresholds unconverged, swapped the components on high-MOI guides and dropped small guides; thresholds now match crispat run to convergence on 105 of 106 Papalexi guides, learning_rate is deprecated and n_iter caps the EM iterations (#1172) @Zethson

🐛 Bug Fixes

• Augur encodes the condition labels after filtering to two conditions, so binary precision, recall and F1 are no longer NaN when the kept labels were not the first two; those metrics now take the first condition as the positive class like R Augur, min_cells defaults to subsample_size and applies per condition like R, data sets with fewer than 1000 features skip feature selection like R, per-cell augur_score is assigned by exact cell type match, and regressors use unstratified folds (#1178) @Zethson
• Dialogue treated signed log10 p-value z-scores as normal z-scores in phase 3, which inflated strong genes, lost the up/down direction and dropped underflowing genes from the Fisher combination (#1177) @Zethson
• Mixscape's plot_barplot no longer fails when the number of target genes is not a multiple of five, and Mixscape takes the control, perturbed and non-perturbed class labels from control, perturbation_type and the new non_perturbed_label instead of hardcoding NT, KO and NP; plot_heatmap and plot_perturbscore use the mixscape_class and pert_key columns they are given, plot_perturbscore keeps a custom palette with before_mixscape=True, and plot_barplot takes a palette (#1191) @kkkklck, @Zethson
• Mixscape takes the minimum number of cells left in a perturbed class from the new min_cells instead of min_de_genes, as Seurat does, which changes results only for a non-default min_de_genes (#1189) @kkkklck, @Zethson
• Milo's plot_nhood_counts_by_cond raised with pandas 3.1, which kept the sample column's name on the per-sample index (#1181) @Zethson
• Guide assignment: assign_by_threshold and assign_to_max_guide accept CSC and sparse-array input, explicitly stored zeros no longer enter the mixture fit, plot_heatmap saves a reusable cell order and only_return_results=True no longer writes to adata.var (#1172) @Zethson

🧰 Maintenance

• Require Python 3.13 (#1172) @Zethson
• Build the docs with the shared scverse theme and configuration from scverse-doc, with the logo blue as its accent for links and highlights (#1179) @Zethson
• The docs take their accents from the scverse-doc registry and colour the sidebar captions in the logo red (#1182) @Zethson
• Group the docs sidebar into Getting started, Guides, Reference and Project like the other scverse core packages, with a new quickstart that replaces the background page and a new glossary (#1180) @Zethson
• Skip the network tests when a download fails with a URLError (#1184) @Zethson
• Keep R from loading the system libpython in the tutorial notebooks on Ubuntu 26.04 (#1185) @Zethson
• Skip PyQt6's stubs in mypy, whose invalid syntax made mypy abort (#1190) @Zethson
• Require LLM usage disclosure in the pull request template (#1165) @Zethson