Paper analyses with the public API

These notebooks run the analyses of the Genome Research article through UniVI’s public API: univi.datasets for the data, the fitted preprocessors in univi.preprocessing, UniVITrainer, UniVIRefiner, and univi.evaluation. They are the place to start if you want to repeat, extend, or adapt an analysis from the paper without reading the archived notebooks’ internal helper code.

Each notebook uses the model settings of the corresponding archived notebook (summarized in Supplemental Table S8) and the preprocessing described in Supplemental Table S7. Because preprocessing and refinement go through UniVI’s reusable components rather than the archived notebooks’ own code, results are close to, but not identical with, the published numbers. To regenerate the published numbers exactly, run the archived notebooks with UniVI 0.4.7 (see Paper reproduction).

Notebook	Figures	Data (univi.datasets)
Paired CITE-seq	2, 3, S1	hao_citeseq_pbmc
Paired 10x Multiome	4, S2	pbmc_multiome_10k
Bridging unpaired RNA and ATAC cohorts	5, S5	pbmc_multiome_bridge
Trimodal TEA-seq, held-out well	6, S6	teaseq_pbmc
AML mosaic integration and mutation heads	7, S7	aml_mosaic
SHARE-seq mouse skin	S3	shareseq_mouse_skin
scNMT-seq mouse gastrulation	S4	scnmt_gastrulation

Training uses the archived notebooks’ epoch limits with early stopping; a GPU is recommended. Every notebook has one cell of settings at the top (for example N_EPOCHS) that you can lower for a quick first run. The benchmarks and sweeps of Figs. 8–10 and Supplemental Figs. S8–S11 depend on other methods and large parameter grids and are only available as archived notebooks.

The datasets page describes what each download contains (files, features, metadata, splits) and links every file on Zenodo.