Tutorials

Each tutorial is a Jupyter notebook that downloads its data, trains a model, and explains every step. Use the Open in Colab badge at the top of a notebook (or the rocket icon in the page header) to run it in the browser, or download it with the button in the header.

Tutorial	You will learn to	Data
Quickstart	load paired data, preprocess without leakage, train, embed, check alignment, predict RNA from ATAC, save a reference	10x Multiome PBMC
CITE-seq: RNA + protein	build a fused embedding, predict protein from RNA and RNA from protein, denoise, read per-cell modality contributions	Hao et al. 2021 CITE-seq
Map query data onto a reference	embed RNA-only and ATAC-only cohorts, transfer labels with confidence scores, spot unfamiliar cells, impute the missing modality	10x Multiome PBMC
Cell-type heads and refinement	attach a classifier to a trained reference, refine with 20% of labels, predict from one modality	10x Multiome PBMC
Imputation, denoising, generation	denoise, evaluate prediction per feature, generate cells from the prior or a cell type, run an in-silico perturbation	10x Multiome PBMC
Custom modalities and likelihoods	combine three modalities, use count and beta-binomial likelihoods, learn per-cell modality gates, try a transformer encoder	synthetic

The quickstart is the best starting point; the others can be read in any order.

Extended and experimental

Notebooks that go beyond the paper, with the checks needed to interpret them built in. See Extended and experimental tutorials for what “experimental” means and what each notebook needs to run.

Tutorial	You will learn to	Data
In-silico chromatin perturbation	close a regulator’s cis or motif peaks and read out predicted RNA per cell type against matched random peak sets	10x Multiome PBMC
Transformer encoders	benchmark MLP, per-modality transformer, and fused transformer encoders, and test the fused model’s attention	10x Multiome PBMC
Unified PBMC atlas from three assays	integrate CITE-seq, Multiome and TEA-seq in one model and validate on a held-out trimodal well	CITE-seq, Multiome, TEA-seq
Prediction uncertainty	attach calibrated intervals to cross-modal predictions and flag cells unlike the reference	10x Multiome PBMC
Cross-modal quality control	find mis-paired barcodes and doublets from cross-modal disagreement, with a controlled false discovery rate	10x Multiome PBMC

For the analyses in the Genome Research article, see Paper reproduction.