Genome Research paper reproduction

Genome Research paper reproduction#

Ashford AJ, Enright T, Somers J, Nikolova O, Demir E. Unifying multimodal single-cell data with a mixture-of-experts β-variational autoencoder framework. Genome Research (2026). doi:10.1101/gr.281431.125

Every figure in the article was produced by a notebook in notebooks/GR_manuscript_reproducibility/. These notebooks are kept exactly as they were run and are the record of the settings used for each analysis; Supplemental Tables S7 (datasets, splits, preprocessing) and S8 (hyperparameters) summarize them.

Figures and notebooks#

Figure	Analysis	Notebook	Data
1	Model and evaluation overview	—	—
2	Paired CITE-seq integration	`Figure__2__CITE_paired` (also covered by the Figure 3 notebook)	`hao_citeseq_pbmc`
3, S1	Cross-modal marker reconstruction, subtype structure	`Figure__3__CITE_paired_biological_latent`	`hao_citeseq_pbmc`
4, S2	Paired Multiome integration, 3D latent views	`Figure__4__Multiome_paired`	`pbmc_multiome_10k`
5, S5	Bridging unpaired RNA and ATAC cohorts; supervised refinement	`Figure__5__Multiome_bridge_mapping_and_fine-tuning`	`pbmc_multiome_bridge`
6, S6	Trimodal TEA-seq, held-out well	`Figure__6__TEA-seq_tri-modal`	`teaseq_pbmc`
7, S7	AML mosaic integration, mutation heads	`Figure__7__AML_bridge_mapping_and_fine-tuning`	`aml_mosaic`
8	Benchmark against 14 methods	`Figure__8__benchmarking_against_pytorch_tools`, `..._against_R_tools`, `..._merging_and_plotting_runs`	`pbmc_multiome_10k`
9	Paired-overlap ablation, computational scaling	`Figure__9__paired_data_ablation_and_computational_scaling_performance` (+ `_compile_plots_from_results_df`)	`pbmc_multiome_10k`
10	Cell-population ablation, modality gating	`Figure_10__cell_population_ablation_MoE` (+ `_compile_plots_from_results_df`)	`pbmc_multiome_10k`
S3	SHARE-seq mouse skin	`Supple_____mouse_skin_SHARE-seq_integration`	`shareseq_mouse_skin`
S4	scNMT-seq mouse gastrulation (RNA, CpG, GpC)	`Supple_____scNMT-seq_mouse_gastrulation_data`	`scnmt_gastrulation`
S8–S11	Hyperparameter, latent-size, dropout sensitivity; scaling	`Supple_____grid-sweep` (+ `_compile_plots_from_results_df`)	`pbmc_multiome_10k`
Supp. Notebook S1	Raw-count decoding and in-silico peak perturbation	`Supple_____Supplemental_Notebook_S1`	`pbmc_multiome_10k`

Notebook file names begin with UniVI_manuscript_GR-. Figures 8–10 and S8–S11 have _compile_plots_from_results_df companions that redraw the panels from saved results tables. The Supplemental Code archive published with the article is a snapshot of this repository at release v0.4.7.

Rerunning an analysis#

Install the release the article used, ideally in the recorded environment:
```
conda env create -f envs/univi_v0.4.7_env.yml
conda activate univi_v0.4.7
```
or pip install univi==0.4.7 in a fresh environment.
Get the data. The first cells of each archived notebook name the files it reads, and the datasets page lists the original data behind every analysis. The Fig. 4 notebook reads the same two files that uds.pbmc_multiome_10k() downloads. The other hosted datasets were exported partway through their archived notebooks (each dataset’s section shows where), so they are the starting point for the public-API notebooks rather than drop-in inputs for the archived ones.
Point the notebook at your files. The first cells of each notebook set paths (for example DATA_ROOT, RNA_PATH); edit them for your machine.
Run the notebook top to bottom. Training used an NVIDIA GPU for the main-text figures (Supplemental Table S8 lists the device per analysis). Neural network training is not bit-for-bit reproducible across hardware and library versions, so expect small numerical differences.

The parameter_files/params_*_GR_fig*.json files and scripts/revision_reproduce_all.sh offer a script-driven route through the same analyses (see scripts).

Running the analyses with the public API#

Paper analyses with the public API has one notebook per analysis (Figs. 2–7, S3, S4) that downloads the data and runs it with UniVI’s reusable components instead of the archived notebooks’ own code. The datasets page describes each downloadable dataset and how it was exported from the archived notebook.

The tutorials and the paper#

The tutorials follow the same workflows with the newer public API (fitted preprocessors, make_loader, UniVIRefiner, reference bundles, dataset downloads) and use the paper’s CITE-seq and Multiome hyperparameters. They are written for learning and reuse; for the published numbers, run the archived notebooks with v0.4.7.

Published figures#

Main text

Figure 2. Paired CITE-seq integration.

Figure 3. Cross-modal reconstruction of CITE-seq markers.

Figure 4. Paired 10x Multiome integration.

Figure 5. Bridging unpaired RNA and ATAC cohorts.

Figure 6. Trimodal TEA-seq.

Figure 7. AML mosaic integration.

Figure 8. Benchmarking.

Supplemental

Supplemental Figure S1. CITE-seq marker recovery at finer annotation.

Supplemental Figure S2. 3D views of the Multiome latent space.

Supplemental Figure S3. SHARE-seq mouse skin.

Supplemental Figure S4. scNMT-seq mouse gastrulation. Processed RNA, CpG, and GpC inputs: Zenodo 10.5281/zenodo.22885463; load with uds.load("scnmt_gastrulation"). See the dataset description and the S4 API notebook.