Biology Atlas

Map / Concepts

concept · intermediate · partly written

How MOFA+, WNN and totalVI join two layers, and how they are benchmarked

All three join by the shared cell or sample, keep each layer's own features, and have been compared mostly by their own authors; independent tests with a trivial baseline are patchy.

The join and the difference

MOFA+, Seurat's weighted-nearest-neighbour method (WNN) and totalVI all join two layers by the cell barcode or sample id and keep each layer's own columns, so the gene-to-protein many-to-many mapping never enters. They differ after the join: a linear factor model, a per-cell weighted neighbour graph, and a deep generative model. All three are one group's choice; no standards body defines any of them. 123

Reported, not yet replicated

MOFA+: a shared set of factors more

Each view (layer) is modelled as the product of cell factor values and feature weights plus noise, which is sparse PCA generalised to several views. Automatic relevance determination lets a factor be active in one view and off in another, so factors split into shared and view-specific. Inference is variational Bayes, with stochastic variational inference for data too large for GPU memory. The readout is a table of variance explained per factor per view. The model is non-convex; the original authors used 25 random restarts and kept the best. 41

Reported, not yet replicated

WNN: a per-cell weighted graph more

Each layer is reduced separately, a nearest-neighbour graph is built inside each, each cell is predicted from its neighbours in both layers, and a softmax gives two non-negative weights per cell that sum to 1. The output is a graph, not an embedding, so embedding-based metrics cannot score it, and it is rated on fewer metrics in the largest independent benchmark. In the authors' vignette progenitors lean on RNA and T cells on protein, which they attribute to the antibody panel; a weight therefore reflects the panel as much as the biology. Antibody tags are named for the CD target, for example CD25 is IL2RA, CD20 is MS4A1 (membrane spanning 4-domains A1, HGNC:7315, with CD20 listed as a previous symbol) and CD28 is CD28 (CD28 molecule, HGNC:1653). 256

Reported, not yet replicated

totalVI: one latent space, two decoders more

It assumes RNA and protein are generated from one latent space of cells, 20-dimensional per cell, with decoders for RNA counts and for protein counts as a background plus foreground mixture. A protein missing in a batch is zero-filled at input and ignored in the loss, and imputed protein quality relies on how well batches mix. Its own documentation lists that it effectively requires a GPU and that the balance between RNA and protein in the representation is hard to read. 3

Reported, not yet replicated

Who benchmarked, and was the trivial baseline run more

The authors' own comparisons exist for each method; the 2021 head-to-head of WNN, MOFA+ and totalVI was run by WNN's authors on one bone-marrow dataset. The NeurIPS 2021 challenge was run by the dataset's makers and included negative controls (zeros, random, mean) and baselines; zeros scored well on sparse ATAC for RMSE, and rerunning the overall winner four times moved its rank by up to 8 places. Makrodimitris and colleagues, independent of the methods' authors, found that concatenating 32 principal components of each layer gave Matthews correlation 0.93 against 0.87 for the best joint method (MOFA+) on RNA plus ATAC PBMC cell-type classification, and that a generalised linear model beat MOFA+, ccVAE and totalVI at imputing protein from RNA. Liu and colleagues (14 methods, 13 paired RNA plus protein datasets, a Registered Report, not by the three methods' authors) found WNN, sciPENN and Multigrate generally better on the example dataset, and no trivial baseline was found in its main text. Hu and colleagues' benchmark (14 prediction and 18 integration algorithms on 47 datasets) was read as an abstract only. 27859

Reported, not yet replicated

What no source settles more

No independent test pits WNN, MOFA+ and totalVI against RNA-only and concatenated-PCA baselines on RNA plus protein with labels not derived from any of them. Rank positions of the three in Liu's figure could not be re-opened and no ordinal is given. 58

Reported, not yet replicated

Where the methods break more

WNN assumes the layers do not define conflicting states, and in the authors' own test added protein noise drove the protein weight toward zero. Its gain depends on the pair: counts of wrongly connected CD8 and CD4 cells were 944 for RNA, 12 for protein and 20 for WNN in cord-blood CITE-seq, and 984, 373 and 322 for RNA, ATAC and WNN in naive cells of RNA plus ATAC PBMC. At 99% unpaired cells it failed to produce results in the MultiVI authors' test. MOFA+ is linear and a large high-variance layer can dominate. totalVI's imputation depends on batch mixing. 213

Reported, not yet replicated

Plain answer

Each of the three methods joins layers through the shared cell or sample, then models the pair differently; in independent tests a plain concatenation of principal components was hard to beat when both layers are present. 82

Reported, not yet replicated

Why it matters

A benchmark score of an integration method is relative to the other methods unless a trivial baseline is in the table. 5

Our inference

Real numbers

Makrodimitris RNA plus ATAC 2026-11-18

Concatenated 32 principal components per layer, Matthews correlation 0.93; best joint method MOFA+ 0.87. 8

Reported, not yet replicated

NeurIPS 2021 challenge size 2026-11-18

More than 280 participants in 172 teams submitted over 2,600 models in three months. 7

Reported, not yet replicated

MOFA on leukaemia patients 2026-11-18

200 patients, 4 views, 10 factors; variance explained 41% drug response, 38% mRNA, 24% methylation, 24% mutations. 4

Reported, not yet replicated

Still open here

Not yet written: Equation.

Linked from

Sources

  1. Argelaguet R, Arnol D, Bredikhin D, Deloro Y, et al.. MOFA+: a statistical framework for comprehensive integration of multi-modal single-cell data. Genome Biol (2020). paper; opened 2026-11-18 all records citing this
  2. Hao Y, Hao S, Andersen-Nissen E, Mauck WM, et al.. Integrated analysis of multimodal single-cell data. Cell (2021). paper; opened 2026-11-18 all records citing this
  3. Gayoso A, Steier Z, Lopez R, Regier J, et al.. Joint probabilistic modeling of single-cell multi-omic data with totalVI. Nat Methods (2021). paper; opened 2026-11-18 all records citing this
  4. Argelaguet R, Velten B, Arnol D, Dietrich S, et al.. Multi-Omics Factor Analysis-a framework for unsupervised integration of multi-omics data sets. Mol Syst Biol (2018). paper; opened 2026-11-18 all records citing this
  5. Liu C, Ding S, Kim HJ, Long S, et al.. Multitask benchmarking of single-cell multimodal omics integration methods. Nat Methods (2025). paper; opened 2026-11-18 all records citing this
  6. HGNC REST records for IL2RA, CCR7, MS4A1, CD28, CRYM, KRAS and NOTCH3. HGNC. database, not peer reviewed; opened 2026-11-18 all records citing this
  7. Multimodal single cell data integration challenge: Results and lessons learned. Proceedings of Machine Learning Research 176 (2022). paper; opened 2026-11-18 all records citing this
  8. Makrodimitris S, Pronk B, Abdelaal T, Reinders M. An in-depth comparison of linear and non-linear joint embedding methods for bulk and single-cell multi-omics. Brief Bioinform (2023). paper; opened 2026-11-18 all records citing this
  9. Hu Y, Wan S, Luo Y, Li Y, et al.. Benchmarking algorithms for single-cell multi-omics prediction and integration. Nat Methods (2024). paper; opened 2026-11-18 all records citing this
Where it sits on the map
History of this record (1)

This record has been changed 1 time, on the dates below. The note kept against each change is written for the people who maintain the atlas, in their own shorthand, so it is not shown here; it is in the record itself.

Status
partly written
Last updated
2026-11-18
Record
content/concepts/statistical-integration-methods-and-benchmarks.json