Skip to content

How scMultiBench works

scMultiBench is a systematic benchmark of single-cell multimodal omics integration methods, not a method itself. It categorises and evaluates 40 integration methods so you can pick the right one for your data.

scMultiBench scenarios, tasks, and methods

(a–b) Multimodal single-cell technologies produce distinct data combinations, organised into four integration scenarios — vertical, diagonal, mosaic, and cross — by how the input data are structured. (c–d) Each scenario feeds seven downstream analytical tasks, each scored by its own evaluation-metric panel. (e) The 40+ integration methods and the scenarios, tasks, and modalities each one covers.

What scMultiBench covers

The study organises integration into four scenarios, defined by how the input data are structured:

  • Vertical — one dataset with fully matched modalities (e.g. CITE-seq, 10x Multiome: RNA + ADT/ATAC from the same cells).
  • Diagonal — multiple datasets with non-overlapping modalities (e.g. scRNA-seq from one experiment, scATAC-seq from another - no shared cells).
  • Mosaic — multiple datasets with partially overlapping modalities, often joined by a bridge (e.g. an RNA-only batch, an ATAC-only batch, and one paired batch).
  • Cross — multiple datasets with matched modalities across batches (e.g. the same CITE-seq assay from several donors or sites).

Across these scenarios it benchmarks the 40 methods on seven downstream tasks: dimension reduction, batch correction, clustering, classification, feature selection, imputation, and spatial registration. The published benchmark spans 64 real and 22 simulated datasets (86 total) covering RNA, ADT (CITE-seq), ATAC, and spatial modalities.

How it evaluates

Each task gets its own metric panel: scIB-style batch-correction and biological-conservation metrics (kBET, iLISI, ASW, graph connectivity, PCR, ARI, NMI, and more), classification accuracy and F1, marker overlap and correlation for feature selection, RMSE for imputation, and label-transfer and alignment scores for spatial registration, plus runtime and peak memory. Results are aggregated into per-method grand rank scores and distilled into a decision tree that recommends top methods per scenario and task (Fig. 6 of the paper; the interactive explorer carries the same rankings).

A key takeaway: no single method wins everywhere. Performance depends on the dataset, task, and metric, with a recurring trade-off between preserving biological signal and correcting batch effects.

When to use scMultiBench

Use scMultiBench when you want to choose or compare integration methods for your own data:

  • Run any of 40+ methods through one typed API (multibench), each in its own environment.
  • Evaluate with scIB and task-specific metrics.
  • Explore rankings, bubble tables, and the interactive Shiny explorer.

Get going