Skip to main content
Back

scTranslation Benchmark Finally Reveals Which AI Cell Biology Models Actually Work

scTranslation is the first systematic benchmark for single-cell multi-omics translation, revealing how feature selection, data quality, and few-shot settings dramatically affect model performance

Paper, Full Access
  • New benchmark: scTranslation is the first comprehensive benchmark for single-cell multi-omics modality translation, covering RNA, ATAC, and Protein modalities.
  • Six model families tested: BABEL, scPair, JAMIE, scDiffusion, MultiODD, and scDiffusion-X evaluated under unified conditions.
  • Key finding: No single model dominates all tasks; few-shot and full-data rankings are poorly correlated, misleading practitioners.
  • Feature selection matters most: Preprocessing pipeline choices have a disproportionate impact on translation quality, often more than model architecture.
  • Practical gap exposed: Diffusion models excel with full data but degrade faster under noise and limited paired-cell settings.
  • Open source: Full code, datasets, and evaluation suite released at github.com/Bunnybeibei/scTranslation.

Biology runs on layers. A cell's identity is not just its RNA, but also its chromatin accessibility, its protein abundance, and the regulatory logic connecting all three. Measuring all of these layers simultaneously in a single cell is the gold standard, but it is also expensive, noisy, and technically demanding. That gap has spawned an entire class of computational tools that try to translate one omics modality into another, inferring what you did not measure from what you did. The problem is that no one has rigorously compared these tools against each other, under the same conditions, on the same data, with the same metrics. A new paper, scTranslation, sets out to fix that.

Multi-omics data layers: ATAC-seq, RNA-seq, and Protein-seq mapped to DNA and 3D protein structures

Why translation, and why now?

Advances in single-cell sequencing have enabled a myriad of modalities, ranging from gene expression to proteomics, as well as technologies jointly measuring combinations of these modalities. But multi-modal profiling presents challenges such as increased complexity, noise, and cost. Multi-omics assays like RNA-protein co-profiling (CITE-seq) and chromatin accessibility-RNA assays yield deeper biological insights but remain costly, technically demanding, and low-throughput, making large, high-quality paired datasets extremely rare.

In contrast, single-modality datasets, particularly scRNA-seq, are abundant in both scale and coverage, offering a rich resource for computational modeling and predictive approaches. This asymmetry is the engine behind the translation idea: if you can reliably predict protein levels from RNA, or chromatin accessibility from gene expression, you effectively multiply the value of every existing single-modality dataset. Machine learning can translate between single-cell omics measurements, enabling the inference of unmeasured modalities.

A field without a scoreboard

Single-cell multimodal omics technologies have empowered the profiling of complex biological systems at a previously unattainable resolution and scale. These biotechnologies have propelled fast-paced innovation and development of data integration methods, leading to a critical need for their systematic categorization, evaluation, and benchmarking. Navigating and selecting the most pertinent integration approach poses a considerable challenge, contingent upon the tasks relevant to the study goals and the combination of modalities present in the data at hand.