scTranslation: Single-Cell Modality Translation

• scTranslation is a benchmark framework for single-cell multi-omics modality translation, defining a clear mapping between paired modality data such as RNA, ATAC, and protein.
• It employs a uniform evaluation protocol using 5-fold cross-validation and metrics like clustering consistency, regression accuracy, and distribution preservation.
• The framework integrates six state-of-the-art models from AE, VAE, and distribution-based families to expose trade-offs between correlation, numerical stability, and global distribution fidelity.

scTranslation is a benchmark framework for single-cell multi-omics modality translation: given a single-cell data matrix $X\in\mathbb R^{S\times M}$ for modality A and a paired matrix $Y\in\mathbb R^{S\times N}$ for modality B, it studies the problem of learning a mapping $f:\mathbb R^M\rightarrow\mathbb R^N$ such that $\hat Y=f(X)\approx Y$. The benchmark supports bidirectional prediction across combinations of RNA, ATAC, and protein modalities, assembles diverse paired datasets, integrates six representative translation methods, and evaluates them under a shared protocol using clustering, regression, and distribution-preservation metrics (Cheng et al., 2 Jun 2026). It is motivated by the fact that simultaneous measurement of multiple omics modalities in single cells is costly, sparse, and technically challenging, while computational translation can impute missing modalities, reduce sequencing cost, denoise sparse measurements, and reveal regulatory relationships.

1. Scope and formal problem

In scTranslation, modality translation is defined on paired single-cell measurements. For a source modality A and target modality B, the goal is to infer modality B from modality A at the single-cell level. The benchmark explicitly supports bidirectional translation, so the same framework is used for RNA$\leftrightarrow$ATAC and RNA$\leftrightarrow$protein tasks (Cheng et al., 2 Jun 2026).

The benchmark is framed by both biological and computational considerations. Multi-omics experiments such as SNARE-seq and CITE-seq provide chromatin, transcriptome, and protein measurements in the same cell, but incomplete modality coverage and measurement noise create a strong incentive for predictive reconstruction. A plausible implication is that scTranslation is intended not merely as a leaderboard, but as a controlled environment for understanding where current translation models succeed or fail under realistic perturbations such as feature restriction, missingness, and data scarcity.

scTranslation also standardizes the comparison space. It includes diverse translation datasets, integrates state-of-the-art models, and provides a comprehensive evaluation metrics; in addition, it assesses model performance under feature selection, feature quality, and few-shot settings, which the benchmark identifies as factors that significantly affect model performance but have rarely been systematically studied before (Cheng et al., 2 Jun 2026).

2. Datasets and preprocessing pipeline

scTranslation assembles eight publicly available single-cell multi-omics datasets spanning multiple species, organs, developmental stages, and scales, described as the “6M” criteria (Cheng et al., 2 Jun 2026).

Dataset	Platform and context	Dimensions
GSE126074_AdBrainCortex	SNARE-seq, adult mouse cortex	$S=10{,}309$; RNA $M=31{,}048$; ATAC $N=244{,}544$
GSE126074_P0	SNARE-seq, neonatal mouse cortex	$S=5{,}081$; RNA $Y\in\mathbb R^{S\times N}$0; ATAC $Y\in\mathbb R^{S\times N}$1
GSM3271041_sciCAR	sci-CAR, human embryonic kidney HEK293T	$Y\in\mathbb R^{S\times N}$2; RNA $Y\in\mathbb R^{S\times N}$3; ATAC $Y\in\mathbb R^{S\times N}$4
Brain	10x Multiome, human fetal cortex	$Y\in\mathbb R^{S\times N}$5; RNA $Y\in\mathbb R^{S\times N}$6; ATAC $Y\in\mathbb R^{S\times N}$7
CL	scCAT-seq, human embryo	$Y\in\mathbb R^{S\times N}$8; RNA $Y\in\mathbb R^{S\times N}$9; ATAC $f:\mathbb R^M\rightarrow\mathbb R^N$0
PBMC	10x Multiome, human peripheral blood	$f:\mathbb R^M\rightarrow\mathbb R^N$1; RNA $f:\mathbb R^M\rightarrow\mathbb R^N$2; ATAC $f:\mathbb R^M\rightarrow\mathbb R^N$3
CITE_BMMC	CITE-seq, human bone marrow	$f:\mathbb R^M\rightarrow\mathbb R^N$4; RNA $f:\mathbb R^M\rightarrow\mathbb R^N$5; Protein $f:\mathbb R^M\rightarrow\mathbb R^N$6
CITE_PBMC	CITE-seq, human PBMC	$f:\mathbb R^M\rightarrow\mathbb R^N$7; RNA $f:\mathbb R^M\rightarrow\mathbb R^N$8; Protein $f:\mathbb R^M\rightarrow\mathbb R^N$9

A uniform preprocessing pipeline is applied to all tasks. For RNA, raw counts are log-normalized and the top 3,000 highly variable genes are selected: $\hat Y=f(X)\approx Y$0 For ATAC, peaks present in fewer than $\hat Y=f(X)\approx Y$1 of cells are removed, followed by TF-IDF transformation: $\hat Y=f(X)\approx Y$2 For protein, all antibody channels are retained and log-transformed (Cheng et al., 2 Jun 2026).

This preprocessing choice fixes an important experimental boundary condition. In particular, the use of 3,000 HVGs for RNA is not incidental: later factor analyses show that translation quality depends materially on the number and quality of selected features, so the benchmark’s default preprocessing is part of the experimental design rather than a generic convenience.

3. Integrated model families

scTranslation evaluates six representative translation methods organized into three modeling paradigms: autoencoder-based, variational autoencoder-based, and distribution-based approaches (Cheng et al., 2 Jun 2026).

Model	Paradigm	Core characteristics
BABEL	AE-based	Two encoders map inputs to a shared latent space $\hat Y=f(X)\approx Y$3; two decoders reconstruct both modalities
scPair	AE-based	Modality-specific encoder-decoder pairs aligned by adversarial training with a discriminator $\hat Y=f(X)\approx Y$4
JAMIE	VAE-based	Separate VAEs produce posteriors $\hat Y=f(X)\approx Y$5; supports partial-modality cells; uses ELBO + Shapley-based feature attribution
scButterfly	VAE-based	Dual-aligned VAE with masked pretraining, translation module, adversarial and cycle-consistency losses
multiDGD	Distribution-based	Latent variables are explicit parameters following a Gaussian Mixture Model; decoder is hierarchical and branched
scDiffusion-X	Distribution-based	Multimodal autoencoder + denoising diffusion model in latent space with dual-cross-attention

BABEL uses two encoders $\hat Y=f(X)\approx Y$6 and $\hat Y=f(X)\approx Y$7 and two decoders $\hat Y=f(X)\approx Y$8 and $\hat Y=f(X)\approx Y$9, with cross-modal reconstruction from a shared latent representation. scPair also employs modality-specific encoders and decoders, but aligns latent spaces adversarially by training an encoder to fool a discriminator that attempts to identify the modality from $\leftrightarrow$0. JAMIE and scButterfly represent the VAE-based family: the former emphasizes latent posterior modeling and Shapley-based feature attribution, while the latter adds masked pretraining, translation modules, dual discriminators, and latent-space augmentation. multiDGD models the latent structure explicitly with a Gaussian Mixture Model and includes probabilistic covariates such as batch and tissue. scDiffusion-X combines multimodal autoencoding with latent diffusion and dual-cross-attention for adaptive cross-modal information flow (Cheng et al., 2 Jun 2026).

The benchmark’s model selection spans complementary inductive biases. AE-based methods are comparatively direct in cross-modal regression; VAE-based methods impose a stronger generative prior and latent alignment structure; distribution-based methods target multimodal uncertainty and global distributional fidelity. This suggests that scTranslation was designed to expose trade-offs among correlation, calibration, clustering structure, and distribution preservation rather than to privilege a single optimization criterion.

4. Evaluation protocol and metrics

All translation tasks in scTranslation use stratified 5-fold cross-validation, and hyperparameters match the original papers (Cheng et al., 2 Jun 2026). Performance is measured along three axes: clustering consistency, regression accuracy, and distribution preservation.

For clustering consistency, the benchmark uses Normalized Mutual Information (NMI), Adjusted Rand Index (ARI), Adjusted Mutual Information (AMI), and Homogeneity (HOM). NMI is defined as

$\leftrightarrow$1

where $\leftrightarrow$2 denotes clusters on translated data and $\leftrightarrow$3 denotes ground-truth labels. Homogeneity is

$\leftrightarrow$4

For regression accuracy, scTranslation uses the Pearson Correlation Coefficient (PCC) and Mean Squared Error (MSE). PCC is computed between predicted $\leftrightarrow$5 and true $\leftrightarrow$6: $\leftrightarrow$7 and

$\leftrightarrow$8

For distribution preservation, the benchmark uses Maximum Mean Discrepancy (MMD) and Local Inverse Simpson’s Index (LISI). MMD is defined as

$\leftrightarrow$9

and LISI is used to quantify both modality mixing (iLISI) and preservation of cell types (cLISI) in local neighborhoods (Cheng et al., 2 Jun 2026).

A central implication of this metric design is that modality translation is treated as a multi-objective problem. High correlation with the target modality is not sufficient if translated cells distort cluster structure or drift from the true global distribution. The benchmark therefore formalizes a recurrent methodological tension: pointwise predictive accuracy and population-level biological plausibility do not always improve together.

5. Empirical findings across translation settings

In bidirectional RNA$\leftrightarrow$0ATAC translation, scButterfly often leads clustering metrics such as NMI, ARI, AMI, and HOM, and achieves the lowest MSE on Brain and AdBrainCortex. JAMIE excels on distributional alignment, attaining the lowest MMD and highest LISI on several datasets, especially PBMC. AE-based BABEL and scPair obtain the highest PCC in many settings, but suffer from poorer scale, reflected in higher MSE, and from distribution mismatch. scDiffusion-X sometimes attains top global distribution metrics on small datasets such as CL, but at the cost of clustering fidelity (Cheng et al., 2 Jun 2026).

In RNA$\leftrightarrow$1protein translation on the CITE-seq datasets, asymmetries become more pronounced. Protein$\leftrightarrow$2RNA is numerically unstable for some models; scPair is reported to yield PCC $\leftrightarrow$3 but MSE in the $\leftrightarrow$4 range. VAE methods, particularly JAMIE and scButterfly, are described as more balanced, with moderate MSE and good MMD/LISI, and JAMIE attains the best LISI on BMMC. BABEL cannot natively handle protein inputs. RNA$\leftrightarrow$5protein is more stable: JAMIE leads on distribution metrics, scPair leads on PCC/MSE, and scButterfly drops on LISI (Cheng et al., 2 Jun 2026).

The benchmark also evaluates three influencing factors using PBMC RNA$\leftrightarrow$6ATAC as an example. Under feature selection, varying the number of HVGs from 500 to 4,000 shows that clustering metrics improve up to approximately 3,000 genes and then decline, while regression metrics such as PCC and MSE steadily improve with more features. Under feature quality perturbation, randomly masking 20–80% of entries causes both clustering and regression performance to degrade, while MMD decreases under high missingness because models regress toward the global mean. Under few-shot learning, replacing normal 80/20 splits with 1-fold training causes most methods to drop modestly, but scDiffusion-X improves clustering in the few-shot regime due to diffusion-based implicit augmentation (Cheng et al., 2 Jun 2026).

These results support several benchmark-level conclusions. No single model dominates across all datasets, directions, and metrics. VAE-based models, especially scButterfly and JAMIE, strike the best balance between structure preservation, numerical accuracy, and distributional alignment. AE-based models achieve high correlation but may distort scale or global distribution. Robustness to feature selection, data sparsity, and data scarcity is therefore not a peripheral issue but part of the core problem definition.

6. Interpretation, limitations, and related directions

scTranslation directly addresses several recurrent misconceptions in modality translation. One is that a single metric can summarize model quality. The reported results contradict this: a model can achieve high PCC while remaining numerically unstable or distributionally misaligned. Another is that one modeling family is uniformly superior. The benchmark instead shows task- and metric-dependent specialization, with VAE-based models more balanced, AE-based models often strongest on correlation, and diffusion-based modeling occasionally advantageous in few-shot clustering (Cheng et al., 2 Jun 2026).

The benchmark also foregrounds practical limitations in current methods. Model usability is uneven, as illustrated by BABEL’s inability to natively handle protein inputs. Numerical stability deserves greater attention, particularly in protein$\leftrightarrow$7RNA translation. Future methods are recommended to explicitly model missingness, incorporate feature-selection priors, and support few-shot adaptation (Cheng et al., 2 Jun 2026). A plausible implication is that benchmark progress will depend as much on robustness engineering and interface generality as on improvements in mean predictive accuracy.

The framework is open-sourced at https://github.com/Bunnybeibei/scTranslation. The repository contains data_preprocessing/, models/, train_and_eval/, metrics/, analysis_plots/, and README.md, with scripts to download, filter, and normalize each dataset; implementations and wrappers for BABEL, JAMIE, multiDGD, scButterfly, scPair, and scDiffusion-X; training pipelines with 5-fold cross-validation; metric implementations; and notebooks for reproducing performance grids and factor analyses (Cheng et al., 2 Jun 2026).

A related but distinct line of research is spatial transcriptome reconstruction. For example, Laplacian Linear Optimal Transport integrates single-cell and spatial transcriptomics data to reconstruct missing information at a whole-genome and single-cell resolution by iteratively correcting platform effects and decomposing each spot into a spatially smooth probabilistic mixture of single cells (Zhu et al., 2024). This is adjacent to scTranslation in its reliance on cross-modal inference, but it addresses a different object: spatial reconstruction from scRNA-seq and spatial measurements rather than paired single-cell modality translation among RNA, ATAC, and protein. This distinction is important because it separates within-cell translation benchmarks from cell-to-location alignment frameworks, even when both use multi-omics integration as their computational substrate.