Research map / Vision-language and multimodal / Cross-modal learning

scMRDR: A scalable and flexible framework for unpaired single-cell multi-omics data integration

NeurIPS 2025 · Vision-language and multimodal › Cross-modal learning · approach: Medical multimodal fusion · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes scMRDR, a single-encoder beta-VAE that disentangles shared and modality-specific latents with isometric, adversarial, and masked-loss regularization to integrate fully unpaired multi-omics single-cell data at scale.

Problem

Integrating unpaired multi-omics single-cell data (e.g., scRNA, scATAC, CITE-seq protein) into a shared latent space is hard: existing methods need paired data, prior correspondences, or global coupling matrices that do not scale and usually handle only two modalities.

Approach

A unified beta-VAE with one encoder-decoder treats all omics cells as a single sample, splitting latents into shared and modality-specific parts; isometric distance regularization preserves intra-omics structure in the shared part, a modality discriminator aligns it adversarially, and a feature-masked ZINB reconstruction loss handles features missing in some modalities.

Key results

• On human kidney two-omics integration, scMRDR achieves overall score 0.66 versus 0.64 for GLUE and 0.56 for scVI and MaxFuse.
• On large-scale mouse motor cortex (69,727 scRNA + 54,844 scATAC cells), scMRDR overall score 0.60 vs 0.59 for Seurat and 0.55 for GLUE (LSI); JAMIE, UnionCom and Pamona failed to run at this scale.
• On triple-omics bone marrow, scMRDR overall score 0.58 vs 0.55 for GLUE, with the masked loss ablation dropping overall score to 0.44.

Measured effects reported in the paper

Surprising findings

• Ablation on the kidney data shows the baseline beta-VAE without regularization (0.59) is the worst configuration, but the gap from scMRDR is modest and the authors note hyperparameter choice has limited influence.
• GLUE's performance depends strongly on preprocessing choice (LSI vs PCA) and degrades on large-scale data.

Limitations

• Adversarial min-max objective and multiple loss weights (beta, lambda, gamma) increase training difficulty and tuning burden.
• Features are mapped to gene level (e.g., scATAC peaks aggregated to gene activity scores), which may lose information.
• Evaluated only on a few datasets; hyperparameter sensitivity results are from a single dataset.

Experimental setup

Datasets: human kidney scRNA-seq and scATAC-seq mouse primary motor cortex scRNA-seq and scATAC-seq human bone marrow scRNA, scATAC, CITE-seq merFISH spatial transcriptomics (mouse primary motor cortex)

Benchmarks: scIB metrics on human kidney two-omics scIB metrics on mouse primary motor cortex large-scale two-omics scIB metrics on human bone marrow triple-omics

Metrics: F1 isolated label NMI ARI cell-type ASW cLISI batch ASW iLISI kBET PCR R2 graph connectivity

Baselines: Seurat v5 Harmony scVI scGLUE JAMIE UnionCom Pamona MaxFuse SIMBA SCOT

Builds on: beta-VAE scVI GLUE UnionCom

Closest papers by meaning

• Semi-supervised Knowledge Transfer Across Multi-omic Single-cell Data (NeurIPS 2024 · similarity 0.86)
• scPilot: Large Language Model Reasoning Toward Automated Single-Cell Analysis and Discovery (NeurIPS 2025 · similarity 0.84)
• sciLaMA: A Single-Cell Representation Learning Framework to Leverage Prior Knowledge from Large Language Models (ICML 2025 · similarity 0.84)
• scMBERT: A Pre-Trained Deep Learning Model for Single-Cell Multiomic Data Representation and Prediction (Student Abstract) (AAAI 2025 · similarity 0.84)
• CellVerse: Do Large Language Models Really Understand Cell Biology? (NeurIPS 2025 · similarity 0.83)
• A versatile informative diffusion model for single-cell ATAC-seq data generation and analysis (NeurIPS 2024 · similarity 0.82)