scDisent: regulatory-aware disentangled representation learning for multi-omic single-cell analysis

In plain English

The model tries to keep what a cell expresses separate from the regulatory program that may be driving it.

What they found

• The authors report the strongest clustering performance among the compared methods.
• The learned mapping remained sparse and recovered lineage-associated programs in perturbation analyses.

The catch

• The record is a preprint and has not completed peer review.
• This summary is based on the abstract; methods and supplementary analyses were not independently rechecked.

Methods at a glance

Study type

Generative representation-learning study

Biological system

Matched single-cell multi-omic measurements

Organism

Benchmark-dependent; not specified in the abstract

Dataset

Matched-modality single-cell benchmark datasets

Sample size

Not reported

Model architecture

Modality-specific encoders, variational disentanglement, sparse directed mapping, and Gumbel-gated causal module

Training objective

Separate expression-associated and regulation-associated latent variables

Code

The abstract links https://github.com/xiguoren/scDisent

Data

Benchmark datasets are described but not enumerated in the abstract.

Tasks and metrics

Tasks: Clustering; Regulatory interpretation; Perturbation analysis

Metrics: Clustering performance; Latent sparsity; Branch separation

Baselines: Competing single-cell integration methods

Principal findings

• The authors report improved clustering and sparse regulator-centered structure.

What remains unresolved

External validation: Not reported

• The record is a preprint and has not completed peer review.
• This summary is based on the abstract; methods and supplementary analyses were not independently rechecked.