scWeave: A deep learning model that bidirectionally translates between gene expression and chromatin structure at...

Area of Science:

• Single-cell genomics
• Computational biology
• Epigenetics

Background:

• Gene expression and chromatin structure are linked but challenging to study together at single-cell resolution.
• Existing single-cell co-assay methods are costly and technically demanding.

Purpose of the Study:

• To develop a computational model for bidirectional translation between single-cell gene expression and 3D chromatin architecture.
• To enable cross-modality inference from single-modality measurements.

Main Methods:

• Developed scWeave, a model using dual autoencoders and translation modules for cross-modality inference.
• Evaluated scWeave on six public co-assay datasets from mouse and human samples.

Main Results:

• scWeave significantly improved prediction accuracy for gene expression from chromatin structure and vice-versa compared to baseline methods.
• The model learned aligned latent representations, enabling matching of cells across modalities.
• scWeave demonstrated generalization to new developmental timepoints and species.

Conclusions:

• scWeave effectively bridges the gap between gene expression and chromatin structure data at single-cell resolution.
• The model reduces the need for costly co-assays, expanding the scope of single-cell multi-omic analyses.
• scWeave provides a powerful tool for inferring unmeasured biological states from single-modality data.