MultiPert: An adversarial alignment and dual attention framework for single-cell multi-omics perturbation prediction.

Area of Science:

• Computational biology
• Systems biology
• Multi-omics data analysis

Background:

• Perturbation response prediction is crucial for understanding cellular mechanisms and identities.
• Current methods are limited to single-cell transcriptomic data, failing to capture cross-layer molecular effects.

Purpose of the Study:

• To introduce MultiPert, a deep learning framework for predicting perturbation responses in single-cell multi-omics data.
• To enable prediction of both gene expression and protein abundance profiles after perturbations.

Main Methods:

• MultiPert utilizes modality-specific encoders with pretraining.
• A dual-attention mechanism integrates perturbation information.
• Adversarial training facilitates cross-modal alignment.

Main Results:

• MultiPert accurately and stably predicts gene expression and protein abundance profiles on human THP-1 and kidney datasets.
• The framework outperforms existing state-of-the-art methods.
• MultiPert generalizes to novel perturbations and reveals regulatory mechanisms of immune checkpoint molecules.

Conclusions:

• MultiPert offers an integrated and interpretable approach for multi-omics perturbation modeling.
• The framework provides a foundation for pathogenesis research and drug discovery.
• It expands the scope of perturbation analysis beyond single-modality data.