CONCERT predicts niche-aware perturbation responses in spatial transcriptomics.

Area of Science:

• Computational Biology
• Genomics
• Systems Biology

Background:

• Spatial perturbation transcriptomics reveals gene expression changes within tissue context.
• Understanding how cellular microenvironments influence perturbation outcomes is crucial.

Purpose of the Study:

• To develop CONCERT, a niche-aware generative model for predicting spatial gene expression changes after perturbations.
• To formalize and evaluate prediction tasks including patch, border, and niche effects.

Main Methods:

• CONCERT utilizes a Gaussian process variational autoencoder to learn spatial kernels and embed perturbation context.
• The model predicts gene expression responses in unperturbed regions, tissue interfaces, and based on microenvironments.
• Evaluation performed on Perturb-map lung datasets against state-of-the-art models.

Main Results:

• CONCERT significantly outperforms existing methods, reducing E-distance by up to 33.77% and improving Pearson correlation by up to 9.10%.
• Case studies demonstrate CONCERT's ability to reconstruct spatial gene expression, perform longitudinal comparisons, and resolve heterogeneity in colitis and stroke models.
• The model accurately predicts responses under variable conditions and in 3D formulations.

Conclusions:

• CONCERT advances spatial perturbation transcriptomics by enabling niche-aware counterfactual predictions and reconstructing missing spatial data.
• The model offers a powerful tool for understanding complex perturbation responses across diverse biological tissues.