Harvard's CONCERT predicts perturbation spread across tissue from a single slide

marinkazitnik · x · 2026-10-25

Marinka Zitnik's Harvard team unveiled CONCERT, a generative model that predicts how perturbation effects propagate across tissue directly on spatial transcriptomics.

First demo: DSS colitis, a standard mouse IBD model where each mouse yields only one time point. From a single day-0 slide, CONCERT predicts the same colon at days 3, 9 and 21 — inflammatory markers like Il1b and Cxcl10 peak at day 9 and decline by day 21, matching measured mice, while distinguishing strong vs. weak responders. This is counterfactual generation along time, producing a longitudinal view from one sample.

Unlike perturbation models built for dissociated cells, CONCERT learns a spatial kernel per perturbation via a Gaussian process variational autoencoder, capturing how far effects extend from perturbed cells and how they depend on the surrounding niche. It supports swapping perturbations, moving perturbed patches into new niches, and works on 2D slides or across 3D sections.

Related event: Harvard's CONCERT Predicts Spatial Spread of Gene Perturbations in Tissue(4 posts)→

Original post →

More from Research

Research channel →