A new paper co-authored by IICD researchers from the labs of Elham Azizi, Dana Pe’er, and David Blei introduces Decipher, a deep generative model that jointly characterizes and visualizes cell-state transitions from healthy to disease conditions. The study, published in Genome Biology, was led by Achille Nazaret, Joy Linyue Fan, and Vincent-Philippe Lavallée, alongside a collaborative team of scientists.
Decipher addresses a critical gap in analyzing single-cell RNA sequencing data by enabling the integration and interpretation of cellular trajectories across normal and perturbed states. The model’s unique architecture allows researchers to align disease and control samples while preserving both global geometry and gene-level trends, offering new insights into complex biological processes such as cancer progression and immune cell dysfunction.
"Decipher introduces a novel hierarchical deep generative architecture that can compare and simultaneously visualize diverging cell trajectories, e.g. from normal and perturbed single-cell datasets. It builds a unified, interpretable latent space that preserves global topology, and pinpoints the precise timing of transcriptional events along these trajectories. This powerful approach unlocks new avenues for discovering early disease drivers, developing early-detection biomarkers e.g. for cancer detection, and identifying coordinated disruptions in transcriptional regulation." — Elham Azizi, PhD
The paper demonstrates Decipher's capabilities across diverse disease contexts, including pancreatic cancer, acute myeloid leukemia, and gastric cancer. In each case, the model not only reconstructed accurate pseudotime trajectories but also revealed disrupted regulatory programs and transcription factor activity involved in disease initiation.
This work marks a significant methodological advancement in computational biology, with the potential to impact how we understand and target early disease states.
Read the full study:
Genome Biology: Joint representation and visualization of derailed cell states with Decipher
For potential users, explore the model and code here:
Azizi Lab Software Page
