PRESCIENT

PRESCIENT simulates cellular differentiation trajectories from time-series single-cell RNA sequencing (scRNA-seq) data using stochastic, physical-time models to predict progenitor cell fate biases and the effects of genetic or perturbational interventions.

Key Features:

• Generative Modeling Framework: Employs a generative neural network to learn an underlying differentiation landscape from time-series scRNA-seq data.
• Stochastic ODE-Based Diffusion Process: Frames differentiation as a diffusion process governed by a stochastic ordinary differential equation with a drift parameter derived from the generative model.
• Robust Fate Prediction and Perturbational Analysis: Predicts progenitor cell fate biases and accounts for cell proliferation dynamics, validated on experimental lineage tracing datasets.
• Simulation of Perturbed Trajectories: Simulates differentiation outcomes under genetic or multi-gene perturbations at varying time points and starting conditions, recovering expected modulators of fate in systems such as hematopoiesis and pancreatic β-cell differentiation.
• Out-of-Sample Prediction Capability: Simulates trajectories for out-of-sample cells not present during training to extend predictive applicability.

Scientific Applications:

• Hematopoiesis and Pancreatic β-Cell Differentiation: Predicts progenitor fate biases to provide insights into hematopoietic and pancreatic β-cell differentiation processes.
• Intervention Analysis: Simulates effects of genetic perturbations to study gene function and potential therapeutic interventions in differentiation.

Methodology:

Integrates time-series scRNA-seq data with a generative neural network to construct a potential energy landscape and models cell differentiation as a stochastic ODE-based diffusion process with a learned drift parameter, incorporating cell proliferation dynamics and physical-time simulations of perturbations.