Learning genetic perturbation effects with variational causal inference

Area of Science:

• Genomics
• Computational Biology
• Systems Biology

Background:

• Perturb-seq enables high-resolution single-cell transcriptomic profiling of genetic perturbations.
• Current computational models struggle with generalizing to unseen perturbations due to overfitting or overly simplistic assumptions.
• Accurate prediction of gene regulatory network (GRN) responses is crucial for functional genomics and identifying therapeutic targets.

Purpose of the Study:

• To develop a novel computational model that accurately predicts transcriptomic responses to genetic perturbations, particularly for unseen perturbations.
• To integrate mechanistic causal inference with deep learning for enhanced extrapolation capabilities in single-cell data analysis.
• To provide a robust tool for interpreting and simulating gene perturbation effects at the single-cell level.

Main Methods:

• Proposed a hybrid approach, Single Cell Causal Variational Autoencoder (SCCVAE), combining a mechanistic causal model with variational deep learning.
• The mechanistic component models perturbations as shift interventions propagating through a learned regulatory network.
• Integrated the mechanistic model within a variational autoencoder framework to generate comprehensive transcriptomic responses.

Main Results:

• SCCVAE demonstrated superior performance in extrapolating and predicting responses to unseen genetic perturbations compared to state-of-the-art methods.
• The model's latent space facilitated the identification of functional perturbation modules for observed perturbations.
• SCCVAE enabled the simulation of single-gene knockdown experiments with varying penetrance.

Conclusions:

• SCCVAE offers a robust and accurate method for predicting gene regulatory network dynamics following perturbations.
• The hybrid approach overcomes limitations of purely deep learning or mechanistic models, enhancing predictive power for novel perturbations.
• This tool advances the interpretation of single-cell transcriptomic data and aids in the discovery of therapeutic strategies.