Learning Genetic Perturbation Effects with Variational Causal Inference.

Area of Science:

• Genomics
• Computational Biology
• Systems Biology

Background:

• Advances in sequencing, like Perturb-seq, enable single-cell transcriptomic profiling of genetic perturbations.
• Existing computational models struggle with generalization: deep learning overfits, while mechanistic models are too simplistic for large-scale data.

Purpose of the Study:

• To develop a hybrid computational model integrating mechanistic causal inference and deep learning for robust prediction of cellular responses to genetic perturbations.
• To improve extrapolation capabilities for unseen perturbations and enhance interpretability of single-cell transcriptomic data.

Main Methods:

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

Main Results:

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

Conclusions:

• SCCVAE offers a robust tool for interpreting and interpolating single-cell perturbational responses.
• The hybrid approach effectively balances mechanistic understanding with deep learning's capacity for complex data.
• This method advances functional genomics and aids in identifying potential therapeutic targets.