Predicting cellular responses to complex perturbations in high-throughput screens.

Area of Science:

• Computational biology
• Genomics
• Pharmacology

Background:

• Single-cell transcriptomics enables high-throughput study of perturbations.
• Exploring all combinatorial perturbations is experimentally infeasible.
• Computational methods are crucial for predicting and prioritizing perturbations.

Purpose of the Study:

• To develop a computational method for predicting single-cell responses to drug and genetic perturbations.
• To create a model that combines interpretability with deep learning for perturbation analysis.
• To enable in silico prediction for unseen dosages, cell types, time points, species, and combinations.

Main Methods:

• Developed the compositional perturbation autoencoder (CPA), a deep learning model.
• CPA integrates linear model interpretability with deep learning flexibility.
• Incorporated chemical representations of drugs for predicting responses to novel compounds.

Main Results:

• CPA accurately predicts single-cell transcriptional responses to unseen drug combinations, outperforming baseline models.
• The model successfully predicted responses to completely unseen drugs by incorporating chemical structures.
• CPA imputed 97.6% of missing genetic combinations in a single-cell Perturb-seq experiment.

Conclusions:

• CPA facilitates efficient experimental design and hypothesis generation through in silico prediction.
• The model accelerates therapeutic applications by enabling accurate single-cell level response prediction.
• CPA offers a powerful tool for analyzing complex perturbation landscapes in biological systems.