iGTP: learning interpretable cellular embedding for inferring biological mechanisms underlying single-cell...

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Deep learning models like Variational Autoencoders (VAEs) facilitate low-dimensional cellular embedding for single-cell transcriptomes.
• However, VAEs often lack biologically meaningful latent spaces without specific structural design.
• Interpretable models are crucial for understanding complex biological systems.

Purpose of the Study:

• To develop a novel interpretable generative transcriptional program (iGTP) framework.
• To model the importance of transcriptional program (TP) space and protein-protein interactions (PPI) within biological states.
• To enhance the interpretability and biological relevance of latent representations in single-cell analysis.

Main Methods:

• Engineered the iGTP framework integrating TP space and PPI networks.
• Validated iGTP using gene ontology, canonical pathways, and PPI databases.
• Integrated the latent layer with a graph neural network (GNN) for perturbation response inference.
• Applied iGTP embeddings with a latent diffusion model for cell type-specific embedding generation.

Main Results:

• iGTP elucidated ground truth cellular responses and outperformed existing methods in functional enrichment.
• The framework successfully modeled TP and PPI importance across diverse biological contexts.
• iGTP effectively inferred cellular responses to perturbations when combined with GNNs.
• Latent diffusion models generated accurate cell embeddings for specific cell types and states using iGTP.

Conclusions:

• iGTP provides a powerful framework for interpretable analysis of single-cell transcriptomes.
• The model offers insights at both PPI and TP levels, advancing biological understanding.
• iGTP shows promise for predicting responses to novel perturbations and generating cell-specific embeddings.