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

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Variational Autoencoders (VAEs) offer low-dimensional representations for single-cell transcriptomes but often lack biological interpretability.
• Existing deep learning models struggle to integrate biological structures like transcriptional programs (TPs) and protein-protein interactions (PPIs) into latent spaces.

Purpose of the Study:

• To develop a novel interpretable generative transcriptional program (iGTP) framework.
• To model the importance of TP space and PPIs across biological states.
• To improve the biological interpretability and downstream task performance of single-cell data analysis.

Main Methods:

• Engineered the iGTP framework integrating TP and PPI information.
• Utilized gene ontology, canonical pathways, and PPI databases for biological context.
• Integrated latent layer with graph neural networks for perturbation response inference.
• Applied iGTP embeddings with latent diffusion models for cell state generation.

Main Results:

• iGTP elucidated ground truth cellular responses and outperformed existing methods in functional enrichment.
• The framework successfully inferred cellular responses to perturbations.
• Generated accurate cell embeddings for specific cell types and states using iGTP TP embeddings and latent diffusion models.

Conclusions:

• iGTP provides insights at both PPI and TP levels for single-cell transcriptomics.
• The framework demonstrates significant potential for predicting responses to novel perturbations.
• iGTP enhances the interpretability and utility of deep learning models in biological research.