Fast Optimization of Robust Transcriptomics Embeddings using Probabilistic Inference Autoencoder Networks for...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell genomics technologies generate large datasets with multiple covariates.
• Dataset integration is crucial but challenging for complex, large-scale single-cell data.
• Existing integration methods struggle with scalability and accuracy for complex datasets.

Purpose of the Study:

• To develop a high-performance computational framework for integrating large and complex single-cell genomics datasets.
• To address limitations of existing methods in handling multi-species, multi-platform, and developmental timepoint data.
• To provide recommendations for effective dataset integration applications.

Main Methods:

• Developed PIANO, a variational autoencoder framework utilizing a negative binomial generalized linear model for enhanced batch correction.
• Implemented code compilation for significantly faster training times compared to existing tools.
• Evaluated PIANO's performance against commonly used integration methods on single-species datasets.

Main Results:

• PIANO demonstrated superior performance in integrating single-species datasets compared to existing methods.
• PIANO successfully integrated multiple atlases across sequencing platforms, developmental timepoints, and species.
• The framework preserved desired biological signals while overcoming challenging integration tasks.

Conclusions:

• PIANO offers a novel, high-performance solution for integrating large, complex single-cell genomics datasets.
• The framework enables superior multi-atlas analyses, addressing scalability and accuracy challenges.
• PIANO provides valuable recommendations for the application of data integration in single-cell genomics research.