GrandPrix: scaling up the Bayesian GPLVM for single-cell data

Area of Science:

• Computational biology
• Single-cell genomics
• Machine learning for biological data analysis

Background:

• Gaussian Process Latent Variable Model (GPLVM) is utilized for dimensionality reduction in single-cell data.
• GPLVM aids in pseudotime estimation using capture time, but current methods are computationally intensive for large datasets.

Purpose of the Study:

• To provide an efficient and scalable GPLVM implementation for modern single-cell datasets.
• To generalize pseudotime inference to include complex biological dynamics such as branching.

Main Methods:

• Developed an efficient implementation of GPLVM for dimensionality reduction and pseudotime inference.
• Applied the model to diverse single-cell datasets including microarray, nCounter, RNA-seq, qPCR, and droplet-based data.
• Extended the model to higher-dimensional latent spaces for simultaneous inference of pseudotime and other structures like branching.

Main Results:

• The new implementation scales to large, droplet-based single-cell datasets.
• The model converges an order of magnitude faster than existing methods while maintaining similar accuracy.
• Demonstrated the ability to infer pseudotime and model branching dynamics, providing insights into cell fate regulation.

Conclusions:

• The efficient GPLVM implementation significantly enhances the scalability of pseudotime inference for single-cell data.
• The generalized model offers a flexible tool for uncovering complex cellular processes, including cell ordering and fate decisions.
• The software is publicly available, facilitating broader adoption in biological research.