LSH-GAN enables in-silico generation of cells for small sample high dimensional scRNA-seq data

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Single-cell RNA sequencing (scRNA-seq) data analysis faces challenges due to limited sample sizes compared to the high dimensionality of genomic features.
• Budgetary constraints and limited patient cohorts often restrict the number of available scRNA-seq samples.
• This scarcity impedes the effectiveness of standard downstream analytical procedures.

Purpose of the Study:

• To address the critical issue of insufficient cell samples in scRNA-seq data analysis.
• To introduce an enhanced generative adversarial network (GAN) model, LSH-GAN, for generating realistic synthetic cell samples.
• To improve the feasibility and performance of downstream analyses on datasets with limited sample sizes.

Main Methods:

• An improved generative adversarial network (GAN) architecture named LSH-GAN was developed.
• The generator training procedure was enhanced using locality-sensitive hashing (LSH) to accelerate sample generation.
• The model was benchmarked against existing methods for realistic sample generation.

Main Results:

• LSH-GAN demonstrated superior performance in generating high-quality, realistic cell samples compared to existing benchmarks.
• The synthetic samples generated by LSH-GAN significantly improved the performance of downstream analysis tasks.
• Specifically, improvements were observed in feature (gene) selection and cell clustering accuracy.

Conclusions:

• LSH-GAN effectively addresses the challenge of small sample sizes in scRNA-seq data analysis.
• The method enhances the utility of scRNA-seq data by enabling more robust downstream analyses.
• LSH-GAN offers a viable solution for researchers facing limitations in obtaining large scRNA-seq datasets.