scRAE: Deterministic Regularized Autoencoders With Flexible Priors for Clustering Single-Cell Gene Expression Data

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Single-cell RNA sequencing (scRNA-seq) data presents high-dimensionality and sparsity challenges, including 'dropout' events, complicating accurate clustering.
• Regularized Auto-Encoders (RAEs) offer a deep learning approach for learning low-dimensional representations but face bias-variance trade-offs in naive formulations, leading to overfitting or under-representation.

Purpose of the Study:

• To address the limitations of existing RAEs in scRNA-seq data clustering.
• To propose a modified RAE framework, scRAE, for improved bias-variance management in latent space representation.
• To enhance the accuracy and robustness of scRNA-seq data clustering.

Main Methods:

• Developed a modified Regularized Auto-Encoder framework named scRAE.
• Incorporated a deterministic Auto-Encoder (AE) with a flexibly learnable prior generator network.
• Jointly trained the AE and the prior generator network to optimize latent space representation.

Main Results:

• scRAE demonstrates improved ability to balance bias and variance in the latent space compared to standard RAEs.
• Extensive experiments on real-world scRNA-seq datasets show the efficacy of scRAE for effective clustering.
• The proposed method leads to more robust and accurate identification of cell populations.

Conclusions:

• The scRAE framework offers a superior approach for clustering scRNA-seq data by effectively addressing the bias-variance trade-off.
• This method enhances the learning of meaningful low-dimensional representations from sparse, high-dimensional single-cell gene expression data.
• scRAE provides a valuable tool for computational biologists and bioinformaticians working with scRNA-seq data.