The Deep Generative Decoder: MAP estimation of representations improves modelling of single-cell RNA data.

Area of Science:

• Computational biology
• Machine learning
• Genomics

Background:

• Single-cell transcriptomics analysis relies on effective low-dimensional representations.
• Current state-of-the-art methods often use variational autoencoders (VAEs) with variational approximations.

Purpose of the Study:

• To present a novel, simpler generative model for learning representations.
• To demonstrate the model's capability in handling complex latent distributions and achieving dimensionality reduction.

Main Methods:

• Developed the Deep Generative Decoder (DGD) model.
• Employed maximum a posteriori estimation for direct computation of model parameters and representations.
• Applied DGD to the Fashion-MNIST benchmark and multiple single-cell datasets.

Main Results:

• DGD successfully learned low-dimensional, meaningful, and structured latent representations.
• The model demonstrated sub-clustering capabilities beyond initial labels.
• DGD achieved significantly smaller representation dimensionality compared to VAEs.
• Showcased flexibility in handling complex parameterized latent distributions.

Conclusions:

• DGD provides a simpler and more effective alternative for dimensionality reduction in single-cell transcriptomics.
• The model's ability to capture complex data structures offers advantages over traditional VAEs.