nPCA: a linear dimensionality reduction method using a multilayer perceptron.

Area of Science:

• Data Science
• Machine Learning
• Bioinformatics

Background:

• Dimensionality reduction techniques are crucial for noise elimination and feature extraction in data analysis.
• Existing methods like Principal Component Analysis (PCA), t-distributed Stochastic Neighbor Embedding (t-SNE), and Autoencoders (AE) have limitations.
• PCA focuses solely on maximum variance, t-SNE is often limited to visualization, and AE/nonlinear methods may discard linear projections.

Purpose of the Study:

• To introduce Neural Principal Component Analysis (nPCA), an unsupervised deep learning approach.
• To develop a method that retains linear projections while improving upon PCA.
• To provide a dimensionality reduction technique suitable for both visualization and downstream analysis.

Main Methods:

• Developed nPCA, an unsupervised deep learning algorithm.
• Evaluated nPCA's performance on 10 public datasets.
• Benchmarked nPCA against classic linear dimensionality reduction methods using 6 pancreatic single-cell RNA sequencing (scRNA-seq) datasets.

Main Results:

• nPCA demonstrated the ability to retain richer information from raw data compared to traditional PCA.
• The method showed competitive performance across diverse public and scRNA-seq datasets.
• nPCA proved effective for both visualization and downstream analytical tasks.

Conclusions:

• nPCA is a promising advancement over PCA for dimensionality reduction.
• The method offers a robust and competitive alternative for various data analysis needs.
• nPCA effectively balances information retention and dimensionality reduction.