Gene-Expression Imputation for Single-Cell Data

Single-cell RNA sequencing captures only a small part of the transcripts of each cell, so 94-98 % of the cells x genes matrix is zero, and many of these zeros are "dropouts" of genes that are really expressed. This project reconstructs the missing values of human blood cells (PBMC) with denoising autoencoders - an MSE network and two count autoencoders with negative-binomial (NB) and zero-inflated NB (ZINB) likelihoods - and compares them honestly with no imputation, the gene mean, kNN smoothing, MAGIC-style diffusion and ALRA. Two corruption tests give a known truth: hidden non-zero entries, and deep cells down-sampled to 10 or 25 % of their reads, whose unused reads are the truth. Free 10x Genomics data (PBMC 10k v3, PBMC 1k v3, FACS-sorted PBMC populations). A complete stand-alone MATLAB R2026a version is included.

Category: Bioinformatics - Transcriptomics and Gene-Expression Analysis (Deep Learning, Denoising Autoencoders, Missing-Value Imputation, Single-Cell RNA-Seq, MATLAB, Python)

Transcriptomics & Gene-Expression Analysis Project B15 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.