Count Matrix

Count matrices are predominantly filled with zeros, often exceeding 90%. This arises from a combination of biological zeros (genes genuinely not expressed in a cell type) and technical zeros (dropout events where lowly expressed transcripts fail to be captured during reverse transcription). Distinguishing between these zero types is the core challenge addressed by imputation and statistical models like negative binomial regression.

The matrix is structured with genes as rows and cell barcodes as columns. Each entry M[i, j] stores the integer count of Unique Molecular Identifiers (UMIs) mapping to gene i in cell j. A typical experiment yields a high-dimensional space of 20,000+ genes across 5,000–50,000 cells, requiring dimensionality reduction techniques like PCA before visualization with t-SNE or UMAP.

The matrix captures not only biology but also technical noise. Ambient RNA inflates background counts in empty droplets. Batch effects introduce systematic offsets between separately processed samples. Quality control metrics derived from the matrix—such as mitochondrial read fraction and ribosomal protein gene fraction—are used to filter damaged or stressed cells before biological interpretation begins.

Not all genes are informative. Highly Variable Genes (HVGs) are the subset of rows exhibiting greater variance than expected from technical noise alone. Selecting the top 2,000–5,000 HVGs focuses downstream analysis on the genes driving biological heterogeneity, dramatically reducing computational load and noise from stochastically expressed transcripts.

Storing a full dense matrix of 30,000 genes × 20,000 cells as 64-bit floats would consume ~4.8 GB of memory. To handle this efficiently, count matrices are stored in sparse formats like Compressed Sparse Column (CSC) or Compressed Sparse Row (CSR). These formats store only non-zero values and their indices, reducing memory footprint by over 90% and enabling scalable analysis in tools like Scanpy and Seurat.