Transcriptomics 101 (cancer)
Under scientific review
This introductory page is under scientific review. Use it for orientation, not clinical decision-making.
Transcriptomics measures RNA abundance and isoform usage. In tumors it reveals pathway activity, immune infiltration, and drug-response signatures.
The Data Problem (The Count Matrix)
The fundamental unit of computational transcriptomics is the Count Matrix. After sequencing RNA and aligning reads to the genome, we create a table where rows are genes (e.g., ~20,000 human genes) and columns are patient samples. The number in the cell is how many times a "read" fell into that gene.
The normalization challenge: You cannot simply compare raw matrices.
- If Sample A had 10 million total reads and Sample B had 20 million, all genes in Sample B will appear more "expressed" (sequencing depth bias).
- If Gene X is 10,000 letters (base pairs) long and Gene Y is 1,000, Gene X will naturally capture 10 times more reads (gene length bias).
Normalization Units
Technologists will encounter three main units in public databases (like TCGA):
- Raw Counts: Strictly used as input for differential analysis software (e.g., DESeq2).
- FPKM / RPKM (Reads Per Kilobase Million): Normalizes by gene length and library depth. It is a legacy metric.
- TPM (Transcripts Per Million): The modern standard. If you sum all TPM values of a sample, the result is always 1,000,000. It makes comparing a gene's proportion across different samples much easier.
Differential Expression Analysis (DEA)
The classic clinical question: "Which genes have altered expression between Tumors that responded to the drug vs Resistant tumors?"
This is a massive statistical problem: testing 20,000 genes simultaneously on just 50 or 100 samples (the "Curse of Dimensionality"). Industry-standard software, primarily written in the R language, such as DESeq2 and edgeR, model the data using a Negative Binomial distribution to handle the extreme variance common in biology data. They output:
- Log2 Fold Change (Log2FC): The magnitude of the change. A log2FC of 2 means the gene is expressed 4 times more in the tumor.
- FDR / p-adj (False Discovery Rate): A rigorously adjusted p-value (usually by the Benjamini-Hochberg method) to punish false positives generated by doing 20,000 statistical tests at once.
Dimensionality Reduction (GSEA)
Instead of focusing on a list of 500 individual altered genes (which is biologically confusing), data scientists use GSEA (Gene Set Enrichment Analysis).
GSEA compares your results against databases of "Biological Pathways" (like KEGG or Hallmark Pathways). Instead of saying "Gene A, B, and C are high", GSEA concludes mathematically: "The entire Glycolysis pathway (Warburg Effect) is enriched in this sample".
Single-cell and spatial
For heterogeneity and microanatomy, see Intro to single-cell and spatial.
Clinical angles
- Fusion transcripts (some assays are DNA; some are RNA-first)
- Immune repertoire and checkpoint-related expression patterns
- Companion diagnostics sometimes use expression (see Biomarkers)