Explain RNA-seq data analysis basics — Biology notes

WHAT is RNA-seq?

WHY do we care? Two cells with the same DNA can be wildly different (a neuron vs a skin cell) because they express different genes. Disease, drug response, and development all show up as changes in expression. RNA-seq is the standard tool to detect those changes.

HOW: the pipeline, step by step

The analysis is a chain. Each step's output is the next step's input.

Figure — Explain RNA-seq data analysis basics

1. Wet-lab → raw reads (FASTQ)

RNA is extracted, reverse-transcribed to cDNA, fragmented, and sequenced. You receive millions of short reads (~50–150 bp) in FASTQ files. Each read carries a base sequence + a per-base quality score.

2. Quality control & trimming

Tools (FastQC) flag low-quality tails and adapter contamination; trimming removes them. WHY? Bad bases cause wrong alignments downstream — garbage in, garbage out.

3. Alignment / mapping

Each read is placed onto a reference genome (aligners: STAR, HISAT2) or matched to a transcriptome (pseudo-aligners: Salmon, kallisto). RNA reads may span exon–exon junctions (introns removed), so aligners must be splice-aware.

4. Quantification → count matrix

Count how many reads fall on each gene. Result: a gene × sample count matrix $C_{gs}$ = number of reads for gene $g$ in sample $s$ .

5. Normalization

Raw counts aren't comparable because samples differ in sequencing depth (total reads) and genes differ in length. We must correct for both.

Formula Deriving CPM, RPKM, TPM from first principles

Problem 1 — depth. Sample A has 40M reads, sample B has 20M. Gene with 100 reads in each is not equally expressed. Fix: divide by library size (per million): $\text{CPM}_{gs}=\frac{C_{gs}}{N_s}\times 10^6,\qquad N_s=\sum_g C_{gs}$

Problem 2 — length. A 4kb gene captures ~2× the reads of a 2kb gene at equal expression (more fragments fit). Divide by gene length in kb: $\text{RPKM}_{gs}=\frac{C_{gs}}{\left(\frac{L_g}{10^3}\right)\left(\frac{N_s}{10^6}\right)}$

Problem 3 — RPKM sums don't match across samples, so proportions aren't comparable. TPM fixes the order: divide by length first, then by the per-sample total of length-normalized values: $\text{TPM}_{gs}=\frac{r_{gs}}{\sum_k r_{ks}}\times 10^6,\quad r_{gs}=\frac{C_{gs}}{L_g}$ Key property: $\sum_g \text{TPM}_{gs}=10^6$ for every sample → columns are directly comparable proportions.

6. Differential expression (DE)

Compare a gene's counts between conditions (e.g. treated vs control) using tools like DESeq2 / edgeR, which model counts with the negative binomial distribution.

Worked examples

Common mistakes (steel-manned)

Recall

Recall Flashcards checkpoint (try before flipping)

What does TPM guarantee about each sample's column? → sums to $10^6$ .
Why negative binomial over Poisson? → overdispersion (variance > mean).
Why FDR correction? → many-genes multiple testing.

Recall Feynman: explain to a 12-year-old

Every cell has the same giant recipe book (DNA). But a cell only cooks some recipes at a time. RNA-seq is like sneaking into the kitchen and counting how many copies of each recipe card are being used right now. If we count more chocolate-cake cards in a sick person than a healthy one, maybe "chocolate cake" (a gene) is involved in the sickness. Because we peeked at different amounts of paper in each kitchen, we first fairly rescale the counts, then compare which recipes changed the most.

Flashcards

What does RNA-seq measure?

The identity and abundance of RNA molecules, i.e. gene expression, via high-throughput sequencing.

What file format holds raw sequencing reads + quality scores?

FASTQ.

Formula linking Phred quality Q and error probability P?

Q=-10\log_{10}P

What does a Phred score of 30 mean?

Error probability

10^{-3}

(1 wrong base in 1000).

Why must aligners for RNA-seq be splice-aware?

Reads can span exon–exon junctions since introns are spliced out.

Define CPM.

Counts per million:

C_{gs}/N_s \times 10^6

, correcting for sequencing depth.

Why do we normalize by gene length?

Longer genes capture more read fragments at equal expression, biasing raw counts.

Key property of TPM across a sample?

The TPM values sum to

10^6

, making columns comparable proportions.

Difference in calculation order between RPKM and TPM?

TPM divides by length first, then by the per-sample total; RPKM divides by depth first.

Why negative binomial for count modeling?

RNA-seq counts are overdispersed (variance > mean); NB adds a dispersion term

\text{Var}=\mu+\phi\mu^2

What is log2FC and what does +1 mean?

\log_2

(treated/control); +1 means expression doubled.

Why use FDR-adjusted p-values?

~20,000 simultaneous tests cause many false positives; BH controls the false discovery rate.

Why are biological replicates essential?

To estimate biological variability/dispersion for reliable DE calls (≥3 per group).

Two things raw counts fail to account for?

Sequencing depth and gene length.

Connections

DNA vs RNA structure
Central Dogma of Molecular Biology
Gene Expression Regulation
Sequence Alignment Algorithms
Multiple Testing & FDR
Negative Binomial Distribution
Differential Gene Expression

Concept Map

Hinglish (regional understanding)

Intuition Hinglish mein samjho

Dekho, har cell mein DNA same hota hai, lekin har cell alag genes ko "on" rakhta hai. RNA-seq ka kaam hai — jitni bhi RNA molecules cell mein active hain, unko count karna. Zyada RNA matlab wo gene zyada "busy" hai (high expression), kam RNA matlab wo gene lagbhag off hai. Isse hum compare kar sakte hain ki treated aur control sample mein kaunse genes badle.

Pipeline simple chain hai: pehle sequencer se short reads milte hain FASTQ file mein (saath quality score, jise Phred score kehte hain, formula $Q=-10\log_{10}P$ ). Phir quality check aur trimming, phir reads ko genome pe align karna (splice-aware aligner, kyunki introns cut ho chuke hote hain), phir har gene ke reads count karke ek count matrix banate hain.

Yahan ek important baat: raw counts ko directly compare mat karo! Kyunki ek sample zyada deeply sequence hua ho sakta hai (zyada total reads) aur lambe genes zyada reads pakadte hain. Isliye normalize karte hain — CPM depth ka correction karta hai, aur TPM depth + length dono. TPM ki khaas baat: har sample ke TPM ka total hamesha $10^6$ hota hai, isliye columns fair comparison dete hain.

Aakhri step differential expression — DESeq2/edgeR negative binomial model use karte hain (kyunki counts mein variance mean se zyada hoti hai, overdispersion). log2 fold change batata hai kitna up/down gaya (+1 = double), aur kyunki hum ~20,000 genes ek saath test karte hain, FDR (Benjamini-Hochberg) correction lagana zaroori hai warna false positives aa jayenge. Bas yahi hai RNA-seq analysis ka core!

Test yourself — Bioinformatics & Computational Biology

Connections