6.4.9
Socho ki har cell ek factory hai. DNA ek master blueprint library hai (har cell mein same hoti hai), lekin kisi bhi moment par har cell sirf kuch blueprints padhti hai. RNA-seq ek tarika hai jisse hum photocopy karke count kar sakte hain saari blueprints (RNA molecules) jo abhi padhi ja rahi hain. Agar koi gene "busy" hai, toh uska bahut saara RNA milta hai; agar "off" hai, toh almost kuch nahi. Toh RNA-seq se hum gene expression measure karte hain — kaun se genes on hain aur kitni strongly — poore genome mein ek saath.
RNA-seq (RNA sequencing) ek technique hai jo high-throughput sequencing use karke ek biological sample mein RNA molecules ki identity aur abundance measure karti hai. Analysis ka end goal usually ek counts per gene per sample ki table hoti hai, jisse hum conditions ke beech differential expression infer karte hain.
WHY do we care? Ek hi DNA wale do cells bilkul alag ho sakti hain (jaise neuron aur skin cell) kyunki wo alag genes express karti hain. Disease, drug response, aur development sab expression mein changes ke roop mein dikhte hain. RNA-seq un changes detect karne ka standard tool hai.
Analysis ek chain hai. Har step ka output agli step ka input hota hai.
RNA extract ki jaati hai, cDNA mein reverse-transcribe ki jaati hai, fragment ki jaati hai, aur sequence ki jaati hai. Tumhe millions of short reads (~50–150 bp) FASTQ files mein milte hain. Har read mein ek base sequence hoti hai aur ek per-base quality score hota hai.
Definition Phred quality scorePhred score Q P Q = − 10 log 10 P Q = 30 P = 1 0 − 3 WHY log? Errors kai orders of magnitude tak pheli hoti hain, aur log se Q
Tools (FastQC) low-quality tails aur adapter contamination flag karte hain; trimming unhe remove karti hai. WHY? Kharab bases downstream mein galat alignments karte hain — garbage in, garbage out.
Har read ko ek reference genome par place kiya jaata hai (aligners: STAR, HISAT2) ya transcriptome se match kiya jaata hai (pseudo-aligners: Salmon, kallisto). RNA reads exon–exon junctions span kar sakti hain (introns remove ho jaate hain), isliye aligners splice-aware hone chahiye.
Count karo ki kitni reads har gene par fall karti hain. Result: ek gene × sample count matrix C g s g s
Raw counts comparable nahi hote kyunki samples mein sequencing depth (total reads) alag hoti hai aur genes ki length alag hoti hai. Hume dono ke liye correct karna padta hai.
Ek gene ke counts ko conditions ke beech compare karo (jaise treated vs control) DESeq2 / edgeR jaise tools use karke, jo counts ko negative binomial distribution se model karte hain.
Intuition Negative binomial kyun, Poisson kyun nahi?Poisson assume karta hai mean = variance. Real RNA-seq counts overdispersed hote hain (variance > mean) kyunki replicates ke beech biological variability hoti hai. Negative binomial ek dispersion parameter ϕ Var = μ + ϕ μ 2
Definition log2 fold change & significance
Fold change : expression ka ratio, control vs treated. log2FC = log 2 expr control expr treated + 1 − 1 Har gene ko ek p-value milta hai, phir ek adjusted p-value (FDR, Benjamini–Hochberg) kyunki hum ek saath ~20,000 genes test karte hain (multiple testing ).
Worked example Example 1 — Phred scoreEk base ki error probability P = 0.01 Q Q = − 10 log 10 ( 0.01 ) = − 10 ( − 2 ) = 20 Yeh step kyun? log 10 ( 0.01 ) = − 2 Q 20
Worked example Example 2 — CPMGene X mein 200 reads hain; sample library size N = 10 , 000 , 000 CPM = 1 0 7 200 × 1 0 6 = 20 Yeh step kyun? Hum "per million reads" par scale karte hain taki zyada depth par sequenced sample artificially zyada expressed na lage.
Worked example Example 3 — TPM (2 genes, 1 sample)Gene A: 300 reads, length 2 kb. Gene B: 300 reads, length 6 kb.
Step 1 — reads per kb: r A = 300/2 = 150 r B = 300/6 = 50 Kyun? Lambi gene ne same reads ko kam true expression par capture kiya → use penalize karo.
Step 2 — total = 150 + 50 = 200 TPM A = 200 150 × 1 0 6 = 750 , 000 TPM B = 250 , 000 Check: sum = 1 0 6
Worked example Example 4 — log2 fold changeControl expression = 50 = 200 log 2 50 200 = log 2 4 = 2 Yeh step kyun? 200/50 = 4 log 2 4 = 2
Common mistake "Raw counts samples compare karne ke liye theek hain."Kyun sahi lagta hai: counts hi direct measurement hote hain, toh unhe compare karna honest lagta hai.
Kyun galat hai: ek sample jo do guna zyada depth par sequenced ho woh har cheez ke liye ~2× counts deta hai — yeh depth artifact hai, biology nahi.
Fix: kisi bhi cross-sample comparison se pehle normalize karo (CPM/TPM ya DESeq2 ke size factors).
Common mistake "RPKM/TPM valuesbetween genes compare karo taaki pata chale kaun sabse zyada hai."
Kyun sahi lagta hai: dono length-normalized hain, toh wo apples-to-apples lagte hain.
Fix: yeh same gene ko samples across compare karne ke liye bane hain. Between-gene comparison mapping/GC effects se biased hai; inhe within-gene use karo.
Common mistake "Ek chota raw p-value matlab gene definitely significant hai."Kyun sahi lagta hai: chota p-value = strong evidence, ek single test mein.
Kyun galat hai: 20,000 genes test karne par, ~1000 sirf chance se p < 0.05 Fix: FDR-adjusted p-values use karo (Benjamini–Hochberg).
Common mistake "Biological replicates skip karo — ek condition per sample kaafi hai."Fix: replicates ke bina tum biological variability/dispersion estimate nahi kar sakte → DE calls unreliable hote hain. Har group mein ≥3 replicates use karo.
Recall Flashcards checkpoint (flip karne se pehle try karo)
TPM har sample ke column ke baare mein kya guarantee karta hai? → 1 0 6
Poisson ki jagah negative binomial kyun? → overdispersion (variance > mean).
FDR correction kyun? → many-genes multiple testing.
Recall Feynman: ek 12-saal ke bacche ko explain karoHar cell mein same bada recipe book hota hai (DNA). Lekin ek cell sirf kuch recipes ek time par banaati hai. RNA-seq jaisa hai jaise kitchen mein ghus ke count karo ki abhi kitni copies kaunse recipe cards use ho rahe hain. Agar kisi beemar insaan mein healthy insaan se zyada chocolate-cake cards milte hain, toh shayad "chocolate cake" (ek gene) bimari mein involved hai. Kyunki humne har kitchen mein alag amount ka paper dekha, hum pehle counts ko fairly rescale karte hain, phir compare karte hain ki kaun se recipes sabse zyada bdale.
"Fun Quality Aligns Counts, Normalize Differences"
F ASTQ → Q C/trim → A lign → C ount → N ormalize → D ifferential expression.
RNA-seq kya measure karta hai? RNA molecules ki identity aur abundance, yaani gene expression, high-throughput sequencing ke zariye.
Raw sequencing reads + quality scores kis file format mein hote hain? FASTQ.
Phred quality Q aur error probability P ko jodne wala formula? Q = − 10 log 10 P
Phred score 30 ka matlab kya hai? Error probability 1 0 − 3
RNA-seq ke aligners splice-aware kyun hone chahiye? Reads exon–exon junctions span kar sakti hain kyunki introns splice out ho jaate hain.
CPM define karo. Counts per million: C g s / N s × 1 0 6
Gene length se normalize kyun karte hain? Lambi genes equal expression par zyada read fragments capture karti hain, raw counts bias hote hain.
Ek sample mein TPM ki key property kya hai? TPM values 1 0 6
RPKM aur TPM ke calculation order mein kya difference hai? TPM pehle length se divide karta hai, phir per-sample total se; RPKM pehle depth se divide karta hai.
Count modeling ke liye negative binomial kyun? RNA-seq counts overdispersed hote hain (variance > mean); NB ek dispersion term add karta hai Var = μ + ϕ μ 2
log2FC kya hai aur +1 ka matlab kya hai? log 2
FDR-adjusted p-values kyun use karte hain? ~20,000 simultaneous tests se bahut false positives aate hain; BH false discovery rate control karta hai.
Biological replicates essential kyun hain? Reliable DE calls ke liye biological variability/dispersion estimate karne ke liye (≥3 per group).
Do cheezein jinhe raw counts account nahi karte? Sequencing depth aur gene length.
RNA-seq measures gene expression
Wet-lab: RNA to cDNA sequenced
Raw reads FASTQ + quality
Reference genome / transcriptome
Gene x sample count matrix
Sequencing depth + gene length bias
