Explain gene prediction methods — Hinglish notes

6.4.7 · Biology › Bioinformatics & Computational Biology

WHY gene prediction ki zaroorat hai?

WHAT: Sequencing se aapko raw DNA milta hai — billions of letters. Lekin raw sequence yeh nahi batata ki kaunse hisse actually proteins ya functional RNA ke liye code karte hain.

WHY it's hard:

Aise koi obvious markers nahi hain jo gene ko junk DNA se alag karein.
Eukaryotes mein, genes interrupted hoti hain: coding pieces (exons) non-coding pieces (introns) se todi jaati hain jo splice out ho jaate hain.
Kisi bhi DNA stretch ke 6 reading frames hote hain (3 forward, 3 reverse), aur inme se zyaadatar meaningless hote hain.

HOW hum attack karte hain: Hum statistical signals aur sequence signals dhundte hain jo real genes chhod jaate hain, phir unhe combine karte hain.

Methods ki do badi families

1. Ab initio (intrinsic) methods

Sirf query sequence aur genes kaisi dikhti hain iske statistical models use karte hain.

2. Homology / evidence-based (extrinsic) methods

Sequence ko known genes, proteins, ya expressed sequences (mRNA/EST/RNA-seq) se compare karte hain.

Sequence signals jo hum dhundte hain (content vs. signal)

Figure — Explain gene prediction methods

Building block 1 — Open Reading Frames (ORFs)

WHY ORFs matter: Prokaryotes mein, genes mostly continuous ORFs hoti hain, isliye akela ORF-finding achha kaam karta hai. Eukaryotes mein, introns ORF ko tod dete hain, isliye ORFs poori kahani ka sirf ek hissa hain.

Deriving: kitne codons, kitna lamba ORF?

DNA triplets mein padha jaata hai. Agar ek ORF $n$ nucleotides span karta hai (start & stop inclusive), toh codons ki sankhya hai:

$C = \frac{n}{3}$

Protein mein amino acids ki sankhya (start codon → Met rakha jaata hai, stop codon → koi amino acid nahi):

$L_{aa} = \frac{n}{3} - 1$

Deriving: probability ki random codon stop ho

$4^{3} = 64$ codons hote hain. Unme se teen (TAA, TAG, TGA) stops hain.

$P (stop) = \frac{3}{64} \approx 0.047$

Stop aane se pehle expected codons ki sankhya (geometric distribution mean):

$E [codons before stop] = \frac{1}{P ( stop )} = \frac{64}{3} \approx 21$

Building block 2 — Markov models & GHMMs (modern predictors ka engine)

k-th order Markov probability derive karna

Ek sequence $S = s_{1} s_{2} \dots s_{n}$ . $k$ -th order Markov model ke under:

$P (S) = P (s_{1} \dots s_{k}) \prod_{i = k + 1}^{n} P (s_{i} ∣ s_{i - 1}, \dots, s_{i - k})$

Yeh product form kyun? Hum probability ka chain rule apply karte hain aur phir Markov assumption lagate hain: har base sirf pichle $k$ bases par depend karta hai, poori history par nahi. Isse badi joint probability tractable ho jaati hai (sirf $4^{k + 1}$ parameters chahiye).

Phir hum ek log-likelihood ratio score compute karte hain jo coding model ko non-coding (null) model se compare karta hai:

$S_{LL R} = lo g \frac{P ( sequence ∣ coding )}{P ( sequence ∣ non-coding )}$

GHMM (Generalized Hidden Markov Model): GENSCAN, AUGUSTUS, GeneMark, Glimmer jaise tools HMMs use karte hain jahan hidden states = {exon, intron, intergenic, 5'UTR, splice site, ...}. Predictor most probable state path dhundta hai (Viterbi algorithm se) jo observed DNA explain kare — woh path hi predicted gene structure hoti hai.

Homology-based methods (dictionary use karna)

Sequence ko protein/nucleotide databases ke against BLAST/BLAT karo → conserved regions likely code karte hain.
EST / cDNA / RNA-seq alignment: expressed sequences directly exons par map hote hain, exon boundaries aur splicing experimentally reveal karte hain.
Tools: GeneWise, Exonerate, spliced aligners (STAR, HISAT).

Prokaryotes vs. Eukaryotes

Worked examples

Common mistakes (Steel-man + fix)

Flashcards

Gene prediction kya hai?

Genomic DNA mein genes (exon/intron boundaries, regulatory signals) ki computational identification, sirf sequence se ya baahri evidence ke saath.

Gene prediction methods ki do main families?

Ab initio (intrinsic, sequence + statistical models use karta hai) aur homology/evidence-based (extrinsic, known sequences se compare karta hai).

Signal sensors aur content sensors mein difference?

Signal sensors short specific motifs detect karte hain (start/stop codons, splice sites); content sensors region-wide statistics measure karte hain (codon bias, GC content, ORF length).

ORF define karo.

DNA stretch jo start codon (ATG) se stop codon tak ho aur beech mein koi in-frame stop na ho.

Teen stop codons?

TAA, TAG, TGA.

Splice donor aur acceptor consensus?

Donor = GT (intron ka 5' end), Acceptor = AG (intron ka 3' end) — "GT...AG rule".

Random codon ke stop hone ki probability, aur expected run length?

3/64 ≈ 0.047; expected ~64/3 ≈ 21 codons before a stop.

n nucleotides ke ORF se amino acids ka formula?

L_aa = n/3 − 1 (stop codon koi amino acid code nahi karta).

Gene finding mein Markov models kyun use hote hain?

Coding DNA mein biased base-neighbor statistics hoti hain (codon usage); k-th order Markov model P(next base | previous k bases) capture karta hai, jo coding vs non-coding ke liye alag hoti hai.

GHMM gene finder mein Viterbi algorithm kya karta hai?

Most probable hidden-state path (exon/intron/intergenic) dhundta hai jo DNA explain kare = predicted gene structure.

Eukaryotic gene finding prokaryotic se zyaada mushkil kyun hai?

Eukaryotic genes introns se split hoti hain aur splice-site + UTR modeling chahiye; prokaryotic genes mostly continuous ORFs hoti hain.

Do ab initio eukaryotic gene finders batao.

GENSCAN aur AUGUSTUS (GeneMark bhi).

Do prokaryotic gene finders batao.

Glimmer aur Prodigal (GeneMarkS bhi).

Ab initio ko homology/evidence ke saath combine kyun karein?

Ab initio novel genes dhundta hai lekin boundaries par galti karta hai; homology/evidence accurate hai lekin unknown genes ke liye andha hai — combine karne se best accuracy milti hai.

Log-odds coding score kya hai aur log kyun use karte hain?

log[P(seq|coding)/P(seq|non-coding)]; log tiny probabilities ke product ko stable sum mein convert karta hai jahan sign coding indicate karta hai.

Recall Feynman: 12-saal ke bachche ko samjhao

Socho ek super-lamba string sirf chaar letters ki — A, T, G, C — NO spaces ke saath. Usme kahin "recipes" (genes) chhipi hain jo cell ko batati hain cheezein kaise banana hai, lekin woh gibberish ke samundar mein chhipi hain. Gene prediction ek smart word-search ki tarah hai: hum jaante hain ki real recipes hamesha ek special "GO" word (ATG) se shuru hoti hain aur ek "STOP" word par khatam hoti hain, aur unke letters aise patterns follow karte hain jo gibberish nahi karta. Computer har chunk score karta hai: "kya yeh real recipe jaisa dikh raha hai, ya random junk?" Hum ek badi cookbook of known recipes mein bhi dekh lete hain double-check ke liye. Jab dono — pattern-clues aur cookbook — agree karein — jackpot, humne ek gene dhundh li!

Connections

Sequence Alignment (BLAST) — homology-based prediction isi par rely karti hai
Hidden Markov Models — GENSCAN/AUGUSTUS ka statistical engine
Central Dogma — transcription/splicing/translation define karte hain ki "gene" kya hai
Genetic Code & Codons — codon bias aur stop-codon statistics ka source
RNA-seq & Transcriptomics — exon boundaries ke liye expression evidence
Genome Annotation Pipelines — MAKER/BRAKER saari methods combine karte hain
Prokaryote vs Eukaryote Genome Organization — kyun methods alag hain