Kuan-Hao Chao

Ab Initio Gene Finding & Structural Annotation

Generalized Hidden Markov Models (GHMMs)

Interactive exploration of explicit biological duration distributions f(d), semi-Markov Viterbi dynamic programming, exon-intron state machines (GENSCAN / AUGUSTUS), and splice signal decoding.

Gene Presets:
🟢 Start Codon: ATG (Initiation)
✂️ Splice Donor: GT (5' Exon-Intron junction)
🔗 Splice Acceptor: AG (3' Intron-Exon junction)
🔴 Stop Codons: TAA, TAG, TGA (Termination)
Ready
Predicted Coding Protein (CDS):
Exons:0
Introns:0
Coding:0 bp
GC Content:0%
Genomic Coordinate Browser Track:Hover feature box to inspect coordinates
1 bpGenomic Coordinates100 bp
Explicit Duration Distribution f(d) vs Standard HMMOvercomes geometric decay
Semi-Markov Dynamic Programming

Unlike standard HMMs that emit single bases, GHMMs (Semi-Markov HMMs) emit entire biological segments of length d.

The joint probability combines:

  • Duration score: log f_q(d) from biological length models
  • Signal score: Splice junctions (GT / AG), start (ATG), stop codons
  • Coding score: 3-periodic in-frame Markov codon potential

Geometric Memoryless Flaw

Standard HMMs enforce geometric state lengths P(d) = pd-1(1-p), which unrealistically favors 1-bp exons. GHMMs allow explicit density curves f(d) peaked at true biological lengths (100–200 bp).

Eukaryotic Splicing Signals

Spliceosomes recognize canonical consensus dinucleotides: GT at the 5' donor site and AG at the 3' acceptor site preceded by a polypyrimidine tract.

Ab Initio Gene Finders

Canonical tools like GENSCAN, AUGUSTUS, and SNAP utilize GHMMs with isochore-specific GC models and phylogenetic conservation to predict protein-coding structures directly from raw DNA.