AI summary by TwinMind
6B. RNA 2: Clustering by Gene or Condition and Other Regulon Data Sources Nucleic Acid ...
MIT OpenCourseWare
44m 46s
Algorithm Improvements for Motif Discovery
- Two different algorithms for improving motif finding: historical method vs 100-fold improvement in speed/accuracy
- First approach: knock out high information content bases by converting to 'X' (non-matching character)
- Prevents Gibbs sampler from finding same motif again
- Disadvantages: may miss overlapping motifs with slight differences
- Second approach: maintain list of previously found motifs, compare emerging motifs during sampling
- If new motif becomes similar to previous winner, reject early (100x speed improvement)
- Allows finding overlapping motifs with different column structures/offsets
- Dual advantage: better discrimination of related but separable consensus elements
MAP Score Components and Mathematical Framework
- Weight matrix (f_jb): actual counts of A,C,G,T at each position j
- Active columns (c) ≤ total width of motif
- GCN4 example: width 10-11 with 10 active columns (c=10)
- Pseudo counts (beta): prevent zeros in limited sampling scenarios
- Background genome frequency (g_b): accounts for base composition bias
- Double-stranded DNA: freq(A) = freq(T)
- Single-stranded RNA: independent background frequencies
- Oversimplified MAP score: over-representation bonus for aligned sites
- Full specificity requires checking presence in rest of genome
Cell Cycle Data Analysis Case Study
- Dataset: 2 cell division cycles, 15 time points, 30 clusters total
- Cluster 2 (S-phase peak): 186 genes, periodic expression pattern
- Functional enrichment: 23/82 DNA synthesis genes overlap (p=10^-16)
- Top motifs discovered: MCB and SCB (algorithmically found, not literature-based)
- Group specificity: MCB motifs concentrated in cluster 2, rare elsewhere
- Positional bias: non-random spike before transcription/translation start (ATG)
Additional Cluster Examples
- Cluster 14: periodic but shifted, two unknown motifs, weaker functional enrichment (p=10^-6)
- Cluster 1: non-periodic (temperature shift artifact), ribosomal protein enrichment (p=10^-30)
- Most impressive functional category match of all 30 clusters
- One known, one unknown motif, both highly specific
Statistical Evaluation Methods
- Hypergeometric distribution: sampling without replacement for functional enrichment and group specificity
- Formula: sum from observed overlap x to maximum possible
- N = total genes (~6000 yeast), S1 = cluster size, S2 = functional category size
- Binomial distribution: positional bias evaluation
- t = total sites, m = sites in most enriched window
- Probability p based on window size vs total promoter region (600bp)
Motif Comparison and Symmetry Analysis
- Comparison scores: correlation coefficient-like, >0.7 statistically significant
- Self-comparison for symmetry detection:
- Reverse complement comparison → dyad symmetry (180° protein dimer)
- Direct repeat comparison → helical repeat in 3D structure
- Examples: PurR, RdgR, CRP show high dyad symmetry scores
- CpxR shows poor reverse complement score but likely direct repeat structure
Validation and Controls
- Test sets: negative controls (random gene selections), positive controls (29 known transcription factors with ≥5 sites)
- Functional categories: 248 categories tested
- Success rates: 21/29 positive controls successful, false negative rate 10-30%
- Ranking methods: MAP score, specificity score, positional bias
- Different rankings reveal different biological insights
- A-rich motifs: high MAP score, poor specificity, astronomical positional bias
Future Directions and Limitations
- Current assumption: motifs act in isolation (one at a time discovery)
- Need to address: motif interactions and co-regulation (2-3+ motifs working together)
- Protein-DNA connection: in vivo cross-linking, one-hybrid assays, proteomics
- Column independence assumption: may need revision based on evidence of interdependence
- Goal: generalize and reduce assumptions to discover new regulatory mechanisms
