Methods blog · Multi-omics & Single-cell Bioinformatics
PromoBotPromoBot: finding plant promoters with hexamers and random triplet pairs
Two feature-selection algorithms, one frequency-based and one driven by a genetic algorithm, feed an SVM that recognises plant promoter sequences with 89% sensitivity and 86% specificity.
#promoter prediction#feature selection#genetic algorithm#SVM#plant genomics
Every gene needs a promoter, the stretch of DNA upstream of the transcription start site where the transcription machinery assembles. As plant genomes were sequenced at an accelerating pace, annotating promoters computationally became a bottleneck: existing predictors were either tuned for animal genomes or struggled to separate genuine promoters from the rest of the sequence. PromoBot was our answer.
The problem: which short words matter?
A promoter is recognisable by its composition: short DNA “words” (motifs) appear more or less often than they would in non-promoter DNA. The challenge is deciding which words to count. With 4 nucleotides there are \(4^6 = 4096\) possible hexamers, and most of them are noise. Feeding all of them to a classifier overfits; choosing too few misses the signal.
Key idea: two complementary feature selectors
We treated a hexamer as a pair of triplets and designed two ways to choose informative pairs:
- FDAFSA, the Frequency Distribution Analyzed Feature Selection Algorithm. For adjacent triplet pairs (ordinary hexamers), compare their frequency distributions in promoters versus non-promoters and keep the hexamers whose frequencies differ most.
- RTPFSGA, the Random Triplet Pair Feature Selecting Genetic Algorithm. Allow the two triplets to be non-adjacent, separated by a gap, which captures spaced motifs that a plain hexamer count misses. The search space of such random triplet pairs (RTPs) is huge, so a genetic algorithm evolves subsets of RTPs whose frequencies best discriminate the two classes.
The selected features from both algorithms are combined and passed to a support vector machine (SVM) with a non-linear kernel, which learns the promoter / non-promoter boundary:
\[f(\mathbf{x}) = \operatorname{sign}\Big(\sum_i \alpha_i y_i \, K(\mathbf{x}_i, \mathbf{x}) + b\Big),\]where \(\mathbf{x}\) is the vector of selected hexamer and RTP frequencies for a sequence.
Data and validation
- Promoters: PlantProm database.
- Non-promoters: plant mRNA, rRNA and tRNA from PlantGDB, plus plant miRNAs from miRBase, so the negatives look like real genomic sequence rather than random strings.
- Protocol: 5-fold cross-validation, with feature selection run inside each training fold to avoid information leakage.
Results
PromoBot performed as well as or better than the five other algorithms we compared against. The lesson that carried into our later work is that careful, biologically motivated feature selection (here, spaced triplet pairs) can matter as much as the classifier itself.
Why it still matters
Spaced k-mer features later became standard in genomics deep learning, through gapped k-mer SVMs and dilated convolutions. PromoBot was an early demonstration that non-adjacent sequence patterns carry regulatory signal.
Read the paper
-
Prediction of plant promoters based on hexamers and random triplet pair analysis
Algorithms for Molecular Biology 6 (1), 19 · 2011