D-Sites: Hybrid TFBS Predictor for Bacterial Genomes

A comprehensive computational tool for predicting transcription factor binding sites (TFBS) in bacterial genomes using hybrid PWM, DNA shape features, and Random Forest classification.

🚀 Quick Start

Installation

git clone https://github.com/pankaj357/D-Sites.git
cd D-Sites
pip install .

This installs the dsites command-line tool and all required dependencies (see requirements.txt for the exact pinned versions used for the published benchmarks).

Basic Prediction

Minimal Command

dsites --fasta <genome.fasta> \
       --gff <annotation.gff> \
       --motif <motif_file> \
       --gene <TF_name> \
       --genome_accession <accession_id>

Complete Example

dsites \
    --fasta <path_to_genome.fasta> \
    --gff <path_to_annotation.gff> \
    --motif <path_to_motif_file> \
    --gene <TF_NAME> \
    --genome_accession <GENOME_ACCESSION> \
    --output results \
    --n_trees 300 \
    --neg_ratio 5 \
    --prob_cutoff 0.5 \
    --pseudocount 0.5 \
    --pad 10 \
    --seed 42 \
    --batch 10000 \
    --up 300 \
    --down 50 \
    --auto_cutoff

If the TF/genome pair is not present in the bundled CollecTF export, provide known binding sequences explicitly (plain text, one sequence per line):

dsites --fasta genome.fna --gff genomic.gff --motif motif.txt \
       --gene FNR --genome_accession NC_003197.2 \
       --known known_binding_sites.txt --output results

Command Breakdown

Required Arguments

--fasta: Genome FASTA file path

--gff: Genome annotation file (GFF3 format, required for promoter-based
       scanning; optional with --full_genome)

--motif: TF motif file (JASPAR or MEME format)

--gene: Transcription factor name

--genome_accession: Genome accession ID, used to look up known sites in the
       bundled CollecTF dataset (scanning always covers ALL contigs in the FASTA)

Optional Arguments with Defaults

--output results: Output directory (alias: --outdir)

--n_trees 300: Number of Random Forest trees

--neg_ratio 5: Negative:Positive ratio (honored exactly)

--min_negatives 0: Optional minimum number of negative examples

--pseudocount 0.5: Count-based PWM pseudocount, p = (count + pc) / (N + 4*pc)

--prob_cutoff 0.5: Probability cutoff

--pad 10: Window padding around known sites

--seed 42: Random seed (runs are fully deterministic for a fixed seed)

--batch 10000: Batch size for parallel scanning

--up 300: Upstream promoter size

--down 50: Downstream promoter size

--full_genome: Slide windows across entire contigs instead of promoter windows

--auto_cutoff: Automatically select probability cutoff via precision-recall curve

📈 Performance

Held-out evaluation (grouped 5-fold site-level cross-validation; thresholds selected from out-of-bag predictions; FIMO compared on identical promoter windows at matched prediction counts — full workflow in evaluation/):

• Recall of known promoter-localized sites: 75% (AmrZ), 76% (GlxR), 96% (CodY)
• Ranking quality comparable to FIMO overall; higher precision and recall than FIMO for CodY (0.154 vs 0.118 precision at N = 500)
• DNA-shape features improve ranking for every factor tested (ablation); largest gain for CodY (~2× PR-AUC vs PWM-only)
• A model trained on 97 E. coli FNR sites (RegulonDB 13.5) transfers to Salmonella Typhimurium LT2 (cross-genome application)

📝 Citation

If you use D-Sites in your research, please cite:

Pankaj et al. D-Sites: A hybrid machine-learning framework for prediction of transcription factor binding sites in bacterial genomes. Evolutionary Bioinformatics (under review).

📄 License

MIT License - see LICENSE for details.

💬 Contact

For questions and support, contact:

• Pankaj: ft.pank@gmail.com
• Kanaka KK: kkokay07@gmail.com