Transcription Factor Binding Site Classification

Binary classification of transcription factor binding sites (TFBS) using k-mer SVM and LS-GKM models, with motif interpretability analysis. Applied to two transcription factors with distinct binding properties: CTCF and SP1.

Tools & Technologies

• Models: Linear k-mer SVM, LS-GKM (gapped k-mer SVM)
• Libraries: scikit-learn, NumPy, pandas, Matplotlib
• Data: ENCODE ChIP-seq narrowPeak datasets
• Scripts: Bash workflows for data preprocessing and model execution

Project Overview

This project explores sequence-based classification of transcription factor binding sites and evaluates whether learned sequence features correspond to known biological motifs. Two TFs with distinct binding characteristics were used as a comparative benchmark:

• CTCF — binds a long, highly specific ~20bp motif; key regulator of genome 3D organization
• SP1 — binds short, degenerate GC-box motifs; broad gene regulatory activity

Key design decisions investigated: input representation (one-hot vs k-mer frequency), negative sample strategy (dinucleotide shuffles vs flanking sequences), and sequence window length.

Data

ChIP-seq narrowPeak datasets from ENCODE:

• SP1: ENCFF769YSM
• CTCF: ENCFF144XRE

Train/test splits were performed by chromosome (Train: chr1–chr19; Test: chr20–chr22, X, Y) to prevent data leakage and ensure generalization to unseen genomic regions.

Results

Model Performance (AUROC)

Model	TF	Window	AUROC
Linear SVM	SP1	30bp	0.56
Linear SVM	SP1	200bp	0.80
Linear SVM	CTCF	50bp	0.83
LS-GKM	SP1	30bp	0.55
LS-GKM	SP1	200bp	0.89
LS-GKM	CTCF	50bp	0.88

Motif Interpretability

Model	TF	Result
Linear SVM	CTCF	Top 50 k-mers significantly enriched for canonical motif vs random (p < 1e-4)
LS-GKM	CTCF	Top 50 k-mers significantly enriched for canonical motif vs random (p < 1e-4)
Linear SVM	SP1	No significant PWM alignment (p = 0.08)
LS-GKM	SP1	No PWM alignment (p = 0.5), but significant CpG enrichment (p = 2e-4)

Key Findings

• k-mer frequency outperforms one-hot encoding as input representation for both TFs
• Dinucleotide-shuffled negatives yield better model discrimination than flanking sequences
• SP1 requires longer sequence windows (200bp) due to its degenerate binding motif; CTCF performs well at 50bp
• LS-GKM consistently outperforms linear SVM across conditions
• CTCF motif is recoverable from both models despite lack of explicit positional modeling, consistent with its strong, specific consensus sequence
• SP1 k-mers reflect GC-box composition rather than a positional motif, consistent with its known binding degeneracy

Conclusions

Model performance is strongly influenced by negative sample choice and input representation. SP1 binding site prediction is inherently more difficult than CTCF due to its degenerate GC-box motif and tolerance for sequence variation. Interpretability analysis reveals TF-specific differences in what the model learns: positional motif recovery for CTCF vs compositional enrichment for SP1.

Setup

pip install scikit-learn numpy pandas matplotlib

LS-GKM requires a separate installation — see the LS-GKM GitHub. Bash scripts for data preprocessing are in the bash_scripts/ directory.