Computational Functional Genomics - Project

VINCA: A CNN-based predictive model for TF-binding

Predict transcription factor (TF) binding (CTCF, REST, EP300) using DNA sequence, ATAC-seq, CpG methylation, and PWM features.

---

Features

• Multi-TF CNN model with sequence + ATAC + methylation + PWM input
• Input encoding: (N × 200 × 9) tensor
• Residual CNN + dilated convolutions + attention
• Reverse complement augmentation (training + inference)
• Chromosome-wise training (no leakage)
• Handles class imbalance using:
  • Negative sampling (NEG_RATIO)
  • Focal Loss
• Trains on chromosomes 1–22 (excluding test chromosomes)
• Generates predictions for:
  • chr3
  • chr10
  • chr17
• Saves:
  • Prediction .tsv.gz files
  • Loss curve
  • Prediction distribution plots

---

Dependencies

• Python 3.10+
• numpy
• pandas
• torch
• scikit-learn
• matplotlib
• tqdm

Features

• Multi-TF CNN model with sequence + ATAC input
• Cross-validation on chromosome 1
• Final model training on chromosomes [1–22]
• Generates predictions for specific chromosomes (3, 10, 17)
• Saves ROC curves and AUC results

---

Running the Model

1. Full Run (Train → Select Best Model → Predict)

Execute the main pipeline:

python3 main.py

2. Fast Run (Prediction Only)

Use this mode when confident about required inputs and/or if you want to skip model selection.

Steps:

1. Clear the checkpoints directory:

rm -r models/checkpoints/*

python3 predict.py

Detailed info on VINCA:

See the model Overview

---