🧬 Exon Start Prediction in Human Genome Using Deep Learning

This project implements deep learning models for predicting exon start sites in the Homo sapiens genome using annotated reference data from Ensembl (GRCh38). It features a custom pipeline for parsing GTF annotations, extracting sequence windows from reference FASTA files, and training convolutional neural networks to distinguish exon starts from other intragenic positions.

Highlights

• Dataset: One-hot encoded 101/201/301bp windows from chromosomes 1–22, X, Y
• Models: Enhanced CNN with SE blocks, attention, inception-style convs
• Performance: Best model achieves 78.3% accuracy on a balanced validation set

Important Files

• preprocess_data.py - script for data preprocessing. It creates the dataset for all chromosomes for the given window_size.``
• utils.py - utility functions for working with data
• train.py - script for model training.
• models.py - file with model definitions
• script.sh - script for downloading data and running the code

More info can be found in Exon Report.pdf.