GenomeInterpretation

Robustness-Driven Feature Attribution in Genomic Deep Learning with Integrated Gradients

Overview

This project develops tools and modeling standards to enhance the interpretability of phenotype-genome models, with the goal of automating biomarker discovery from genome sequences.

We train CNN models on one-hot encoded DNA sequences and apply gradient-based interpretation methods (e.g., Integrated Gradients). This repository contains experiments to validate and improve these interpretation approaches in biological contexts.

Project Structure

GenomeInterpretation/
├── synthetic/          # Synthetic data experiments
├── koo/                # Explainability/interpretation research
├── sporulation/        # Sporulation phenotype analysis
├── phenotype/          # General phenotype prediction
├── slurm_scripts/      # SLURM job submission scripts
└── toy_single_arch.py  # Main synthetic experiment entry point

Components

1. Synthetic Data Experiments

Controlled experiments with synthetic sequences to validate interpretation methods.

Script	Description
toy_single_arch.py	Main entry point for synthetic experiments
synthetic/code/gc_baseline_sweep.py	GC-content baseline sanity check

Documentation: synthetic/docu/

2. Explainability Research (koo/)

Task-based pipeline for model interpretation and robustness analysis.

Script	Description
koo/code/task3_step1.py	Step 1 of interpretation pipeline
koo/code/task3_step2.py	Step 2 of interpretation pipeline
koo/code/task3_step3.py	Step 3 of interpretation pipeline
koo/code/task3_step4.py	Step 4 of interpretation pipeline
koo/code/gc_baseline_classifier.py	GC-content baseline classifier

Documentation: koo/README2.md

3. Sporulation Analysis

Sporulation phenotype prediction and multi-model analysis.

Script	Description
sporulation/code/bugphyzz_visual.py	BugPhyzz data visualization
sporulation/code/codon_msa.py	Codon-level MSA analysis
sporulation/code/figure_four_panel.py	Multi-panel figure generation
sporulation/code/rashomon_*.py	Rashomon set analysis (multiple variants)
sporulation/code/run_streme.py	STREME motif discovery

Documentation: sporulation/docu/

4. Phenotype Prediction (phenotype/)

General-purpose phenotype prediction framework. All scripts in phenotype/code/ are actively used.

Key scripts include training, evaluation, GFF processing, and genome feature analysis.

Environment

conda env create -f environment.yml
conda activate genome

Requires: Python 3.12, PyTorch 2.3, CUDA 12.1

Contact

yichen.han AT campus.lmu.de