Explainable Subcellular Localization Predictor

This is a lightweight, interpretable machine-learning pipeline designed to predict protein subcellular localization from primary amino-acid sequences and reveal the key sequence motifs driving each classification. We combined classical feature engineering with a Random Forest classifier and Shapley Additive exPlanations (SHAP) to generate transparent, biologically meaningful insights.

Note

This project has won the 2025 ISCB YBS Student Challenge at the 2025 joint international conference on Intelligent Systems for Molecular Biology (ISMB) and the European Conference on Computational Biology (ECCB) for its innovative approach to AI in bioinformatics.

Features

• Random Forest classifier trained on curated UniProtKB/Swiss-Prot sequences across 16 compartments
• SHAP values assign per-feature contributions for each protein prediction
• Minimal dependencies and fast inference on standard workstations
• Easily swap in new classifiers or add custom features

Installation

1. Clone the repository:

   git clone https://github.com/cytronicoder/explainable-localization-predictor.git
   cd explainable-localization-predictor

2. Create and activate a Conda environment (recommended):

   conda env create --file environment.yml
   conda activate eslp

3. To update or add dependencies, modify environment.yml and run:

   conda env update --file environment.yml --prune

Requirements

• Python 3.8 or higher
• scikit-learn
• SHAP
• pandas
• numpy
• matplotlib (for optional plotting)

Feature Engineering

• Amino-acid composition (20 dimensions)
• Dipeptide frequencies (400 dimensions)
• Physicochemical properties (molecular weight, isoelectric point)
• Entropy measures for disorder prediction

Results

Sample performance on the held-out test set:

• Overall weighted F1: 0.88
• Compartment AUCs: 0.93–0.97
• Key biological insights:
  • Hydrophobicity (GRAVY) drives membrane vs. soluble distinction
  • Isoelectric point correlates with cytosolic adaptation
  • Sequence length penalties reveal targeting constraints

License

This project is licensed under the MIT License. See LICENSE for details.

Contributing

Contributions are welcome! Please open issues for bug reports or feature requests, and submit pull requests against the main branch.