🧬 Protein-Language-Model-Steering🧬

Overview

This repository includes the code for the paper Where to Edit? : Complementary Protein Property Control from Weight and Activation Spaces.

We note that for SAE-based steering, we use SAEFold from Parsan et al., 2025. Their repository is available here. We are not able to include it in our repository due to licensing issues.

---

Table of Contents

• Features
• Installation
• Quick Start
• Running Notebooks
• Directory Structure
• Third-Party Models and Licenses
• Documentation
• Contributing
• License

---

Features

• Steering and fine-tuning protein language models
• Utilities for dataset preparation and analysis
• Visualization tools for model interpretability
• Benchmarking scripts for evaluation
• Modular and extensible design for integration of new models or datasets

---

Installation

1. Clone the repository

   git clone https://github.com/Ulton321/Protein-Language-Model-Steering.git
   cd Protein-Language-Model-Steering

2. Install dependencies

   pip install -r requirements.txt

   Or, using Anaconda:

   conda create -n plm-steering python=3.8
   conda activate plm-steering
   pip install -r requirements.txt

3. (Optional) Install Jupyter Notebook

   pip install notebook

   Or for JupyterLab:

   pip install jupyterlab

---

Quick Start

1. Launch Jupyter Notebook

Navigate to the repository folder and start Jupyter:

jupyter notebook

Or for JupyterLab:

jupyter lab

2. Open the Notebook

In your browser, open the notebook you wish to run (e.g., notebooks/plm_steering_demo.ipynb).

3. Run Notebook Cells

• Select each cell and press Shift + Enter to execute.
• Follow the instructions in the notebook to perform data preparation, model training, steering, and evaluation.

4. Modify Parameters as Needed

• Most notebooks allow customization of paths, hyperparameters, and options.
• Read any comments or markdown cells for guidance.

---

Running Notebooks: Full Instructions

1. Prepare Data

   • Place your protein sequence files in the data/ directory.
   • Supported formats: .fasta, .csv, or as specified in each notebook.

2. Configure Notebook

   • Update any file paths and parameters in the first cell or as instructed.

3. Execute Cells in Order

   • Start from the top and run all cells sequentially.
   • If you encounter errors, check for missing dependencies or review the cell’s instructions.

4. Save Results

   • Output files (e.g., steered sequences, model checkpoints) will be saved in the results/ or checkpoints/ folder.

5. Visualize Outputs

   • Most notebooks include visualization cells (plots, tables). Run these to inspect your results.

---

Directory Structure

Protein-Language-Model-Steering/
│
├── data/                # Input/Output protein sequences
├── notebooks/           # Jupyter Notebooks for workflows
├── results/             # Output files and figures
├── checkpoints/         # Saved models
├── requirements.txt     # Python dependencies
├── LICENSE              # Project license
├── LICENSES/            # Third-party licenses
├── README.md            # This file
└── docs/                # Additional documentation

---

Third-Party Models and Licenses

This project uses the following third-party models:

• Esm3 - MIT License
• Prollama - Apache 2.0 License
• ProGen2 - BSD License

Please refer to the respective license files in the LICENSES/ directory for details.

How to add third-party licenses:

1. Find the license text in the original repository of each model.
2. Copy the license text into a new file in the LICENSES/ directory.
3. Name the file clearly (e.g., LICENSE.modelA.txt).
4. Reference each license and model in this section of the README.

---

Documentation

See the docs/ directory for:

• Getting Started Guide
• Model architecture overview
• Data format documentation
• Benchmarking protocols

---

Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

---

License

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

---

Empowering protein research through deep learning and open science.