GFP Variant Design

CNN ensemble model for protein activity prediction using ESM-2 embeddings.

Installation

Prerequisites

• Python 3.12.1
• pip

Note: All versions of Python and packages are pinned to the exact versions used during analysis for reproducibility.

Setup

1. Clone the repository:

git clone https://github.com/Arcadia-Science/2025-GFP-variant-design.git
cd 2025-GFP-variant-design

2. Install dependencies:

pip install -e .

3. Download the ESM-2 embeddings from https://zenodo.org/records/17088257, and place the file in the data directory.

Notebooks

01_train_model.ipynb

Trains a CNN ensemble model on protein sequence embeddings and activity scores.

02_generate_sequences.ipynb

Generates protein sequence variants and predicts their activities using the trained model.

03_create_figures.ipynb

Creates visualizations from the results.

04_analyze_experimental_data.ipynb

Extracts fluorescence readings from raw plate reader data found in the experimental_data directory, analyzes the data, and creates visualizations for the protein variants we analyzed in the lab.

05_visualize_fluorescence.ipynb

Generates figures showing dual-channel microscopy images of E. coli expressing GFP variants, providing a representative view of fluorescence properties within the broader dataset.

Scripts

create_embeddings_from_seq_score.py

Generates embeddings and metadata from seq_and_score.csv with progress tracking.

Usage

Run notebooks in order: 01_train_model.ipynb → 02_generate_sequences.ipynb → 03_create_figures.ipynb

Generate Embeddings from Original Dataset

Converts seq_and_score.csv to full embeddings dataset with ESM-2 embeddings and metadata columns.

Some systems need certificates for embedding generation:

apt update && apt install -y ca-certificates && update-ca-certificates

Usage:

python create_embeddings_from_seq_score.py

Performance Notes:

• Original embeddings (~50k sequences) were generated on an H100 GPU in ~9 hours
• Uses GPU-to-CPU memory offloading, allowing the ESM-2 15B model to run on GPUs with limited VRAM by storing model parameters in CPU memory and transferring them to GPU as needed during computation
• Can run on smaller GPUs (e.g., A10G ~35 hours) but will take significantly longer
• Requires significant computational resources and CUDA-compatible GPU