Integrating a New Model into FoldBench

Welcome, contributor! This guide outlines the process for integrating your protein structure prediction model into the FoldBench benchmarking platform.

Platform Requirements

Before you begin, ensure your environment meets the following system requirements:

OS: Linux
Containerization: Apptainer
Package Management: Conda
Hardware: NVIDIA GPU

Preparing Your Algorithm

To add your model, create a new directory inside FoldBench/algorithms/ using your algorithm's name. This directory must contain the following four files:

1. 📦 `container.def`

This Apptainer definition file specifies the complete environment for your model. It should install all necessary system packages, Python libraries (e.g., via pip), and set up any required environment variables.

2. 📝 `preprocess.py`

This script prepares the input data for your model. It must contain a PreProcess.preprocess() method.

Input: The script should read from the benchmark's standard input file: ./alphafold3_inputs.json.
Function: Convert the data from the standard JSON format into the specific file format that your model requires for inference.
Output: The preprocessed data should be saved to your algorithm's dedicated input directory: ./outputs/input/{algorithm_name}/.

3. 🚀 `make_predictions.sh`

This is the main inference script that runs your model. It will be executed from within the Apptainer environment.

Input: It should read the preprocessed data from ./outputs/input/{algorithm_name}/.
Function: Execute your model's prediction command-line interface.
Output: The prediction artifacts (e.g., .cif or .pdb files) must be written to the prediction directory: ./outputs/prediction/{algorithm_name}/.

4. ✨ `postprocess.py`

This script standardizes your model's output for evaluation. It must contain a PostProcess.postprocess() method and perform two key tasks:

Generate Prediction Summary: Create a summary file named prediction_reference.csv in the evaluation directory: ./outputs/evaluation/{algorithm_name}/prediction_reference.csv. This CSV file is required for the benchmark and must include the following columns: pdb_id, seed, sample, ranking_score, and prediction_path.
Format for Evaluation: Convert your model's raw output files (located in ./outputs/prediction/{algorithm_name}/) into a format compatible with our evaluation tools (OpenStructure and DockQ).

Running the Benchmark

Once you have prepared your four files, you can test the entire workflow using our provided scripts.

Step 1: Build Environments

Navigate to the root FoldBench/ directory and run the build script. This command builds the Apptainer image for your algorithm.

cd FoldBench

# Build the Apptainer image for your algorithm
./build_apptainer_images.sh

# Create the conda environment for evaluation
conda env create -f environment.yml

Step 2: Run Inference and Evaluation

Activate the conda environment and execute the main run script. This will automate the preprocessing, prediction, postprocessing, and scoring for all registered algorithms.

conda activate foldbench
./run.sh

Once your model runs successfully, please submit a pull request to add it to our platform. We look forward to your contribution!