Protein Structure Prediction

Overview

An Open OnDemand Batch Connect app that provides a web-based interface for running protein structure prediction jobs using AlphaFold 2 and AlphaFold 3 on the ICDS Roar cluster. The app simplifies the process of submitting and monitoring AlphaFold jobs by providing a user-friendly interface and automated job management.

This app uses the Batch Connect basic template with Slurm. It executes a two-phase workflow: a CPU phase for MSA generation and a GPU phase for structure prediction, with the GPU job submitted as a dependency of the CPU job.

• Upstream project: AlphaFold by DeepMind
• Batch Connect template: basic
• Scheduler: Slurm
• Container runtime: Singularity

How it looks

Launch form

Model selection, partition & working directory (left)	JSON input & terms of service (right)

Progress after job submission

Supporting Materials

Conference Materials

• GOOD25 Conference Talk Abstract
• GOOD25 Conference Presentation

News Articles

• Leveraging AlphaFold in Graduate Research
• OSC News: Inaugural GOOD Conference Draws Strong Attendance from 10 Countries

> Presented as a talk at the Global Open On Demand Conference 2025, Harvard University
> Date: March 19, 2025, 4:00 PM – 4:25 PM (25 min)
> Title: AlphaFold accessibility: an optimized open-source OOD app for Protein Structure Prediction
> Speakers: Vinay Saji Mathew [Pennsylvania State University] , William Lai [Cornell], Matt Hansen [Pennsylvania State University] > > Track: Application Track [featuring AI OnDemand]
> Location: Tsai Auditorium (CGIS S010)

Features

Multiple Prediction Engines

• AlphaFold 2:

  • Supports AlphaFold v2.3.2 for protein structure prediction
  • Handles both monomer and multimer predictions
  • Uses full database configuration for maximum accuracy
  • Automated MSA generation and template search

• AlphaFold 3 (New!):

  • Latest version of AlphaFold with improved accuracy
  • Supports protein-protein, protein-DNA/RNA, and protein-ligand complexes
  • Enhanced diffusion-based structure prediction
  • Requires acceptance of Google's terms of service

Job Management

• Two-phase execution:
  • CPU phase for MSA/templates
  • GPU phase for prediction (set as a dependency)
• Real-time job status monitoring
• Detailed progress tracking
• Automatic error handling and recovery

User Interface

• Flexible Input Formats:
  • FASTA sequence input for AlphaFold 2
  • JSON format input for AlphaFold 3 (following official specifications)
• GPU allocation selection
• Working directory customization
• Real-time progress visualization
• Direct access to output files

Output Files

• AlphaFold 2:

  • PDB structure files (ranked by confidence)
  • Multiple Sequence Alignment (MSA) files
  • Detailed prediction metrics and confidence scores
  • Comprehensive log files

• AlphaFold 3:

  • CIF structure files
  • Ranking scores for multiple predictions
  • Detailed model outputs and metrics
  • Complete execution logs

Prerequisites

Open OnDemand

• Slurm scheduler
• Has been tested to work with OOD v3 & v4

Database Setup

Both AlphaFold versions require genetic databases that must be set up before using the app:

• AlphaFold 2: Download using script from AlphaFold 2 repository
• AlphaFold 3: Additional databases required. Setup instructions available here

Singularity Containers

The app uses Singularity containers for execution:

• AlphaFold 2: Download from Sylabs
• AlphaFold 3: Requires official container from Google (subject to terms of use). Weights needed for running AlphaFold 3 have to be requested from Google here

Installation

1. Clone this repository into your Open OnDemand apps directory
2. Configure paths in template/alphafold_env.sh
3. Ensure all required databases are properly set up
4. Verify GPU compute capabilities.

Configure for your site

Edit form.yml.erb and update these values for your cluster:

In before.sh.erb, the app sources alphafold_env.sh to set environment variables for database paths, container paths, and working directories. You must configure this file for your site.

Configuration

form.yml attributes

Usage

1. Access the Open OnDemand dashboard
2. Navigate to "Interactive Apps"
3. Select "Protein Structure Prediction"
4. Choose prediction engine (AlphaFold 2 or 3)
5. Fill out the form:
   • For AlphaFold 2: Enter protein sequence in FASTA format
   • For AlphaFold 3: Provide input in JSON format
   • Select GPU allocation
   • Choose working directory
6. Accept terms of service (required for AlphaFold 3)
7. Submit the job

Input Format Examples

AlphaFold 2 (FASTA)

The app accepts protein sequences in FASTA format.

Example:

>sequence_name
MVKVGVNGFGRIGRLVTRAAFNSGKVDIVAINDPFIDLNYMVYMFQYDSTHGKFHGTVKA
ENGKLVINGNPITIFQERDPSKIKWGDAGAEYVVESTGVFTTMEKAGAHLQGGAKRVIIS

AlphaFold 3 (JSON)

{
"name": "example_complex",
"sequences": [
{
"protein": {
"id": "protein_chain_A",
"sequence": "MVKVGVNG..."
}
}
],
"modelSeeds": [1, 2, 3]
}

Output Files

The app generates the following output structure:

working_directory/
└── run_YYYYMMDD_HHMMSS/
├── input/
│ ├── [structure files] # Predicted structures
│ ├── [prediction data] # Detailed predictions
│ └── msas/ # Multiple sequence alignments
├── logs/ # Job logs
├── CPU-SLURM/ # CPU phase files
└── GPU-SLURM/ # GPU phase files

Monitoring Jobs

The app provides real-time monitoring of:

• MSA generation progress
• Template search status
• Structure prediction progress
• Model relaxation status

Troubleshooting

Common issues and solutions:

1. Job fails in CPU phase:

   • Check available disk space
   • Verify database paths
   • Examine CPU phase logs

2. GPU phase errors:

   • Verify GPU allocation
   • Check memory requirements
   • Review GPU phase logs
   • For AlphaFold 3: Ensure GPU compute availability.

Contributing

For bugs or feature requests, open an issue.

References

• AlphaFold 2 -- protein structure prediction by DeepMind
• AlphaFold 3 -- latest version with protein-ligand complex support
• Open OnDemand -- the HPC portal framework
• OOD Batch Connect app development docs

License

MIT (see LICENSE file)

Acknowledgements

• AlphaFold by DeepMind Technologies Limited
• Singularity container by prehensilecode
• The research project is generously funded by Cornell University BRC Epigenomics Core Facility (RRID:SCR_021287), Penn State Institute for Computational and Data Sciences (RRID:SCR_025154) , Penn State University Center for Applications of Artificial Intelligence and Machine Learning to Industry Core Facility (AIMI) (RRID:SCR_022867) and supported by a gift to AIMI research from Dell Technologies.
• Computational support was provided by NSF ACCESS to William KM Lai and Gretta Kellogg through BIO230041

Contact

For questions or issues, please contact:

• Technical support: vinaysmathew@psu.edu
• ICDS support: icds@psu.edu