Troubleshooting

​

Problem

​

Solutions

1. Increase Docker resources: Ensure Docker has sufficient memory (at least 16GB) and CPU cores allocated
2. Check disk space: The build requires significant disk space (50GB+)
3. Use a stable network connection: Large dependencies need to be downloaded
4. Clear Docker cache if you encounter corruption:
   docker system prune -a
   

​

Problem

​

Solutions

1. Ensure you have requested and received access via the official form
2. Verify you received the download link directly from Google
3. Check that the model parameters directory is correctly mounted in Docker
4. Confirm file permissions allow read access

​

Problem

​

Solutions

1. Ensure you have enough disk space (several TB required for full databases)
2. Verify network connectivity to database sources
3. Check the database directory structure matches the expected format
4. Use --run_data_pipeline=false to skip database searches during testing

​

Problem

​

Solutions

1. Reduce complex sizes: Try simpler or smaller protein complexes first
2. Check GPU availability: Ensure --gpus all flag is set
3. Monitor GPU memory: Use nvidia-smi to check available memory
4. Use appropriate hardware: AlphaFold 3 requires GPUs with at least 16GB VRAM
5. Close other GPU processes: Ensure no other programs are using the GPU

​

Problem

​

Solutions

export XLA_FLAGS="--xla_disable_hlo_passes=custom-kernel-fusion-rewriter"

​

Problem

​

Solutions

1. Validate JSON syntax: Use a JSON validator to check for syntax errors
2. Check required fields: Ensure name, sequences, dialect, and version are present
3. Verify sequence format: Check protein sequences use valid amino acid codes
4. Review ligand definitions: Ensure SMILES strings are properly formatted
5. Check bonded atom pairs: Verify indices match the actual atoms

{
  "name": "my_protein",
  "sequences": [
    {
      "protein": {
        "id": ["A"],
        "sequence": "MKFLKFSLLTAVLLSVVFAFSSCGDDDDTYPYDVPDYASL"
      }
    }
  ],
  "modelSeeds": [1],
  "dialect": "alphafold3",
  "version": 1
}

​

Problem

​

Solutions

1. This is expected behavior: Data pipeline can take hours or even days for complex targets
2. Run on CPU-only machine: Use --run_data_pipeline=true --run_inference=false on a separate machine
3. Check database accessibility: Ensure all database files are readable
4. Monitor progress: Check log files for activity
5. Skip for testing: Use --run_data_pipeline=false and provide MSAs manually

​

Problem

​

Solutions

1. Check the output directory: Verify the path is correctly mounted
2. Review error logs: Look for error messages in the console output
3. Verify disk space: Ensure sufficient space for output files
4. Check permissions: Output directory must be writable

​

Problem

​

Solutions

1. Check input quality: Verify sequences are correct and complete
2. Review confidence metrics: Low pLDDT scores indicate uncertain regions
3. Inspect ranking scores: Compare multiple model seeds
4. Validate ligand geometry: Use tools like RDKit to check SMILES
5. Try different model seeds: Generate multiple predictions with different seeds
6. Review covalent modifications: Ensure bonded atom pairs are correctly specified

​

Problem

​

Solutions

1. Check SMILES definition: Ensure stereochemistry is explicitly defined
2. Verify CCD codes: Use correct Chemical Component Dictionary codes
3. Review bonded atom pairs: Double-check atom indices
4. Consult glycan modeling guide: See the paper on modeling glycans with AlphaFold 3

​

Problem

​

Solutions

1. Use absolute paths: Always use full paths, not relative paths
2. Check Docker permissions: Ensure Docker has access to the directories
3. Verify paths exist: Create directories before mounting
4. Check SELinux/AppArmor: May need to adjust security settings on Linux

docker run -it \
    --volume $HOME/af_input:/root/af_input \
    --volume $HOME/af_output:/root/af_output \
    --volume $HOME/models:/root/models \
    --volume $HOME/databases:/root/public_databases \
    --gpus all \
    alphafold3 \
    python run_alphafold.py \
    --json_path=/root/af_input/fold_input.json \
    --model_dir=/root/models \
    --output_dir=/root/af_output

​

Problem

​

Solutions

1. Install nvidia-docker2: Required for GPU access
2. Use —gpus all flag: Must be included in docker run command
3. Check NVIDIA drivers: Ensure host has proper NVIDIA drivers installed
4. Test GPU access: Run nvidia-smi inside the container
5. Restart Docker daemon: Sometimes required after installing nvidia-docker2