Installation and Compatibility Issues - Multiple Environments Tested

[arvinhm]

I'm experiencing significant installation and runtime issues with IgFold that prevent successful antibody structure prediction. Despite the promising results shown in the Nature Communications paper, the practical implementation has proven extremely challenging.

Issues Encountered:

1. Model Loading Failure:

• Models appear to load successfully but igfold.models list remains empty (length 0)
• Error: ValueError: min() arg is an empty sequence during prediction
• Debug shows: "Successfully loaded X IgFold models" but models attribute contains 0 models

2. Dependency Conflicts:

• PyTorch/torchvision version incompatibilities
• transformers version conflicts causing import errors
• pytorch-lightning dependency resolution issues

3. Configuration Errors:

• __init__() missing 1 required positional argument: 'config' when attempting manual model loading
• Missing antiberty module imports
• VGG16_Weights import errors from torchvision

Environments Tested:

• ✅ Python 3.8, 3.9
• ✅ Multiple PyTorch versions (1.7.1, 1.11.0, 2.4.0, 2.7.1)
• ✅ Fresh conda environments
• ✅ Exact versions from requirements.txt
• ✅ Official Jupyter notebook specifications
• ✅ Multiple installation methods (pip, conda, mixed)

Steps Attempted:

1. Fresh environment with exact package versions from paper
2. Using official Jupyter notebook specifications
3. Manual model loading with config fixes
4. Downgrading all dependencies to compatible versions
5. Cache clearing and complete reinstalls
6. Different Python versions and environments

System Information:

• OS: Linux (WSL2)
• Hardware: CPU-only (no GPU requirements mentioned in docs)
• Memory: Sufficient (8GB+ available)

Expected vs Actual Behavior:

Expected: Successful antibody structure prediction as demonstrated in paper
Actual: Complete failure at model loading/prediction stage despite successful import

Impact:

This significantly impacts reproducibility of the research and practical adoption of the method. While the scientific contribution is valuable, the implementation barriers prevent researchers from utilizing this tool effectively.

Suggestions:

1. Provide Docker container with working environment
2. Pin exact dependency versions that are known to work
3. Add comprehensive troubleshooting guide
4. Test installation procedures on fresh systems
5. Consider updating dependencies to maintain compatibility

Workaround:

Currently recommending users utilize AlphaFold3 server or ColabFold as alternatives for antibody structure prediction.

Would appreciate guidance on resolving these issues or official acknowledgment of compatibility problems with current dependency versions.