Add rescore_docking.py script

[Jnelen]

Add Easy-to-Use Docking Rescoring Script for AEV-PLIG

This PR introduces a simple and user-friendly docking rescoring script for AEV-PLIG, allowing users to efficiently rescore an entire docking run with minimal input.

Key Features:

• Accepts a protein file and a directory of docked ligands as input.
• Supports SDF format (recommended for best accuracy, matching reference values exactly).
• Adds Mol2 (TRIPOS) format support, which is natively handled by RDKit.
• Enables automatic conversion of PDBQT and other formats via OpenBabel, making it compatible with various docking tools.
• OpenBabel conversions, while not always perfect, have worked reliably in my testing, typically introducing only minimal deviations.

This implementation makes rescoring large docking runs effortless while maintaining flexibility in supported input formats.

Would love any feedback or suggestions!

[isakvals]

Very curious to see whether AEV-PLIG can work in this setting as it hasn't been trained on negative examples. Future work for me includes retraining AEV-PLIG on redocked structures in an attempt to improve rescoring and VS performance.

[Jnelen]

Very curious to see whether AEV-PLIG can work in this setting as it hasn't been trained on negative examples. Future work for me includes retraining AEV-PLIG on redocked structures in an attempt to improve rescoring and VS performance.

Thanks for the insight—you make a great point, and to be honest, I hadn’t fully considered the impact of AEV-PLIG not being trained on negative examples. My goal with the script was really to provide a general and flexible workflow for rescoring docking runs using AEV-PLIG, so hopefully it’s still useful for initial testing and experimentation.

Excited to see how things develop as you explore retraining—if this script can help along the way or be adapted in future iterations, I’d be glad to contribute further!

[Jnelen]

Very curious to see whether AEV-PLIG can work in this setting as it hasn't been trained on negative examples. Future work for me includes retraining AEV-PLIG on redocked structures in an attempt to improve rescoring and VS performance.

Thanks for the insight—you make a great point, and to be honest, I hadn’t fully considered the impact of AEV-PLIG not being trained on negative examples. My goal with the script was really to provide a general and flexible workflow for rescoring docking runs using AEV-PLIG, so hopefully it’s still useful for initial testing and experimentation.

Excited to see how things develop as you explore retraining—if this script can help along the way or be adapted in future iterations, I’d be glad to contribute further!

An additional feature that might be interesting in the context of using AEV-PLIG for virtual screening could be some sort of confidence or uncertainty metric. It might help users get a better feel for how reliable a given score is, especially when ranking or filtering large sets of compounds. Could be a nice complement to the raw scores. Either way, really looking forward to seeing how things evolve!