Using deeprank2 without training

[Verdandy]

Hi, I wonder if there are any pre-trained model which can be use to score the AF predicted structures?

[gcroci2]

Currently, there isn’t a general pre-trained model available for scoring AlphaFold (AF) predicted structures. Most models for structure evaluation are highly specialized for specific use cases, and there are no immediate plans to develop a universal foundational model that can assess any given structure.

For your use case, I’d recommend identifying a dataset that includes both correct structures and incorrect ones with varying docking scores. This will help the model learn to distinguish good predictions from poor ones. DeepRank2 can then be used to train on these structures and applied during inference to score AF-predicted ones.

Alternatively, you could rely on more classical software like HADDOCK to generate docking models and compare them with AlphaFold-generated structures. This comparison could provide additional insights into the reliability of AF predictions in your specific context.