This PR adds scripts and benchmarks for predicting binding affinities from PDBbind using interaction fingerprints (via RdkitGridFeaturizer) and random forests / fully-connected networks.
Notes on the benchmark:
The benchmarking runs are pretty slow because of the large fingerprint sizes (easy to get OOM errors here) and the combinatorics of different things to try, so I changed some inputs to give smaller fingerprint vectors and tested all combinations of R^2 and RMSE metrics, RF vs fully-connected network models, and Voxelized ECFP vs Ligand ECFP fingerprints.
Unsurprisingly, optimized AtomicConvs outperform all of the above combinations. Optimizing the ECFP settings could almost certainly boost performance further, but I stuck with default settings.
This PR adds scripts and benchmarks for predicting binding affinities from PDBbind using interaction fingerprints (via RdkitGridFeaturizer) and random forests / fully-connected networks.
Notes on the benchmark:
CCing @rbharath @mufeili