matthewtwarren / aescore Goto Github PK

Add functionality to fine tune a model pretrained on a separate dataset.

Add option to use only protein-ligand atom pairs/triplets when computing radial/angular terms of AEVs. Requires modification of TorchANI module or moving TorchANI classes and functions into AEScore.

When using a trained model to perform inference there is no option to use scaling for labels. If scaling has been used during training, this will result in inconsistencies between the predicted outputs and the actual label values.

Could be resolved by storing a scaler object during training which can then be loaded and applied during inference.

Using consensus scoring (-c argument), the predictions (training and validation) are uncorrelated and with true values for all models other than the first, despite the first model performing well. In contrast, the training and validation losses for all repeats are similar.

e.g. training results
correlation between predicted_0 and true = 0.95, rmse = 0.69
correlation between predicted_1 and true = 0.008, rmse = 2.58