Updates.txt

6/28/2022
PepSep models (PepSep1.hdf5, PepSep6.hdf5) are replaced with newer versions. Old versions are renamed to PepSep1_v1.hdf5 and PepSep6_v1.hdf5, respectively.

Changes in the models:
1. We increased perturbation of peptide ligands c2 used for the training; the average RMSD is now 1.07 Å (instead of 0.12 Å in the first version). Perturbations were done by changing φ- and ψ-backbone angles and translation of coordinates within [-0.7, 0.7] Å along x, y, z directions.

2. Complexes with peptide ligands having polar residues only as hot-spots were removed (e.g. if a native peptide ligand has two hot-spots which are ASP and GLU): the previous model incorporated too many buried polar residues into designs, so we did an attempt to reduce their rates. Number of complexes removed from the training, validation, and test sets are 1629 (1.7%), 6 (0.2%) and 16 (1.3%), respectively.

3. The attention block of the model was replaced by a bidirectional attention block.

4. Class- and sample- weighting schemes were changed. Regularization of LSTM layers was added.

5. We found out that the changes from the point above (4) and some other changes in the architecture improved performance of the main block and the reverse prediction block became redundant. Therefore, we omitted the reverse prediction block during training of the new model.

Databases
A File "database_complexes.json" containing lists of complexes used for training and testing of the model is added. The file consists of python dictionaries which names specify a dataset (training/validation/test/benchmark). Names of complexes consists of name of crystal structure, label of chain with protein ligand, number of the first residue of 6-residue protein ligand fragment and label of chain with a binding site.


7/27/2022
iNNterfaceDesign/iNNterfaceDesign_scripts/4.amn_sampling.py is edited: an error in postprocessing of PepSeP6 outputs is fixed.

10/4/2022
A version of PepSep6 trained on hetero-oligomeric PPIs in the last 5 steps of optimization is added as PepSep6_v2.hdf5. It performs better on hetero-oligomeric PPIs than other versions of the model.
PepBB_C models are replaced with newer versions.