Persistent spectral based ensemble learning (PerSpect-EL) for protein-protein binding affinity prediction
This manual is for the code implementation of paper "Persistent spectral based ensemble learning (PerSpect-EL) for protein-protein binding affinity prediction".
Platform: Python>=3.6, MATLAB 2016B
Python Packages needed: math, numpy>=1.19.5, scipy>=1.4.1, scikit-learn>=0.20.3, GUDHI 3.0.0
Before the representation, the atom coordinates from each protein-protein complex needs to extracted accordingly to its element specific combinations e.g. mut_binding_C_C.pts. The atom coordinates in *.pts can be generated from the original TopNetTree code 10.24433/CO.0537487.v1.
For each protein-protein complex, the relevant element specific atom groups are used to construct the simplicial complexes to generate the Hodge Laplacians.
def hodge(args):
# this function generates the eigenvalues from the L0 hodge laplacians in wild and mutation types from Vietoris Rips Complexes.
# parameter "args" is foldername format (PDBID, chainid, wildtype, resid, mutanttype) e.g. (1AK4 D A 488 G)
# Runs the "binding_hodge.m" and "mutation_hodge.m" to generate the hodge laplacians for each complex.
def alpha(args):
# this function generates the eigenvalues from the L1 and L2 hodge laplacians in wild and mutation types from alpha complexes.
# parameter "args" is foldername format (PDBID, chainid, wildtype, resid, mutanttype) e.g. (1AK4 D A 488 G)
# Outputs numpy array of eigenvalues which can be used to generate persistent attributes.
def read_eig(args):
# this function takes the eigenvalues from "alpha" and "hodge" functions and compute the persistent spectral attributes.
# parameter "args" is foldername format (PDBID, chainid, wildtype, resid, mutanttype) e.g. (1AK4 D A 488 G)
# Outputs the persistent attributes in a numpy array after passing through the subfunction "compute_stat"
def compute_stat(dist):
# this function takes a distribution of eigenvalues of a particular filtration parameter and computes the persistent spectral attributes.
# Outputs the persistent attributes of that filtration parameter. The above functions can be found in feature/skempi_feature.py and feature/run_all.py.
binding_hodge.m --> Computes L0 Hodge Laplacian by constructing Vietoris Rips Complex from the atom coordinates between the binding sites.
mutation_hodge.m --> Computes L0 Hodge Laplacian by constructing Vietoris Rips Complex from the atom coordinates between the mutation site and its neighborhood.
computeVRcomplex.m --> Constructs the VR complex from atom coordinates.Each persistent attribute is put into a base learner (1D-CNN) for training.
spectcnn_ab_prelu_v6 --> Base Learner for AB-Bind S645 (with non-binders)
spectcnn_ab_prelu_Nout_v6 --> Base Learner for AB-Bind S645 (without non-binders)
spectcnn_skempi_prelu_v6 --> Base Learner for SKEMPI 1131
spectcnn_homology_test --> Base Learner for AB-Bind S645 Blind Homology Test Prediction
The outputs from Base Learners are then combined and input into the Meta Learner for final prediction.
spectnettree_ab_prelu_v6 --> Meta Learner for AB-Bind S645 (with non-binders)
spectnettree_ab_prelu_Nout_v6 --> Meta Learner for AB-Bind S645 (without non-binders)
spectnettree_skempi_prelu_v6 --> Meta Learner for SKEMPI 1131
spectnettree_homology_test --> Meta Learner for AB-Bind S645 Blind Homology Test Prediction
Software: Jackal, PDB2PQR, SPIDER2, MIBPB
Relevant code and software can be found in https://doi.org/10.24433/CO.0537487.v1.
The auxiliary features for AB-Bind S645 is in models/X_ab_aux.npy.
The auxiliary features for SKEMPI-1131 is in models/X_skempi_aux.npy.
If you use this code in your research, please cite our paper:
- JunJie Wee and Kelin Xia, "Persistent spectral based ensemble learning (PerSpect-EL) for protein-protein binding affinity prediction", Briefings In Bioinformatics, bbac024 (2022)