LP-PDBBind: Developing More Generalizable Scoring Functions with Better Utilization of the PDBBind Dataset

This repository contains all the code for creating LP-PDBBind of PDBBind dataset, building the BDB2020+ dataset, prepared dataset files, and scripts for retraining AutoDock vina, IGN, RFScore and DeepDTA models

Authors

• Jie Li jerry-li1996@berkeley.edu
• Xingyi Guan nancy_guan@berkeley.edu
• Oufan Zhang oz57@berkeley.edu
• Kunyang Sun kysun@berkeley.edu
• Yingze Wang ericwangyz@berkeley.edu
• Dorian Bagni dorianbagni@berkeley.edu
• Teresa Head-Gordon thg@berkeley.edu

Results summary

BDB2020+ benchmark

mPro benchmark

EGFR benchmark

Compiled datasets

LP-PDBBind

The LeakProof of PDBBind 2020 (LP-PDBBind) is given in dataset/LP_PDBBind.csv. The new_split column corresponds to which category the data belongs to in LP-PDBBind. Additionally, CL1, CL2, CL3 and covalent are boolean columns indicating whether the data is in the corresponding clean levels, and whether the data is covalent or not. Therefore, the following python code reads in the dataset and selects all data in the train set, and satisfy CL1 and non-covalent:

import pandas as pd
df = pd.read_csv('dataset/LP_PDBBind.csv', index_col=0)
df_train = df[(df['new_split'] == 'train') & df.CL1 & ~df.covalent]

Only the meta-information (including pdbid, category, clean levels, whether is covalent, ligand SMILES strings, protein sequence, binding affinity values, etc.) are present in the csv file. Users need to download the prepared structure files (.pdb format for the proteins and .sdf/.mol2 format for the ligands) from the PDBBind website.

BDB2020+

The compressed BDB2020+ dataset (dataset/BDB2020+.tgz) can be decompressed by:

tar -xzf dataset/BDB2020+.tgz

which will create the BDB2020+ folder. Inside the folder there will be BDB2020+.csv and the dataset folder. The BDB2020+.csv contains the meta-information of the dataset, and the dataset folder contains the prepared structure files for each protein-ligand complex.

SARS-COV-2 main protease dataset

The compressed SARS-COV-2 main protease dataset (dataset/mpro.tgz) can be decompressed by:

tar -xzf dataset/mpro.tgz

which will create the mpro folder. Inside the folder there will be info.csv and multiple folders named by the pdbids. The info.csv contains the meta-information of the dataset, and the other folders contain the prepared structure files for the protein-ligand complexes.

EGFR

The compressed EGFR dataset (dataset/EGFR.tgz) can be decompressed by:

tar -xzf dataset/EGFR.tgz

which will create the EGFR folder. Inside the folder there will be info.csv and three folders protein, ligand and ligand_addH. The info.csv contains the meta-information of the dataset, and the other folders contain the prepared structure files for the protein-ligand complexes.

Dataset creation scripts

Inside the dataset_creation folder there are multiple jupyter notebooks for the creation of LP-PDBBind and BDB2020+ dataset.

Required packages

• pandas
• rdkit
• biopython

Notebook descriptions

• prepare_info.ipynb: Preparing the basic PDBBind dataset information. It downloads the PDBBind PDB files directly from RCSB to retain the header information, and extracts category, release year and resolution from the PDB files. Additionally, it extracts the SMILES strings for the ligands and the sequence of the proteins from the PDB files.

• calc_similarities.ipynb: Extracting protein types from the categories, calculate pairwise ligand fingerprint similarities and pairwise protein sequence similarities.

• create_splitting.ipynb: Creating the LP-PDBBind splitting of PDBBind dataset based on protein and sequence similarity.

• CL_filters: Scripts for defining the three clean levels (CL1-CL3), and finding out covalent binders in PDBBind.

• prepare_BDB2020+_dataset.ipynb: Script for cleaning up the raw matched data in BindingDB and prepare the BDB2020+ dataset.

Model retraining

Please refer to the model_retraining folder for the scripts for retraining and retrained model weights for AutoDock vina, IGN, RFScore and DeepDTA.