Overview

This project is based on our paper "Mitigating Activity Cliff-induced Discrepancies in Deep Learning of Compound-Protein Interaction Affinity Prediction". GGAP-CPI stands for protein Graph and ligand Graph network with Attention Pooling for Compound-Protein Interaction prediction.

Table of Contents

• Introduction
• Dataset
• Model Training
• Model Inference
• Benchmark Results
• Use Your Own Data
• Citation

Introduction

Machine learning-based protein-ligand binding affinity prediction is crucial for drug virtual screening. Structure-free Compound-Protein Interaction (CPI) prediction methods leverage millions of bioassay measurements, offering greater flexibility than crystal structure-dependent methods. However, activity cliffs—minor chemical modifications leading to substantial bioactivity changes—pose significant challenges that are not well-explored in CPI prediction.

In this study, we present CPI2M, a large-scale CPI benchmark dataset containing approximately 2 million endpoints across four activity types (Ki, Kd, EC50, and IC50), specifically annotated for activity cliff data. Additionally, we developed GGAP-CPI, a deep learning model that mitigates the impact of activity cliffs through protein feature fusion and activity type-based transfer learning.

Our evaluation across three simulated scenarios (general, unknown proteins, and transfer learning) shows that GGAP-CPI outperforms 12 target-specific and 7 general CPI methods, excelling in distinguishing bioactivity differences among activity cliff molecules. It also demonstrates comparable performance on CASF-2016 and LIT-PCBA benchmarks, highlighting its potential for practical virtual screening.

This study proposes effective strategies for investigating activity cliffs in CPI predictions, enhancing our understanding of structure-activity relationship discontinuities and aiding future CPI methodology development.

Dataset

Dataset Summary

Dataset	Activity Type	Num.	Num. Mol.	Num. Prot.	Avg. Bioactivity	Std. Bioactivity	% AC
CPI2M-main (train, internal validation)	Ki	341,244	124,345	418	6.50	1.42	25.39
	Kd	4,337	3,212	21	6.90	1.60	34.03
	EC50	88,302	61,095	178	5.80	1.56	25.08
	IC50	751,941	419,985	1115	6.15	1.47	30.60
CPI2M-few (external validation)	Ki	65,529	41,365	2373	6.20	1.60	-
	Kd	55,017	14,667	1564	5.79	1.35	-
	EC50	42,301	28,818	1506	6.00	1.48	-
	IC50	148,929	94,883	4562	5.69	1.43	-

We also incorporate MoleculeACE for activity cliff estimation, CASF-2016 and LIT-PCBA for virtual screening estimation. We have organized these datasets and all of them can be downloaded by:

Description of CPI2M

• Source: EquiVS (ChEMBL29, BindingDB, PubChem, Probe&Drugs, IUPHAR/BPS), and Papyrus (ChEMBL30, EXCAPE, literature)
• Structure: CPI2M-main for model training and internal evaluating, CPI2M-few for external evaluating.
• Preprocessing: Including multistep filtering and duplicate cleaning for activity value, unit, ligand, and protein data.

Model

Model training

Available training dataset includes: CPI2M-main (noted as 'ki', 'kd', 'ec50', 'ic50', 'integrated'), MoleculeACE ('MolACE_CPI_ki', 'MolACE_CPI_ec50'), and PDBbind ('PDBbind_all', only for reproducing GGAP-CPI-ft on CASF-2016).

Please run the following command for model training:

sh run_bash/run_CPI.sh KANO_Prot {DATA_NAME} train {SEED}

parameters include: 1. training dataset; 2. mode (e.g., train); 3. random seed.

Model Inference

We provide pretrained GGAP-CPI model on CPI2M-main dataset with different activity type predictor(GGAP-CPI-pKi, -pKd, -pEC50, -pIC50, and -pAC). We generally recommand you to use GGAP-CPI-pKi, GGAP-CPI-pIC50 or GGAP-CPI-pAC (AC means pretrained on "integrated" activity data) for inferencing on your own data.

Available inference dataset includes: CPI2M-few ('ki_last', 'kd_last', 'ec50_last', 'ic50_last'), CASF-2016 ('PDBbind_CASF'), and LIT-PCBA ('UNIPROT_ID_LITPCBA').

Taking GGAP-CPI-pIC50 for example, please run the following command for model inference on MoleculeACE-pEC50 dataset:

sh run_bash/run_CPI.sh KANO_Prot MolACE_CPI_ec50 inference {SEED} ic50/2

parameters include: 1. inference dataset; 2. mode (e.g., inference, finetune); 3. random seed; 4. pretrained model path (only for KANO_Prot);

Benchmark Results

The performances of GGAP-CPI and 19 baseline methods are evaluated on CPI2M-main, CPI2M-few, MoleculeACE, CASF-2016, and LIT-PCBA datasets. For your convience, we add the benchmarking result files for each of them in "benchmark_result" folder.
Also, to reproduce results for GGAP-CPI, you can run the code:

Please note that there should be mild performance differences with different devices and package versions.

Use Your Own Data

To train GGAP-CPI and other baseline models on your own .CSV data, which should at least include columns ['smiles', 'Uniprot_id', 'label']. Please run the following commands:

# preprocess data
python process_data.py --dataset {DATA} --task {CPI or QSAR} --split {random or ac} --train_ratio {RATIO} --seed {SEED}
# train
sh run_bash/run_CPI.sh {MODEL} {DATA_NAME} train {SEED}
# optional: finetune, inference, ...

GGAP-CPI is also applicable for classification tasks such as binder/nonbinder classification and drug-target interaction prediction. Please replace run_CPI.sh by run_CPI_cls.sh for model training and testing.

Citation

TBD