Bioactivity Deep Learning for Complex Structure-Free Compound-Protein Interaction Prediction

Yaowen Gu, Song Xia, Qi Ouyang, Yingkai ZhangJournal of Chemical Information and Modeling 2025 · DOI

TL;DR

This paper introduces CPI2M, a large-scale benchmark dataset for compound-protein interaction (CPI) prediction, featuring activity cliff annotations and diverse bioactivity types. They also present GGAP-CPI, a novel deep learning model that leverages integrated bioactivity learning and advanced attention mechanisms to achieve state-of-the-art performance in predicting binding affinities, even in the presence of challenging activity cliffs.

What Problem Does This Paper Address?

Predicting how strongly a small molecule (compound) will bind to a protein is crucial for drug discovery, particularly in virtual screening. Traditional methods often rely on the 3D structure of the protein-ligand complex, which is not always available. Structure-free CPI methods aim to overcome this by using machine learning on compound and protein sequence/graph representations. However, existing methods face two major challenges:

  1. Data Heterogeneity: Bioactivity data comes from various assay types (e.g., IC50, Ki) with different reliability and experimental conditions, making direct comparison difficult.
  2. Activity Cliffs (ACs): These are pairs of structurally similar compounds that exhibit vastly different binding affinities. ACs pose a significant challenge for machine learning models that rely on the "chemical similarity principle," leading to inaccurate predictions and hindering model training, especially for virtual screening.

This paper addresses the need for a more comprehensive dataset and a robust model that can handle these challenges effectively.

Methods

The authors developed a novel deep learning model called GGAP-CPI with the following key components:

  • Pretrained Ligand Encoder (KANO): Leverages a knowledge graph-enhanced molecular contrastive learning approach to learn rich atom-level and molecule-level embeddings from ligand structures.
  • Pretrained Protein Encoder (ESM-2): Utilizes a protein language model (ESM-2) to generate residue embeddings, combined with a graph convolutional network (GCN) to learn protein graph representations.
  • Multihead Cross-Attention Pooling: This novel module aggregates the ligand and protein embeddings by using attention mechanisms to learn complex interactions between ligand atoms and protein residues, generating a comprehensive CPI complex representation.
  • Integrated Bioactivity Learning: The model is trained on a diverse set of bioactivity data (Ki, Ka, EC50, IC50) simultaneously, allowing it to learn from different assay types and mitigate the impact of ACs.
  • Feed-Forward Decoder: A two-layered decoder takes the aggregated complex embedding to predict the final binding affinity/bioactivity.

Datasets

  • CPI2M: A newly constructed benchmark dataset containing approximately 2 million bioactivity data points across four activity types (Ki, Ka, EC50, IC50). Crucially, it includes annotations for activity cliffs (ACs), addressing a significant gap in existing benchmarks. CPI2M is derived from the integration of EquiVS and Papyrus datasets.
  • MoleculeACE: Used for transfer learning validation.
  • CASF-2016, MerckFEP, DUD-E, DEKOIS-v2, and LIT-PCBA: Standard benchmarks used for evaluating virtual screening performance.

Key Insights and Ideas

  • Activity Cliffs are a Major Hurdle: The study strongly emphasizes that ACs significantly impact the performance of CPI models, often leading to larger prediction errors than for non-AC compounds.
  • Integrated Bioactivity Learning is Key: Training on multiple bioactivity types simultaneously, rather than relying on a single type or complex conversion, helps the model generalize better and mitigate AC effects.
  • Protein Context Matters for ACs: Incorporating protein information (via embeddings and graph representations) is crucial for better handling ACs, as it can capture binding conformation differences that are often the root cause of ACs.
  • Pretrained Models are Powerful: The use of pretrained ligand (KANO) and protein (ESM-2) encoders significantly boosts performance, highlighting the value of transfer learning from large-scale pretraining.
  • Attention Mechanisms Capture Complex Interactions: The multihead cross-attention pooling module effectively learns intricate relationships between ligand and protein features, contributing to GGAP-CPI's superior performance.

Key Results

  • State-of-the-Art Performance: GGAP-CPI consistently outperforms 12 target-specific and 7 general CPI baselines across various scenarios, including general prediction, rare protein prediction, transfer learning, and virtual screening.
  • Improved AC Prediction: GGAP-CPI shows significant improvements in AC-specific metrics (RMSEcliff) compared to baselines, demonstrating its ability to handle challenging AC cases. For example, it achieved notable improvements in RMSEcliff on Ki, Ka, and IC50 data sets.
  • Competitive Virtual Screening: GGAP-CPI demonstrates competitive ranking and scoring power on established virtual screening benchmarks like CASF-2016 and MerckFEP, and strong screening power on DUD-E, DEKOIS-v2, and LIT-PCBA.
  • Uncertainty Estimation and Interpretability: GGAP-CPI can quantify prediction uncertainty and provide structural insights by highlighting key protein pocket residues through attention visualization, enhancing its applicability in real-world drug discovery.

Limitations and Future Work

  • AC Annotation Granularity: The current CPI2M dataset primarily identifies potential ACs rather than specific AC molecule pairs, which may under-represent the nuanced effects of ACs. Future work aims to refine the benchmark by incorporating pairwise AC annotations.
  • Model Design for ACs: While GGAP-CPI mitigates AC effects, more sophisticated designs in model architecture and training strategies could further improve AC handling.
  • Interpretability for Key Substructures: Further research is needed to develop better interpretability methods for differentiating key substructures responsible for ACs.
  • Fair Model Comparisons: The authors acknowledge the ongoing challenge of achieving perfectly fair model comparisons due to differences in similarity cutoff criteria, training protocols, and evaluation pipelines across studies.

Summary updated

BibTeX
@article{gu2025bioactivity,
  author = {Yaowen Gu and Song Xia and Qi Ouyang and Yingkai Zhang},
  title = {{Bioactivity Deep Learning for Complex Structure-Free Compound-Protein Interaction Prediction}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2025},
  doi = {10.1021/acs.jcim.5c00741},
}
Figure