An open unified deep graph learning framework for discovering drug leads

Haifeng Hu, Yueming Yin, Haifeng Hu 0004, Zhen Yang, Jitao Yang, Chun Ye, Jiansheng Wu, Wilson Wen Bin GoharXiv 2023 · DOI · arXiv

Abstract

Drug discovery is traditionally a fragmented process involving hit screening, property prediction, and lead optimization. This paper introduces GAFSE (Generative Adversarial Feature Subspace Enhancement), a unified deep graph learning framework that bridges these stages. By leveraging adversarial learning and graph reconstruction, GAFSE improves prediction accuracy and provides actionable insights for molecular optimization.

Introduction: The Fragmentation of Drug Discovery

The transition from identifying a biological target to developing a viable drug candidate is often hindered by a "clumsy" relay race between three distinct stages: hit screening, molecular property prediction (ADMET), and lead optimization. Currently, these stages rely on disparate models with incompatible architectures and data formats, leading to high research overhead. GAFSE addresses this by treating molecules as graphs within a single, consistent mathematical framework.

The GAFSE framework unifies hit discovery, lead optimization, and property analysis into a single pipeline.

The GAFSE Framework

The GAFSE framework utilizes Deep Graph Learning (DGL) to represent molecules as graphs where atoms are nodes and bonds are edges. Its architecture is built on three pillars:

  1. Adversarial Feature Subspace Enhancement (AFSE): This component injects adversarial perturbations into the molecular embedding space. This forces the model to learn "smoother," more generalizable representations that perform better on unseen chemical space. The authors provide a theoretical convergence guarantee, proving that this approach is Lipschitz continuous.
  2. Attentive Graph Reconstruction Networks (AGRNs): A novel decoder that uses graph attention to reconstruct molecules from embeddings, identifying key atomic positions that influence bioactivity.
  3. Unified Loss Function: The model optimizes for classification/regression, reconstruction, and adversarial validity simultaneously.
Detailed architecture of GAFSE showing the interaction between the embedding model, adversarial enhancement, and reconstruction networks.

Experimental Setup and Datasets

The framework was benchmarked against industry standards:

  • GPCR Benchmark: 33 tasks for hit screening and bioactivity regression.
  • ADMET Benchmark: 25 classification tasks covering Absorption, Distribution, Metabolism, Excretion, and Toxicity.
  • AID1706 (PubChem): High-throughput screening for SARS-CoV-2 3CLPro inhibitors.

Key Results and Insights

GAFSE consistently outperformed state-of-the-art (SOTA) models:

  • Screening: Improved Enrichment Factor (EF10%) by 1.29% and ranking index ($\tau_B$) by 8.07% on the GPCR benchmark.
  • Property Prediction: Outperformed ADMETlab 2.0, increasing AUC by 1.73% and MCC by 10.38%.
  • Optimization: Achieved a 98.5% average score in validity, uniqueness, and novelty for COVID-19 inhibitors.

A significant advantage of GAFSE is its ability to identify "Activity Cliffs"—small structural changes that cause large shifts in biological activity. By generating Matched Molecular Pairs (MMPs), it provides a clear roadmap for medicinal chemists to optimize leads.

Conclusion and Future Directions

GAFSE represents a major step toward a unified drug discovery pipeline. Future work will focus on pre-training on larger molecular databases, evolving the framework to handle functional group (substructure) modifications, and refining multi-objective optimization to balance complex drug requirements.

Summary updated

Abstract

Computational discovery of ideal lead compounds is a critical process for modern drug discovery. It comprises multiple stages: hit screening, molecular property prediction, and molecule optimization. Current efforts are disparate, involving the establishment of models for each stage, followed by multi-stage multi-model integration. However, this is non-ideal, as clumsy integration of incompatible models increases research overheads, and may even reduce success rates in drug discovery. Facilitating compatibilities requires establishing inherent model consistencies across lead discovery stages. Towards that effect, we propose an open deep graph learning (DGL) based pipeline: generative adversarial feature subspace enhancement (GAFSE), which first unifies the modeling of these stages into one learning framework. GAFSE also offers standardized modular design and streamlined interfaces for future expansions and community support. GAFSE combines adversarial/generative learning, graph attention network, graph reconstruction network, and optimizes the classification/regression loss, adversarial/generative loss, and reconstruction loss simultaneously. Convergence analysis theoretically guarantees model generalization performance. Exhaustive benchmarking demonstrates that the GAFSE pipeline achieves excellent performance across almost all lead discovery stages, while also providing valuable model interpretability. Hence, we believe this tool will enhance the efficiency and productivity of drug discovery researchers.
BibTeX
@misc{hu2023open,
  author = {Haifeng Hu and Yueming Yin and Haifeng Hu and Zhen Yang and Jitao Yang and Chun Ye and Jiansheng Wu and Wilson Wen Bin Goh},
  title = {{An open unified deep graph learning framework for discovering drug leads}},
  year = {2023},
  eprint = {2301.03424},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2301.03424},
}
Figure