Bridging the gap between target-based and cell-based drug discovery with a graph generative multi-task model

Fan Hu, Dongqi Wang, Huazhen Huang, Yishen Hu, Peng Yin

Introduction

A significant hurdle in drug discovery is the discrepancy between a compound's efficacy against a specific protein target (in vitro) and its actual performance within a living cell (in vivo). Many "hits" identified through target-based screening fail in later stages because they cannot effectively navigate the complex cellular environment, despite passing traditional ADME (Absorption, Distribution, Metabolism, and Excretion) filters. This paper introduces MATIC, a novel AI framework designed to bridge this gap by simultaneously predicting target inhibition and cellular activity.

The Challenge of the "In Vitro to In Vivo" Gap

Traditional drug discovery often begins with target-based screening to find molecules that bind to disease-related proteins, such as the SARS-CoV-2 3CL protease. However, these molecules frequently fail to show antiviral activity in cells. The authors demonstrate that standard ADME filters are often insufficient for predicting this transition; in their study, a standard filter achieved an AUC of only 0.40, performing worse than random chance. There is a critical need for computational methods that can integrate both protein binding and cellular efficacy during the design phase.

Methods: The MATIC Framework

The authors developed a two-part framework to address this challenge:

  1. MATIC (Multi-task Predictor): This component utilizes a multi-task architecture based on Graph Attention Networks (GAT). It incorporates "shared" and "task-specific" experts, with a gating mechanism that dynamically weighs information for two distinct tasks: 3CL inhibition (target-based) and antiviral activity (cell-based).
  2. Multi-property Generator: To design novel compounds, the framework employs Monte Carlo Tree Search (MCTS) to identify rational substructures contributing to dual efficacy. These fragments are then processed by a Variational Autoencoder (VAE), fine-tuned via Reinforcement Learning (RL) to generate complete molecules optimized for both properties.
Overview of the MATIC framework, illustrating the multi-task predictor and multi-property generator architectures.

Datasets and Training

The model was trained on a curated SARS-CoV-2 dataset containing 7,458 compounds with validated $IC_{50}$ or $EC_{50}$ values. The generative model was pre-trained on the ChEMBL database to learn fundamental chemical rules before being specialized for dual-efficacy drug design. Performance was further validated using three independent external test sets.

Key Insights

  • Synergy in Multi-Tasking: Training on both target inhibition and cellular activity simultaneously improved performance on each individual task. The shared architecture allows the model to leverage "hidden" correlations between protein binding and cellular penetration.
  • Interpretability through Atom Attention: Visualization of the model's attention mechanisms revealed that it prioritizes different molecular features for different tasks. For target inhibition, it focuses on reactive "warheads," while for cellular activity, it emphasizes hydrophobic regions necessary for membrane permeability.
  • Global Structural Understanding: Unlike models relying solely on local chemical similarity, MATIC uses global message passing to understand how the entire molecular structure contributes to its biological function.

Results

MATIC demonstrated superior predictive performance compared to single-task models and standard multi-task architectures like MMOE. Specifically, it achieved an AUC of 0.88 for target inhibition and 0.832 for antiviral activity.

Performance comparison of MATIC against single-task and standard multi-task models across various metrics.

The generative component also proved highly effective, with a 96% success rate in producing "dual-positive" molecules. These generated compounds maintained high novelty (0.97) and diversity (0.88). In a practical application, the team screened 500,000 commercial compounds, identifying 331 high-probability candidates for further validation.

Conclusion and Future Directions

While MATIC successfully identifies compounds with dual efficacy, the precise biological mechanisms behind why certain target-effective drugs fail in cells remain partially opaque. Future research will focus on modeling entire protein networks and integrating more complex biological data to move beyond "in silico" predictions toward a more comprehensive theoretical understanding of cellular drug behavior.

Summary updated

BibTeX
@article{hu2022bridging,
  author = {Fan Hu and Dongqi Wang and Huazhen Huang and Yishen Hu and Peng Yin},
  title = {{Bridging the gap between target-based and cell-based drug discovery with a graph generative multi-task model}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2022},
  doi = {10.1021/acs.jcim.2c01180},
}
Figure