Bridging the Gap between Target-Based and Cell-Based Drug Discovery with a Graph Generative Multitask Model

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

TL;DR

This paper introduces MATIC, a novel graph-based multi-task deep learning model that bridges the gap between in vitro target inhibition and in vivo cell activity for drug discovery. By leveraging multi-task learning and a generative approach, MATIC can predict and design novel drug candidates with dual efficacy, addressing a critical bottleneck in traditional drug development.

What Problem Does This Paper Address?

Drug discovery is a vital but challenging process. A common approach is target-based screening, where researchers identify molecules that inhibit a specific disease-related protein in vitro. However, many promising in vitro candidates fail in vivo due to issues like poor cell permeability or different cellular environments. This disconnect between in vitro and in vivo efficacy leads to a high attrition rate in drug development. Phenotypic screening, which tests compounds directly in cells, can identify in vivo active compounds but often struggles with identifying the specific molecular target and can lead to off-target effects. The paper aims to bridge this gap by developing a computational method that can simultaneously predict and optimize for both target inhibition (in vitro) and cell activity (in vivo).

Methods

The core of the proposed framework is the MATIC (Multi-task predictor) model. This is a graph-based deep learning model that employs a multi-task learning architecture. It consists of:

  • Expert Networks: These include task-specific experts for 3CL inhibition and antiviral activity, as well as a shared expert network to capture common knowledge across both tasks.
  • Gating Network: This network learns to weigh the outputs of the expert networks to create a final, task-specific molecule representation.
  • Tower Network: This part of the model performs the final prediction for each task.

The model uses Graph Attention Networks (GATs) to process molecular structures and generate embeddings. To further bridge the gap, the authors developed a multi-property molecular optimization approach. This involves:

  • Monte Carlo Tree Search (MCTS): Used to extract single-property substructures (those good for either 3CL inhibition or antiviral activity) from existing molecules.
  • Multi-property Substructure Generation: Combining these single-property substructures to create novel substructures with potential for dual efficacy.
  • Graph-based Variational Autoencoder (VAE): Trained via reinforcement learning, this component takes the multi-property substructures and completes them into full molecules, guided by rewards from the MATIC predictor.

Datasets

The primary dataset used for training and evaluation was a carefully curated SARS-CoV-2 dataset. This dataset was compiled from various public databases and publications, containing compounds with validated 3CLpro inhibition (in vitro) and antiviral activity (in vivo) labels. The authors also tested their model on three independent datasets to evaluate its generalizability.

Key Insights and Ideas

  • The disconnect is real and significant: The paper visually demonstrates the gap between in vitro and in vivo efficacy, showing that compounds effective against the target protein don't always translate to cell-based activity. Traditional ADME filters are shown to be insufficient for accurately predicting in vivo effectiveness.
  • Multi-task learning is crucial: By training a single model to predict both in vitro and in vivo activity, MATIC can leverage shared information and learn complementary features that would be missed by single-task models. This allows it to identify compounds that might be overlooked otherwise.
  • Learned features reveal task-specific preferences: The model's interpretability analysis shows that the features learned for 3CL inhibition and antiviral activity are distinct, focusing on different molecular properties and atom attentions. This highlights that optimizing for one task doesn't automatically optimize for the other.
  • Generative models can create novel dual-efficacy compounds: The reinforcement learning-based generative approach, guided by the MATIC predictor, successfully produces novel molecules that are predicted to be effective in both in vitro and in vivo assays. This is a significant step towards designing drugs with desired multi-property profiles.

Key Results

  • The proposed MATIC model achieved superior performance compared to single-task models and another multi-task model (MMOE) across various metrics like AUC, Accuracy, Precision, Recall, and F1-score on the SARS-CoV-2 dataset. For example, MATIC achieved an F1-score of 0.744 for 3CL inhibition and 0.696 for antiviral activity.
  • A simulated screening using traditional ADME filters resulted in a low success rate (38 out of 167 compounds positive in vivo), with an AUC of 0.4, highlighting the limitations of this approach.
  • The generative model produced molecules with high validity (99%) and success rates (96%), meaning most generated molecules were chemically valid and predicted to have both 3CL and antiviral activity. These generated molecules also showed improved novelty (0.97) and diversity (0.88) compared to other methods.
  • Visualization of generated molecules showed that they incorporated substructures known to interact with the 3CLpro target, while also exhibiting improved properties for cell permeability (higher LogP values).

Limitations and Future Work

  • Theoretical basis for in vitro-to-in vivo translation: The paper acknowledges that the theoretical basis for why in vitro effective compounds translate to in vivo effectiveness is still obscure.
  • Complex disease modeling: For more complex diseases, modeling protein families and networks might be necessary, which is beyond the scope of this study.
  • Molecular properties as sole determinants: The authors suggest that molecular properties might not be the sole determinants of efficacy, and compensatory mechanisms within protein networks could play a role.
  • Future work: The authors plan to further explore these challenges and extend their in silico findings into theoretical frameworks and biological validation. They also aim to investigate the compensatory mechanisms among protein networks.

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 Multitask Model}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2022},
  doi = {10.1021/acs.jcim.2c01180},
}
Figure