MolDesigner: Interactive Design of Efficacious Drugs with Deep Learning

Kexin Huang, Tianfan Fu, Dawood Khan, Ali Abid, Ali Abdalla, Abubakar Abid, Lucas M. Glass, Marinka Zitnik, Cao Xiao, Jimeng SunarXiv 2020 · OpenReview

Abstract

The efficacy of a drug depends on its binding affinity to the therapeutic target and pharmacokinetics. Deep learning (DL) has demonstrated remarkable progress in predicting drug efficacy. We develop MolDesigner, a human-in-the-loop web user-interface (UI), to assist drug developers leverage DL predictions to design more effective drugs. A developer can draw a drug molecule in the interface. In the backend, more than 17 state-of-the-art DL models generate predictions on important indices that are crucial for a drug's efficacy. Based on these predictions, drug developers can edit the drug molecule and reiterate until satisfaction. MolDesigner can make predictions in real-time with a latency of less than a second.

Introduction

A successful drug candidate must possess both high binding affinity for its target and favorable pharmacokinetic (PK) properties, commonly assessed through ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) metrics. Traditionally, obtaining these metrics requires expensive and time-consuming wet-lab testing. While deep learning (DL) models have significantly improved prediction accuracy, existing tools are often fragmented and lack the real-time interactivity necessary for creative lead optimization.

MolDesigner addresses these challenges by providing a unified, human-in-the-loop web interface. It allows drug developers to draw molecular structures and receive instantaneous feedback, transforming DL models from static prediction pipelines into interactive design assistants.

System Architecture and Workflow

The MolDesigner platform is built on a robust tech stack designed for speed and flexibility. The backend utilizes the DeepPurpose library to manage a suite of over 50 DL models. These include Message Passing Neural Networks (MPNN), Convolutional Neural Networks (CNN), and Deep Neural Networks (DNN) for molecular analysis, alongside CNNs and DNNs for processing target protein sequences.

The MolDesigner input interface allows users to enter target protein sequences and draw molecular structures using an integrated editor.

The workflow is highly intuitive:

  1. Input: The user provides a target protein sequence and draws a candidate molecule.
  2. Prediction: The backend executes the models, returning predictions for binding affinity and 16 ADMET properties.
  3. Iteration: The user modifies the chemical structure and immediately observes the impact on the drug's predicted efficacy and safety.

Data and Model Training

The system's predictive power is derived from training on diverse, high-quality datasets:

  • Drug-Target Interaction (DTI): Models for binding affinity were trained using the BindingDB dataset.
  • ADMET Properties: 16 distinct datasets were curated to predict critical indices, including solubility, lipophilicity, Caco-2 permeability, human intestinal absorption (HIA), bioavailability, blood-brain-barrier (BBB) permeability, and various CYP enzyme inhibitions.

Results and Interactive Insights

MolDesigner achieves sub-second latency, which is critical for maintaining the "creative flow" of drug design. By consolidating 17 crucial indices into a single view, the tool helps researchers avoid the "whack-a-mole" problem—where optimizing for binding affinity might inadvertently compromise safety or solubility.

The prediction dashboard displays real-time results for binding affinity and a comprehensive suite of 16 ADMET properties.

Key performance highlights include:

  • Speed: Overall prediction latency is consistently less than one second per iteration.
  • Scale: Integration of over 50 state-of-the-art models (providing 5 models for binding affinity and 3 models for each of the 16 PK properties).
  • Usability: Features such as screenshotting and flagging allow for easy tracking of the design history during optimization.

Conclusion

MolDesigner represents a shift toward interactive, AI-augmented drug discovery. By acting as a "spellchecker" for chemists, it leverages the speed of deep learning to validate structural hypotheses in real-time. Future work will focus on real-world validation with domain experts, particularly in the context of emerging therapeutic targets like COVID-19.

Summary updated

Abstract

The efficacy of a drug depends on its binding affinity to the therapeutic target and pharmacokinetics. Deep learning (DL) has demonstrated remarkable progress in predicting drug efficacy. We develop MolDesigner, a human-in-the-loop web user-interface (UI), to assist drug developers leverage DL predictions to design more effective drugs. A developer can draw a drug molecule in the interface. In the backend, more than 17 state-of-the-art DL models generate predictions on important indices that are crucial for a drug's efficacy. Based on these predictions, drug developers can edit the drug molecule and reiterate until satisfaction. MolDesigner can make predictions in real-time with a latency of less than a second.
BibTeX
@misc{huang2020moldesigner,
  author = {Kexin Huang and Tianfan Fu and Dawood Khan and Ali Abid and Ali Abdalla and Abubakar Abid and Lucas M. Glass and Marinka Zitnik and Cao Xiao and Jimeng Sun},
  title = {{MolDesigner: Interactive Design of Efficacious Drugs with Deep Learning}},
  year = {2020},
  eprint = {2010.03951},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2010.03951},
}
Figure