Qsarna: An Online Tool for Smart Chemical Space Navigation in Drug Design

Marcin Cieslak, Jan Leski, Olga Krzysztynska-Kuleta, Justyna Kalinowska-Tluscik, Tomasz DanelJournal of Chemical Information and Modeling 2025 · DOI · OpenReview

TL;DR

Drug discovery is a costly and time-consuming process. This paper introduces Qsarna, a free, web-based platform that integrates machine learning (ML) for activity prediction with molecular docking to streamline virtual screening. Qsarna aims to democratize advanced drug discovery tools for researchers, enabling faster identification of promising drug candidates.

What Problem Does This Paper Address?

Bringing a new drug to market is an incredibly long and expensive journey, often costing billions of dollars and taking over a decade. The high failure rate in clinical trials means pharmaceutical companies and researchers are constantly seeking more efficient ways to discover potential drug candidates. Virtual screening, which uses computational methods to sift through vast libraries of molecules, has become a crucial tool for this early stage. However, existing virtual screening tools often fall into two camps: they are either highly specialized (focusing on just docking or just QSAR models) or they are commercial, limiting access for academic researchers. This fragmentation forces researchers to manually combine results from different tools, potentially missing valuable insights. Furthermore, many advanced tools require significant computational expertise, creating a barrier for medicinal chemists who may not have a deep background in computational science. This paper addresses the need for an integrated, user-friendly, and freely accessible platform that combines multiple computational approaches for drug discovery.

Methods

Qsarna offers a comprehensive, end-to-end virtual screening workflow. Its core functionalities include:

  • Molecular Docking: Utilizes Smina (based on AutoDock Vina) for structure-based virtual screening. Users can customize docking parameters and visualize ligand-protein interactions directly within the platform.
  • Ligand-Based Activity Prediction (QSAR): Employs an AutoML tool to build predictive models (random forests, support vector machines, and artificial neural networks) using various molecular descriptors and fingerprints. These models can predict continuous activity values (like IC50) or binary activity classes, enabling virtual screening of large, unannotated compound libraries.
  • Fragment-Based Generative Design (CRET): This novel module allows for the creation of entirely new molecules. It connects known molecular fragments using linkers, guided by their positions within a binding pocket, to generate diverse and synthesizable chemical entities. This approach helps explore novel chemical spaces beyond existing libraries.
  • ADMET Prediction: Integrates predictors for Absorption, Distribution, Metabolism, Excretion, and Toxicity (ADMET) properties, allowing for multiparametric optimization of drug candidates.

The platform is built on a Django framework with a PostgreSQL database and leverages a Celery queue for handling computationally intensive tasks, ensuring scalability and flexibility.

Datasets

  • ChEMBL LSS: Used for training QSAR models for MAO-A and MAO-B inhibition.
  • MolPort Catalog: Screened for potential MAO inhibitors.
  • DUD-E Database: Used for evaluating the performance of the molecular docking module against decoy compounds.
  • Publicly available data: Used for training ADMET predictors.

Key Insights and Ideas

  • Integration is Key: The paper strongly advocates for the synergistic combination of different computational methods (docking, QSAR, generative design) to overcome the limitations of individual approaches and improve the reliability of drug candidate identification.
  • Democratizing Access: By providing a free, web-based platform, Qsarna significantly lowers the barrier to entry for advanced virtual screening, empowering researchers who may lack specialized computational resources or expertise.
  • Generative Design for Novelty: The inclusion of a fragment-based generative module is crucial for exploring uncharted chemical spaces and discovering novel chemotypes, moving beyond simply screening existing compound libraries.
  • User-Centric Design: The emphasis on an intuitive graphical user interface and onboarding tutorials is vital for making complex computational tools accessible to a broader scientific audience.

Key Results

  • The Qsarna platform successfully identified three new MAO-B inhibitors with nanomolar potency (down to 1.37 nM IC50) in a case study, which were confirmed by experimental assays.
  • The QSAR models trained for MAO-A and MAO-B achieved high performance, with ROC AUC values of 0.92 and 0.88, respectively.
  • The molecular docking module performed comparably to commercial software (Glide) on two out of three tested targets.
  • The generative fragment-linking module (CRET) was shown to create diverse and synthesizable libraries of novel compounds.

Limitations and Future Work

  • Limited Resources: The platform is provided free of charge with monthly limits on analyzed compounds and calculations due to resource constraints. While academic users can request increases, this remains a practical limitation.
  • Experimental Validation: While the platform streamlines computational discovery, experimental validation remains a critical downstream step.
  • Future Work: The authors suggest that Qsarna will expedite drug discovery efforts by automating computational screening procedures and supporting researchers who are not deeply familiar with in silico drug design. The platform's hybrid approach is particularly valuable for targets with limited experimental data.

Summary updated

BibTeX
@article{cieslak2025qsarna,
  author = {Marcin Cieslak and Jan Leski and Olga Krzysztynska-Kuleta and Justyna Kalinowska-Tluscik and Tomasz Danel},
  title = {{Qsarna: An Online Tool for Smart Chemical Space Navigation in Drug Design}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2025},
  doi = {10.1021/acs.jcim.5c00720},
}
Figure