SPARK: Harnessing Human-Centered Workflows with Biomedical Foundation Models for Drug Discovery

Bum Chul Kwon, Simona Rabinovici-Cohen, Beldine Moturi, Ruth Mwaura, Kezia Wahome, Oliver Njeru, Miguel Shinyenyi, Catherine Wanjiru, Sekou L. Remy, William Ogallo, Itai Guez, Parthasarathy Suryanarayanan, Joseph A. Morrone, Shreyans Sethi, Seung-gu Kang, Tien Huynh, Kenney Ng, Diwakar Mahajan, Hongyang Li, Matan Ninio, Shervin Ayati, Efrat Hexter, Wendy D. CornellIJCAI 2024

Introduction

Biomedical foundation models, trained on vast datasets of small molecules and proteins, offer a transformative path for drug discovery. However, their complexity often renders them "black boxes" to the biologists and chemists who possess the necessary domain expertise. SPARK addresses this friction point by providing a web-based platform that integrates these models into a human-centered, iterative workflow. By wrapping intricate neural architectures in an intuitive interface, SPARK allows researchers to steer molecule generation and bridge the gap between raw AI output and actionable biological insights.

System Architecture and Methods

SPARK is built on a modular microservices architecture using the Caikit open-source framework. This design allows for the seamless integration of various state-of-the-art generative models, such as PaccMann (for anticancer sensitivity) and MolFormer (a chemical language model).

The platform's frontend is designed for high-fidelity scientific visualization. It utilizes the NGL Viewer for interactive 3D protein structures and D3.js for generating dynamic density charts. These charts allow researchers to visualize the distribution of critical molecular properties—such as solubility, toxicity, and drug-likeness—across a generated library of candidates.

Interactive Workflow and Steerability

A core tenet of the SPARK platform is that drug discovery is an iterative process rather than a "one-shot" task. The system provides researchers with "knobs" to adjust model behavior, including sampling temperature and specific inclusion criteria.

The SPARK user interface components, including protein visualization (A), parameter tuning (B), metric filtering (C), and result analysis (D-F).

As illustrated in the figure, the workflow begins with target protein selection (A). Researchers can then define model parameters (B) and set thresholds for metrics of interest (C). The resulting molecules are presented as detailed cards alongside statistical distributions (D, E), enabling "visual triangulation" to identify candidates that are both chemically valid and physically plausible for docking. To further assist non-AI experts, an integrated chatbot powered by IBM watsonx Assistant (F) provides natural language explanations for technical terms and model logic.

Case Study: Chagas Disease

The utility of SPARK was demonstrated through a use case involving Chagas Disease, a neglected tropical disease. Researchers targeted a specific kinase (protein 5U2E) and initiated a generation cycle. While the first run produced candidates, visual and statistical analysis revealed potential toxicity issues.

Leveraging the platform's interactivity, the researchers switched to the PaccMannRL model and tightened filters for "synthetic accessibility." This refinement led to a more promising set of candidates. The system's ability to export these findings into shareable reports facilitates the cross-disciplinary collaboration essential for drug development.

Conclusion and Future Work

SPARK represents a significant step toward making AI-driven drug discovery accessible to domain experts. By prioritizing steerability and transparency, the platform empowers researchers to guide AI models toward biologically relevant chemical space. While the current outputs are speculative and require rigorous "wet lab" validation, future work aims to expand the library of integrated models and further refine the human-AI collaborative interface to address a broader range of neglected diseases.

Summary updated

BibTeX
@inproceedings{kwon2024spark,
  author = {Bum Chul Kwon and Simona Rabinovici-Cohen and Beldine Moturi and Ruth Mwaura and Kezia Wahome and Oliver Njeru and Miguel Shinyenyi and Catherine Wanjiru and Sekou L. Remy and William Ogallo and Itai Guez and Parthasarathy Suryanarayanan and Joseph A. Morrone and Shreyans Sethi and Seung-gu Kang and Tien Huynh and Kenney Ng and Diwakar Mahajan and Hongyang Li and Matan Ninio and Shervin Ayati and Efrat Hexter and Wendy D. Cornell},
  title = {{SPARK: Harnessing Human-Centered Workflows with Biomedical Foundation Models for Drug Discovery}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure