EAGER: NAIRR Pilot: VizFold: Visualizing & Interpretable Protein Folding in AlphaFold and Similar Protein Models.

Advancements in artificial intelligence (AI) are transforming scientific discovery, enabling breakthroughs that were once thought unattainable. In biology, AI-driven models like AlphaFold have revolutionized protein structure prediction, achieving unprecedented accuracy in determining 3-dimensional structures from amino acid sequences. This capability has opened the door to applications ranging from drug discovery to protein engineering.

Despite these successes, the inner workings of such models remain a black box, hindering their broader scientific utility and the development of fundamental theories. This project aims to illuminate the internal activity and their correspondence to bio-physical principles by developing advanced visualization and interpretability tools. These tools will allow researchers to gain unprecedented insights into how these AI models operate, fostering a deeper understanding of protein folding processes.

By providing open-source software and an interactive Science Gateway, the project will democratize access to these capabilities, empowering a wider range of scientists to explore and refine protein folding theories and showcasing the role of AI in advancing fundamental scientific knowledge.

VizFold tackles the critical challenge of understanding and interpreting the mechanisms underpinning protein folding predictions made by AI models like AlphaFold. While these models excel at accurately predicting protein structures, the processes and principles they use remain opaque, limiting their broader applicability and the development of scientific theories.

To bridge this gap, the project focuses on three key objectives. First, it will develop visualization tools to examine activation propagation within the specialized architectural components, including attention mechanisms and structure modules unique to AlphaFold-like models. Second, it will enhance interpretability by applying probing techniques, dimensionality reduction and Layerwise Relevance Propagation, to elucidate how these models encode folding processes.

Third, it will build an interactive Science Gateway powered by Cybershuttle, enabling researchers to visualize and analyze model outputs in real-time while offloading computational tasks to high-performance infrastructure. By achieving these goals, VizFold will advance computational biology, support non-expert users, and lay the foundation for a new physical-chemical understanding of protein folding, paving the way for future innovations in AI-driven science.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.