FoldMark: Safeguarding Protein Structure Generative Models with Distributional and Evolutionary Watermarking

The preprint operates within the dominant paradigm of protein structure prediction and design, heavily influenced by recent advancements in machine learning, particularly deep learning models like AlphaFold. It acknowledges the transformative impact of these models while addressing emerging concerns within this paradigm.

This preprint presents FoldMark, a novel technique for watermarking protein structures generated by AI models. It addresses specific problems arising within the current successful protein structure prediction paradigm, namely biosecurity risks and intellectual property concerns. The research doesn't challenge the fundamental principles of the existing paradigm but extends it by adding a security layer. This aligns with Kuhn's description of "normal science," where research aims to solve puzzles within the established framework. The development of FoldMark is akin to developing a new experimental technique or refining an existing one, rather than proposing a fundamental shift in how we understand protein structures. Therefore, it represents an advancement within the current paradigm of protein structure prediction and design, solidifying its classification as "normal science." I considered 'Model Drift' as a secondary classification possibility, driven by the preprint's focus on mitigating risks associated with the increasing power of PGMs. However, 'Normal Science' appears more fitting as FoldMark primarily builds upon existing methods and doesn't deviate substantially from the core paradigm.