Quantifying the Role of OpenFold Components in Protein Structure Prediction

This is a research paper (arXiv) presenting an interpretability methodology applied to an existing system (OpenFold), not a deployable software project. With 0 stars, 2 forks, zero velocity, and 152-day age, it has no adoption as a tool or library. The contribution is methodological—systematic ablation to understand component importance—which is a novel analytical approach rather than a novel algorithmic breakthrough. However, this is fundamentally a theoretical/reference analysis rather than a reusable component or framework. The low defensibility reflects: (1) no working codebase or package, (2) the methodology is reproducible by any lab with OpenFold access, (3) the findings are insights about an existing system, not a new capability. Frontier risk is low because this is analytical work on existing models—OpenAI/Anthropic don't need to replicate it, they can run similar ablations on their own systems. The paper likely contains pseudocode or experimental setup descriptions, not production software. Integration surface is reference_implementation because downstream use would require reading the paper and reimplementing the ablation study, not consuming a package or API.