The sweet spot in protein design-Where deep learning meets first principles

This Perspective explores recent methods and prospective ideas for developing hybrid AI-physics-based pipelines for protein and antibody de novo design. We argue that the highest-confidence candidates emerge where deep learning and first-principles models agree, a "sweet spot" that balances generative flexibility with thermodynamic realism. For example, although interface confidence scores such as ipTM, pDockQ2, or ipSAE are widely used to rank generated designs, we show that they are not well suited to rank similar sequences, which suggests the need to combine them with physics-based methods to improve design filtering and ranking. Furthermore, we describe a generalizable framework for implementing antibody design pipelines that combine AI with physics-based modeling and scoring methods and also showcase MadraX, a differentiable and AI-compatible implementation of the FoldX force field. In addition, we classify three tiers of AI-physics integration, from post hoc filtering to full embedding of differentiable physics inside deep learning models. Finally, we discuss the future of the protein design community and underline the need to support current initiatives for community wide blind assessments of the growing number of de novo design pipelines.