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

Area of Science:

• Computational Biology
• Biophysics
• Artificial Intelligence

Background:

• De novo protein and antibody design increasingly utilizes AI, but requires robust validation.
• Current AI design methods often lack thermodynamic realism, limiting confidence in generated candidates.
• Physics-based scoring methods are crucial for filtering and ranking AI-generated designs, especially for similar sequences.

Purpose of the Study:

• To explore hybrid AI-physics pipelines for de novo protein and antibody design.
• To identify the optimal integration of AI and physics-based models for high-confidence design.
• To propose a framework for antibody design combining AI and physics-based modeling.

Main Methods:

• Review of recent AI methods and prospective ideas for hybrid pipelines.
• Analysis of confidence scores (ipTM, pDockQ2, ipSAE) for ranking designs.
• Development of a generalizable framework for AI-physics antibody design pipelines.
• Showcasing MadraX, a differentiable FoldX force field implementation.
• Classification of three tiers of AI-physics integration.

Main Results:

• The "sweet spot" for high-confidence candidates lies in the agreement between deep learning and first-principles models.
• Standard AI confidence scores are insufficient for ranking similar sequences, necessitating physics-based integration.
• A framework for AI-physics antibody design pipelines was described.
• MadraX offers a differentiable and AI-compatible physics-based tool.
• Three tiers of AI-physics integration were classified, from post hoc filtering to full embedding.

Conclusions:

• Hybrid AI-physics pipelines offer a promising direction for de novo protein and antibody design.
• Integrating AI flexibility with physics-based realism is key to achieving high-confidence designs.
• Community-wide blind assessments are needed to validate and advance de novo design pipelines.