Machine learning in biological physics: From biomolecular prediction to design.

Area of Science:

• Biological Physics
• Computational Neuroscience
• Biomolecular Engineering

Background:

• Machine learning (ML) is increasingly used in biological physics, but often separate from theoretical modeling.
• Early computational neural networks, like Hopfield networks, originated from physical modeling of neuronal processes.

Purpose of the Study:

• To advocate for and demonstrate the synergistic integration of physical modeling and machine learning in biological physics.
• To explore the connections between various ML approaches and physical models, particularly through energy representations.

Main Methods:

• Discussing the historical link between physical modeling and early neural networks.
• Analyzing modern ML methods (Potts models, Boltzmann machines, transformers) through a shared energy representation framework.
• Reviewing recent successful applications of integrated physical modeling and ML in biomolecular research.

Main Results:

• Established connections between diverse ML techniques and physical modeling principles.
• Highlighted successes in protein structure prediction, molecular dynamics, and evolutionary modeling.
• Demonstrated ML's revolutionary impact on protein engineering and design, including the generation of novel protein sequences.

Conclusions:

• The integration of physical modeling and machine learning provides a more successful framework for tackling complex biological physics challenges.
• This combined approach has led to significant advancements in understanding and manipulating biomolecular systems.
• Future work includes using learnable physical models for generating unique synthetic protein sequences with desired structures.