Steering generative models for protein design: Aligning and conditioning strategies.

Area of Science:

• Computational Biology
• Artificial Intelligence
• Protein Engineering

Background:

• Generative artificial intelligence (AI) models learn data distributions to create novel samples.
• Proteins are ideal for generative AI due to diverse data representations (sequences, structures, functions).
• Generative models have shown success in designing functional proteins and enzymes.

Purpose of the Study:

• To review and categorize strategies for steering generative AI models toward specific protein properties.
• To address the limitation of generative models focusing on probable data modes, potentially missing valuable low-probability regions.
• To enable the design of proteins with user-specified characteristics.

Main Methods:

• Categorization of steering strategies based on model parameter modification.
• Distinguishing between methods that modify model parameters (e.g., reinforcement learning, supervised fine-tuning).
• Identifying approaches that keep model parameters fixed (e.g., conditional generation, retrieval-augmented strategies, Bayesian guidance, tailored sampling).

Main Results:

• Generative models for protein design can be steered toward desired properties.
• Two main categories of steering strategies exist: modifying model parameters or keeping them fixed.
• Various techniques are emerging to guide AI in exploring under-explored regions of protein property space.

Conclusions:

• Steering strategies are crucial for overcoming the limitations of standard generative models in protein design.
• These methods allow for the targeted generation of proteins with specific, potentially rare, valuable properties.
• The development of these steering techniques is advancing the field of AI-driven protein engineering.