MIT News breed · Wetenschap

New AI Framework Designs Novel Proteins Beyond Natural Sequences

Researchers have developed a new AI framework, PottsMPNN, that can design novel proteins with sequences not found in nature. This approach moves beyond simply replicating existing protein sequences, focusing instead on generating structurally feasible proteins and accurately predicting their stability.

Protein function is dictated by its structure, which in turn is determined by its amino acid sequence. Current protein design methods often involve generating sequences that fit a predetermined structure. However, nature shows that multiple sequences can fold into the same structure, and a single sequence can adopt different structures. The challenge for AI in protein design is to recognize this diversity and generate useful sequences.

Amy E. Keating, a professor of biology and biological engineering, stated that measuring success by replicating evolved sequences is not the optimal metric for protein design. The new PottsMPNN framework incorporates physical principles governing protein structure and stability, improving sequence generation and mutation effect prediction. This means the model better understands the relationship between amino acid identity and protein stability.

PottsMPNN's ability to design proteins with sequences unlike any native protein is significant, especially for entirely novel structures where no natural sequence exists for comparison. The key metrics for success are the likelihood of generated sequences folding into desired structures, the model's understanding of the sequence-energy landscape, and its predictive accuracy for mutation effects on stability.

Machine learning has accelerated biological research, enabling reliable generation of protein structures and sequences. PottsMPNN builds on recent advancements, incorporating 'noise' (variations) during training to reduce mimicry of native sequences and increase structural diversity. It also uses a pairwise distribution to model interactions between amino acids and incorporates evolutionarily related sequences to teach the model how different sequences can achieve the same fold.

While incorporating evolutionary information might seem like a reliance on natural sequences, PottsMPNN demonstrates that decreasing dependence on them improves structural compatibility and energy prediction for novel proteins. This advancement opens possibilities for designing new-to-nature proteins for various applications and provides a stronger foundation for future biological engineering.

AI-samenvatting op basis van de bron.