AI enables design of proteins not found in nature
MIT's new model revolutionizes protein sequence generation.
1New approach in protein design
Traditional protein design methods focused on determining target structures first, then generating compatible amino acid sequences. MIT biologists developed PottsMPNN, a novel machine learning framework that incorporates physical principles governing protein folding. Published in PNAS, this model significantly expands the diversity of possible protein sequences. The approach allows creation of entirely new proteins not seen in nature, with potential applications in drug development and biotechnology.
2Overcoming natural sequence limitations
The research team, led by graduate student Foster Birnbaum, introduced 'noise' during training to reduce the model's tendency to mimic natural sequences. This innovation enables generation of novel sequences that can adopt desired structures. The method proves particularly valuable for designing proteins targeting disease-causing molecules. By moving beyond evolutionary constraints, researchers can explore previously inaccessible regions of protein sequence space.
3Modeling amino acid interactions
PottsMPNN employs pairwise distribution to capture interactions between amino acids, allowing more accurate modeling of physical interactions among all 20 possible amino acids at any two positions. This significantly improves protein stability predictions and mutation impact assessments. Such capabilities are crucial for designing therapeutic proteins that must maintain stability under physiological conditions while performing specific biological functions.
4Incorporating evolutionary knowledge
The team trained the model using sets of evolutionarily related sequences to teach how different sequences can adopt identical folded structures. Professor Amy E. Keating notes this represents a major advancement in protein design. As the model reduces dependence on native sequences, it demonstrates improved performance in structural compatibility and energy predictions, even for completely novel protein architectures.
5Revolution in biological engineering
PottsMPNN enables design of structurally feasible sequences unlike any natural protein. Lead author Birnbaum acknowledges the potentially transformative - even 'scary' - impact this could have on biological engineering. The technology opens new frontiers in drug development, enzyme design, and biomaterial production. It particularly shines when targeting completely novel protein structures without natural analogs.
6Surpassing the 2022 benchmark
Until now, a 2022 model remained the gold standard in protein design. The MIT team investigates why PottsMPNN outperforms its predecessor. Key advantages include better understanding of sequence-energy landscapes and more accurate mutation effect predictions. These improvements stem from the model's incorporation of fundamental physical principles governing protein folding and stability.
7Task-specific fine-tuning potential
Birnbaum highlights opportunities for further model refinement for specific applications. Previous experience shows such customization can enhance performance, like improving mutation outcome predictions. This adaptability makes the framework valuable for personalized medicine approaches, particularly in treating rare genetic disorders. The technology could enable development of patient-specific protein therapeutics.
8Dawn of a new era in biology
Professor Keating emphasizes these methods advance the field toward designing useful novel proteins for diverse applications while laying groundwork for future breakthroughs. Birnbaum expresses optimism about 21st century biology progress. Beyond medicine, these protein design innovations could transform industrial biotechnology and sustainable material production, potentially addressing global challenges in healthcare and environmental sustainability.
Original article: MIT News — AI — Yapay zeka ile doğada olmayan proteinler tasarlanıyor