A new generative artificial intelligence framework, named APEX GO, has demonstrated the ability to design and optimize novel peptide antibiotics, according to research published in Nature Machine Intelligence. This approach marks a shift from screening existing molecules to creating entirely new ones with enhanced antimicrobial properties. The AI model successfully generated optimized peptide sequences that showed significant antimicrobial activity in laboratory tests and preclinical mouse models.
Antimicrobial resistance (AMR) presents a substantial global health challenge, driving the need for new therapeutic strategies. Peptide antibiotics, which mimic natural immune defenses, offer a promising avenue, but their experimental optimization has been a resource-intensive process. APEX GO integrates a transformer-based variational autoencoder with Bayesian optimization to design these peptides. Unlike traditional methods that search through existing databases, APEX GO modifies template peptides to create novel sequences. This framework includes design and diversity constraints to maintain similarity to templates while allowing for sequence innovation. This work represents the first experimental validation of generative Bayesian optimization for peptide design in both in vitro and in vivo settings.
The researchers used ten de-extinct peptides as starting templates. APEX GO then generated optimized derivatives with improved antimicrobial characteristics. From these generated sequences, 100 peptides were synthesized and subjected to comprehensive laboratory evaluations. These included assessments of their antimicrobial activity against clinically relevant Gram-negative pathogens, their mechanism of action, secondary structure, and cytotoxicity. The study reported an 85% "ground-truth experimental hit rate" and a 72% success rate in enhancing antimicrobial activity. These results reportedly outperformed previous methods for antibiotic discovery and optimization.
In preclinical mouse models involving Acinetobacter baumannii infections, several of the AI-optimized molecules demonstrated potent anti-infective activity. Notably, derivatives of mammuthusin-3 and mylodonin-2 showed efficacy comparable to or exceeding that of polymyxin B, a widely used antibiotic for resistant infections. These findings suggest APEX GO's potential as a tool to accelerate antibiotic discovery and combat AMR.
The APEX GO framework represents a departure from earlier AI models. For instance, an earlier AI model from the de la Fuente lab, named APEX, could predict whether a given peptide was likely to have antimicrobial properties. APEX GO builds on this by not only predicting but also generating new sequences. The AI model was trained on extensive datasets of peptide sequences and their associated antimicrobial activities. By learning patterns linked to potent antibacterial effects, the AI can explore vast chemical spaces. This ability to create, evaluate, and optimize molecules de novo is a key advancement.
The AI-driven workflow involves iterative stages where the model proposes precise edits to a starting peptide, predicts the impact of these changes on antimicrobial activity, and progresses towards more effective versions. This dynamic adaptation allows the generative model to "evolve" peptide antibiotics computationally, significantly speeding up the optimization process compared to traditional trial-and-error methods.
While the study showcases promising results with AI-generated peptides, the researchers emphasize that these are early-stage candidates. Further optimization for safety, stability, and in-body duration will be necessary before human clinical use. However, the study provides a proof-of-concept for AI's role in guiding researchers toward molecules worth pursuing, potentially transforming the future of drug discovery.