AI Meets Peptides: Revolutionizing Therapeutic Discovery with Machine Learning
This month, we want to draw your attention to a review article by Goles et al. covering the field of artificial intelligence applied to peptide therapeutics (https://doi.org/10.1093/bib/bbae275).
This article provides a comprehensive exploration of the role of machine learning (ML) and artificial intelligence (AI) in advancing peptide drug discovery, with a strong focus on de novo design and optimization. The paper highlights several advanced ML techniques, including classifier methods for activity prediction, numerical property estimations, and deep generative models (DGMs) like generative adversarial networks (GANs) and variational autoencoders (VAEs) for sequence generation. Deep generative models are central to the de novo design of therapeutic peptides. VAEs, for instance, allow the encoding of peptide data into latent spaces to generate novel sequences resembling training data, while GANs iteratively refine the quality of generated sequences through adversarial training. Newer approaches, such as diffusion models, offer even higher fidelity in sequence generation by leveraging iterative, reversible transformations to fine-tune peptide designs. The use of predictive models alongside these generative techniques facilitates the assessment of physicochemical properties and binding affinities, streamlining the identification of promising candidates before experimental testing.
A unified AI-driven pipeline proposed in the article integrates several critical steps to optimize therapeutic peptide discovery. The process begins with functional classification models, which categorize peptide activity and evaluate properties. This is followed by affinity prediction models that assess peptide interactions with target proteins using deep learning architectures like graph convolutional networks. Generative methods then explore peptide sequence space, creating candidates with specific biological activities and favorable structural properties. The peptide properties are then optimized in silico for stability, toxicity, and pharmacokinetics. By
incorporating iterative validation and reinforcement learning, the pipeline allows continuous refinement of predictive and generative models using experimental feedback.
Finally, the authors address the significant challenges that remain in advancing the field of predictive peptide design. One primary issue is the lack of centralized, high-quality peptide datasets, as fragmented and inconsistent data often lead to misclassification and hinder model training. Additionally, post-translational modifications (PTMs), critical to many of natural peptides' functional roles in vivo, are frequently overlooked in current design pipelines. Moonlighting effects, where peptides exhibit multiple biological activities, further complicate the prediction of specific properties. Addressing these challenges will enable the rapid and precise design of highly specialized peptides that can tackle a wide range of indications with unprecedented efficiency and specificity.
Open access article: https://doi.org/10.1093/bib/bbae275
Goles M., Daza A., Cabas-Mora G., Sarmiento-Varón L., Sepúlveda-Yañez J., Anvari-Kazemabad H., Davari M.D., Uribe-Paredes R., Olivera-Nappa A., Navarrete M.A., Medina-Ortiz D., Peptide-based drug discovery through artificial intelligence: towards an autonomous design
of therapeutic peptides, Briefings in Bioinformatics, Volume 25, Issue 4, July 2024.
Ewa Lis, Ph.D.
Founder/ CEO, Koliber Biosciences
Member, BPF Scientific Advisory Board
https://linkedin.com/in/ewa-lis-6b99528
Read previous editions of the BPF Journal Club series: https://www.boulderpeptide.org/journal-club