PeptimAIze: Revolutionazing Peptide Design through Multi-Property Optimization and Explainable AI

Description du projet

Modèle innovant d’apprentissage profond pour la conception de peptides

Les interactions protéine-protéine (IPP) jouent un rôle essentiel dans de nombreux processus physiologiques et pathologiques de l’organisme humain, et sont des cibles prometteuses pour le traitement de maladies telles que les infections bactériennes et le cancer. L’utilisation thérapeutique et la découverte de peptides sont toutefois entravées par des propriétés physico-chimiques défavorables et par des méthodes inefficaces et gourmandes en ressources. Soutenu par le programme Actions Marie Skłodowska-Curie, le projet PeptimAIze va développer un modèle d’apprentissage profond spécialisé dans la génération de peptides thérapeutiques destinés à inhiber les IPP. À cette fin, le projet entend combler les lacunes des approches actuelles d’apprentissage profond en matière de conception de peptides en développant un modèle offrant une meilleure interprétabilité. Ce modèle permettra de surmonter les limites actuelles tout en étant plus efficace et plus accessible.

Objectif

PeptimAIze aims to develop a deep learning model capable of generating therapeutic peptides for inhibiting protein-protein interactions (PPIs). Many PPIs are essential to key physiological and pathological processes, making these interactions ideal targets for selective intervention in several human diseases, including cancer and bacterial infections. However, the therapeutic potential of peptides is often limited by unfavorable physicochemical properties, such as low solubility and poor bioavailability. Current peptide discovery approaches are time- and resource-intensive, generating significant waste through iterative and labour-intensive screening of compound batches across multiple optimisation rounds to experimentally achieve the desired physicochemical properties. While deep learning and generative models have recently emerged for on-demand peptide design, most focus on a single aspect, typically binding affinity, disregarding physicochemical properties and experimental validation. PeptimAIze addresses these gaps by developing a multi-objective deep learning model based on protein language processing that designs peptides by co-optimising binding to the target and the physicochemical properties necessary for optimal administrability and bioavailability. The model will incorporate explainable AI techniques to enhance interpretability and reveal which amino acids in the peptide sequence are associated with the tuning of each considered property. A key step will be the wet-lab validation of the model on targets responsible for vital bacterial functions (Doc/Phd) and cancer development (p53/MDM2 and Bcl-xL/Bak), demonstrating real-world lab applicability. To maximise impact and usability, the model will be deployed through an intuitive, user-friendly app, enabling researchers in academia and industry to design peptides with tailored properties.