A hybrid protocol for peptide development: integrating deep generative models and physics simulations for biomolecular design targeting IL23R/IL23

Abstract

Recent advances in machine learning have revolutionized molecular design; however, a gap remains in integrating generative models with physics-based simulations to develop functional modulators, such as stable peptides, for challenging targets like the interleukin-23 receptor (IL23R) and its associated cytokine, interleukin-23 (IL23). The IL23R/IL23 axis plays a critical role in autoimmune diseases, and current therapies have largely been limited to antibody-based approaches. To address this gap, we employed a hybrid computational approach that combines Long Short-Term Memory (LSTM) networks for peptide generation, a Gated Recurrent Unit (GRU)-based classifier for anti-inflammatory property prediction, and molecular dynamics (MD) simulations to assess structural dynamics, binding interactions, as well as key properties such as binding affinity and stability. Using this hybrid framework, we identified novel inhibitory peptides, particularly P4, with an IC₅₀ of 2 μM. Systematic experimental validation established its inhibitory activity, elucidated its binding mechanism, confirmed its specificity toward the IL23R, and demonstrated its ability to disrupt IL23R/IL23 interaction. This integrated approach highlights the significant potential of combining deep learning and simulations to accelerate the identification of peptide-based therapeutics targeting key protein targets.