Designing peptide inhibitors that target specific protein-protein interactions (PPIs) is challenging. Existing methods fall into two categories: structure-based and sequence-based. Structure-based approaches, such as Rosetta FlexPepDock, leverage existing structural templates but can be computationally expensive and limited by template biases. Sequence-based methods, using Recurrent Neural Networks (RNNs) like Variational Autoencoders (VAEs), offer a robust framework for sequence analysis and optimization but often lack sufficient training data. This study proposes a multi-step approach combining a GRU-based VAE with Metropolis-Hastings (MH) sampling for peptide sequence generation, followed by Rosetta FlexPepDock for rapid binding assessment and MD simulations for refined binding energy calculation. This integrated pipeline aims to efficiently generate high-affinity peptide binders for specific protein surfaces, addressing limitations of existing structure-based and sequence-based approaches.

Computational peptide design has seen advancements in antibiotics and biomaterials, but designing inhibitors for PPIs remains difficult. Structure-based methods like Rosetta FlexPepDock are effective but computationally intensive and potentially biased by structural templates. Methods to identify binding pockets and design inhibitors to occupy them exist, but finding optimal solutions remains challenging. Molecular Dynamics (MD) simulations provide a dynamic view of peptide-protein interactions, refining binding poses and calculating affinity. Sequence-based methods, particularly RNN-based VAEs, are promising for sequence analysis and optimization, with successful applications in molecule and antimicrobial peptide design. However, limited data availability for peptide-protein binding remains a constraint for sequence-based approaches. This paper addresses these challenges by integrating deep learning with structure-based modeling and simulation.

The study employs a two-pronged computational methodology. The first pipeline uses a GRU-based VAE and MH sampling to generate peptide sequences, reducing the search space. Rosetta FlexPepDock rapidly assesses binding affinity, ranking peptides based on interface energy (I_sc), root-mean-square deviation of interface atoms (rmsALL_if), and buried surface area (I_bsa). High-ranked peptides undergo further evaluation via MD simulations and MM/GBSA calculations for more accurate binding energy estimation. The second pipeline iteratively fine-tunes the VAE-MH model using Rosetta FlexPepDock scores from previously generated peptides, creating target-specific models. This iterative process enriches the generation of peptides with better binding characteristics. For experimental validation, peptides are synthesized and their binding affinity is measured using competitive fluorescence polarization (FP) or homogeneous time-resolved fluorescence (HTRF) assays. MD simulations are used to analyze binding modes and identify key interactions.

The study successfully designed peptide inhibitors targeting β-catenin and NEMO. For β-catenin, using N-terminal extensions, two out of four tested peptides displayed improved potency compared to the parent peptide, although the improvement was modest (around twofold). A library screen for N-terminal extensions did not yield peptides with significantly higher affinity. In contrast, using C-terminal extensions, a remarkable 15-fold improvement in β-catenin binding affinity was observed for the best peptide (CAL-2, IC₅₀ = 0.010 ± 0.06 µM) compared to the parent peptide (IC₅₀ = 0.15 ± 0.04 µM). For NEMO, two out of four designed peptides showed substantially improved binding compared to the parent peptide (NBD), highlighting the applicability of the methodology to different protein targets. MD simulations revealed key interactions, such as π-cation and π-π stacking interactions between the C-terminally extended peptides and β-catenin. For the N-terminal extensions, the interaction with β-catenin appears to be less strong.

The results demonstrate the power of integrating deep learning with structure-based modeling and simulations. The iterative fine-tuning of the VAE-MH model proves effective in generating target-specific peptide inhibitors, overcoming data scarcity challenges. The success of C-terminal extensions over N-terminal extensions suggests that the design space and interaction possibilities differ significantly. The findings underscore the potential of this integrated computational approach for designing potent peptide inhibitors targeting various proteins. Although the N-terminal extension strategy yielded limited improvement, the substantial enhancement in affinity achieved with C-terminal extensions and the successful application to NEMO highlight the versatility and effectiveness of this methodology. The ability to identify potent peptide inhibitors without extensive modifications, like unnatural amino acids, is a key strength of this approach.

This study successfully integrated deep learning and molecular dynamics simulations to design target-specific peptide inhibitors. The iterative VAE-MH model, coupled with Rosetta and MM/GBSA calculations, proved effective in generating high-affinity binders for β-catenin and NEMO. Future research could explore expanding the training datasets, incorporating other deep learning architectures, and applying this methodology to other PPI targets. The significant improvement in affinity observed for the C-terminal extensions warrants further investigation into the structural features and interaction mechanisms responsible for the enhanced binding.

The study's limitations include the computational cost of MD simulations, which restricts the number of peptides that can be thoroughly evaluated. The accuracy of binding energy predictions relies on the accuracy of the force fields and scoring functions used. The experimental validation was performed only for a subset of the computationally designed peptides. Finally, the applicability of this method might be limited to proteins with readily available structural information.