Combination of Coevolutionary Information and Supervised Learning Enables Generation of Cyclic Peptide Inhibitors...

Area of Science:

• Computational chemistry
• Biochemistry
• Machine learning in drug discovery

Background:

• Designing cyclic peptide inhibitors computationally often requires extensive experimental data, which is challenging to obtain.
• Developing effective inhibitors for tumor-associated proteases is crucial for cancer therapy.

Purpose of the Study:

• To demonstrate an enhanced computational design pipeline for cyclic peptide inhibitors using machine learning.
• To validate the efficacy of computationally designed inhibitors against a tumor-associated protease, particularly with small experimental datasets.

Main Methods:

• Sequential combination of Random Forest Regression, pseudolikelihood maximization Direct Coupling Analysis, and Monte Carlo simulation.
• In vitro experimental validation of designed cyclic peptides.
• Crystal structure analysis of cyclic peptide-protease complexes.

Main Results:

• The enhanced computational method effectively improved the design pipeline for cyclic peptide inhibitors, even with small experimental datasets.
• In silico-evolved cyclic peptides demonstrated superior potency compared to previously developed inhibitors.
• Crystal structures revealed large interaction surfaces, constrained backbones, and multiple interactions contributing to high binding affinity and selectivity.

Conclusions:

• A hybrid computational approach significantly enhances the design of potent cyclic peptide inhibitors for tumor-associated proteases.
• This strategy overcomes limitations of small experimental datasets in computational drug design.
• The findings offer a promising avenue for developing novel cancer therapeutics.