(718g) Developing a Data-Driven Model for Entropy Change upon Peptide Binding

Peptides are small and diverse biomolecules that play key roles in all living organisms. Given their many favorable features such as tunable specificity and biocompatibility, peptides are frequently studied as diagnostic tools, therapeutic treatments, and drug delivery agents. Peptide performance relies on their capacity to attach to other molecules and trigger specific responses, but engineering high-affinity and stability for target analytes is not trivial. Like all thermodynamic processes governed by changes of Gibbs free energy (ΔG), the binding of peptides is determined both by enthalpy (ΔH) and entropy (ΔS). Despite many efforts to study these molecules, computational peptide design research faces the challenge of inadequate accounts of ΔS that often lead to flawed predictions of ΔG. This deficiency is particularly important for peptides given that they lack stable tertiary structure resulting in high flexibility. This work describes the development of a data-driven model that predicts entropy changes upon peptide binding with the objective of improving ΔG predictions and, eventually, guiding computational peptide design overall.

In this study, we assembled a protein-peptide database containing 180 complexes with both experimentally determined thermodynamic properties from Isothermal Titration Calorimetry (ITC) and computational interface data from Rosetta’s Interface Analyzer (RIA) software. Each complex in the database is based off an experimental structure and is composed of a peptide between 5 to 30 amino acids in length bound to a single-domain protein that has no post-translational modifications or non-canonical amino acids. The thermodynamic parameters required to calculate entropy changes from ITC experiments were changes in Gibbs free energy, dissociation constants (kD), enthalpy changes, and binding temperatures. RIA data collected included predicted interaction energies, buried surface areas, and shape complementarity values amongst others. This presentation will discuss the use of this protein-peptide database in the development a data-driven machine learning model that can rapidly predict entropy changes and give insight into the most important features characterizing ΔS throughout peptide binding. This model will then be used to inform the design of peptide recognition elements targeting interleukin-6 (IL6) and insulin for continuous in-line biosensors, intended for use in the biomanufacturing industry.