Sifting a Massive Virtual Library of Peptide Ligands for an Optimal Binder to a Given Receptor

Abstract

The identification of an optimal protein ligand that binds to a target is a difficult problem because the library contains more than a mole of ligands if only 18 residue sites are allowed to vary across all the naturally occurring amino acids. Such a library is far too large to specify explicitly. Instead, the so-called massive virtual library (MVL) is specified indirectly through a set of class rules. An efficient search method through the MVL, which incorporates the principles of sequence design, protein docking and statistical mechanics, has been recently introduced:∗ (i) a random sub-library is created according to the user defined pruning criteria, (ii) each member in the sub-library is docked to the target using AutoDock and ranked according to the binding free energies that are calculated with both AutoDock's scoring function and CHARMM non-bonded interaction energies, (iii) using the calculated free energies, Boltzmann weighted probabilities are assigned to each sequence, (iv) the weights are then used to select the next-generation MVL, and (v) iterate back to step (i) using the current MVL. The algorithm concludes when convergence occurs between the results from two subsequent rounds. For an 8 residue peptide design that binds to Deoxyribonuclease I, the convergence is achieved in as few as 16 iterations. The generated sequence was found experimentally to have high binding affinity selectively towards the desired target.∗S. Quirk, S. Zhong, and R. Hernandez, Proteins: Struct. Func. Bioinfo. 76, 693-705 (2009).