Attractors in Sequence Space: Peptide Morphing by Directed Simulated Evolution

Abstract

Certain antimicrobial peptides (AMPs) and mitochondrial targeting peptides (mTPs) share common features such as a positive net charge, helical propensity, and the ability to interact with lipid membranes1. While a primary function of AMPs is to disrupt membrane integrity, many mTPs interact with the mitochondrial membranes and the translocator complex, leaving the membranes intact when proteolytically degraded by matrix-located protease(s)2. We present a computational approach that may help rationalize the delicate balance of these effects and other sequence-activity relationships. We implemented and applied a directed evolutionary strategy for stepwise peptide morphing of one peptide into another (MoPED, Morphing of Peptides by Evolutionary Design). For the prospective application, we chose cationic peptides as representatives of both peptide classes and converted an AMP (Protonectin, start sequence)3 into an mTP (target or “attractor” sequence). Novel peptides were generated based on a chemical similarity index (Grantham substitution matrix).4 The target sequence served as an attractor point for evolutionary exploration of sequence space. We synthesized and tested the individual peptides that were generated during the morphing process. The designed peptides showed systematic loss of membranolytic potential with increasing distance from the start sequence. The results of biophysical and bacterial growth experiments confirmed the applicability of MoPED to designing chemically motivated peptide derivatives. Receiver operating characteristic (ROC) analysis advocates the inducible peptide α-helicity as a semi-quantitative indicator of membranolytic antimicrobial activity. According to this mutation model, the pair-wise amino acid similarity values are used to obtain pseudo-probabilities for each residue position in a peptide (Figure 1). This concept results in a non-symmetric residue substitution matrix, so that forward and reverse mutations can have different transition probabilities, P(i→j)≠P(j→i). The width σ of the approximately bell-shaped distribution of offspring around the parent is a so-called strategy parameter (adaptive memory) during peptide evolution, which itself underlies selection, thereby enabling automated fitness-dependent adaptation of function-altering and neutral mutations.9,10 Distance-based residue mutation. The idealized graph shows pseudo-probability functions P for residue mutations i→j, with widths σ=1.0 and 1.4, centered at the parent residue i (d=0). The chance for observing the transition i→j decreases with the distance between residues i and j, dij. The choice of σ influences the residue diversity (entropy) of the mutated sequence. MoPED builds on the SME principle. The two main differences to the original algorithm are: The stochastic algorithm is equipped with an attractor sequence (“target” or “end” sequence), which means that starting from the initial parent sequence offspring is generated in an iterative mutation-selection cycle until the target sequence has been obtained. The algorithm performs a directed stochastic walk in sequence space. New sequences were generated by help of the Grantham matrix, which captures the pair-wise amino acid exchange based on composition and physicochemical properties, specifically residue polarity and molecular volume.4 Grantham distances (so-called “biochemical distances”) are in statistically significant agreement with observed evolutionary transition probabilities.4,11 The distance concept pursued in this study does not represent the only solution to computing physicochemical similarities, as there are several related approaches, e.g. the Grantham matrix derived measure of Miyata et al.12 and the experimentally obtained exchangeability matrix approach of Yampolsky and Stoltzfus.11 The results of a MoPED run will critically depend on the substitution matrix used, particularly for skewed residue compositions.13 In MoPED, stepwise mutations morph the start sequ