Machine learning-enabled discovery and design of membrane-active peptides

Abstract

Antimicrobial peptides are a class of membrane-active peptides that form a critical component of innate host immunity and possess a diversity of sequence and structure. Machine learning approaches have been profitably employed to efficiently screen sequence space and guide experiment towards promising candidates with high putative activity. In this mini-review, we provide an introduction to antimicrobial peptides and summarize recent advances in machine learning-enabled antimicrobial peptide discovery and design with a focus on a recent work Lee et al. Proc. Natl. Acad. Sci. USA 2016;113(48):13588-13593. This study reports the development of a support vector machine classifier to aid in the design of membrane active peptides. We use this model to discover membrane activity as a multiplexed function in diverse peptide families and provide interpretable understanding of the physicochemical properties and mechanisms governing membrane activity. Experimental validation of the classifier reveals it to have learned membrane activity as a unifying signature of antimicrobial peptides with diverse modes of action. Some of the discriminating rules by which it performs classification are in line with existing "human learned" understanding, but it also unveils new previously unknown determinants and multidimensional couplings governing membrane activity. Integrating machine learning with targeted experimentation can guide both antimicrobial peptide discovery and design and new understanding of the properties and mechanisms underpinning their modes of action.

Keywords: Antimicrobial peptides; Cell-penetrating peptides; Machine learning; Membrane-active peptides; Quantitative structure activity relationship models.

Figure 1

Schematic illustration of a support vector machine (SVM) classifier operating in a feature space of dimensionality m = 2. Data points x are represented in this space by their m = 2-dimensional feature vetors. The maximum margin hyperplane defines a (m−1)-dimensional surface that maximally separates the two classes “hits” (red crosses) and “misses” (blue circles). Mathematically, the hyperplane is defined by the locus of points x satisfying w.x − b = 0, where w is the (possibly non-unit) surface normal, ŵ = w/||w|| is the corresponding unit vector, and ŵ.x = b/||w|| is the offset of the hyperplane from the origin. The margin is defined by the two parallel hyperplanes satisfying w.x − b = +1 and w.x − b = −1, and the SVM is trained by maximizing the width of the margin $\widehat{\mathbf{w}}.({\mathbf{x}}_{+}-{\mathbf{x}}_{-})=\frac{2}{\Vert \mathbf{w}\Vert }$, where x₊ is the closest “hit” to the hyperplane separator and x₋ the closest “miss”. For data that are not linearly separable, the hard margin formulation (Eqn. 2) is supplanted by the soft margin (Eqn. 3) version that allows for classification errors.

Figure 2

Physical interpretation of the distance to hyperplane σ predicted by the trained SVM classifier. (a) A scatterplot of σ against in vitro minimum inhibitory concentration (MIC) for 478 AMPs active against Staphylococcus aureus reveals no significant correlation (ρ_Spearman = −0.06 [−0.15, 0.03], p = 0.19). (b) Schematic illustration of a surface possessing negative Gaussian curvature (NGC) wherein principal curvatures of opposing sign give rise to a saddle-shaped topography. (c) Illustration of the Pm3m and Im3m cubic phase space groups that are rich in NGC. (d) A scatterplot of σ against average NGC |〈K〉| induced in artificial mimics of bacterial cell membranes for 16 peptides selected for synthesis and experimental characterization reveals a strong and statistically significant positive correlation (ρ_Spearman = 0.65 [0.23, 0.89], p = 0.006). Panels a, c, and d are adapted from Lee et al. Proc. Natl. Acad. Sci. USA 113 48 13588–13593 (2016) [1].

Figure 3

Helical wheel plot showing the relative residue locations along the α-helical backbone. Our classifier favors positive classification of peptides in which residues separated by 2, 9, and 30 positions are of opposing physicochemical character (i.e., those located on opposing faces of the helix; red bold lines) and those separated by 4 positions tend to have similar character (i.e., those located on the same face; green bold line). Facially amphipathic peptides are therefore scored highly by the classifier over these four features, and it has learned these patterns as a discriminating rule with which to distinguish membrane activity. Image adapted from Lee et al. Proc. Natl. Acad. Sci. USA 113 48 13588–13593 (2016) [1].

Figure 4

Directed search of sequence space and adherence of Pareto optimal candidates to the saddle splay selection rule. (a) Projection of the 242,110 peptide candidates considered in our directed traversal of sequence space into the minimum sequence homology to any known AMP measured by the Jukes-Cantor distance and classifier distance to hyperplane σ. Highlighted are the 85 Pareto optimal candidates within the three dimensional search space of [σ, degree of α-helical structure, sequence homology to a known AMP] (orange diamonds), and the 13 Pareto optimal candidates subject to the additional condition that the 12 descriptors employed by the classifier (Table 1) lie no more than 10% outside the range of the training data (green diamonds). Peptides from families with other putative primary functions are situated close to the Pareto frontier and are positively classified by the SVM (σ > 0) suggesting that they possess membrane activity as part of a multiplexed functionality (colored stars). (b) Optimal peptide candidates identified in a guided traversal of sequence space by the SVM classifier obey the previously identified saddle splay selection rule governing a trade-off between peptide hydrophobicity and proportion of arginine and lysine residues. N_K and N_R respectively denote the number of lysine and arginine residues in the peptide. The mean hydrophobicity is defined as the mean value of the Eisenberg consensus hydrophobicity averaged over all residues in the peptide [94]. The 13 physicochemical restricted Pareto optimal peptides identified within our directed search of 242,110 peptide candidates (green diamonds) fall precisely on the saddle splay selection rule trend defined by the α-helical AMPs extracted from the AMP database (black circles) [9]. We plot all 299 α-helical peptides harvested from the database clustered into 31 bins according to mean hydrophobicity in order to smooth the distribution and improve visual clarity. Panels a and b are adapted from Lee et al. Proc. Natl. Acad. Sci. USA 113 48 13588–13593 (2016) [1].