Encodings and models for antimicrobial peptide classification for multi-resistant pathogens

Sequence based encodings

Amino acid composition

An approach to overcome the limitations of sparse encoding and hence making the resulting feature space more dense, is the representation of the amino acid sequence as its respective composition. Here, the final feature vector contains at each position the proportion of an amino acid in relation with the sequence length (Fig. 1). For instance, one can divide a peptide into chunks including both termini and calculate the local amino acid composition [22]. The amino acid composition differs from one class to the another and, for instance, cell penetrating peptides require hydrophobic residues at the N-terminus, which could be approximated well by the features gained from the local composition [1]. Additional performance has been achieved by introducing a technique called distance frequency, which calculates the distance between amino acids of similar properties and bins the occurrence according to the gap length. Matsudo et al. (2005) used both encodings to predict the subcellular location by means of SVMs [22]. Commonly, amino acid composition is applied to distinguish between different classes of peptides, i.e., antimicrobial and non-antimicrobial peptides [23] or to classify antiviral, antitumor, antibacterial, and antifungal peptides [24]. The former introduces quantitative matrices as a novel descriptor, which encodes the propensity of each amino acid at a certain position. This encoding has been employed in addition to local sparse encoding for analysing as well as predicting antimicrobial peptides in general. In contrast, the latter study applied increment of diversity (ID) to classify unknown peptides to the respective classes. To ensure a well-performing classifier, the ID is not only based on the amino acid composition, but is rather used along with the dipeptide and the pseudo-amino acid composition, which will be introduced hereinafter. Dubchak et al. (1995) proposed an encoding, which describes the composition (C), transition (T) and distribution (D) of similar, hence in terms of physicochemical properties, amino acids along the peptide sequence [25]. C refers to the composition of the respective residues, T denotes the frequency of the transition from one group to another and finally, D reflects the distribution of properties within 0, 25, 50, 75 and 100% of the sequence. The CTD-descriptor has been employed to predict protein folding classes [25].

Pseudo-amino acid composition

Sparse encoding and the amino acid composition do not take into account the sequence order effect. This effect considers the vast amount of possible amino acid combinations as the sequence length increases. That is, for a peptide of length 6, there are already 20⁶ = 64,000,000 different sequence arrangements. In terms of antimicrobial activity, Cherkasov et al. (2009) pointed out that, albeit having very similar amino acid compositions, some peptides were virtually inactive [26]. Thus, the pseudo-amino acid composition (PseAAC) has been introduced to consider the effect of the sequence order [27]. The PseAAC computes the correlation between different ranges among a pair of amino acids, which leads to a 20 + λ dimensional vector (Fig. 2a). The first 20 components are the composition of the 20 natural occurring amino acids, whereas the 20 + 1 to 20 + λ components describe the correlation according to the respective sequence order level. For the most contiguous (λ = 1) and the second-most contiguous (λ = 2) amino acids, the PseAAC results in a 22-D (dimensional) vector. Thus, for λ = 1 the sequence order for adjacent amino acids are taken into account. The correlation function incorporates several physicochemical properties, such as the hydrophobicity and amino acid side chain mass. To verify that this method leads to a lower loss of information compared to the usual amino acid composition, several similarity measures have been employed. These include the prediction of subcellular locations of proteins, membrane protein types, as well as their particular locations [27]. To improve prediction accuracy, the PseAAC has been used by several studies, e.g., [28, 29], and [30], in combination with other types of encodings. For instance, in order to predict AMPs and additional efficiencies towards, e.g., cancer cells and HIV, PseAAC was applied in a two-level approach: first, it was used to encode peptide sequences to distinguish between AMPs and non-AMPs and second, to determine additional effects. Both classifications have been conducted by means of fuzzy k-nearest neighbors [28]. Moreover, additional physicochemical properties have been used to enhance the discriminative power of PseAAC [28]. Chen et al. (2016) tried to unveil novel anticancer peptides by enhancing the default dipeptide composition with PseAAC [29]. This approach considers long range interactions between amino acid pairs along with the dipeptide composition. The latter might reflect structural interactions, such as hydrogen bridge bonds between spatial close amino acids to form alpha helices [31]. An extension to the interaction of multiple encodings, including PseAAC, has been conducted by Meher et al. [30]. They used PseAAC in addition to structural and physicochemical encodings in order to distinguish between AMPs and non-AMPs. Again, an SVM was used to conduct the classification [30].

Reduced amino acid alphabet

Sparse encoding, amino acid composition, and PseAAC consider, more or less, the actual amino acid sequence to encode a peptide. Therefore, the encoding might not reflect sequence variations well and this might negatively contribute to the classifier performance. In order to improve generalization, also considering mutations, one could make use of the reduced amino acid alphabet. Here, similar amino acids are grouped together, based on physicochemical, such as hydrophobicity and hydrophilicity [9] or structural properties, e.g., the backbone structure (Fig. 2b) [32]. The reduced amino acid alphabet has been employed in combination with the n-gram model to ease the classification of protein sequences. The n-gram model counts the occurrences of n-mers for an alphabet of size m, leading to a mⁿ dimensional, sparse representation of the initial sequence (Fig. 3). Nevertheless, despite the preceding alphabet reduction, the increased dimensionality is again a major drawback of the n-gram model. Thus, single value decomposition [33] has been applied to reduce the number of features to efficiently train an artificial neural network (ANN). Finally, the ANN is used to assign the query proteins to the respective protein families [9]. Comparable to the n-gram model, the n-peptide composition leads, in particular for an increasing n, to an inflation of the feature space. Yu et al. (2004) used the n-peptide model to predict the subcellular location of proteins in Gram-negative bacteria [34]. For this purpose, the dipeptide, amino acid, as well as the partitioned amino acid composition have been leveraged. For the latter, the sequence is split into equal-length segments and these segments are used to train several SVMs. The assignment of the respective subcellular location is then based on a majority vote of all classifiers [34]. Furthermore, the reduction of the amino acid alphabet, based on structural properties, has been used as the initial step to construct more complex features. These complex features consist of compositional, positional, position-shifted, and correlated features, which are combined through several boolean functions, such as matches and/or matchesAtPosition. The ultimate goal of the study was the prediction of AMPs and their selectivity for different kinds of bacteria and to this end, the complex features are further reduced by means of a filter-based feature selection [35, 36]. Another study uses the k-nearest neighbor clustering to aggregate amino acids into five classes using their amino acid index, i.e., amino acids with the respective highest (T), high (H), medium (M), low (L), and lowest (B) values of a particular physicochemical property are clustered together. This encoding (AAIndexLoc) is extended by the five-level dipeptide composition, which extends the aforementioned clustering by aggregating pairs of amino acids, such as TT, TH, and so forth. Along with these descriptors, Tantoso et al. (2008) employed the amino acid composition, for both termini and the middle part of the peptide, which leads to a dataset of 70 features for an SVM to predict subcellular location [37].

Physicochemical properties

One of the important encodings in AMP prediction, if not the most important one, is the translation of an amino acid to a particular physicochemical property, which have been determined in various wet lab experiments (Fig. 4a). The amino acid index database (AAindex) has been established as a unified source for these descriptors [38]. The AAindex is grouped into three parts, whereby the AAindex1 contains the just mentioned biochemical properties (one for each amino acid) and the AAindex2 aggregates different substitution matrices, such as the PAM250 or the BLOSUM62. The AAindex3 provides protein contact potentials, hence empiric values for spatial close amino acids, such as the Gibbs free energy change, to indicate preferred interactions between residue pairs [39]. The AAindex database, as a consistent source for numerical amino acids indices, has proven its usefulness in several studies. An example is the prediction of transmembrane protein segments [40]. Deber et al. (2001) used, among others, the hydrophobicity scale introduced by Kyte and Doolittle [41], as a reference to their experimental derived values of hydrophobicity [40]. The program annotates α-helical regions in the query sequence, based on the respective hydrophobicity and helix tendency thresholds [40]. So called z-descriptors have been employed as part of the prediction of cell-penetrating peptides [42]. These type of peptides reveal an important property, as they are capable to introduce macromolecules into the cell, which is especially interesting for the pharmaceutical industry [43]. The z-descriptors are derived from the principal components of physicochemical properties by means of partial least squares (PLS) projections [44]. PLS leads to a subset of five final features, capable to describe the 20 proteinogenic as well as 67 additional amino acids. The first three components can be considered as lipophilicity, volume (steric bulk), and polarity, respectively, whereas the fourth and the fifth component are not clearly derivable [44]. These properties are appropriate for the cell-penetrating peptide prediction, due to the intrinsic properties, which are the polarity (positively charged residues are advantageous) as well as the the amphi- and hydrophobicity [42]. However, Hansen et al. (2008) pointed out, that the method benefits from averaging z-descriptors, because that allows to compare sequences with varying length [42]. Nevertheless, to deal with varying protein or peptide sequence lengths, interpolation techniques have been introduced [45]. Sequence interpolation refers to a method, which connects multiple points, that is amino acid indices, via different linear and nonlinear functions. In order to obtain a continuous feature vector, the amino acid sequence is first mapped to the respective physicochemical property, followed by the actual smoothing, employing one of the interpolation functions [45, 46]. Physicochemical representations of peptides have been utilized to classify AMPs and non-AMPs [47]. To this end, Torrent et al. (2011) investigated the different characteristics of antimicrobial peptides, such as the isoelectric point, in-vivo aggregation, and hydrophobicity with respect to their discriminative power [47]. A peptide is described by its different characteristics and the particular averages were fed into an ANN to obtain the class to which the query peptide belongs [47]. In addition, the physicochemical property encoding is employed by various web servers for peptide retrieval, i.e., database queries, as well as for classification. Two examples are AVPpred [48] for antiviral peptide prediction and DBAASP for structure and activity of AMPs [49]. Moreover, this encoding has been used as part of several other studies to predict antimicrobial effects of synthetic peptides [50] or to find substructures with antimicrobial potency in larger proteins [51]. In order to take into account that some traits of AMPs are dependent on particular parts within the sequence, such as a positively charged N-terminus, further studies elucidated the physicochemical property dependence with respect to different sequence sections. One of these studies divided AMPs into datasets for both termini, calculated the physicochemical representation, and finally uses an SVM for classification on the best performing feature subset [52]. Another study leverages pattern changes of amino acid characteristics along a peptide sequence for the prediction of antimicrobial peptides by means of random forests (RF) [53]. An alternative approach, which leverages hydrophobicity values, is designated as the d-descriptor [54]. This encoding is founded on sequence moments, a two dimensional extension of sequence profiles. The amino acid sequence is squeezed between the y- (N-terminus) and the x-axis (C-terminus) with gradually bending of the single amino acids and subsequent vector summation. The length of the vectors arise from the respective property and the angle results from the amino acids orientation in the 2D space. Finally, the sequence moments are mapped to scalar values, which is named the d-descriptor. Juretić et al. (2009) used the latter in order to estimate the therapeutic index, the ratio of hemolytic and antimicrobial activity [54]. Finally, owing to the high dimensional feature vectors, if one uses all possible amino acid indices, several studies, such as [52], performed statistical analysis in order to reduce the features before the accomplishment of the actual experiments. Other studies used techniques such as PCA to obtain the aforementioned z-descriptors as well as factor analysis in order to describe all amino acids with only five factors [55]. Recently, Boone et al. (2018) proposed a classification method by means of the rough set theory [56]. To this end, physicochemical properties have been used to encode the samples and afterwards the algorithm finds suitable boundaries to differentiate between antimicrobial and non-active peptides [56].

Structure based encodings

The secondary structure of a protein or peptide, respectively, is mainly determined by its primary structure, i.e., the order of the amino acids [75]. Moreover, the peptide structure has a strong correlation with antimicrobial activity [76]. Thus, for the prediction of antimicrobial activity, it is reasonable to use sequence-based encodings, but, since the secondary structure cannot be completely derived from the primary structure, it is also conclusive to develop structure-based encodings. In addition, the employing of both descriptors simultaneously, allows the classifier a better generalization and thus improves the overall accuracy [77]. The following section introduces several applications of structure-based encodings.

General structural encodings

Unlike QSAR-based methods, general structural encodings map structure information derived from the whole peptide, to a numerical representation. The peptide structure is described by means of their total 3D shape. This is contrary to QSAR, because instead encoding an amino acid sequence from a molecular viewpoint, the whole peptide structure is considered (Fig. 5a). For instance, Cui et al. (2008) predicted the secretion of proteins into the bloodstream [86]. They used features including physicochemical properties as well as structural information, such as secondary structure, and solvent accessibility. The final prediction has been facilitated by an SVM [86]. Chang et al. (2015) employed conditional random fields (CRF) for probability prediction of critical regions along an AMP sequence [87]. CRFs are an algorithm similar to hidden Markov models, but more variables, such as the surrounding context, can be incorporated. In the present case, several structural descriptors along with physicochemical properties have been used for the prediction. The structure-based encodings encompasses the assignment of predicted secondary structure, conserved protein domains, predicted antimicrobial regions [88] as well as the aggregation tendency [87]. Dybowski et al. (2010) proposed a stacked classifier model to predict the HIV-1 tropism [89]. To this end, the authors trained two independent RFs, whereby the first used hydrophobicity values and the second used the hulls of the electrostatic potentials of the V3 loop, a short peptidic sequence of the viral gp120 protein, as descriptors. The electrostatic hull has been determined in order to acknowledge even subtle differences between different co-receptor tropisms as well as to wrap superimposed shapes of the peptides sub-structure. A third RF combined the output of the other models for the final class assignment [89]. Due to high computational effort during the calculation of the electrostatic potential, Heider et al. (2014) presented an extension of this method [90]. The authors leveraged, that the current model achieves good performance with a constant dielectric value and ionic strength, thus simplifying the calculation of the potential to Coulomb’s law. Finally, the electrostatic potential has been calculated based on the cluster centers. The centroids are determined by all points within a certain distance to the C_α-atoms of the V3 loop [90]. As part of another study, the authors increased the prediction power by means of multiple RFs, combined to an ensemble classifier. The respective classifiers used physicochemical as well as structural properties to predict resistance against a novel HIV-1 maturation inhibitor.

The structural encoding is based on the aforementioned electrostatic potential. In addition, a genetic algorithm has been implemented to find an optimal subset of the physicochemical properties [17]. However, Bozek et al. (2013) pointed out, that the structural encoding of the V3 loop exhibits limitations, since only two physicochemical properties has been used for description [91]. To this end, they proposed a novel encoding, which incorporates structural variations as consequence of sequential rearrangements. Thus, based on the template structure, spheres, whose centers are depicted by reference atoms, are used to enclose spatial related residues of different loop variants. Afterwards, the averaged physicochemical properties of all residues within these regions are used to determine HIV-1 co-receptor usage [91]. In contrast, Sander et al. (2007) introduced an distance distribution approach in order to improve co-receptor tropism based on V3 loops [92]. This method calculates the euclidean distances between each atom type. Afterwards, the respective distances are used to obtain the underlying distribution. Finally, the feature vector is obtained by sampling from this distribution, leading to a final size of each possible combination times samples [92]. Nevertheless, HIV-1 is a very complex organism and hence, several strategies have been tackled in order to combat the virus, such as the aforementioned relation between the V3 loop and tropism as well as between mutations, structure and drug resistance [93]. To this end, another encoding has been developed to describe protein structure based on Delaunay triangulation (Fig. 5b). In essence, the Delaunay triangulation states that, if three points or vertices, respectively, are connected via edges, no further vertex must be located within the circumcircle of these three vertices. This encoding facilitates to encode the complete protein shape by finding the optimal edges between representative points, such as C_α-atoms. Thus, it is able to incorporate information about spatial close residues, which might be lost by a descriptor based on the primary structure only. Finally, the feature vector consists of 210 entries, derived from the adjacency matrix of all amino acid pairs. The respective values are resulting from the averaged distance among these pairs [94]. Albeit this encoding has been mainly employed in the context of computational HIV research, it might work as well for antimicrobial peptides, owing to very good classification results of several studies [95]. To sum up, structural encodings are an appropriate extension to sequence-based encodings since antimicrobial activity is determined by the three-dimensional composition of the residues [96] and in addition, the combination of sequence- and structure-based encodings increases discriminating power [97].

Alternative encodings

There are further encodings, which do not really fit into the proposed categories, i.e., sequence or structural encodings. One of these encodings, which are summarized in Table 3, is the Chaos Game Representation (CGR). In general, the CGR is a visual encoding of a sequence, generating a fractal. The sequence can be obtained, e.g., from random numbers or from biological sequences, such as bases (DNA) and amino acids (proteins). In the case of the former, numbers from 1 to 3 denoting a vertex of a triangle. The algorithm works as follows: firstly, a starting point s is determined and afterwards, one of the numbers is randomly selected as the target vertex t. The next point is located on the half way between s and t. By repeating this procedure, the so called Sierpinski triangle will be generated. The Sierpinski triangle is special about its recursively defined sub-structures, which are also triangles [98]. In the case of the DNA, t is not selected by chance, but rather by the successive base. Here, adenine (A), thymine (T), guanine (G) and cytosine (C) are the labels of a square. After conducting the algorithm, the resulting fractal shows lower order, but still exhibits notably patterns, originated from the underlying sequence. Moreover, points which are close in the CGR do not have to be necessarily adjacent in the sequence, which means that the CGR might introduce novel distance metrics of subsequences [98]. However, with respect to AMPs, CGR has been applied as part of a variety of studies in order to deal with amino acid sequences. As such, Basu et al. (1997) classified similar amino acids to 12 different groups, each representing a target vertex for the CGR algorithm [99]. In addition, the resulting dodecagon has been divided in 24 grids and the amount of points per grid has been used to predict the affiliation to protein families [99]. A further study reduced the amount of vertices to 8, whereby the grouping happened according to the respective physicochemical properties [100]. Moreover, He et al. (2016) extended the illustration to three dimensions, which results in a cube, rather than a planar octagon [100]. The study investigated how this encoding could be employed for multiple sequence alignments. To this end, the authors introduced a method, which computes the euclidean distance between amino acid pairs of two encoded proteins. Finally, the similarity of two proteins is denoted by the sum of the distances [100]. Recently, one study used CGR in a 10D space, using a hypercube for the prediction of anticancer peptides [101] as well as for protein-protein interactions [102].

Another method, which does not fit into the proposed sections has been introduced by Loose et al. (2006) in order to design novel AMPs [103]. In this study, the authors considered AMPs as a corpus of sentences and the goal was to examine, whether antimicrobial activity is described by a certain grammar. To this end, a linguistic model has been derived from active peptides and successfully employed for the design of AMPs [103].

String kernel

Support vector machines (SVM) are capable to efficiently distinguish between binary input data by projecting the data to a higher input space, using kernel techniques [112]. Moreover, these kernel techniques allow a linear separation of a nonlinear classification problem, which is also known as the kernel trick [113]. One type of these kernels are string kernels, which are employed to measure sequence similarity [112]. In essence, the idea of string kernels implies that strings are mapped to a numerical representation in order to be used as input for an SVM. Thus, it is basically another encoding of an amino acid sequence, i.e., a method to map the string representation of peptide sequences to high dimensional feature vectors. Hence, several studies proposed corresponding methods, such as Leslie et al. (2002), who extended the spectrum kernel, in order to incorporate sequence variations, to the mismatch kernel [112]. The former generates all possible subsequences of length k and counts the occurrences of these k-mers within the query sequences, leading to a similarity metric based on shared k-mers [112]. This encoding is similar to the k-peptide composition, for instance the dipeptide composition (k = 2), which has been introduced earlier. The mismatch kernel on the other hand, considers a certain distance, hence mismatches, between two k-mers and takes into account, that similar sequences might have similar properties. Owing to the nature of spectrum kernels, further investigations revealed important and meaningful motifs. As a case study, the authors predicted homolog proteins [114]. Furthermore, string kernels have been applied to predict tumor suppressors, among others. Here, small molecules are encoded in their 1D, 2D, and 3D representations. In 1D, mismatch kernels have been employed to measure the similarity between the atomic sequences [115]. Another study investigated the performance of combined as well as weighted mismatch and structure derived similarity score kernels [116]. For these kernels, each entry in the feature vector is obtained from structure alignments between the input peptide and a peptide database [117]. The encoding incorporates the similarity to further peptides, whereby conserved peptides are depicted with higher scores. Boisvert et al. (2008) proposed an extension of the string kernel, which allows a gap between two k-mers [118]. Thus, the distant segment kernel takes into account the co-occurrence of remote sequence segments. The authors used this kernel in order to predict HIV-1 co-receptor tropism and achieved higher levels of accuracy compared to other methods [118]. Moreover, several string kernels have been employed and compared to predict linear B-cell epitopes [119]. These include the already introduced spectrum and mismatch kernel as well as the local alignment kernel, obtained from local alignment scores, and the subsequence kernel, which measures sequence similarity, similar to the mismatch kernel, albeit gaps within k-mers are taken into account. A third kernel measures the frequency of amino acid pairs (see dipeptide composition), which is due to a bias towards certain dipeptides in B-cell epitopes [119]. Toussaint et al. (2010) recognized that dealing with the sequence only might result in a loss of information [120]. For this reason, the aim of their study was the optimization of existing string kernels such that these involve physicochemical properties [120]. This kernel has been used by another study in conjunction with the penalization of non-adjacent segments, which finally has led to the generic string kernel for small molecules [121]. The authors applied this kernel in a subsequent study in order to detect antimicrobial peptides. All possible peptides with a specific length have been generated by means of source-to-sink graphs. In these graphs, all vertices are k-mers and all edges are weighted according to the antimicrobial activity, computed by means of the generic spectrum kernel. Finally, the detection of the most active peptide corresponds to the detection of the longest path within the graph [122].

Deep learning

Machine learning algorithms based on artificial neural networks, especially deep learning models, have the advantage of incorporating automated encoding, i.e., feature generation. In general, the encoding results from several, successive connected layers, which work as filters for particular parts of the input [5]. However, these models require a large number of training examples in order to generalize well. Fortunately, owing to advances in next-generation sequencing technologies, biological sequences, such as peptides and proteins, are publicly available in vast amounts [123]. Several studies made use of that and showed how deep neural networks perform well on biological problems. For instance, Asgari et al. (2015) proposed a method called protein-vectors, which splits a sequence into k-mers to learn the context of these word representations [124]. Here, amino acid sequences are encoded as a distributed representation of k-mers, which were employed for protein family classification or the prediction of disordered proteins. This approach is derived from natural language processing and uses the context, hence the adjacent residues, for the central k-mers (“words”) syntactic and semantic description. The realization is carried out through building a sufficient large training corpus of protein sequences (“sentences”) by breaking all available sequences into overlapping k-mers. Afterwards, neural networks are used to find optimal, numerical representations, i.e., feature vectors, of the input sequences by means of the skip-gram model. By using these vectors, the authors showed that this framework encodes physicochemical properties well and high levels of accuracy have been achieved in the family classification task [124]. Jiménez et al. (2017) utilized deep learning to predict protein-binding sites [125]. To this end, the structures of proteins are encoded as three-dimensional objects, whereby a cubic segmentation in so-called voxels, which are 3D pixels, takes place beforehand. The encoding of each of these cubes is based on the contained atoms. In order to incorporate physicochemical properties, the input is further upscaled to 8 property channels [125]. A similar approach has been elaborated by Amidi et al. (2018) to predict enzyme classes [126]. Again, protein structures are encoded as voxels and are used as input for a convolutional neural network (CNN), but in contrast to Jiménez et al., the orientation of the protein has been considered. The authors point out, that the structure orientation in the Protein Data Bank (PDB) does not capture the dynamic of the protein and consequently used the proteins barycenter as origin and the first principal components for the orientation of the coordinate system. Overall, the model achieves good accuracy [126]. Another study uses position-specific scoring matrices (PSSM) as 2D input for CNNs, hence mimicking images by regarding the respective substitution probabilities as pixel densities. The studies goal is the automated partitioning of efflux proteins families [127]. This class of proteins provide an important tool for multi-resistant pathogens, because they allow them to convey molecules out of the cell, thus lowering the overall concentration of antibiotics [128]. Two further publications deal with alignment-free comparison of sequences, using CNNs. Both methods encode the input sequences as two-dimensional one-hot matrices, leveraging the convolutional layers for unveiling of latent features. Seo et al. (2018) employed this approach in order to predict protein families [129]. However, Zheng et al. (2018) extended this approach by training of two identical neural networks (siamese neural networks), which allows to compare sequences with respect to their dissimilarity [130]. These two methods, as well as the earlier introduced ProtVec [124], have in common that they aggregate amino acid sequences of varying lengths to a fixed-length numeric vector of lower dimension. Since this feature reduction keeps intrinsic properties of the proteins, these algorithms might serve as potential encodings for AMPs. Similar to this CNN based dimension reduction are autoencoders. Autoencoders are applied to learn a dense representation of the input, i.e., to extract representative characteristics in order to reproduce the input as good as possible. For instance, Wang et al. (2017) employed stacked autoencoders to predict protein-protein interactions [131].

Databases

Piotto et al. (2012) presented YADAMP (yet another database of antimicrobial peptides) [132]. The authors collected the data sets, i.e., AMPs, from various, published studies. Potential hits can be limited, e.g., by specifying certain physicochemical properties and/or target organisms. Respective results provide more details with respect to activity and structural properties [132]. CAMP (collection of antimicrobial peptides) obtains AMP sequences and structures from well-known protein databases, such as UniProtKB [133]. Active peptides have been filtered out via keyword search. By providing several links to further web services, CAMP is a comprehensive resource for AMPs as well as active peptides in general [133, 134]. Wang et al. (2016) published the third update for the antimicrobial peptide database (APD3) [135]. Besides its focus on natural occurring AMPs, this database stores various active peptides, e.g., anti-HIV, spermicidal, and for wound healing. A web form lets the user specify custom query parameters, such as physicochemical properties [135]. Pirtskhalava et al. (2016) extended the database of antimicrobial activity and structure of peptides to the second version (DBAASPv.2) [49]. The service provides, among further details, potency values against several pathogens, described by inhibition coefficients. Moreover, the authors conducted molecular modeling for unveiling unknown structures of AMPs [49]. Finally, a comprehensive data repository of antimicrobial peptides (DRAMP) has been set up by Fan et al. (2016) [136]. They included additional features, hence similarity search, sequence alignment, and conserved domain search, besides established tools, which already have been introduced by other [136]. More information about web services for AMP retrieval can be found in two recent studies, published by Porto et al. [137] and Gabere et al. [138].

Packages

As mentioned before, many of the sequence-based encodings have been implemented in user-friendly packages, using, e.g., R² or Python.³ Interpol is an R-package for normalizing peptide sequences to a uniform length, using different interpolation methods and descriptors of the AAindex database [45]. Cao et al. (2013) developed propy, which provides Python access to methods for amino acid composition, autocorrelation and pseudo-amino acid composition (PseAAC), among others [139]. In contrast, protr, implemented by Xiao et al. (2015), provides similar methods for the R programming language [140]. In addition, all methods can be accessed through a public web server. However, the web interface lacks the possibility of passing custom method parameters and is hence only recommended for ad-hoc calculations [140]. Ofer et al. (2015) released ProFET, i.e., protein feature engineering toolkit, a Python-based distribution with a variety of ready-to-use amino acid encodings [141]. Among default encodings, which have been implemented by others, this package offers also reduced amino acid alphabet, autocorrelation, amino acid propensities, as well as transformed CTD features [141]. modlAMP is a Python library specifically developed for antimicrobial peptides. Besides a selective choice of encodings, Müller et al. (2017) added methods for the whole prediction pipeline, i.e., sequence retrieval, visualization, and machine learning algorithms [142]. Moreover, performant model parameters can be obtained automatically via a grid search [142]. In contrast, POSSUM (position-specific scoring matrix-based feature generator for machine learning) is a toolkit, which facilitates the representation of amino acids with PSSM derived encodings [143]. Wang et al. (2017) published POSSUM as a public web server as well as a Perl/Python-based tool, executable via the command line [143]. PyBioMed is another Python library foremost aiming cheminformaticians, owing to the fact, that many molecular encodings are implemented, e.g., topological descriptors, applicable in QSAR studies [144]. Nevertheless, Dong et al. (2018) rounded out this package with a variety of amino acid encodings and additional tools, such as sequence and structure retrieval [144]. Recently, Chen et al. (2018) published iFeature, which is accessible as a Python package and web server [145]. This tool adds functionality in order to encode amino acids based on AAindex entries as well as structure-based encodings, such as accessible surface area and main-chain-torsional angles. Moreover, algorithms for clustering, feature selection, and dimensionality reduction are available [145].