Plant Antimicrobial Peptides: State of the Art, In Silico Prediction and Perspectives in the Omics Era

Abstract

Even before the perception or interaction with pathogens, plants rely on constitutively guardian molecules, often specific to tissue or stage, with further expression after contact with the pathogen. These guardians include small molecules as antimicrobial peptides (AMPs), generally cysteine-rich, functioning to prevent pathogen establishment. Some of these AMPs are shared among eukaryotes (eg, defensins and cyclotides), others are plant specific (eg, snakins), while some are specific to certain plant families (such as heveins). When compared with other organisms, plants tend to present a higher amount of AMP isoforms due to gene duplications or polyploidy, an occurrence possibly also associated with the sessile habit of plants, which prevents them from evading biotic and environmental stresses. Therefore, plants arise as a rich resource for new AMPs. As these molecules are difficult to retrieve from databases using simple sequence alignments, a description of their characteristics and in silico (bioinformatics) approaches used to retrieve them is provided, considering resources and databases available. The possibilities and applications based on tools versus database approaches are considerable and have been so far underestimated.

Introduction

Proteins and peptides play different roles depending on their amino acids (aa) constitution, which may vary from tens to thousands residues.¹ Peptides are conventionally understood as having less than 50 aa.² Proteins, on the contrary, would be any molecule presenting higher amino acid content and both—proteins and peptides—present a plethora of variations in plants. Despite that, plant proteomes have been much more studied than peptidomes. It is well-known that the biochemical machinery necessary for the synthesis and metabolism of peptides is present in every living organism. From the variations of this machinery, a wide structural and functional diversity of peptides was generated, justifying the growing interest in their study.

In eukaryotes, peptides are prevalent in intercellular communication, performing as hormones, growth factors, and neuropeptides, but they are also present in the defense system.³ Besides plants and animals, several pathogenic microorganisms, peptides can serve as classical virulence factors, which disrupt the epithelial barrier, damage cells, and activate or modulate host immune responses. An example of this performance is represented by Candidalysin,⁴ a fungal cytolytic peptide toxin found in the pathogenic fungus Candida albicans that damages epithelial membranes, triggers a response signaling pathway, and activates epithelial immunity. There are also reports of defense-related fungal peptides. For example, the Copsin, a peptide-based fungal antibiotic recently identified in the fungus Coprinopsis cinerea⁵ kills bacteria by inhibiting their cell wall synthesis. Regarding bacterial peptides, certain species from the gastrointestinal microbial community can release low-molecular-weight peptides, able to trigger immune responses.⁶ There are additionally peptides that act like bacterial “hormones” that allow bacterial communities to organize multicellular behavior such as biofilm formation.⁷ Some peptides are known for their medical importance, as defensins that present antibacterial, antiviral, and antifungal activities. For example, human alpha- and beta-defensins present in the saliva may potentially impede virus replication, including SARS-CoV-2,⁸ besides other roles as protection against intestinal inflammation (colitis).⁹

Considering the roles of plant peptides, they can also be multifunctional, and have been classified into 2 main categories¹⁰ (Supplementary Figure S1): (1) Peptides with no bioactivity, primarily resulting from the degradation of proteins by proteolytic enzymes, aiming at their recycling, and (2) bioactive peptides, which are encrypted in the structure of the parent proteins and are released mainly by enzymatic processes.

The first group is innocuous regarding signaling, regulatory functions, and bioactivity. So far, it has been reported that some of them may play a significant role in nitrogen mobilization across cellular membranes.¹¹ The second group of bioactive peptides has a substantial impact on the plant cell physiology. Some peptides of this group can act in the plant growth regulation (through cell-to-cell signaling), endurance against pathogens and pests by acting as toxins or elicitors, or even detoxification of heavy metals by ion-sequestration.

Comprising bioactive peptides, additional subcategorization has been proposed regarding their function. Tavormina et al¹² divided them into 3 groups (Supplementary Figure S1) based on the type of precursor:

• Derived from functional precursors: originated from a functional precursor protein;

• Derived from nonfunctional precursors: originated from a longer precursor that has no known biological function (as a preprotein, proprotein, or preproprotein);

• Not derived from a precursor protein: some sORFs (small Open Read Frames; usually <100 codons) are considered to represent a potential new source of functional peptides (known as “short peptides encoded by sORFs”).

A more intuitive classification of bioactive peptides was further proposed by Farrokhi et al¹⁰ according to their intracellular role (Supplementary Figure S1):

• Phytohormone peptides: the characteristic feature of these peptides is the regulation of fundamental plant physiological processes. They can be classified into those with (1) signaling roles in non-defense functions or those with (2) signaling roles in plant defense. Concerning the first group (Supplementary Figure S1), the peptide CLE25 (CLAVATA3/EMBRYO-SURROUNDING REGION-RELATED 25) is one of the representatives. This peptide transmits water-deficiency signals through vascular tissues in Arabidopsis thaliana, affecting abscisic acid biosynthesis and stomatal control of transpiration in association with BAM (BARELY ANY MERISTEM) receptors in leaves.¹³ Another example is the PLS (POLARIS) peptide that acts during early embryogenesis but later activates auxin synthesis, also affecting cytokine synthesis and ethylene response.¹⁴ Regarding the second group, it includes peptides with signaling roles in plant defense, comprising at least 4 subgroups, including SYST (systemin) (Supplementary Figure S1). The SYST peptides were identified in Solanaceae members, like tomato and potato¹⁵ (acting on the signaling response to herbivory). The SYST leads to the production of a plant protease inhibitor that suppresses insect’s proteases.¹⁶ Stratmann¹⁷ suggested that in plants, SYSTs act to stimulate the jasmonic acid signaling cascade within vascular tissues to induce a systemic wound response.

• Defense peptides or antimicrobial peptides (AMPs): to be fitted into this class, a plant peptide must fulfill some specific biochemical and genetic prerequisites. Regarding biochemical features, in vitro antimicrobial activity is required. Concerning the genetic condition, the gene encoding the peptide should be inducted in the presence of infectious agents.¹⁸ In practice, this last requirement is not ever fulfilled as some AMPs are tissue-specific and are considered as part of the plant innate immunity, while other isoforms of the same class appear induced after pathogen inoculation.¹⁹

Plant AMPs are the central focus of the present review, comprising information on their structural features (at genomic, gene, and protein levels), resources, and bioinformatic tools available, besides the proposition of an annotation routine. Their biotechnological potential is also highlighted in the generation of both transgenic plants resistant to pathogens, and new drugs or bioactive compounds.

Overall Features of Plant AMPs

Antimicrobial peptides are ubiquitous host defense weapons against microbial pathogens. The overall plant AMP characterization regards the following variables (Figure 1): electrical charge, hydrophilicity, secondary and 3-dimensional (3D) structures, and the abundance or spatial pattern of cysteine residues.²⁰ These features are primarily related to their defensive role(s) as membrane-active antifungal, antibacterial, or antiviral peptides.

Figure 1.

Plant antimicrobial peptide features considering DNA sequence level, protein structure, and physicochemical properties. LTP indicates lipid transfer protein.

Regarding the nucleotide sequence, plant AMPs are hypervariable and this genetic variability is considered crucial to provide diversity and the ability to recognize different targets. For their charges, AMPs can be classified as cationic or anionic (Figure 1). Most plant AMPs have positive charges, which is a fundamental feature for the interaction with the membrane lipids of pathogens.²¹ Concerning hydrophilicity, AMPs are generally amphipathic, that is, they exhibit molecular conformation with both hydrophilic and hydrophobic domains.²² With respect to their 3D structure, AMPs can be either linear or cyclic (Figure 1). Some linear AMPs adopt an amphipathic α-helical conformation, whereas non-α-helical linear peptides generally show 1 or 2 predominant amino acids.²³ In turn, cyclic AMPs, including cysteine-containing peptides, can be divided into 2 subgroups based on the presence of a single or multiple disulfide bonds. A peculiar feature of these peptides is a cationic and amphipathic character, what improves their functioning as membrane-permeabilizing agents.²³

Considering the secondary structures, AMPs may exhibit α-helices, β-chains, β-pleated sheets, and loops (Figure 1). Wang²⁴ classified plant AMPs into 4 families (α, β, αβ, and non-αβ), based on the protein classification of Murzin et al,²⁵ with some modifications. Antimicrobial peptides of the “α” family present α-helical structures,¹ whereas AMPs from the “β” family contains β-sheet structures usually stabilized by disulfide bonds.^26,27 Some plant AMPs show an α-hairpinin motif formed by antiparallel α-helices that are stabilized by 2 disulfide bridges.²⁸ Such AMPs present a higher resistance to enzymatic, chemical, or thermal degradation.²⁹ Antimicrobial peptides from the “αβ” family having both “α” and “β” structures are also stabilized by disulfide bridges. An example of AMP presenting “αβ” structures are defensins, usually with a cysteine-stabilized αβ motif (CSαβ), an α-helix, and a triple-stranded antiparallel β sheet stabilized mostly by 4 disulfide bonds.³⁰ Finally, AMPs that do not belong to the “αβ” group exhibit no clearly defined “α” or “β” structures.²⁶

Plant AMPs are also classified into families considering protein sequence similarity, cysteine motifs, and distinctive patterns of disulfide bonds, which determine the folding of the tertiary structure.³¹ Therefore, plant AMPs are commonly grouped as thionins, defensins, heveins, knottins (linear and cyclic), lipid transfer proteins (LTP), snakins, and cyclotides.^27,31 These AMP categories will be detailed in the next sections, together with other groups here considered (Impatien-like, Macadamia [β-barrelins], Puroindoline (PIN), and Thaumatin-like protein [TLP]) and the recently described α-hairpinin AMPs. The description includes comments on their structure, pattern for regular expression (REGEX) analysis (when available), functions, tissue-specificity, and scientific data availability.

Thionin

Thionins are composed by 45 to 48 amino acid residues with a molecular weight around 5 kDa, considering the mature peptide. They are synthesized with a signal peptide together with the mature thionin and the so-called acidic domain.³² To date, there is no experimental information available about possible functions of the acidic domain, even though it is clearly not dispensable as shown by the high conservation of the cysteine residues.³³

The thionin superfamily comprises 2 distinct groups of plant peptides α/β-thionins and γ-thionins with distinguished structural features.³⁴ The α/β thionins have homologous amino acid sequences and similar structures.³⁵ Besides, they are rich in arginine, lysine, and cysteine.³⁶ In turn, γ-thionins have a greater similarity with defensins, and some authors classify them within this group.³⁷ However, compared with the defensins, they present a longer conserved amino acid sequence.³¹

Regarding the cysteine motif, it can be divided into 2 subgroups, one with 8 residues connected by 4 disulfide bonds called 8C and the other with 6 residues connected by 3 disulfide bonds called 6C.³⁸ The general designation of thionins has been proposed as a family of homologous peptides that includes purothionins. The first plant thionin was isolated in 1942 from wheat flour and labeled as purothionin.³⁹ Since then, homologues from various taxa have been also identified, like viscotoxins (Viscum album) and crambins (Crambe abyssinica).⁴⁰ They have also been isolated from different plant tissues like seeds, leaves, and roots.^41,42

Thionins have been tested for different elicitors: Gram-positive^43,44 or Gram-negative bacteria,^45,46 yeast,^38,43 insect larvae,⁴⁷ nematode,³³ and inhibitory proteinase.⁴⁸ Thionins are hydrophobic in nature, interact with hydrophobic residues, and lyse bacteria cell membrane. Their toxicity is due to an electrostatic interaction with the negatively charged membrane phospholipids, followed by either pore formation or a specific interaction with a membrane.³⁸ It has been reported that they are able to inhibit other enzymes possibly through covalent attachment mediated by the formation of disulfide bonds, as previously observed for other thionin/enzyme combinations.⁴⁸

Thionin representatives with known 3D structures determined by X-ray crystallography are crambin (PDB ID: 1CRN), α1- and β-purothionins (PDB ID: 2PHN and 1BHP), β-hordothionin (PDB ID: 1WUW), and viscotoxin-A3 (PDB ID: 1OKH). The first to be determined was the mixed form of crambin.^35,49 It showed a distinct capital Γ shape with the N terminus forming the first strand in a β-sheet. The architecture of this sheet is additionally strengthened by 2 disulfide bonds.⁵⁰ After a short stretch of extended conformation, there is a helix-turn-helix motif. In crambin, there is a single disulfide involved in stabilizing the helix-to-helix contacts. At the center of this motif, there is a crucial Arg10 that forms 5 hydrogen bonds to tie together the first strand, the first helix, and the C terminus.⁵⁰

Defensin

The first plant defensins were isolated from wheat⁵¹ and barley grains,⁵² initially called γ-hordothionins. Due to some similarities in cysteine content and molecular weight, they were classified as γ-thionins. Later, the term “γ-thionin” was replaced by “defensin” based on the higher number of primary and tertiary structures of these proteins and also on their antifungal activities more related to insect and mammalian defensins than to plant thionins.⁵³ Plant defensins belong to a diverse protein superfamily called cis-defensin⁵⁴ and exhibit cationic charge, consisting of 45 to 54 aa with 2 to 4 disulfide bonds.^53,55 Plant defensins share similar tertiary structures and typically exhibit a triple-stranded antiparallel β sheet, enveloped by an α-helix and confined by intramolecular disulfide bonds¹ (Figure 2A). This motif is called cysteine-stabilized αβ (CSαβ).⁵⁶ The CSαβ defensins were classified into 3 groups based on their sequence, structure, and functional similarity.

Figure 2.

Three-dimensional structure of plant antimicrobial peptide representatives. Yellow arrows correspond to β-sheets; α-helices are represented in purple and disulfide bonds in red. (A) Defensin NaD1 of Nicotiana alata (PDB ID: 1MR4). (B) Lipid transfer protein TaLTP1.1 isolated from wheat (PDB ID: 1GH1). (C) Hevein of Hevea brasiliensis (PDB ID: 1HEV). (D) Knottin Ep-AMP1 of Echinopsis pachanoi (PDB ID: 2MFS). (E) MiAMP1 peptide of Macadamia integrifolia (PDB ID: 1C01). (F) Snakin-1 of Solanum tuberosum (PDB ID: 5E5Q). (G) Thaumatin from Thaumatococcus daniellii (1RQW). (H) Zeamatin of Zea mays (1DU5, chain A). (I). Theoretical 3-dimensional model of Impatiens balsamina IbAMP1. (J) Helical hairpin structure of the novel antimicrobial peptide EcAMP1 from seeds of barnyard grass (Echinochloa crus-galli) (PDB ID: 2L2R). PDB indicates protein data bank.

Defensins are known for their antimicrobial activity at low micromolar concentrations against Gram-positive and Gram-negative bacteria,⁵⁷ fungi,⁵⁸ viruses, and protozoa.⁵⁹ In addition, they present protein inhibition, insecticidal, and antiproliferative activity, acting as an ion-channel blocker, being also associated with the inhibition of pathogen protein synthesis.⁶⁰ Instead, plant defensins act in the regulation of signal transduction pathways and induce inflammatory processes, in addition to wound healing, proliferation control, and chemotaxis.⁶¹

In general, plant defensins do not present high toxicity to human cells, having in vivo efficacy records, with relevant therapeutic potential, and can be applied in treatments associated with traditional medicine.⁶² Cools et al⁶³ reported that a peptide derived from a plant defensin (HsAFP1) acted synergistically with caspofungin (an antimycotic) (in vivo and in vitro) against the formation of Candida albicans biofilm on polystyrene and catheter substrate, indicating that the HsAFP1 variant presented a strong antifungal potential in the proposed treatment.

Other biotechnological applications of defensins are described, as in the case of EcgDf1, which was isolated from a legume (Erythrina crista-galli), heterologously expressed in Escherichia coli and purified. EcgDf1 inhibited the growth of various plant and human pathogens (such as Candida albicans and Aspergillus niger and the plant pathogens Clavibacter michiganensis ssp. michiganensis, Penicillium expansum, Botrytis cinerea and Alternaria alternate).⁶⁴ Due to these features, EcgDf1 is a candidate for the development of antimicrobial products for both agriculture and medicine.⁶⁴

Lipid transfer protein

Non-specific lipid transfer proteins (ns-LTPs) were first isolated from potato tubers⁶⁵ and are actually identified in diverse terrestrial plant species. They comprise a large gene family, are abundantly expressed in most tissues, but absent in most basal plant groups as chlorophyte and charophyte green algae.⁶⁶ They generally include an N-terminal signal peptide that directs the protein to the apoplastic space.⁶⁷ Some LTPs have a C-terminal sequence that allows their post-translational modification with a glycosylphosphatidylinositol molecule, facilitating the integration of LTP on the extracellular side of the plasma membrane.

The ns-LTPs are small proteins which were thus named because of their function of transferring lipids between the different membranes carrying lipids (non-specifically, the list includes phospholipids, fatty acids, their acylCoAs, or sterols). They consist of approximately 100 aa and are relatively larger in size than other AMPs, such as defensins.

Depending on their sizes, LTPs may be classified into 2 subfamilies: LTP1 and LTP2, with relative molecular weight of 9 and 7 kDa, respectively.^68,69 The limited sequence conservation turned this classification inadequate. Thus, a modified and expanded classification system was proposed, presenting 5 main types (LTP1, LTP2, LTPc, LTPd, and LTPg) and 5 additional types with a smaller number of members (LTPe, LTPf, LTPh, LTPj, and LTPk).⁶⁶ The new classification system is not based on molecular size but rather on (1) the position of a conserved intron, (2) the identity of the amino acid sequence, and (3) the spacing between the cysteine (Cys) residues (Supplementary Figure S2). Although this latter classification system is the most recent, the conventional classification of LTP1 and LTP2 types has been maintained by most working groups.

Lipid transfer protein nomenclature has been confusing and without consistent guidelines or standards. There are several examples where specific LTPs receive different names in different scientific articles. The lack of a robust terminology sometimes turns it quite difficult, extremely time-consuming, and frustrating to compare LTPs with different roles/functions.⁶⁷ Therefore, an additional nomenclature was also proposed by Salminen et al,⁶⁷ naming LTPs as follows: AtLTP1.3, OsLTP2.4, HvLTPc6, PpLTPd5, and TaLTPg7, with the first 2 letters indicating the plant species (eg, At = Arabidopsis thaliana, Pp = Physcomitrella patens); LTP1, LTP2, and LTPc indicating the type; while the last digit (here 3-7) regards the specific number given to each gene or protein within a given LTP type. For the sake of clarity, the authors recommend the inclusion of a point between the type specification (LTP1 and LTP2) and the gene number. For LTPc, LTPd, LTPg, and other types of LTP defined with a letter, the punctuation mark was not recommended. This latter classification system is currently recommended as it comprises several features of LTPs and is more robust than the previous classification systems.

Lipid transfer proteins are small cysteine-rich proteins, having 4 to 5 helices in their tertiary structure (Figure 2B), which is stabilized by several hydrogen bonds. Such a folding gives LTPs a hydrophobic cavity to bind the lipids through hydrophobic interactions. This structure is stabilized by 4 disulfide bridges formed by 8 conserved cysteines, similar to defensins, although bound by cysteines in different positions. The disulfide bridges promote LTP folding into a very compact structure, which is extremely stable at different temperatures and denaturing agents.^70-72 These foldings provide a different specificity of lipid binding at the LTP binding site, where the LTP2 structure is relatively more flexible and present a lower lipid specificity when compared with LTP1.³⁴

The first 3D structure of an LTP was established for TaLTP1.1 based on 2D and 3D data of 1H-NMR, purified from wheat (Triticum aestivum) seeds in aqueous solution.^73,74 Currently, several 3D structures of LTPs have been determined, either by nuclear magnetic resonance (NMR) or X-ray crystallography; in their free, unbound form or in a complex with ligands.

Hevein

The heveins were first identified in 1960 in the rubber-tree (Hevea brasiliensis), but its sequence was determined later, whereas a similarity was detected to the chitin-binding domain of an agglutinin isolated from Urtica dioica (L.)⁷⁵ with 8 cysteine residues forming a typical Cys motif.⁷⁶ The primary structure of the hevein consists of 29 to 45 aa, positively charged, with abundant glycine (6) and cysteine (8-10) residues,⁷⁶ and aromatic residues.^31,77 The chitin-binding domain is a determinant component in the identification of hevein-like peptides whose binding site is represented by the amino acid sequence SXFGY/SXYGY, where X regards any amino acid.^76,78 Most heveins have a coil-β1-β2-coil-β3 structure that occurs by variations with the secondary structural motif in the presence of turns in 2 long coils in the β3 chain.³¹ Antiparallel β chains form the central β sheet of the hevein motif with 2 long coils stabilized by disulfide bonds (Figure 2C).

Although the presence of chitin has not been identified in plants, there are chitin-like structures present in proteins that exhibit a strong affinity to this polysaccharide isolated from different plant sources.⁷⁹

The presence of 3 aromatic amino acids in the chitin-binding domain favors chitin binding by providing stability to the hydrophobic group C-H and the π electron system through van der Waals forces, as well as the hydrogen bonds between serine and N-acetylglucosamine (GlcNAc) present in the chitin structure.^76,77 This domain is commonly found in chitinases of classes I to V, in addition to other plant antimicrobial proteins, such as lectins and PR-4 (Pathogenesis-Related protein 4) members.^80,81 It may also occur in other proteins that bind to polysaccharide chitin,⁸⁰ such as the antimicrobial proteins AC-AMP1 and AC-AMP2 of Amaranthus caudatus (Amaranthaceae) seeds which are homologous to hevein but lack the C-terminal glycosylated region.⁸² Plant chitinases (class I) have the hevein-like domains, called HLDs. Due to the similar structural epitopes between chitinases and heveins, they are responsible for the cross reactive syndrome (latex-fruit syndrome).^83,84

Among the several classes of proteins mentioned, the proteins with a high degree of similarity to hevein are chitinases I and IV.⁷⁶ Chitinases are known to play an essential role in plant defense against pathogens,⁸⁵ also inhibiting in vitro fungal growth,⁸⁶ especially when combined with β-1,3-glucanases.⁸⁷ It also interferes with the growth of hyphae, resulting in abnormal ramification, delay, and swelling in their stretching.⁸¹ However, it has been shown that heveins have a higher inhibitory potential than chitinases and that their antifungal effect is not related only to the presence of chitinases⁸⁸; Pn-AMP1 and Pn-AMP2 AMPs with hevein domains have potent antifungal activities against a broad spectrum of fungi, including those without chitin in their cell walls.^88,89 Modes of action of chitinases usually include degradation and disruption of the fungal cell wall and plasma membrane due to its hydrolytic action, causing extravasation of plasma particles.^21,89 Therefore, heveins have good antifungal activity, and only a few are active against bacteria, most of them with low activity.

Another role of hevein chitinases regards the antagonistic effect in triggering the aggregation of rubber particles in the latex extraction process in rubber trees. Unlike heveins, other chitinases inhibit rubber particle aggregation. However, its action in conjunction with other proteins (β-1,3-glucanase) increases the effect of β-1,3-glucanase on rubber particle aggregation.⁹⁰ A study by Shi et al⁹¹ found that the interaction of the protein network related to the antipathogenic activity released by lutoids (lysosomal microvacuole in latex) is essential in closing laticiferous cells (cells that produce and store latex), not only providing a physical barrier, but a biochemical barrier used by laticiferous cells affected by pathogen invasion.

Knottins (cysteine-knot peptides)

Knottins are part of the cysteine-rich peptides (CRPs) superfamily, sharing the Cysteine-knot motif and therefore resembling other families as defensins, heveins, and cyclotides.⁹² Their structure was initially identified by crystallography of carboxypeptides isolated from potato, showing the Cysteine-knot motif with 39 aa and 6 cysteine residues.⁹³ They are also called “Cysteine-knot peptides,” “inhibitor Cysteine-knot peptides,” or even “Cysteine-knot miniproteins” because their mature peptide presents less than 50 aa, forming 3 interconnected disulfide bonds in the Cysteine-knot motif, characterizing a particular scaffold.⁹² This conformation confers thermal stability at high temperatures. For example, the cysteine-stabilized β-sheet (CSB) motif derived from knottins presents stability at approximately 100°C with only 2 disulfide bonds.⁹⁴ The knottins may have linear or cyclic conformation. However, both exhibit connectivity between the cysteines at positions 1-4C, 2-5C, and 3-6C, forming a ring at the last bridge⁹² (Figure 2D).

Knottins have different functions, such as signaling molecules,⁹⁵ response against biotic and abiotic stresses,⁹⁶ root growth,⁹⁷ symbiotic interactions as well as antimicrobial activity against bacteria,⁹⁸ fungi,⁹⁹ virus,¹⁰⁰ and insecticidal activity,¹⁰¹ among others. Knottins antimicrobial activity has been attributed to the action of functional components of the plasma membrane, leading to alterations of lipids, ion flux, and exposed charge.⁹⁹ The accumulation of peptides on the surface of the membrane results in the weakening of the pathogen membrane,¹⁰² resulting in transient and toroidal perforations.⁹⁹

Macadamia (β-barrelins)

In the course of a large-scale survey to identify novel AMPs from Australian plants,^103,104 an AMP with no sequence homology was purified. Its complementary DNA (cDNA) was cloned from Macadamia integrifolia (Proteaceae) seeds, containing the complete peptide coding region. The peptide was named MiAMP1, being highly basic with an estimated isoelectric point (pI) of 10 and a mass of 8 kDa.

The MiAMP1 is 102 aa long, including a 26 aa signal peptide in the N-terminal region, bound to a 76 aa mature region with 6 cysteine residues.¹⁰⁵ Its 3D structure was determined using NMR spectroscopy,¹⁰⁴ revealing a unique conformation among plant AMPs, with 8 beta-strands arranged in 2 Greek key motifs, forming a Greek key beta-barrel (Figure 2E). Due to its particularities, MiAMP1 was classified as a new structural family of plant AMPs, and the name β-barrelins was proposed for this class.¹⁰⁴ This structural fold resembles a superfamily of proteins called γ-crystallin-like characterized by the precursors βγ-crystallin.¹⁰⁶ This family includes AMPs from other organisms, for example, WmKT, a toxin produced by the wild yeast Williopsis mraki.¹⁰⁷

The MiAMP1 exhibited in vitro antimicrobial activity against various phytopathogenic fungi, oomycetes, and gram-positive bacteria¹⁰³ with a concentration range of 0.2 to 2 μM generally required for a 50% growth inhibition (IC50). In addition, the transient expression of MiAMP1 in canola (Brassica napus) provided resistance against blackleg disease caused by the fungus Leptosphaeria maculans,¹⁰⁸ turning MiAMP1 potentially useful for genetic engineering aiming at disease resistance in crop plants.

There are few scientific publications with Macadamia-like peptides, maybe because they prevail in primitive plant groups (eg, lycophytes, gymnosperms to early angiosperms as Amborella and Papaver), being apparently absent in derived angiosperms (eg, Asteridae, including Brassicaceae as Arabidopsis thaliana). On the contrary, they have been identified in some monocots (as Zantedeschia, Zea, and Sorghum).¹⁰⁹ In fact, peptides similar to MiAMP1 appear to play a role in the defense against pathogens in gymnosperms, including species of economic importance (as Pinus and Picea) thus deserving attention for their biotechnological potential.¹⁰⁹

Impatiens-like (Ib-AMPs)

Four closely related AMPs (Ib-AMP1, Ib-AMP2, Ib-AMP3, and Ib-AMP4) were isolated from seeds of Impatiens balsamina (Balsaminaceae) with antimicrobial activity against a variety of fungi and bacteria, and low toxicity to human cells in culture. These AMPs are the smallest isolated from plants to date, consisting of only 20 aa in length. The Ib-AMPs are highly basic and contain 4 cysteine residues that form 2 disulfide bonds. Interestingly, they have no significant homology with other AMPs available in public databases. Sequencing of cDNAs isolated from I. balsamina revealed that all 4 peptides are encoded within a single transcript. Concerning the predicted precursor of Ib-AMP protein, it consists of a pre-peptide followed by 6 mature peptide domains, each one of them flanked by propeptide domains ranging from 16 to 35 aa in length (Supplementary Figure S3). This primary structure with repeated domains of alternating basic peptides and acid propeptide domains has, to date, not been reported in other plant species.¹¹⁰

Patel et al¹¹¹ conducted an experiment to purify Ib-AMP1 from seeds of Impatiens balsamina. After purification, this peptide had its secondary structure tested by Circular Dichroism (CD). The results revealed a peptide that may include a β-turn but do not show evidences for either helical or β-sheet structure over a range of temperature and pH. Structural information from 2D 1H-NMR was obtained in the form of proton-proton internuclear distances inferred from nuclear overhauser enhancements (NOEs) and dihedral angle restraints from spin-spin coupling constants, which were used for distance geometry calculations. Owing to the difficulty in obtaining the correct disulfide connectivity by chemical methods, the authors had built and performed 3 separate calculations: (1) a model with no disulfides; (2) another with predicted disulfide bonds; and (3) a model with alternative connectivity disulfide, as assigned from the Nuclear Overhauser Effect spectroscopy (NOESY) NMR spectra. As a result, 2 hydrophilic patches were observed at opposite ends and opposite sides of the models, whereas in between them a large hydrophobic patch was identified. However, the study did not conclude which of the 3 models would be the most likely representative of Ib-AMP1, reporting only that cysteines are necessary for maintaining the structure.

Based on the experiment performed by Patel et al,¹¹¹ the present work built 3 different models: Model 1: without disulfide bonds, and the other 2 models with different disulfide connections—Model 2: NMR prediction by Patel et al¹¹¹ 6-Cys;16-Cys and 7-Cys;20-Cys, and Model 3: Disulfide Bond partner prediction by DiANNA 7-Cys;16-Cys and 6-Cys;20-Cys. Calculations have shown that although the peptide is small, the cysteines constrain part of it to adopt a well-defined main chain conformation. From residue 4 to 20 (except 11), the main chain is well-defined, whereas residues 1 to 3 in the N-terminal region present few restrictions and appear to be more flexible (Supplementary Figure S4). Analyzing the RMSD (root mean square deviation), we observed that all the models lost the initial conformation and, among them, Model 3 was the most stable. Models 1 and 2 showed a similar pattern (Supplementary Figure S5), as in the models of Patel et al,¹¹¹ although Model 1 was the most flexible.

Little is known about Impatiens-like AMPs mode of action. Lee et al¹¹² investigated the antifungal mechanism of Ib-AMP1 noting that when oxidized (bound by disulfide bridges), there occurs a 4-fold increase in antifungal activity against Aspergillus flavus and Candida albicans, as compared with reduced Ib-AMP1 (without disulfide bridges). Confocal microscopy analyses have shown that Ib-AMP1 can either bind to the cell surface or penetrate cell membranes, indicating an antifungal activity by inhibiting a distinct cellular process, rather than ion channel or membrane pore formation. Fan et al¹¹³ reported the Ib-AMP4 antimicrobial activity dependent of β-sheet configuration to enable insertion into the lipid membrane, thus killing the bacteria through a non-lytic mechanism.¹¹⁴

Current approaches aim to make changes in Ib-AMP to improve its antimicrobial activity. As an example, synthetic variants of Ib-AMP1 were fully active against yeasts and fungi, where the replacement of amino acid residues by arginine or tryptophan improved more than twice the antifungal activity.¹¹⁵ Another study involving AMP modification generated a synthetic peptide without the disulfide bridges (ie, a linear analog of Ib-AMP1), which showed an antimicrobial specificity 3.7 to 4.8 times higher than the wild-type Ib-AMP1.¹¹⁶

Puroindoline

Puroindolines are small basic proteins that contain a single domain rich in tryptophan. These proteins were isolated from wheat endosperm, have a molecular mass around 13 kDa, and a calculated isoelectric point higher than 10. At least 2 main isoforms (called PIN-a and PIN-b) are known, which are encoded by Pina-D1 and Pinb-D1 genes, respectively. These genes share 70.2% identical coding regions but exhibit only 53% identity in the 3′ untranslated region.¹¹⁷

Both PIN-a and PIN-b contain a structure with 10 conserved cysteine residues and a tertiary structure similar to LTPs, consisting of 4 α-helices separated by loops of varying lengths, with the tertiary structure joined by 5 disulfide bonds, 4 of which identical to ns-LTPs.¹¹⁷

The conformation of the 2 PIN isoforms was studied by infrared and Raman spectroscopy. Both PIN-a and PIN-b have similar secondary structures comprising approximately 30% helices, 30% β-sheets, and 40% non-ordered structures at pH 7. It has been proposed that the folding of both PINs is highly dependent on the pH of the medium. The reduction of the disulfide bridges results in a decrease of PINs solubility in water and to an increment of the β-sheet content by about 15% at the expense of the α-helix content.¹¹⁸ No high-resolution structure for any of the PIN isoforms is available, bringing challenges to understanding the function of their hydrophobic regions, with some evidence coming only from partially homolog peptides.¹¹⁷ However, Wilkinson et al¹¹⁹ proposed a theoretical model for several sequences of this AMP.

Puroindolines are proposed to be functional components of wheat grain hardness loci, control core texture, besides antifungal activity.^120-123 Although the biological function of PINs is unknown, their involvement in lipid binding has been proposed. While LTPs bind to hydrophobic molecules in a large cavity, PINs interact only with lipid aggregates, that is, micelles or liposomes, through a single stretch of tryptophan residues. This stretch of tryptophan residues is especially significant in the main form, PIN-a (WRWWKWWK), while it is truncated in the smaller form, PIN-b (WPTWWK).^124-126

Puroindolines form protein aggregates in the presence of membrane lipids, and the organization of such aggregates is controlled by the lipid structure. In the absence of lipids, these proteins may aggregate, but there is no accurate information on the relationship between aggregation and interaction with lipids. The antimicrobial activity of PINs is targeted to cell membranes. Charnet et al¹²⁷ indicated that PIN is capable of forming ion channels in artificial and biological membranes that exhibit some selectivity over monovalent cations. The stress and Ca²⁺ ions modulate the formation and/or opening of channels. Puroindolines may also be membranotoxins, which may play a role in the plant defense mechanism against microbial pathogens.

Morris¹²⁸ reported that the PIN-a and PIN-b act through similar but somewhat different modes, which may involve “membrane binding, membrane disruption and ion channel formation” or “intracellular nucleic acid binding and metabolic disruption.” Natural and synthetic mutants have allowed the identification of PINs as key elements for antimicrobial activity.

Snakin

Snakins are CRPs first identified in potato (Solanum tuberosum).^129,130 Due to their sequence similarity to GASA (Gibberellic Acid Stimulated in Arabidopsis) proteins, the snakins were classified as members of the snakin/GASA family.¹³¹ The genes that encode these peptides have (1) a signal sequence of approximately 28 aa, (2) a variable region, and (3) a mature peptide of approximately 60 residues, with 12 highly conserved cysteine residues. These cysteine residues maintain the 3D structure of the peptide through disulfide bonds, besides providing stability to the molecule when the plant is under stress^{129,130,132,133} (Figure 2F; Supplementary Figure S6).

Snakins may be expressed in different parts of the plant, like stem, leaves, flowers, seeds, and roots,^134-137 both constitutive or induced by biotic or abiotic stresses. In vitro activity was observed against a variety of fungi, bacteria, and nematodes, acting as a destabilizer of the plasma membrane.^129,138,139 Moreover, they were reported as essential agents in biological processes such as cell division, elongation, cell growth, flowering, embryogenesis, and signaling pathways.^140-143

Alpha-hairpinins

As reported by Nolde et al,¹⁴⁴ alpha-hairpin emerged as a new AMP with unusual motif configuration. These peptides prevail in plants and their structure was resolved based on NMR data obtained from the EcAMP-1 peptide isolated from Barnyard grass seeds (Echinoa crus-galli).¹⁴⁴ Some α-hairpinins comprise trypsin inhibitors with helical hairpin structure and this group was recently proposed as a new plant AMP family.¹⁴⁵ Similar to other AMPs, the amino acid sequences of α-hairpinins are variable. They share the conserved cysteine motif (CX3CX1-15CX3C) that form a helix-loop-helix fold and may have 2 disulfide bridges C1-C4 and C2-C3.¹⁴⁶ Its structural stability is maintained by forming hydrogen bonds, so that the side chains have a relatively stable spatial orientation.¹⁴⁷

As reviewed by Slavokhotova et al,¹⁴⁸ members of alpha-hairpin family have been described in both mono and dicot groups, including species as Echinochloa crus-galli and Zea mays (both Poaceae, monocot), Fagopyrum esculentum (Polygonaceae, eudicot), and Stellaria media (Caryophyllaceae, eudicot). Several transcripts with α-hairpinin motif exhibit similarities to snakin/GASA genes and are sometimes positioned within this family.

Although the α-hairpinins structure has been published, its mechanism of action is still not resolved (Figure 2J, PDB ID: 2L2R). However, studies indicate they present a potential DNA binding capacity.¹⁴⁹

Cyclotide

The term cyclotide was created at the end of the past century to designate a family of plant peptides with approximately 30 aa in size and a structural motif called cyclic cysteine knot (CCK).¹⁵⁰ This motif is composed by a head-to-tail cyclization that is stabilized by a knotted arrangement of disulfide bridges, with 6 conserved cysteines, connected as follows: C1-2, C3-6, C4-5.¹⁵¹ Cyclotides are generally divided into 2 subfamilies, Mӧbius and Bracelets, based on structural aspects. In addition to CCKs, 2 loops (between C1-2 and C4-5) have high similarity between both subfamilies, while the other 2 loops (between C2-3 and C3-4) exhibit some conservation within the subfamilies^152,153 (Supplementary Figure S7).

To date, several cyclotides were identified in eudicot families such as Rubiaceae,¹⁵⁴ Violaceae,¹⁵⁵ Fabaceae,¹⁵⁶ and Solanaceae,¹⁵⁷ in addition to some monocots of Poaceae family.¹⁵⁸ In general, cyclotides may act in defense against a range of agents like insects, helminths, or mollusks. In addition, they can also act as ecbolic (inducer of uterus contractions),¹⁵⁴ antibacterial,¹⁵⁹ anti-HIV,¹⁰⁰ and anticancer factors.¹⁶⁰ All these characteristics added to the stability conferred by the CCK motif turn these peptides into excellent candidates for drug development.^161,162

Thaumatin-like protein

Thaumatins or TLPs belong to the PR-5 (Pathogen-related protein) family and received this name due to its first isolation from the fruit of Thaumatococcus daniellii (Maranthaceae) from West Africa.¹⁶³ Thaumatin-like proteins are abundant in the plant kingdom,¹⁶⁴ being found in angiosperms, gymnosperms, and bryophytes,¹⁶³ being also identified in other organisms, including fungi,^165,166 insects,¹⁶⁷ and nematodes.¹⁶⁸

Thaumatin-like proteins are known for their antifungal activity, either by permeating fungal membranes¹⁶⁹ or by binding and hydrolyzing β-1,3-glucans.^170,171 In addition, they may act by inhibiting fungal enzymes, such as xylanases,¹⁷² α-amylases, or trypsin.¹⁷³ Besides, the expression of TLPs is regulated in response to some stress factors, such as drought,¹⁷⁴ injuries,¹⁷⁵ freezing,¹⁷⁶ and infection by fungi^177,178 viruses, and bacteria.¹⁷⁹

As to the TLP structure, this protein presents characteristic thaumatin signature (PS00316): G-x-[GF]-x-C-x-T-[GA]-D-C-x(1,2)-[QG]-x(2,3)-C.^180,181 Most of the TLPs have molecular mass ranging from 21 to 26 kDa,¹⁶³ possessing 16 conserved cysteine residues (Supplementary Figure S8) involved in the formation of 8 disulfide bonds,¹⁸² which help in the stability of the molecule, allowing a correct folding even under extreme conditions of temperature and pH.¹⁸³ Thaumatin-like proteins also contain a signal peptide at the N-terminal, which is responsible for targeting the mature protein to a particular secretory pathway.¹⁶³ The tertiary structure presents 3 distinct domains, which are conserved and form the central cleft, responsible for the enzymatic activity of the protein, being located between domains I and II.¹⁸⁴ This central cleft may be of an acidic, neutral, or basic nature depending on the binding of the different linkers/receptors. All plant TLPs with antifungal activity have an acidic cleft known as motif REDDD due to 5 highly conserved amino acid residues (arginine, glutamic acid, and 3 aspartic acid; Supplementary Figure S8), being very relevant for specific receptor binding and antifungal activity.^169,185,186

Crystallized structures were determined for some plant TLPs, such as thaumatin¹⁸⁷ (Figure 2G), zeamatin¹⁶⁹ (Figure 2H), tobacco PR-5d¹⁸⁵ and osmotin,¹⁸⁶ the cherry allergen PruAv2,¹⁸⁸ and banana allergen Ba-TLP,¹⁸⁴ among other TLPs.

Some TLPs are known as small TLPs (sTLPs) due to the deletion of peptides in one of their domains, culminating in the absence of the typical central cleft. These sTLPs exhibit only 10-conserved cysteine residues, forming 5 disulfide bonds, resulting in a molecular weight of approximately 16 to 17 kDa. They have been described in monocots, conifers, and fungi, so far.^163,189,190 Other TLPs exhibit an extracellular TLP domain and an intracellular kinase domain, being known as PR5K (PR5-like receptor kinases)¹⁹¹ and are present in both monocots and dicots. For example, Arabidopsis contains 3 PR5K genes, while rice has only 1.¹⁶³

Bioinformatic Tools and Databases for Plant AMPs

Databases

With the rapid growth in the number of available sequences, it is unfeasible to handle such amount of data manually. Thus, AMP sequences (as well as their biological information) have been deposited in large general databases, such as UniProt and TrEMBL, which contain sequences of multiple origins.^192,193 In this sense, the construction of databases that deal specifically with AMPs was an important step to organize the data.

During the past decade, several databases were built to support the deposition, consultation, and mining of AMPs. Thus, these databases can be classified into 2 groups: general and specific.¹⁹⁴ The specific databases can be divided into 2 subgroups: those containing only 1 specific group (defensins or cyclotides) and those containing data from a supergroup of peptides (plant, animal, or cyclic peptides) (Supplementary Table 1). In general, both types of databases share some characteristics such as the way that the data are available or the tools to analyze AMPs.

The Collection of Antimicrobial Peptides (CAMPR3) is a database that comprises experimentally validated peptides, sequences experimentally deduced and still those with patent data, besides putative data based on similarity.^195-197 The current version includes structures and signatures specific to families of prokaryotic and eukaryotic AMPs.¹⁹⁷ The platform also includes some tools for AMP prediction.

The antimicrobial peptide database (APD)¹⁹⁸ collects mature AMPs from natural sources, ranging from protozoa to bacteria, archaea, fungi, plants, and animals, including humans. AMPs encoded by genes that undergo post-translational modifications are also part of the scope, besides some peptides synthesized by multienzyme systems. The APD provides interactive interfaces for peptide research, prediction, and design, statistical data for a specific group, or for all peptides available in the database.

The LAMP (Database Linking Antimicrobial Peptides) comprises natural and synthetic AMPs, which can be separated into 3 groups: experimentally validated, predicted, and patented. Their data were primarily collected from the scientific literature, including UniProt and other AMP-related databases.¹⁹⁹

The Database of Antimicrobial Activity and Structure of Peptides (DBAASP)²⁰⁰ contains information about AMPs from different origins (synthetic or non-synthetic) and complexity levels (monomers and dimers) that were retrieved from PubMed using the following keywords: antimicrobial, antibacterial, antifungal, antiviral, antitumor, anticancer, and antiparasitic peptides. This database is manually curated and provides information about peptides that have specific targets validated experimentally. It also includes information on chemical structure, post-translational modifications, modifications in the N/C terminal amino acids, antimicrobial activities, cell target and experimental conditions in which a given activity was observed, besides information about the hemolytic and cytotoxic activities of the peptides.²⁰⁰ Due to the diversity of AMPs and the need to accommodate the most representative subclasses, several databases were established, focusing on specific types, sources, or features. There are several ways to classify AMPs, and they can range from biological sources such as bacterial AMPs (bacteriocins), plants, animals, and so on; biological activity: antibacterial, antiviral, antifungal, and insecticide; and based on molecular properties, pattern of covalent bonds, 3D structure and molecular targets.^201,202

The “Defensins Knowledgebase” is a database with manual curation and focused exclusively on defensins. This database contains information about sequence, structure, and activity, with a web-based interface providing access to information and enabling text-based search. In addition, the site presents information on patents, grants, laboratories, researchers, clinical studies, and commercial entities.^203,205

The CyBase is a database dedicated to the study of sequences and 3D structures of cyclized proteins and their synthetic variants, including tools for the analysis of mass spectral fingerprints of cyclic peptides, also assisting in the discovery of new circular proteins.²⁰⁵

The PhytAMP is a database designed to be solely dedicated to plant AMPs based on information collected from the UniProt database and from the scientific literature through PubMed.²⁰⁶

PlantPepDB is a database with manual curation of plant-derived peptides, mostly experimentally validated at the protein level. It includes data on the physical-chemical properties and tertiary structure of AMPs, also useful to identify their therapeutic potential. Different search options for simple and advanced compositing are provided for users to perform dynamic search and retrieve the desired data. Overall, PlantPepDB is the first database that comprises detailed analysis and comprehensive information on phyto-peptides from a wide functional range.²⁰⁷

Biological data banks (DBs) are organized collections of data of diverse nature that can be retrieved using different inputs. The management of this information is done through various software and hardware resources, whose retrieval and organization can be performed in a quick and efficient way.²⁰⁸ Considering biological data, information can be classified into (1) primary (sequences), (2) secondary (structure, expression, metabolic pathways, types of drugs, etc), and (3) specialized, for example, containing information on a species or on a class of protein.²⁰⁹ Within this third group, some references to AMPs can be mentioned, such as CAMPR3¹⁹⁶ and APD¹⁹⁸ that compile sequence data and structure retrieved from diverse sources, and also the Defensin knowledgebase²⁰³ and the CyBase²⁰⁵ which are dedicated to specific classes of peptides (defensins and cyclotides, respectively), in addition to PhytAMP,²⁰⁶ a specific database of plant AMPs (Supplementary Table 2).

Retrieving and annotating sequences from databases

The first step to infer the function of a given sequence (annotation) is to retrieve it in databases. For this purpose, 3 approaches have been used mostly: (1) local alignments, especially by using Basic Local Alignment Search Tool (BLAST)²¹⁰ and FASTA²¹¹; by searching for specific patterns using (2) REGEX or (3) Hidden Markov Model (HMM).¹⁹⁴

The first approach has been widely used, since most of the information is available in databases as sequences, together with tools to align them, whereas the BLAST is the primary tool for doing so.²¹² This tool splits the sequence into small pieces (words), comparing it with the database. However, this approach has a limitation. Small motifs may not be significantly aligned as they comprise small portions of the sequences that can be smaller than 20% of the total size.^31,194 Due to the high variability of AMPs, only few highly conserved sequences can be identified using this type of inference. To reduce the effects of local alignment limitations, other strategies based on the search for specific patterns were introduced, such as REGEX²¹³ (Supplementary Table 1) and HMM.²¹⁴

The REGEX is a precise way of describing a pattern in a string where each REGEX position must be set, although ambiguous characters (or wildcards) can also be used. For example, if we want to find a match for both amino acid sequences CAIESSK and WAIESK, we can use the following expression: [CW]AIES{1,2}K, this expression would find a pattern starting with the letter “C” or “W,” followed by an “A,” an “I,” and an “E,” 1 or 2 “S,” and ending with a “K.”

The HMMs are well-known for their effectiveness in modeling the correlations between adjacent symbols, domains, or events, and they have been extensively used in various fields of biological analysis, including pairwise and multiple sequence alignment, base-calling, gene prediction, modeling DNA sequencing errors, protein secondary structure prediction, noncoding RNA (ncRNA) identification, protein and RNA structural alignments, acceleration of RNA folding and alignment, fast noncoding RNA annotation, and many others. Using HMM, a statistic profile is included in the model, which is calculated from a sequence alignment, and a score is determined site-to-site, with conserved and variable positions defined a priori.^194,215

Predicting antimicrobial activity

The design of new AMPs led to the development of methods for the discovery of new peptides, thus allowing new experiments to be done by researchers. In this sense, the new challenge lies in the construction of new prediction models capable of discovering peptides with desired activities.

The APD DB has established a prediction interface based on some parameters defined by the entire set of peptides available in this database. These values are calculated from natural AMPs to consider features like length, net charge, hydrophobicity, amino acid composition, and so on. If we take as an example the net load, the AMPs deposited in the APD range from –12 to +30. This is the first parameter incorporated into the prediction algorithm. However, most AMPs have a net load ranging from –5 to +10, which then becomes the alternative prediction condition. Therefore, the same method is applied to the remaining parameters. The prediction in APD is performed in 3 main steps. First, the sequence parameters will be calculated and compared. If defined as an AMP, the peptide can then be classified into 3 groups: (1) rich in given amino acids, (2) stabilized by disulfide, and (3) linear bridges. Finally, sequence alignments will be conducted to find 5 peptides of higher similarity.^198,216,217

The advent of machine learning (ML) methods has promoted new possibilities for drug discovery. In ML inferences, both a positive and a negative dataset are usually required to train the predictive models. The positive data, in this case, regard preferably experimentally validated AMPs that can be collected in databases, whereas negative data are randomly selected protein sequences that do not have AMP characteristics.^197,218 Machine learning methods based on support vector machine (SVM), random forest (RF), and neural networks (NN) have been the most widely used. SVM is a specific type of supervised method of ML, aiming to classify data points by maximizing the margin between classes in a high-dimensional space.^219,220 Random forest is a non-parametric tree-based approach that combines the ideas of adaptive neighbors with bagging for efficient adaptive data inference. Neural networks is an information processing paradigm inspired by how a biological nerve system process information. It is composed of highly interconnected processing elements (neurons or nodes) working together to solve specific problems.^221-223

Evaluating proteomic data

Regarding the use of AMPs in peptide therapeutics, as an alternative to antimicrobial treatment, new efficient and specific antimicrobials are demanded. As aforementioned, AMPs are naturally occurring across all classes of life, presenting high active potential as therapeutic agents against various kinds of bacteria.²²⁴ The identification of novel AMPs in databases is primarily dependent on knowing about specific AMPs together with a sufficient sequence similarity.²²⁵ However, orthologs may be divergent in sequence, mainly because they are under strong positive selection for variation in many taxa,²²⁶ leading to remarkably lower similarity, even in closely related species. In this scenario, where alignment tools present limited use, 1 strategy to identify AMPs is related to proteomic approaches.

Proteins and peptides are biomolecules responsible for various biochemical events in living organisms, from formation and composition to regulation and functioning. Thus, understanding of the expression, function, and regulation of the proteins encoded by an organism is fundamental, leading to the so-called “Proteomic Era.” The term “proteome” was first used by Marc Wilkins in 1994 and it represents the set of proteins encoded by the genome of a biological system (cell, tissue, organ, biological fluid, or organism) at a specific time under certain conditions.²²⁷ Protein extraction, purification, and identification methods have significantly advanced our capacity to elucidate many biological questions using proteomic approaches.^228,229 Due to the wide diversity of proteomic analysis, methods makes the choice of the correct approach dependent on the type of material and compounds to be analyzed.^213,230 Two main tools are used to isolate proteins: (1) the 2-dimensional electrophoresis (2-DE) associated with mass spectrometry (MS) and (2) liquid chromatography associated with MS, each one with its own limitations.^230-232 Obtaining native proteins is a challenge in proteomics or peptidomics, due to high protein complexity in samples, as the occurrence of post-translational modifications. Alternative strategies applied to extraction, purification, biochemical, and functional analyses of these molecules have been proposed, favoring access to structural and functional information of hard-to-reach proteins and peptides.²³³

Based on 2D gel, Al Akeel et al²³⁴ evaluated 14 spots obtained from seeds of Foeniculum vulgare (Apiaceae) aiming at proteomic analyses and isolation of small peptides. Extracted proteins were subjected to 3 kDa dialysis, and separation was carried out by DEAE-ion exchange chromatography while further proteins were identified by 2D gel electrophoresis. One of its spots showed high antibacterial activity against Pseudomonas aeruginosa, pointing to promising antibacterial effects, but requiring further research to authenticate the role of the anticipated proteins. For AMPs, 2DE is challenging due to the low concentration of the peptide molecules captured by this approach, their small sizes, and their ionic features (strongly cationic). In addition, the limited number of available specific databases and high variability turn their identification through proteolysis techniques and mass spectrometry, matrix-assisted laser desorption/ionization (MALDI-MS) difficult. In addition, the partial hydrophobicity characteristics and surface charges facilitate peptide molecular associations, making analysis difficult by any known proteomic approaches.²³² In addition, peptides are most often cleaved from larger precursors by various releasing or processing enzymes.²³⁵ Furthermore, profiles generated do not represent integral proteome, as 2DE has limitations to detect proteins with low concentration, values of extreme molecular masses, pIs, and hydrophobic proteins, including those of membranes.²³⁶ Due to these limitations, multidimensional liquid chromatography–high-performance liquid chromatography (MDLC-HPLC) has been successfully employed as an alternative to 2D gels. Techniques and equipments for the newly developed separation and detection of proteins and peptides, such as nano-HPLC and multidimensional HPLC, have improved proteomics evaluation.²³⁷ Molecular mass values obtained are used in computational searches in which they are compared with in silico digestion results of proteins in databases. In silico approaches, usually by the action of trypsin as a proteolytic agent, may generate a set of unique peptides whose masses are determined by MS.^238,239 These methodologies are widely adopted for large-scale identification of peptide from MS/MS spectra.²⁴⁰ Theoretical spectra are generated using fragmentation patterns known for specific series of amino acids. The first 2 widely used search engines in database searching were SEQUEST²⁴¹ and MASCOT (Matrix Science, Boston, MA; www.matrixscience.com).²⁴² They rank peptide matches based on a cross-correlation to match the hypothetical spectra to the experimental one.

MASCOT is widely used for peptidomics and proteomics analysis, including AMP identification in many organisms, or to evaluate the antibacterial efficacy of new AMPs. Evaluating new AMP against multidrug-resistant (MDR) Salmonella enterica, Tsai et al²⁴³ used 2D gel electrophoresis and liquid chromatography–electrospray ionization–quadrupole-time-of-flight tandem MS to determine the protein profiles. The protein identification was performed using the MASCOT with trypsin as cutting enzyme, whereas NCBI nr protein was set as a reference database. The methodology used in this study indicated that the novel AMP might serve as a potential candidate for drug development against MDR strains, confirming the usability of MASCOT. In a similar way, Umadevi et al²⁴⁴ described the AMP profile of black pepper (Piper nigrum L.) and their expression on Phytophthora infection using label-free quantitative proteomics strategy. For protein/peptide identification, MS/MS data were searched against the APD database²⁴⁵ using an in-house MASCOT server, established full tryptic peptides with a maximum of 3 missed cleavage sites and carbamidomethyl on cysteine, besides an oxidized methionine included as variable modifications. The APD database was used for AMP signature identification,²⁴⁵ together with PhytAMP²⁰⁶ and CAMPR3.¹⁹⁷ To enrich the characterization parameters, isoelectric point, aliphatic index, and grand average of hydropathy were also used²⁴⁶ (GRAVY) (using ProtParam tool), besides the net charge from PhytAMP database. Based on label-free proteomics strategy, they established for the first time the black pepper peptidomics associated with the innate immunity against Phytophthora, evidencing the usability of proteomics/peptidomics data for AMP characterization in any taxa, including plant AMPs, aiming the exploitation of these peptides as next-generation molecules against pathogens.²⁴⁴

Other tools use database searching algorithms, such as X!TANDEM,²⁴⁷ Open mass spectrometry search algorithm (OMSSA),²⁴⁸ ProbID,²⁴⁹ RADARS,²⁵⁰ and so on. These search engines are based on database search but use different scoring schemes to determine the top hit for a peptide match. General information on database search engines, their algorithms, and scoring schemes were reviewed by Nesvizhskii et al.²⁵¹ Despite its efficient ability to identify peptides, database searching presents several drawbacks, like false positive identifications due to overly noisy spectra and lower quality peptides score (related to the short size of peptides). So, the identification is strongly influenced by the amount of protein in the sample, the degree of post-translational modification, the quality of automatic searches, and the presence of the protein in the databases.^252,253 In this scenario, the knowledge about the genome from a specific organism is important to allow the identification of the exact pattern of a given peptide. If an organism has no sequenced genome, it is not searchable using these methods.^235,240 Once the sequences are obtained, bioinformatic tools can be used to predict peptides structure and estimate bioactive peptides.²⁵⁴

More recently, an interactive and free web software platform, MixProTool, was developed, aiming to process multigroup proteomics data sets. This tool is compiled in R (www.r-project.org), providing integrated data analysis workflow for quality control assessment, statistics, gene ontology enrichment, and other facilities. The MixProTool is compatible with identification and quantification outputs from other programs, such as MaxQuant and MASCOT, where results may be visualized as vector graphs and tables for further analysis, in contrast to existing softwares, such as GiaPronto.²⁵⁵ According to the authors, the web tool can be conveniently operated, even by users without bioinformatics expertise, and it is beneficial for mining the most relevant features among different samples.²⁴

Antimicrobial peptide modeling

The central tenet of structural biology is that structure determines function. For proteins, it is often said the “function follows form” and “form defines function.” Therefore, to understand protein function in detail at the molecular level, it is mandatory to know its tertiary structure.²⁵⁶ Experimental techniques for determining structures, such as X-ray crystallography, NMR, electron paramagnetic resonance, and electron microscopy, require significant effort and investments.²⁵⁷

All methods mentioned have their own limitations, and the gap between the number of known proteins and the number of known structures is still substantial. Thus, there is a need for computational framework methods to predict protein structures based on the knowledge of the sequence.²⁵⁶ In addition, in recent years, there has been impressive progress in the development of algorithms for protein folding that may aid in the prediction of protein structures from amino acid sequence information.²⁵⁸

Historically, the prediction of a protein structure has been classified into 3 categories: comparative modeling, threading, and ab initio. The first 2 approaches construct protein models by aligning the query sequences with already solved model structures. If the models are absent in the Protein Data Bank (PDB), the models must be constructed from scratch, that is, by ab initio modeling, considered the most challenging way to predict protein structures.²⁵⁶

In the case of comparative modeling methods, when inserting a target sequence, the programs identify evolutionarily related models of solved structures based on their sequence or profile comparison, thus constructing structure models supported by these previously resolved models.²⁵⁹ This approach comprises 4 main steps: (1) fold assignment, which identifies similarity between the target and the structure of the solved model; (2) alignment of the target sequence to the model; (3) generation of a model based on alignment with the chosen template; and (4) analysis of errors considering the generated model.²⁶⁰

There are several servers and computer models that automate the comparative modeling process, with SWISS-MODEL and MODELER figuring as the most used.^261,262 Although automation makes comparative modeling accessible to experts and beginners, some adjustments are still needed in most cases to maximize model accuracy, especially in the case of more complex proteins.²⁶² Therefore, some caution must be taken regarding the generated models, considering the resolution and quality of the model used, as well as homology between the model and the protein of interest.

Threading modeling methods are based on the observation that known protein structures appear to comprise a limited set of stable folds, and those similarity elements are often found in evolutionarily distant or unrelated proteins. The most used servers based on this approach are MUSTER,²⁶³ SPARKS-X,²⁶⁴ RaptorX,²⁵⁹ ProSa-Web,²⁶⁵ and most notably the I-TASSER.²⁶⁶ In some cases, the incorporation of structural information to combine the sequence used in the search with possible models allows the detection of similarity in the fold, even in the absence of an explicit evolutionary relation.

The prediction of structures from known protein models is, at first sight, a more straightforward task than the prediction of protein structures from available sequences. Therefore, when no solved model is available, another approach is recommended, namely, the ab initio modeling. This method is intended to predict the structure only from the sequence information, without any direct assistance from previously known structures. The ab initio modeling aims to predict the best model, based on the minimum energy for a potential energy function by sampling the potential energy surface using various searchable information.^267,268 Such approaches turn it challenging to produce high-resolution modeling, essential for determining the native protein folding and its biochemical interpretation. On the contrary, later resolved structures and comparisons with previously predicted proteins point to a higher successful modeling generated by ab initio methods than those generated by pure energy minimization methods, classical or even pure methods.²⁵⁶

Among the most used servers and programs for ab initio modeling, we highlight the ROSETTA,²⁵⁷ QUARK,²⁶⁹ and TOUCHSTONE II.²⁶⁷ The accuracy of the models calculated by many of these methods is evaluated by CAMEO (Continuous Automated Model EvaluatiOn)²⁷⁰ and by CASP (Critical Assessment of protein Structure Prediction).²⁵⁸ Probably the first reasonably accurate ab initio model was built in CASP4. Since then, sustained progress was achieved in ab initio prediction, but mainly for small proteins (120 residues or less). In CASP11, for the first time, a novel 256-residue protein with a sequence identity with known structures lower than 5% was constructed with high precision for sequences of this size.²⁷¹

In CASP12, a significant improvement was reported in 4 areas: contact prediction, free modeling, template-based modeling, and estimating the accuracy of models. The authors report that this improvement is due to the accuracy of modeling and alignment methods, as well as increased data availability for both sequence and structure.²⁵⁸

Due to the number of AMPs deposited in the PDB (to date approximately 1099 structures), comparative modeling is the most used. However, when it comes to de novo peptide design, the most recommended choice would be ab initio²⁷² or a hybrid approach that uses more than 1 modeling method.²⁷³

Molecular dynamics simulation

After the generation of a model, the AMP stability should be evaluated using molecular dynamics (MD). Molecular dynamics comprises the application of computational simulations that predict the changes in the positions and velocities of the constituent atoms of a system under given time and condition. This calculation is done through a classical approximation of empirical parameters, called “force field.”²⁷⁴ If, on one hand, this approximation makes the dynamics of a system containing thousands of atoms numerically accessible, it obviously limits the nature of the processes that can be observed during the simulations. No quantum effect is visualized in a MD simulation; just as no chemical bond is broken, no interactions occur between orbitals, resonance, polarization, or charge transfer effects.²⁷⁵ However, the molecules go beyond a static system. Thus, MD is a computational technique that can be used for predicting or refining structures, dynamics of molecular complexes, drug development, and action of molecular biological systems.²⁷⁶ Molecular dynamics simulation is widely used for protein research, aiming to extract information about the physical properties of individual proteins. The results of such simulations are then compared with experimental results. As these experiments are generally carried out in solvents, it is necessary to simulate molecular systems of protein in water. These simulations have a variety of applications, such as determining the folding of a structure to a native structure and analyzing the dynamic stability of this structure.²⁷⁷

The use of MD to simulate protein folding processes is one of the most challenging applications and should be relatively long (in the order of microseconds to milliseconds) to allow observing a single fold event. In addition, the force field used must correctly describe the relative energies of a wide variety of shapes, including unfolding and poorly folded shapes that may occur during the simulation.²⁷⁵ The considerable application potential led to the implementation of MD simulation in many software packages, including GROMACS,^278-280 AMBER,²⁸¹ NAMD,²⁸² CHARMM,²⁸³ LAMMPS,²⁸⁴ and Desmond.²⁸⁵ In addition to the above mentioned, there are other simulation types available, such as the Monte Carlo Method, Stochastic Dynamics, and Brownian Dynamics.²⁸⁰

In the last decades, MD simulation has become a standard tool in theoretical studies of large biomolecular systems, including DNA or proteins, in environments with near realistic solvents. Indeed, simulations have proven valuable in deciphering functional mechanisms of proteins and other biomolecules, in uncovering the structural basis for disease, and in the design and optimization of small molecules, peptides, and proteins.²⁸⁶ Historically, the computational complexity of this type of computation has been extremely high, and much research has focused on algorithms to achieve unique simulations that are as long or as large as possible.²⁷⁸

Plant-pathogen interaction and molecular docking

The interplay between a given pathogen (eg, virus, bacteria, fungus) must be studied through a holistic approach. Host-pathogen relationships are very complex and occur at diverse conceivable levels, including the cellular/molecular level of both, pathogen and host, under given environmental conditions. A most approximate understanding of these interactions at every level is the ultimate goal of “systems biology” (SB). It comprises a holistic approach, integrating distinct disciplines, as biology, computer science, engineering, bioinformatics, physics, and others to predict how a given system behaves under given conditions and what is the role of its parts. Systems biology stands out because it is capable of correlating omics data for the understanding of plant-pathogen interaction. The construction of a plant-pathogen interaction network includes the reconstruction of metabolic pathways of these organisms, identification of the degree of pathogenicity, besides the expression of genes and proteins from both plant and pathogen. The networks can be classified into 5 types: (1) regulatory; (2) metabolic; (3) protein-protein interaction; (4) signaling and regulatory; and (5) signaling, regulatory, and metabolic.²⁸⁷ Each of these networks can be plotted according to computational approaches.

Also, further studies are required to contemplate the construction of evolutionary in silico models and the characterization of these molecular targets in vitro.^288,289 Studies of protein-protein interactions to understand the regulatory process are essential²⁹⁰ and new computational methods are necessary for this purpose with more optimized algorithms, also to remove potential false positives. Thus, in-depth studies on the orientation of molecules and their linkages to the formation of a stable complex are of great importance for understanding plant-pathogen studies and also to develop new drugs.²⁹¹

Molecular docking

The understanding of the regulatory principles by which protein receptors recognize, interact, and associate with molecular substrates or inhibitors is of paramount importance to generate new therapeutic strategies.²⁹² In modern drug discovery, docking plays an important role in predicting the orientation of the binder when it is attached to a protein receptor or enzyme, using forms and electrostatic interactions, van der Walls, Coulombic, and hydrogen bond as parameters to quantify or predict a given interaction.^293,294 Molecular docking aims at exploring the predominant mode(s) of binding of a molecule (protein or ligand) when it binds to a protein with a known 3D structure based on a scoring function that has 3 main functions: the first is to determine the binding mode and the binding site of a protein, the second is to predict the absolute binding affinity between protein and ligand (or other protein) in lead optimization, and the third is virtual screening, which can identify potential drug leads for a given protein target by searching a large ligand or protein in database.²⁹⁵

Protein-protein interactions are essential for cellular and immune function. In many cases, due to the absence of an experimentally determined structure of the complex, these interactions must be modeled to obtain an understanding about their structure and molecular basis.²⁹⁶ Few studies on plant-pathogen interactions include docking approaches and most studies focus on drug development for medical purposes. Drug research based on structure is a powerful technique for the rapid identification of small molecules against the 3D structure of available macromolecular targets, usually by X-ray crystallography, NMR structures, or homology models.

Due to abundant information on protein sequences and structures, the structural information on specific proteins and their interactions have become crucial for current pharmacological research.²⁹⁷ Even in the absence of knowledge about the binding site and limited backbone movements, a variety of algorithms have been developed for docking over the past 2 decades. Although the ZDOCK,²⁹⁶ the rDOCK,²⁹⁸ and the HEX²⁹⁹ have provided results with high coupling precision, the complexes provided are not very useful for designing inhibitors for protein interfaces due to constraints on rigid body docking.²⁹⁴ In this context, more flexible approaches have been developed which generally examine very limited conformations compared with rigid body methods. These docking methods predict that binding is more likely to occur in broad surface regions and then defines the sites in complex structures of high affinity.³⁰⁰ The best example is the HADDOCK software,²⁹⁷ which has been successful in solving a large number of precise models for protein-protein complexes. A good example of its use is the study of the complex formed between plectasin, a member of the innate immune system, and a precursor lipid of bacterial cell wall II. The study identified the residues involved in the binding site between the 2 proteins, providing valuable information for planning new antibiotics.³⁰¹

However, the absolute energies associated with intermolecular interaction are not estimated with satisfactory accuracy by the current algorithms. Some significant issues as solvent effects, entropic effects, and receptor flexibility still need to be addressed. However, some methods, such as MOE-Dock,³⁰² GOLD,³⁰³ Glide,³⁰⁴ FlexX,³⁰⁵ and Surflex³⁰⁶ which deal with lateral chain flexibility, have proven to be effective and adequate in most cases. Realistic interactions between small molecules and receptors still depend on experimental wet-lab validation.^294,307

Despite the current difficulties, there is a growing interest in the mechanisms and prediction of small molecules such as peptides, as they bind to proteins in a highly selective and conserved manner, being promising as new medicinal and biological agents.³⁰⁸ While both “small molecule docking methods” and “custom protocols” can be used, short peptides are challenging targets because of their high torsional flexibility.³⁰⁷ Protein-peptide docking is generally more challenging than those related to other small molecules, and a variety of methods have been applied so far. However, few of these approaches have been published in a way that can be reproduced with ease.^309-311 Although it is difficult to use peptide docking, a recent focus of basic and pharmacological research has used computational tools with modified peptides to predict the selective disruption of protein-protein interactions. These studies are based on the involvement of some critical amino acid residues that contribute most to the binding affinity of a given interaction, also called hot-spots.^312,313

Despite the number of docking programs, existing algorithms still demand improvements. However, approaches are being developed to improve all issues related to punctuation, protein flexibility, interaction with plain water, among other issues.³¹⁴ In this context, the CAPRI (Critical Assessment of Predicted Interactions) is a community that provides a quality assessment of different docking approaches. It started in 2001 and since then has aided the development and improvement of the methodologies applied for docking.³¹⁵ An evaluation was carried out for CAPRI in 2016, resulting in an improvement in the integration of different modeling tools with docking procedures, as well as the use of more sophisticated evolutionary information to classify models. However, adequate modeling of conformational flexibility in interacting proteins remains an essential demand with a crucial need for improvement.³¹⁴ Different docking programs are currently available,²⁹⁴ and new alternatives continue to appear. Some of these alternatives will disappear, just as others will become the top choices among field users.

Molecular docking technique is not often used for AMPs, due to its standard mechanism of action based on the classical association with the external membrane of the pathogen. Despite that, some AMPs have the ability to bind other proteins and/or enzymes, a feature still scarcely studied. In such cases, molecular docking can be useful. An example of success is the study performed by Melo et al,⁴⁷ where they showed the specific binding of a trypsin to a cowpea (Vigna unguiculata) thionine, revealing that this interaction occurs in a canonical manner with Lys11, located in an extended exposed loop. Therefore, further application of docking may bring new evidences about the antimicrobial mechanisms revealing other molecular targets of interest.

It is clear that the combination of data bank information with bioinformatic tools (especially those allowing the identification of patterns, rather than sequence order) is able to revolutionize the identification of AMPs and prediction of their activity. The data may come from genomic, transcriptomic, or proteomic databases, or a combination of different information sources (eg, genomic and transcriptomics, transcriptomics and proteomics).

Supplementary Figure S9 brings a schematic flowchart describing the steps for mining, annotation, and structural/functional analysis of AMPs, in addition to some wet-lab analyses that can be integrated to assess/confirm candidate AMPs.

Similar bioinformatic approaches have been actually used to identify potential peptide candidates with anti-SARS-CoV-2 activity, especially those potentially able to interact with the spike protein and proteases involved in viral penetration.^316,317

Concluding Remarks and Perspectives

As emphasized, plant AMPs show greater diversity and abundance, when compared with other kingdoms. It can be speculated that plants shelter many yet undescribed AMP classes, given their vast abundance and isoform diversity.

The genomic and peptidic structure of AMPs can be variable, with few key residues conserved, which turns their identification, classification, and comparison challenging even in the omics age. Nevertheless, advances in the generation of new bioinformatics tools and specialized databases have led to new and more efficient approaches for both the identification of primary sequences and molecular modeling, besides the analysis of the stability of the generated models.

Despite the large availability of omics data and bioinformatics tools, most new plant peptides have been discovered by wet-lab approaches regarding single candidates. High throughput in silico methods have the potential to transform this scenario, revealing many new candidates, including some new or “non-canonical” peptides. It may be also speculated that a myriad of new peptides may exist considering even smaller peptides, still less considered and more difficult to identify. Finally, in silico approaches shall in future studies be mandatory to define the design of wet-lab studies, turning the identification more efficient and requiring reasonably less time to track, identify, and confirm new candidate AMPs.

Considering the actual pandemic scenario of COVID-19, plant AMPs may be regarded as an important source of antiviral drug candidates, especially considering that some AMP categories present not only antiviral effects but also a wide spectrum antimicrobial activity, act as anti-inflammatory, and also induce the immune response.