Improved Methods for Classification, Prediction and Design of Antimicrobial Peptides

Abstract

Peptides with diverse amino acid sequences, structures and functions are essential players in biological systems. The construction of well-annotated databases not only facilitates effective information management, search and mining, but also lays the foundation for developing and testing new peptide algorithms and machines. The antimicrobial peptide database (APD) is an original construction in terms of both database design and peptide entries. The host defense antimicrobial peptides (AMPs) registered in the APD cover the five kingdoms (bacteria, protists, fungi, plants, and animals) or three domains of life (bacteria, archaea, and eukaryota). This comprehensive database (http://aps.unmc.edu/AP) provides useful information on peptide discovery timeline, nomenclature, classification, glossary, calculation tools, and statistics. The APD enables effective search, prediction, and design of peptides with antibacterial, antiviral, antifungal, antiparasitic, insecticidal, spermicidal, anticancer activities, chemotactic, immune modulation, or anti-oxidative properties. A universal classification scheme is proposed herein to unify innate immunity peptides from a variety of biological sources. As an improvement, the upgraded APD makes predictions based on the database-defined parameter space and provides a list of the sequences most similar to natural AMPs. In addition, the powerful pipeline design of the database search engine laid a solid basis for designing novel antimicrobials to combat resistant superbugs, viruses, fungi or parasites. This comprehensive AMP database is a useful tool for both research and education.

1. Introduction

There are at least two good reasons for our current focus on host defense antimicrobial peptides (AMPs). First, AMPs have remained potent for millions of years. Therefore, AMPs constitute useful templates for developing a new generation of antimicrobials to meet the growing antibiotic resistance problem worldwide. Second, AMPs are key components of the innate immune system universally required for the survival of both invertebrates and vertebrates. Thus, research in this direction improves our understanding of innate immunity and its relationships with the adaptive immune system in vertebrates [1-6].

Lysozyme, discovered by Alexander Fleming in 1922, is now recognized as the first antimicrobial peptide. However, there was little research on AMPs until the discoveries of cecropins, defensins, and magainins in the 1980s [7-9]. Since then, AMPs have been identified from a variety of living species. Select AMPs identified during 1922-2012 are listed in the discovery timeline page of the antimicrobial peptide database (APD) [10, 11]. In earlier days when the number of AMPs was limited, these peptides were handled in review articles. With a rapid increase in the number of such peptides, it became impractical to continue to manage them manually. As a consequence, several databases have been established to categorize these peptides [10-31]. AMSDb appears to be the first such database available online in 1998 [12]. The information format of this database is identical to the SWISS-Prot (UniProt) [32]. It contains 895 antimicrobial peptides, proteins, and their precursors from plants and animals. Unfortunately, AMSDb is no longer updated. To meet the need of better databases with a broad scope, two general databases were published side by side in 2004. ANTIMIC reported more than 1700 entries [13], while a new version of ANTIMIC called DAMPD [14] contains 1232 entries. In 2004, the first version of the APD [10] reported 525 peptide entries. These peptides were manually collected from the literature with the aid of public search engines such as Pub-Med, Swiss-Prot, and PDB [32-34]. The peptide number reached 1228 entries in the second version of the APD [11] and there are 2329 peptide entries in the current version.

Since the publication of APD and ANTIMIC, several specialized databases have been established to emphasize certain aspects of natural, synthetic, or recombinant AMPs from a special peptide family (circular peptides, defensins, and thiopeptides) or source (e.g., bacteria, plants, shrimps, amphibians) [15-28]. For example, defensin knowledgebase is dedicated to defensins only, while DADP contains only polypeptides from frogs. More recently, the CAMP [29], YADAMP [30], and LAMP [31] were also built. Table 1 lists major databases dedicated to AMPs. Among these databases, the APD [10, 11] stands out. This article highlights the unique aspects of the APD as well as new developments since the publication of the second version in 2009.

Table 1.

A chronological list of the databases for antimicrobial peptides^a

Year	Database	URL (http://)	Scope	Country	Ref
1998	AMSDb	www.bbcm.univ.trieste.it/~tossi/amsdb.html	Plant/animal AMPs	Italy	[12]
2002	SAPD	oma.terkko.helsinki.fi:808 0/~SAPD/	Synthetic AMPs	Finland	[25]
2003, 2004	Peptaibols	www.cryst.bbk.ac.uk/peptaibol/home.shtml	Fungal peptaibols	England	[26]
2004, 2009	APD	aps.unmc.edu/AP/	AMPs	USA	[10- 11]
2004, 2012	DAMPD	apps.sanbi.ac.za/dampd/	AMPs	South Africa/Sau di Arabia	[13- 14]
2006	PenBase	penbase.immunaqua.com	Shrimp AMPs	France	[15]
2006, 2008	Cybase	research1t.imb.uq.edu.au/ cybase/	Circular proteins	Australia	[18]
2006, 2010	BAGLE	bioinformatics.biol.rug.nl/ websoftware/bagel/bagel_ start.php	Bacterial AMPs	Netherland	[21]
2007	AMPer	marray.cmdr.ubc.ca/cgi- bin/amp.pl	Like AMSDb	Canada	[24]
2007, 2010	BACTIBAS E	bactibase.pfba-lab- tun.org/main.php	Bacteriocins	Canada/Tu nisie	[17]
2007	Defensins	defensins.bii.a-star.edu.sg/	Defensins	Singapore	[16]
2008	RAPD	faculty.ist.unomaha.edu/chen/rapd/index.php	Recombinan t AMPs	USA	[20]
2009	PhytAMP	phytamp.pfba-lab- tun.org/main.php	Plant AMPs	Tunisie/Ca nada	[19]
2010	CAMP	www.bicnirrh.res.in/antimicrobial	AMPs	India	[29]
2012	YADAMD	yadamp.unisa.it/	AMPs	Italy	[30]
2012	DADP	split4.pmfst.hr/dadp/	Amphibian AMPs	Croatia	[22]
2012	THIOBASE	db- mml.sjtu.edu.cn/THIOBA SE/	Bacteria thiopeptides	China	[23]
2012	EnzyBase	biotechlab.fudan.edu.cn/d atabase/EnzyBase/home.p hp	Cleaving enzymes	China	[27]
2013	LAMP	biotechlab.fudan.edu.cn/d atabase/lamp/guide.php	AMPs	China	[31]
2013	MilkAMP	/milkampdb.org	Milk AMPs	Canada	[28]

^aAdapted from the APD website (http://aps.unmc.edu/AP/links.php) [10, 11].

2. Database design and search functions

2.1. Criteria for peptide collections

In terms of peptide registration, the APD database [11] follows a set of self-defined criteria. First, the peptide must have a known amino acid sequence, at least partially. Second, the peptide should have demonstrated antimicrobial activity. Third, the peptide contains less than 100 amino acids (this has recently been expanded to 200 amino acids so that some important antimicrobial proteins could be collected). Fourth, the peptide originates primarily from natural sources, including bacteria, protozoa, fungi, plants, and animals. Only a small set of synthetic peptides of general interest was collected. Also, the APD emphasizes unique sequences. Therefore, peptides from different species currently occupy the same entry in this database if they share the same amino acid sequence. At present, there are 46 such entries in the APD, which were “found in multiple species” (the quoted phrase can be searched in the additional information field). Since in silico-predicted peptides may not be truly antimicrobial peptides, they are not registered into the APD at this stage. By following the above criteria, the APD database provides a well-defined set of peptides to the research community. Indeed, the APD is a well-recognized resource in the field of AMPs. For example, the web hits were ~15,000 per year prior to 2008 [11]. Since the publication of the second version in 2009, there is a dramatic increase in database use. For example, the web hits reached 86,000 in 2012 alone.

2.2. A flexible database design

The design of any database is to facilitate information search. Users can conduct a simple search by using peptide name and amino acid sequence in single-letter code. Different from other databases in Table 1, the power of the APD search engine can be ascribed to two important features. First, the search engine is composed of a pipeline of search functions. Second, the modular design of the APD enables continued expansion and development. These features greatly facilitate information search at an advanced level. For example, we obtained 268 defensins using the word “defensin” as a search term. The number of defensins rapidly reduced to 19 when the word “monkey” is also used. Only seven peptides were found when a combination of “defensin”, “monkey” and “theta” are used.

2.3. Database search functions

To make it easier for the APD users, Table 2 lists major search functions, peptide information and examples. Most of these search functions are self-explanatory. The name field of the APD, however, has been substantially expanded and deserves some description. It consists of the following elements:

Table 2.

Search functions of the antimicrobial peptide database

Search Function	Peptide Information	Examples
APD ID	A unique 5-digit number for each database entry	AP00310
Name	Peptide name or synonyms	LL-37 (LL37, FALL-39)
AMP sequence	Amino acid sequence in single-letter code	LLGDFFRKSKEKIGKEFKRI VQRIKDFLRNLVPRTES
Name	Life kingdoms	Bacteria, plants, fungi, protists, animals
Name	Life domains	Bacteria, archaea
Name	Classes	Fish, reptiles, amphibians, birds, insects,
Name	Peptide family	Defensins, cathelicidins, histatins, cecropins, magainins,
Source species	Location where the peptide is found	Neutrophils; Homo sapiens
Length	The number of amino acids	37 (for LL-37)
Net Charge	At pH 7	+6 (for LL-37)
Hydrophobi c%	Sum of L, I, V. M, A, F, W, C divided by peptide length	35% (for LL-37)
Name	Chemical modification type	See Table 5
Structure	(1) Known 3D (α, β, αβ, non-αβ);  (2) Partial known (bridged, rich);  (3) unknown	Helix for LL-37
Structural method	X-ray; NMR; CD	NMR (for LL-37)
PDB ID	Self explained	2K6O (for LL-37)
Activity	Known antimicrobial activity	Gram+/Gram−; Gram+; Gram−; viruses; HIV-1; fungi
Name	Binding target	See Table 6
Additional info	Mechanism of action	Magainin: forming pores
Additional info	Synergy	LL-37 and lysozyme
Additional info	Animal model	Mouse
Author or Pub year	Search author or publication year separately	Any

Peptide name + family name + peptide source kingdom + post-translational modification + peptide binding molecules.

In the beginning, it gives peptide name, including synonyms and even the outdated names. In the case of human cathelicidin LL-37, the word LL37 is also used in the literature and FALL-39 is an outdated name. To help users to understand the AMP nomenclature, the major methods used to name AMPs are summarized in the APD website (aps.unmc.edu/AP/naming.php). These include the peptide property-based method, the source-based method, and a third method that uses both peptide features and source information. For examples, please visit the APD website.

After the peptide name, the peptide family name is also given in the NAME field. Selected AMP families are tabulated in Table 3. Using the peptide family name, one can obtain a list of AMPs from the same family. For example, there are 268 defensins from a variety of sources and 185 brevinins from amphibians.

Table 3.

Select antimicrobial peptide families in the APD^a

Peptide family	Count	Peptide family	Count
Defensins	268	Aureins	12
Cathelicidins	78	Maximins	30
Histatins	12	Brevinins	185
Neuropeptides	20	Temporins	105
Chemokines	26	Ranatuerins	49
Ribonucleases	6	Dermaseptins	55
		Caerins	29
Cyclotides	151	Maculatins	7
		Uperins	12
Lantibiotics	51	Magainins	5
Microcins	13	Cecropins	24

^aPeptide counts in this and subsequent tables were obtained from the APD on November 30, 2013.

Following the family name, the peptide is further annotated in the NAME field based on the source domains or kingdoms. The five kingdoms of life are bacteria, protists (protozoa + algea), fungi, plants, and animals [35], while the three domains of life are bacteria, archaea, and eukaryote [36]. The peptide counts in each kingdom are listed in Table 4. Selected classes in each life domain are also given in the NAME field, allowing users to focus only on the AMPs of their interest.

Table 4.

Antimicrobial peptides from the three domains and five kingdoms of life^a

Domain	Peptide count	Class	Peptide count
Bacteria	209	Insects	216
Archeae	2	Spiders	33
Eukaryota	2082	Molluscs	27
		Worms	14
Kingdom	Peptide count	Crustaceans	32
Bacteria	208	Birds	36
Protists	7	Reptiles	10
Fungi	12	Fish	79
Plants	301	Amphibians	929
Animals	1761	Ruminants	44
		Humans	102

The importance of post-translational modifications (PTMs) is only secondary to the peptide sequence itself [37]. Because PTMs could influence both structure and function of the peptide, it is necessary to annotate sequence modification information in the same location. Table 5 contains 23 types of PTMs in the APD. To our knowledge, the APD is the only AMP database that contains extensive information on peptide chemical modifications. In addition, the effect of chemical modification on a peptide net charge is considered in the APD.

Table 5.

Post-translational modifications of natural antimicrobial peptides

Search key	Post-translational modification	Peptide count
XXA	Amidation	448
XXB	Chromophore/ion-binding moieties	4
XXC	Backbone cyclization	176
XXD	D-amino acids	17
XXE	Acetylation	11
XXF	Carboxylic-acid-containing unit	8
XXG	Glycosylation	12
XXH	Halogenation (Cl, Br)	8
XXJ	Sidechain-backbone cyclization	15
XXK	Hydroxylation	9
XXL	Lipidation	9
XXM	Methylation	3
XXN	Nitrolation	0
XXO	Oxidation	10
XXP	Phosphorylation	3
XXQ	N-terminal cyclic glutamate	15
XXR	Reduction	2
XXS	Sulfation	1
XXT	Thioether bridge	46
XXU	Rana Box via a single S-S bond	269
XXW	Dehydration	21
XXY	Citrullination	1
Structure searcha	Disulfide bridges	551

^aThis number was obtained by searching for disulfide bond-containing AMPs classified as “Bridge”, “β structure”, and “αβ structure” families, respectively. The “bridged” AMPs are known to have disulfide bonds but unknown 3D structure. Beta structures without disulfide bonds were excluded by including “c” as a sequence search term. For the αβ structures, only the AMPs with a packed 3D fold were counted.

How AMPs kill pathogens is an important question to ask. The information for binding targets of AMPs is also annotated in the APD (Table 6). In addition to membranes, AMPs can bind to DNA, heat shock proteins, carbohydrates, and lipid II [1-6].

Table 6.

Binding targets of antimicrobial peptides

Search keya	Binding target	Count
BBBh2o	Self aggregation in water	15
BBBm	Oligomers in membranes	4
BBII	Ions	16
BBW	Lipid II	17
BBL	LPS	54
BBr	Receptors	3
BBMm	Membranes	81
BBN	Nucleic acids	11
BBS	Sugars/carbohydrates	44

^aSearch by entering the code into the name field of the APD [10, 11].

3. Classification of AMPs based on peptide activity, 3D structure and chain bonding pattern

There are a variety of approaches for classifying AMPs. Some of these methods are summarized on the classification page of the APD website (aps.unmc.edu/AP/class.php). For example, the peptides may be classified based on the biosynthesis machinery. Some peptides are synthesized by a multiple enzyme system, while the majority of AMPs are gene-coded. The expression and degradation of gene-coded AMPs are elegantly regulated because either over or under expression of AMPs could cause problems [1-5]. AMPs can also be classified based on molecular targets (e.g., membrane targeting and cell-penetrating peptides) [6]. In the following, we first describe structure and activity-based classification schemes in the APD and then introduce a universal classification scheme for antimicrobial peptides.

3.1. Antimicrobial activity

As key effector molecules of innate immunity, AMPs are able to control invading pathogenic microbes, including bacteria, viruses, fungi, and parasites [1-4]. It is natural to classify these host defense peptides based on their functions, including antibacterial, antiviral, antifungal, insecticidal, and spermicidal activities. In addition, some AMPs also possess other functional roles such as anticancer, wound healing and immune modulation [4]. The APD database has annotated 17 types of peptide activities or functions (Table 7). Several newly annotated activity types are unique in this database, making the APD most comprehensive in terms of activity annotation.

Table 7.

Biological activities of host defense antimicrobial peptides

Year created	Activitya	Count
2003	Antibacterial (G+/G−)	1909
2003	Antifungal	850
2003	Antiviral	138
2003	Anticancer	158
2003	Hemolytic	284
2008	Anti-HIV	92
2009	Anti-G+	360
2009	Anti-G−	172
2009	Antiparasitic	59
2009	Insecticidal	22
2009	Spermicidal	9
2011	Chemotactic	47
2012	Anti-protist	4
2013	Antioxidant	10
2013	Anti-inflammatory	2
2013	Wound healing	7
2013	Enzyme inhibitor	5

^aSome newly defined search functions can be searched in the “additional information” field of the APD by entering the words in the table. These include antioxidant, anti-inflammatory, and wound healing, and enzyme inhibitor.

3.2. Three-dimensional structure of AMPs

According to the APD, only a small population of AMPs (13%) has a known 3D structure, primarily determined by solution nuclear magnetic resonance (NMR) spectroscopy [10]. In addition, X-ray diffraction was also used to solve the structures of some AMPs with a folded structure in water. The structural information is well annotated in the APD database, including structural class, method for structural determination, structural regions, key residues, and membrane-mimetic models for structural determination. In addition, users can directly view the 3D structure via the link to the PDB [33]. The AMP structures are usually classified into α-helical, β-sheet, and extended structures [4, 38]. A more general classification approach has been proposed recently [6]. In this approach, the AMP structures are classified into four families: α, β, αβ, and non-αβ based on the types of secondary structures. Peptides in the α family contain α-helical structure (Figure 1A) as the major secondary structure. In contrast, AMPs in the β family are characterized by at least a pair of two β-strands in the structure (Figure 1B). The αβ family contains both α and β structures (Figure 1C), whereas the non-αβ family has neither α nor β structure (Figure 1D). This structural classification scheme is now executed in the APD. Typical examples and peptide counts from different families are provided in Table 8. While the α-helical family is the largest with 328 entries, the non-αβ family is the smallest with merely 9 entries. Table 8 also shows that the lysine/arginine (K/R) ratios in these structural families differ. While lysines are dominant in the α-helical family, arginines are preferred in the β-family as well as the non-αβ family. Not surprisingly, AMPs with both α and β structures have a moderate K/R ratio of ~1.2. These ratios might become useful as indicators for classifying a newly discovered peptide into a particular structural family.

Figure 1. Classification of the 3D structures of antimicrobial peptides into four families [6].

Shown are representatives from each family: (A) α-helical structure of human cathelicidin LL-37 (PDB entry: 2K6O) [55]; (B) the β-sheet structure of plant kalata B1 (PDB entry: 1JNU) [56]; (C) the αβ structure of human β-defensin-1 (HBD-1) (PDB entry: 1IJV) [57]; and (D) the non-αβ structure of cattle indolicidin (PDB entry: 1G89) [58].

Table 8.

Classification of 3D structures of antimicrobial peptides

Structurea	K/R ratio	Peptide count	Examples
α	13.65/5.26=2.59	329	Cecropin, dermcidin, LL-37, magainin
β	5.63/10.7=0.53	97	Human alpha defensins (HNP-1, HNP- 4, and HD-5), plant kalata B1
α β	8.47/7.05=1.2	81	Drosomycin, Human beta defensins (HBD-1, HBD-4), PhD1
Non-αβ	4.85/10.19=0.48	9	Indolicidin, tritrpticin, drosocin, nisin A

^aFor AMPs without 3D structures, additional annotations were made in the APD: (1) unknown, no 3D structure; (2) bridge, disulfide-linked, usually beta-structure; (3) rich, rich in certain amino acids.

3.3. A universal classification of AMPs based on peptide bonding patterns

Because only a small number of AMPs has a 3D structure, we herein propose a systematic classification approach that is independent of 3D structure, peptide source, or activity. This classification is framed based on the connection mode of polypeptide chains. Class I includes linear AMPs (Figure 2A), which may be chemically modified (amidation, sulfate, phosphate, bromide, or glycosylation) at side chains or even backbones. However, such modifications (Table 5) for class I AMPs do not lead to chain connections between different amino acids. Class II covers all AMPs with chemical bonds between different peptide side chains (Figure 2B). These include lantibiotics (thioether rings) and the defensin family (disulfide bonds). Broadly, it can be any type of chemical connections between two amino acids. When two or more peptides work together, they belong to this class as long as any of the polypeptide chain contains a sidechain-sidechain connection. Class III AMPs must possess a chemical bond between peptide side chain and backbone (Figure 2C). The typical members are lassos where the carboxyl group of residue E8 or D9 is covalently linked to the N-terminal amine group. It can be any type of chemical bonding between the side chain of one amino acid and the backbone of another amino acid (see Table 9). Lastly, class IV is composed of circular peptides where a peptide bond is formed between the amino and carboxylic ends of the peptide backbone (Figure 2D). These circular peptides may (or may not) contain additional modifications such as disulfide bonds. Examples are enterocin AS-48 from bacteria, cyclotides from plants and θ-defensins from primates [37].

Figure 2. Classification of antimicrobial peptides based on the connection patterns of the polypeptide chain.

(A) linear polypeptide chains (e.g. LL-37 and magainins); (B) sidechain-linked peptides such as defensins and lantibiotics; (C) polypeptide chains with side chain to backbone connection (e.g. lassos); and (D) circular peptides with a seamless backbone (e.g. cyclotides).

Table 9.

A universal classification of antimicrobial peptides

Class	Chain linkage	Subclass	Link type	Class symbol	Examples
I	Linear & open chainsa	1. One chain	None	UCLL1a	LL-37, magainins
		2. Two chains	None	UCLL2	Enterocin L50
II	Sidechain- Sidechain	1. One chain	Cβ-S-S-Cβ (Disulfide- bond)	UCSS1aa	Defensin-like
			Cβ-S-Cβ (thioether)	UCSS1ba	lantibiotics
		2. Two chains	Inter-chain Cβ-S-S-Cβ	UCSS2a	Distinctin, halocidin, centrocin
			Intra-chain Cβ-S-Cβ	UCSS2b	Lacticin-3147, Smb
III	Sidechain- Backbone	One chain	CO-NH amide	UCSB1a	Microcin J25, Lariatins
			CO-O ester	UCSB1b	Fusaricidin A
			Cβ-S-Cα	UCSB1c	Thuricin CD
IV	Backbone- Backbone	One chain	CO-NH amide	UCBB1aa	AS-48, subtilosin A, cyclotides, θ- defenins

^aFurther classifications are provided in Tables 10-13.

Each class of AMPs can be further classified. For class I peptides, they can be classified into two subclasses based on the number of polypeptide chains (Table 9). Single-chain linear AMPs are further classified based on chemical modifications. Unmodified AMPs include “amino acid rich” and “not amino acid rich” families. Modified peptides are further divided into two types based on modification sites (side chain or backbone). These systematic classifications for class 1 AMPs are summarized in Table 10 with examples. Likewise, class II AMPs with connections between side chains can be further classified based on the number of polypeptide chains as well as the type of chemical bonds (Table 9). A further classification of the single-chain disulfide-bonded AMPs (e.g. defensins or defensin-like) based on the number of S-S bonds is provided in Table 11. It is also possible to further classify single-chain lantibiotics based on the number of thioether bonds (Table 12). A new type of sidechain-sidechain connection will constitute a new subclass. In the same vein, class III AMPs can be further separated into different types based on the bond type (Table 9). This chemical bond-based classification is also extended to class IV. Circular AMPs are classified based on the additional types and number of chemical bonds in the polypeptide chain (Table 13). This systematic classification system covers all AMPs and should complement with the existing classification systems proposed for AMPs from different life domains [39-43].

Table 10.

Classification of class 1 linear antimicrobial peptides (UCLL1)

Subclass	Modification sitea	Modification type	Sub-type	Peptide examples
UCLL1A	None	None	Not-AA-richb	LL-37
			AA-Rich (25%)	Pro-rich; Arg-rich PR-39
UCLL1B	Sidechain	Group attachment	Hydroxylation; halogenation; phosphorylation ; glycosylation; lipidation; sulfation	Piscidin 4 (hydroxylated Trp); datucin, MccC7
		Sidechain cyclization	cyclic glutamate	Heliocin
UCLL1C	Backbone	End capping	Amidation; acetylation; other attachments	Aurein 1.2; temproin A
		Configuration change	D-amino acids,	Gramicidin; bombinin H4
		Backbone transformed	Dehydrated;	Cypemycin (Linaridins)
			Heterocyclic rings	Thiopeptides in ThioBase

^aPost-translational modification (PTM) is a broad concept that includes all types of functional groups attached to the peptide chain via covalent bond formation. A detailed list of PTMs is provided in Table 5. Some common examples are N-terminal acetylation, C-terminal amidation, phosphorylation, glycosylation, aromatic halogenation, and sulfation. In the extreme case, even the peptide backbone is modified, leading to dehydrated or heterocycles. However, all these modifications are limited to a single amino acid and do not lead to a polypeptide chain connection between different amino acids as observed in the other three major classes of AMPs (Table 9).

^bAA = Amino acids.

Table 11.

Sidechain-sidechain connected antimicrobial peptides: further classification of single-chain peptides containing disulfide bonds (UCSS1a)

Type	S-S bond count	Sub-typea	Examples
I	1	A	Brevinin, esculentin (Rana box)
		B	Thanatin
		C	Bactenecin
II	2	A	Ec-AMP1, lasiocepsin, Glycocin F
		B	Protegrin, polyphemusin, CXCL1, LEAP-2
III	3	A	NK-lysin, caenopore-5
		B	HNP-1, HBD-1, big defensins
IV	4	B	ASABF, NaD1, drosomycin
V	5	B	PhD1, WAMP-1a, Ec-CBP
VI	6	B	Copsin

^aThe peptides can further be classified into sub-types based on 3D structure (A: α-helical; B: β-sheet-containing (β and αβ families); C: non-αβ; D: unclassified due to an unknown 3D structure).

Table 12.

Sidechain-sidechain connected antimicrobial peptides: further classification of single-chain lantibiotics containing thioether bonds (UCSS1b)

Type	Number of linkage	Examples
I	1	Not found
II	2	Bovicin HJ50
III	3	Epilancin 15X, Lacticin 481
IV	4	Cinnamycin, Actagardine A
V	5	Nisin, Microbisporicin, Subtilin, Ericin A, Paenibacillin
VI	6	Paenicidin A
VII	7	Geobacillin I

Table 13.

Classification of circular antimicrobial peptides (UCBB1a)

Type	Additional Links	Examples
A	None	Bacterial enterocin AS-48
B	Sidechain-sidechain (Cβ-S-S-Cβ)	Plant cyclotides, primate θ-defensins
C	Sidechain-backbone (Cβ-S-Cα)	Bacterial subtilosin

4. Peptide Prediction

Based on the information content used in the prediction programs, the prediction methods of AMPs have been classified into five types [6]. The first type uses only mature peptide sequences, while the second method involves only the precursor sequences. The third prediction type considers both mature and precursor sequences. The fourth method employs the sequence similarity of the modifying enzymes. Finally, the fifth prediction uses genomic information. It is possible that each prediction above can be achieved in different ways. For example, based on the mature AMP sequences in the APD [10, 11], numerous prediction methods have been developed. In the Lata method [44], two data sets were utilized: antimicrobial and non-antimicrobial. While it is easy to download the positive data set from the APD, it is difficult to get a true negative data set because the activities of the sequences in the negative data set have not been validated by experiments. Yet, the program is set up with a good predictive ability. A recent prediction method iAMP-2L [45] considers multiple functions of AMPs annotated in the APD. Different from all other prediction protocols (reviewed in ref. [6]), a unique prediction method is programmed in the APD. This method does not require a negative data set, but is coupled with the database. In the following, we describe an upgraded version of this APD method.

The original prediction method in the APD made predictions based on some known rules [10]. Hence, the method was referred to as knowledge-based prediction. For example, AMPs are usually cationic. A peptide with a negative net charge was predicted as “less likely to be an antibacterial peptide”. This simple prediction has its limitations because the database does contain anionic AMPs. To overcome this shortcoming, we have updated the prediction interface based on the parameter space defined by the whole peptide set in the APD. The parameters for antimicrobial peptides are better defined due to a four-fold increase in peptide number from the original 525 to the current 2329. Peptide parameters such as length, net charge, hydrophobic percentage, and amino acid composition can all be calculated. These parameters constitute the parameter space of natural AMPs.

In terms of net charge, the known AMPs occupy a very broad range. The AMP with the most negative net charge is chrombacin (net charge −12). Two AMPs, sheep cathelicidin OaBac11 and fish histone-derived Oncorhyncin II, possess the highest net charge of +30. Thus, the boundary conditions for net charge are defined as

$$-12<\text{net charge}<+30.$$

The above boundary condition can be incorporated into the APD program to make database-based predictions. This expansion enables the prediction of a broader range of peptide sequences. Because the majority of the AMPs (97.4%) have a net charge between −5 and +10, it may be useful to define this range as the core region. The small number of AMPs outside the core region may be called the minor region. This core region may be used as an alternative condition for prediction.

The hydrophobic content (i.e., the sum of hydrophobic amino acids divided by the total number of amino acids in a peptide) is another important parameter that determines peptide properties. In the APD, hydrophobic amino acids include alanines (Ala), valines (Val), leucines (Leu), isoleucines (Ile), methionines (Met), phenylalanines (Phe), tryptophans (Trp), and cysteins (Cys) [10]. Based on the database sorting function, we identified the AMPs with the lowest and highest hydrophobic contents. Sheep anionic peptide SAAP (sequence: DDDDDD) contains no hydrophobic residues in the sequence, leading to a hydrophobic content of 0%, while gramicidins have the highest hydrophobic content of 93%. Thus, the boundary conditions for peptide hydrophobic contents are defined as

$$0\%\le \text{hydrophobic content}<93\%.$$

The peak of this hydrophobic distribution is located between 40-50% [46]. This leads to another set of boundary conditions for our database-based prediction. We can also define the core region based on the hydrophobic content. The AMPs in the core region (98.6%) possess a hydrophobic content between 10% and 80%.

The length of the peptides in the current APD ranges from 5 to 174. The lower limit is real, while the upper limit is arbitrary since it is defined by the scope of peptides collected into the database (<200 amino acids). However, the majority of AMPs (92.9%) are less than 60 amino acids in length, leading to a definition of the core length region of 5-60. We can anticipate that these boundary conditions will be fully determined when a sufficient number of representative natural AMPs have been identified and registered into the APD.

During this study, we have executed these new database-derived boundary conditions in the prediction interface of the APD (Figure 3). This interface makes predictions based on sequence similarity. In the first step, the prediction program will calculate the peptide parameters based on the input sequence. The calculated peptide parameters will then be compared with the APD parameter space. If one or more calculated parameters fall outside the database-defined parameter space, the users will be informed that “your input is less likely to be an antibacterial peptide”. If all the parameters fall within the defined parameter space, the database will conduct a second tier of prediction by broadly classifying input peptides into several classes: rich in amino acids (>25% for any amino acid), helical, and disulfide-linked. With the execution of the universal classification proposed in Table 9, a more accurate prediction will be realized. As the third tier of our prediction, the database compares the input sequence with all the peptides in the database by performing sequence alignment. Five peptides with most similar sequences will be provided in the output. Because we use database-derived parameters for prediction, we refer to this upgraded method as the APD-based prediction (the November 2013 version). Compared to the original prediction [10], the upgraded version is able to handle a broader range of peptide sequences. In addition, the chance of identifying the most similar sequences in the APD also increases substantially as a consequence of a four-fold increase in natural compounds.

Figure 3. Prediction of antimicrobial peptides based on the antimicrobial peptide database.

The prediction consists of three steps. As the first step, the program will determine whether the input sequence is in the database-defined parameter space (such as charge and hydrophobic content). If identical, the users will be informed. If one or more calculated parameters of the input peptide are out of the boundaries, it is predicted as “your sequence is less likely to be an antibacterial peptide”. Second, the input sequence will be classified into three families: rich in amino acids such as histatins and tryptophans, disulfide-linked peptides, and linear. Third, sequence alignments will be conducted to find five peptides that are most similar to the query sequence.

The identification of most similar AMPs is a useful feature. For example, O’Shea did not find similar sequences by searching the BLAST database [47], but were able to do so using the APD. Based on the sequence similarity of a novel bacteriocin with plant Ib-AMP3, these authors named the new bacteriocin as bactofensin. The similarity also inspired the authors to test possible antimicrobial activities listed for Ib-AMP3. In addition, the authors can also check whether the new peptide has a similar 3D structure. Thus, the output from the APD prediction programs can guide users to design new experiments to test the structure and activity of the newly identified peptide based on the knowledge annotated for the most similar candidates in the database. Such a prediction of sequence, structure and activity at multiple levels requires careful annotation of AMP information in the APD.

5. Peptide design

The APD [10, 11] also provides a useful platform for identification of useful antimicrobials to combat difficult-to-kill pathogens such as human immune-deficiency virus (HIV) and methicillin-resistant Staphylococcus aureus (MRSA) [46]. Both database screening and database-guided design have been conducted. By screening a representative set of AMPs selected from the APD, we found several potent anti-HIV or anti-MRSA peptides [48, 49]. New peptides were also obtained by modifying, shuffling, or hybriding natural sequences. Mathematically, a known peptide sequence can be shuffled into multiple sequences. Experimentally, we found that sequence shuffling could lead to all the possibilities: less active, equally active, and more potent sequences [49]. An MIT group developed a large-scale hybrid approach by combining sequence segments of 10 residues (i.e., grammars). This grammar approach can generate new sequences, which may, or may not, be bactericidal [50]. A complete different approach in the form of combinatorial libraries can also be pursued [51]. In principle, the amino acid at each position of the peptide sequence can be changed into other amino acids. In practice, it is necessary to bias the choice of amino acids in order to obtain active peptides [52]. This is because the amino acid use in natural AMPs is biased. The APD enabled us to identify the frequently occurring amino acids for AMPs from a variety of life domains [10, 53]. For example, the frequently occurring residues (≥8.5%) are leucines, glycines, and lysines based on the average percentages of all the 2329 peptides in the current APD. We demonstrated previously that these three amino acids contained sufficient information for designing antibacterial peptides [11].

Another important approach is de novo design (reviewed in ref. [6]). We have recently developed a novel database approach [54]. A flow chart for this approach is provided as Figure 4. This flow chart contains two major tiers of information filters. The first tier consists of an activity filter that enables one to obtain a set of peptides with desired activity. Table 3 lists 17 types of peptide activities, each of which contains a set of model peptides. In our design, we selected a group of peptides with activity against Gram-positive bacteria. This set of peptides formed the templates for extracting useful parameters for designing anti-MRSA peptides. The second tier contains numerous filters (F1, F2, to Fn), each defines one parameter for the peptide (P1, P2, to Pn). In determining these parameters, we followed the most probable principle, which projected the maximum for each parameter. Because the most probable parameters were used, the peptides assembled in this manner had a good chance to be antimicrobial. This is indeed the case. The designed peptide DFTamP1 rapidly killed MRSA USA300, a community-associated staphylococcal pathogen. It also showed some bacterial selectivity since DFTamP1 did not kill Gram-negative bacteria E. coli, P. aeruginosa, or Gram-positive B. subtilis. This success opens a new avenue to designing peptides with various types of activities (Table 7). Because this new method differs from all existing de novo approaches, it was referred to as ab initio design [54].

Figure 4. Ab initio peptide design based on the database filtering technology (DFT).

The DFT tech developed recently [54] is composed of two layers of filters. The first layer filter enables the identification of a set of antimicrobial peptides with the desired activity from the antimicrobial peptide database (see Table 7). This set of peptides is then used as templates to extract useful parameters for peptide design by utilizing the second layer filters (F1, F2, F3, …, to Fn). These peptide parameters (P1, P2, P3, …, to Pn) are combined to generate a single or limited number of peptides.

6. Concluding Remarks and Future Studies

The antimicrobial peptide database was constructed in 10 years ago. It is an original construction in terms of both database design and peptide entries. Each peptide entry in the APD was manually collected from the literature using the public search engines such as PubMed, PDB, and Swiss-Prot. By following a set of rules for data registration, the APD presents a well-defined set of natural AMPs. To achieve a more complete sampling of natural AMPs, the database is extensively annotated and regularly updated. In addition, the pipeline design led to a powerful search engine. This unique database, therefore, constitutes the basis for developing new methods for peptide classification, prediction and design. The APD is the first to adopt both five-kingdom and three-domain classifications, allowing users to search the AMP information from any kingdom (bacteria, protists, fungi, plants, and animals) or classes (e.g. insects, spiders, molluscs, crustaceans, reptiles, amphibians, fish, and birds) (Table 4). Once a domain is defined in the NAME field, the APD behaves like a specialized database (e.g., plant AMPs, bacteriocins, and amphibian peptides). The APD also executed a new structure classification scheme based on the types of secondary structures (α, β, αβ, and non-αβ) in a variety of 3D structures of AMPs (Figure 1) [6]. Needless to say, the structures in each family can be further grouped based on the number of secondary structures (e.g., α-helix and β-strand). Due to a limited number of known 3D structures, we have proposed a universal classification scheme here based on peptide chain bonding patterns (Figure 2). Since the information on peptide source, activity, and 3D structure is not required, this systematic classification (Tables 9-13) complements to the existing classification methods for AMPs in a defined life kingdom such as bacteria and plants [39-43]. It also offers an approach to unifying the classification of antimicrobial peptides. This classification is general and can be applied to other biologically active peptides.

There are various prediction methods for AMPs (reviewed in ref. [6]). The APD is unique in that the prediction is highly coupled with the database. The upgraded version of the APD makes predictions in three steps by following the similarity principle. Each step deals with a specific question. The first tier asks whether the peptide parameters of the input sequence fall within the database parameter space. Based on the amino acid composition analysis, the second tier asks which peptide class the input sequence belongs to. The third tier determines five most similar sequences based on sequence alignment with all the peptides in the database. It is clear why we have been strict in following a set of rules in registering AMPs. Our practice allows us to more accurately map the parameter space for natural AMPs. When a large number of predicted or artificial sequences are included, such parameters could deviate from nature’s parameters, thereby influencing the prediction quality. In addition, users can get an idea of the structural type and functional space of the input sequence by viewing the similar sequences already in the APD. For example, the input sequence is most likely to form a helix-bundle structure stabilized by three disulfide bonds if the best match is a saposin-like protein. If the sequence matches human cathelicidin LL-37, it is likely to have multiple functions, ranging from antimicrobial, wound healing, to immune modulation. Like LL-37, the peptide may also have a broad-spectrum activity to kill bacteria, fungi, viruses, and parasites. This information will guide the users to validate both structure and activity of a new peptide.

Finally and importantly, the construction of this well-annotated database also enabled us to develop novel approaches for designing peptides with desired properties. Based on the database, we have tested two general approaches: peptide screening [48,49] and database-guided design [46, 54]. In particular, we demonstrated the first ab initio design based on the database by developing the database filtering technology [54]. This approach is not limited to the development of anti-MRSA peptides and can be applied to the design of peptides with other types of activities (Table 7) as well. It is also desirable that the designed peptides only kill a specific species. Our detailed annotations of AMP targeting organisms into the database set the stage for this effort. In addition, other database filters such as peptide selectivity and stability to proteases can be created as well. Taken together, the APD is a powerful engine for research and education in the field of innate immunity and drug discovery.