Get Our e-AlertsSubmit Manuscript
BioDesign Research / 2022 / Article

Review Article | Open Access

Volume 2022 |Article ID 9783197 |

Suchetana Gupta, Noora Azadvari, Parisa Hosseinzadeh, "Design of Protein Segments and Peptides for Binding to Protein Targets", BioDesign Research, vol. 2022, Article ID 9783197, 25 pages, 2022.

Design of Protein Segments and Peptides for Binding to Protein Targets

Received29 Dec 2021
Accepted16 Mar 2022
Published16 Apr 2022


Recent years have witnessed a rise in methods for accurate prediction of structure and design of novel functional proteins. Design of functional protein fragments and peptides occupy a small, albeit unique, space within the general field of protein design. While the smaller size of these peptides allows for more exhaustive computational methods, flexibility in their structure and sparsity of data compared to proteins, as well as presence of noncanonical building blocks, add additional challenges to their design. This review summarizes the current advances in the design of protein fragments and peptides for binding to targets and discusses the challenges in the field, with an eye toward future directions.

1. Introduction

Peptides, short stretches of amino acids (AAs), often smaller than 50 residues, play key roles in our cellular function (Figure 1, Table 1). Many of these peptides act as hormones that transfer messages in our body; metabolic hormones such as insulin and neuropeptides such as oxytocin are two examples of peptide hormones. Some protein fragments (also called peptides, hereafter) are in charge of modulating our cell signaling cascades such as signal peptides that guide proteins to their proper cellular locations, or proline-rich peptides that interact with SH3 domains in multiple signaling pathways. Some peptides play a role in defense; antimicrobial peptides (AMPs) and many antibiotics such as gramicidin S and lantibiotics and some toxins represent this class. This diversity in function suggests that designing peptides can open the door for many applications, from tuning cell signaling to generating novel antibiotics.

Type of peptideSubclassesGeneral featuresExample

Cyclic peptides
Peptides with the N- and C-termini connected via a bond (N-to-C amide bond, disulfide bond, thioether bond, …)
UnicyclicCan be chemically synthesized, can include ncAAs, often <20 AA, often lack clear secondary structure elements, often protease resistantCyclosporin (11mer N-to-C cyclized peptide) binds to cyclophilin (PDB 1CWA [1])
Multicyclic: where there are additional internal bridges such as disulfide bonds in a cyclic peptideG7-B1 (bicyclic 11mer) binds to Grb7-SH2 domain (PDB 5EEL [2])

Lariat peptides
Cyclic peptides with linear segments flanking from one or both termini
LariatsCan be chemically synthesized, can include ncAAsOxytocin binds to oxytocin receptor (NMR structure of unbound oxytocin: 2MGO [3])

Lasso peptides
A ring with a C-terminal tail that goes through the ring [4]
Class I: 2 disulfide bonds between the tail and the ringA member of ribosomally assembled posttranslationally modified peptides (RiPPs), often hard to synthesize, heat resistantRP 71955 (PDB 1RPC [5] shows structure in solution)
Class II: no disulfide bondCaulosegnin I (PDB 2LX6 [6] shows structure in solution)
Class III: one disulfide bond between the tail and the ringBI-32169 (PDB 3NJW [7] shows structure in solution)

Constrained peptides
Peptides that are stabilized by multiple internal cross-linking bridges
Disulfide constrained peptides such as conotoxinsVery stable, can be chemically synthesized or expressed, often resistant to loop swappingA-conotoxin binds to nicotinic acetylcholine receptor (PDB 2C9T [8] in complex with AChBP homolog)
Peptides with multiple noncanonical crosslinkers such as lantibioticsOften naturally synthesized and posttranslationally modified, often include a large variety ofNisin (PDB 1WCO [9] shows Nisin in complex with lipid-II)

Linear fragmentsHelical fragmentsCan be expressed or synthesized, originally no ncAAs, often require additional crosslinkers or scaffolds to stabilize the given conformationHelical fragment of p53 bound to MDM2 (PDB 1YCR [10])
-Strand fragmentsC-terminal of p53 bound to SIRT1 (PDB 4ZZJ [11])
Loop fragmentsProline rich peptides (PDB 1GBR [12])

MiniproteinsCan be classified based on the order of secondary structure elements, or presence or absence of disulfide bondsLarger peptides often around 20-50 amino acids, stabilized by a packed core, can have disulfide bonds as well, often expressed but in the case of shorter ones can be synthesizedMiniprotein anti-ACE binder (PDB 7JZM [13])

In addition to their natural roles, peptides have been used as a therapeutic modality complementary to antibodies and small molecules [1417]. Similar to antibodies, peptides can bind to flat protein surfaces with high affinities and selectivities. And similar to small molecules, they can cross the cell membrane to access intracellular targets. Thus, they offer a unique opportunity to target the so-called undruggable space of disease-related targets that are currently not accessible by antibodies or small molecules. While current methods for obtaining these therapeutic peptides often require a library-screening step, the ability to design peptides with desired properties to guide these libraries is of high interest.

Despite their important role both as natural molecules and as future therapeutics, design of functional peptides has been a challenging endeavor. In addition to challenges faced by the field of protein design in general, peptide design has its own unique limitations. One of the key difficulties in peptide design arises from the practical limitations in fully modeling the structural features of peptides. The abundance of noncanonical amino acids (ncAAs) in peptide sequences (Figure 1) necessitates additional parameterization of these amino acids for the force fields used for computational design. Additionally, these ncAAs confer unique structural features to the peptide that are not present in the large database of protein structures. Many peptides also include crosslinkers and/or are cyclized. The presence of these crosslinkers and cyclizations adds restraint to the backbone of the peptide and leads to peptide bonds or angles that deviate from the equilibrium seen in natural proteins. Unusual puckerings of the proline ring, and peptide bonds that are out of plane are examples of such deviations [18, 19]. Thus, the heuristic knowledge gained from studying protein structures is often not enough for modeling peptide structures, and learning methods trained on protein databases will not be generalizable to all peptides.

Another challenge in designing peptides is that the one-to-one structure-function link that is often very prominent in functional proteins is blurry for peptides. In fact, many functional peptides are intrinsically disordered or sample multiple conformations [20, 21]. This conformational plasticity complicates the design process in two ways. First, it adds complexity to the already challenging design process. Second, since these conformational ensembles are often harder to experimentally validate, the number of examples one can learn from to create novel peptides is more limited compared to stable proteins with one conformation. Finally, due to their small size, many peptides lack secondary structures that are the fundamental units of designing proteins [22].

Designing functional peptides that bind to protein surfaces creates an additional challenge to the design: correct modeling of the binding competent orientation. Sampling this conformation is often done via docking or molecular dynamics simulations. However, these methods often suffer from long run times and/or low accuracy. Part of this challenge is due to insufficient sampling and ranking of generated docked models. Another contributor is the difficulty in modeling water molecules at the interface. It is well known that many peptides bind to their protein targets via water-mediated hydrogen bonds [23]. However, despite current advances in computational modeling of water molecules [24], accurate modeling of water molecules at the interface is still a computationally challenging task.

Despite these challenges, researchers have succeeded in designing peptides covering a wide range of functions: peptides that can bind to target proteins (topic of this review), catalytic peptides [25], and peptide-based higher-order assemblies [26, 27]. These methods can be broadly categorized into three strategies (Figure 2): (a) structure-guided, (b) de novo, and (c) learning-based methods. It should be noted that the lines between these methods are at times blurry and they are mainly used for structuring this review.

To provide a general overview of these design methods and the progress of the field, in the following sections we will focus the discussion on the strategies used for the design of peptides for their most common application, binding to targets, along with some representative examples.

2. Structure-Guided Design of Functional Peptides

One of the most successful strategies to design functional peptides is to start the design from an existing fragment (often ≥4 residues) or motif, that is known to have the function of interest, and improve upon that. These improvements are often minimal and retain the overall sequence of the peptide motif as well as its structure. Integrin-binding RGD loop is an example of such motifs [28]. We refer to these methods that are inspired by naturally occurring peptide motifs structure-guided design.

The motif is often extracted from natural proteins using available structures from PDB. The process of motif identification and extraction is slow and can be highly accelerated through the use of computational methods. For example, Tsai et al. [29] recently created TP-DB (therapeutic peptide design database) that allows the users to search helical motifs that match a given sequence pattern. Their pattern-specific search engine is based on features derived from nearly 1.7 million helical structures extracted from PDB. These identified helical motifs can provide a starting point for design of peptides starting from helical motifs. In another example, Alam et al. [30] used an algorithm called FlexPepBind as a screening tool in order to find potential substrates of histone deacetylase 8 (HDAC8). FlexPepBind uses the structural model of a peptide-protein complex to calculate the binding specificity of the peptide to its receptors. After calibrating the results of FlexPepBind over an existing set of peptides with known affinities for HDAC8, they successfully used FlexPepBind to fish out new substrates of HDAC8 by running it on all the peptides in human proteome that are reported to have acetylation sites. HADC8 showed reactivity towards 20 peptides from a set of 26 peptides that were experimentally tested.

After finding the motif, it is improved upon in one of several ways (Figure 3): (a) mutagenesis, (b) crosslinking, and (c) scaffolding. It is important to note that while motifs are important contributors to binding, in most cases the binding affinity of the protein partner from which the motif is extracted is better due to the contributions of interacting residues outside of the motif. Thus, as shown in some of the examples reviewed, in order to generate a peptide that can compete with the native binder, a combination of approaches is required. For example, a scaffolded motif will undergo a series of mutagenesis studies to generate binders that can compete with the native binder. These mutagenesis studies can be done in a high throughput way using library-based methods.

2.1. Designing Binders Using Motif Mutagenesis

Naturally occuring protein-protein interactions are evolved to perform a function. This function does not always correlate with high binding affinity. In fact, in many cases, promiscuous or low affinity transient binding events are required to carry out biological processes [3133]. Thus, a known peptide-protein interaction can be strengthened by making mutations to the peptide at the binding interface (Table 2). Mutagenesis often starts with a change in sequence and then assessment of that change. The change can be made purely randomly or it can be guided by biochemical knowledge of the interface such as higher shape complementarity of the peptide at the interface, or added hydrogen bond interactions. The assessment step ranges from visual inspections to computational scoring methods to high throughput experimental screening.

Starting pointTargetResults

The structure of the APY hairpin bound to the receptorEphA4 receptorKd as low as 30 nM [47]
C-terminal binding domain of the peptideC5a ReceptorSeries of agonists with a range of ki values as low as 1 μM [48]
Structure of CDK2-cyclinCyclinIC50 values as low as 2.1 0.4 μM [39]
Helix segment from HIV-1 gp41 structuregp41Ki values < 0.2 nM after mutating to new a/B AA [49]
pDI inhibitor discovered through phage display [50], binding motif from p53MDM2/MDMXKD of 0.91 nM for MDM2 and 2.31 nM for MDMX after mutagenesis including ncAAs and crosslinking [51],  nM and  nM against MDM2 and MDMX after adding a res that binds in a newly identified pocket [46]
Helical motif of SARS-COV2 bound to Human ACE-2ACE-2Showed better computational energies for designed binder [44]
Short helical segment from TRF1TRFH-Fbx4G structureubiquitin E3 ligase SCFFbx4kD as low as 23.3 ± 12.8 μM and IC50 of 31.3 μM after mutagenesis [52]

One powerful example of using mutagenesis to improve the binding and selectivity of peptides is the case of helical peptides that bind to the BCL-2 family of proteins. Modifications of this helical motif using a combination of computational mutagenesis, rational mutagenesis, and experimental screening have resulted in a series of potent and selective inhibitors for different members of this family such as Bcl-xL [34], Bfl-1 [35], and Mcl-1 [36]. Mackenzie et. al. showed that statistical information of recurring tertiary structural motifs (TERMs) [37] and their associated energies (dTERMen) can be used to predict the binding energies of the previously designed peptides against BCL-2 family members [38].

To automate the process of incorporating noncanonical amino acids into peptides for improved properties, Andrews et al. developed REPLACE (REplacement with Partial Ligand Alternatives through Computational Enrichment) [39]. This process starts with the structure of the starting peptide in complex with the target of interest. Each amino acid of interest is removed from the peptide and a variety of functional groups are instead docked into the pocket and ranked based on their calculated score function. The best ranking group will be then selected as the new mutation. This process allowed for obtaining new peptides that retained inhibitory activity against CDK2 but with overall less charge, thus a higher chance of being permeable.

In order to identify improved sequences for an existing motif at a binding site starting, Pearce et al. used EvoDesign [40]. In brief, EvoDesign for interface uses evolutionary information from other protein-protein interfaces to design better interfaces. It aligns a library of nonredundant protein interfaces [41] to the interface of interest using iAlign, a computational method developed for aligning and comparing protein-protein interfaces [42]. A multiple sequence alignment (MSA) is then generated from the best scoring aligned interface. Added to this MSA is the evolutionary profile of the scaffold itself that can be used as a guide for designing the peptide. The information from these MSA profiles is used as an energy term that is added to their main energy function, EvoEF, which is then used for optimizing sequences for best binding. The authors recently used EvoDesign in conjunction with an improved energy function, EvoEF2 [43], to design a peptide that blocks SARS-CoV-2 spike protein interaction with human ACE-2 receptor [44].

Rooklin et al. showed that focusing the mutagenesis studies to the unoccupied pockets at the interface between motif and target can result in better binding affinities. To assist with the choice of residues for mutagenesis, they developed AlphaSpace [45], a fragment-centric algorithm for identification and analysis of binding pockets at protein interfaces. At the core of the AlphaSpace algorithm are AlphaSpheres. These spheres are calculated after tessellation of the protein structure. Filtering, clustering, and ranking AlphaSpheres ultimately results in selection of protein pockets. If a pocket is part of a protein-protein interface, the shape complementarity of the residue at the pocket and the pocket can be calculated. If the difference between the volume of the residue and the pocket is large enough to accommodate at least one methyl group, it will be reported and the residue can be mutated to better fit the pocket. In a more recent study, the authors used AlphaSpace to identify secondary binding pockets on protein surfaces near the primary binding sites of MDM2 and MDMX. After identification, they performed a local docking to find out which side chains can best fit that pocket. The original peptide binder was extended to include the newly identified side-chain. The extended peptide had improved binding affinity over the original design (around one order of magnitude). Constraining the original helix and optimizing the linker resulted in final peptides with binding affinities of 4 8 and 50 16 nM against MDM2 and MDMX, respectively (original affinities were around 1 μM) [46].

2.2. Stabilizing Motifs Using Crosslinkers

Binding affinity is determined by two factors: the energy of binding at the interface (enthalpic contribution) and the entropic cost of binding as a result of moving from a free state to a bound conformation. The entropic cost is especially prominent for cases where the peptide can adopt multiple conformations in the free form. To reduce this entropic cost and thus enhance binding, many researchers have used crosslinkers that hold peptides in the desired binding competent conformation (Table 3) [53].

Starting pointTargetResults

Linear peptide known to bind in the pocketTLE1Kd as low as  nM [75]
Originally designed antagonistC5a receptorIC50 as low as 0.2 nM (cyclization combined with mutagenesis) [76]
Mutagenesis study of native protein-protein interactionsCK2 subunit interactionIC50 as low as 3.0 nM [77]
Antibody loopsInfluenza hemagglutininKd as low as  nM and IC50 as low as 30 nM (after mutagenesis) [74]
Original helical motifsNHR2-binding (N2B) motif of E-proteinsKd as low as 53 ± 20 μM [78]
Structure of B-hairpin at the binding siteCdiIKd of 13 ± 2 μM [79]
Structure of bovine immunodeficiency virus Tat with TAR RNAHIV TARKd as low as 1 nM (disulfide cyclization and mutations) [80]
Structure of LXXLL motif (NR box)Estrogen receptorKD as low as 0.075 μM for ER and 0.155 μM for ER [81]
Crystal structure of 14-3-3ζ in complex with ESp14-3-3Kd as low as 0.25 0.01 after stapling [82]
Structure of similar CYFIP1 to homologus WASF1CYFIP1Showed inhibition in cells [83]
EED binding domain of EZH2EEDKd as low as 264 nM after stapling [84]
Helical domain from Ras interaction with sosRasnucleotide-free Ras with a KD of μM and GDP-bound Ras with a KD of μM [85]
Previously known binder, StRIP3 [86]Rab8aKd as low as 7.8 μM after double stapling and mutagenesis [87]
Helical domain of HIF1a interacting with p300p300μM [88], kd as low as  nM [89]
Loop identified by LoopFinderStonin2 and Eps15 interactionsKd as low as μM and IC50 as low as μM by crosslinking the identified hot loop with Kd of μM and μM [68]
LC3 interacting regionLC3BKd of μM after stapling and mutational optimization [69]
10mer hot segment identified at the interfaceTLR4Synergistic activation of TLR signaling was observed for cyclized peptides [71]
Short helical segment from TRF1TRFH-Fbx4G structureUbiquitin E3 ligase SCFFbx4kD as low as μM and IC50 of 31.3 μM after mutagenesis [52]
BH3 domain mimicsBCL-2, Bcl-xLDissociation constants as low as  nM [90], Kd of 38.8 nM [91], IC50 as low as  nM with stapling and -amino acid [92]
BCL9 helix-CateninKi as low as μM [93]
Helix segment from HIV-1 gp41 structuregp41Kd as low as  nM after stapling [94], half life enhanced by an order of magnitude after double stapling [95], IC50 as low as 10 nM and Kd of 17 nM after stapling peptides containing the core binding residues [96]
Structure of p53 helical domain interacting with MDM2/MDMXMDM2/MDMXIC50 as low as 2 nM for MDM2 and  nM for MDMX [97], MDM2 ki as low as  nM [98], IC50 as low as  nM for MDM2 and as low as  nM for MDMX using a photoinduced cycloaddition [99], Kd of  nM for MDM2 [100]

One of the most successful examples of protein crosslinkers are those that keep peptide fragments in a helical conformation, or staples. This helix stapling method has successfully been used over the years to generate peptide binders for many targets including MDMX and BCL-2 family of proteins [5456]. There are now well-defined guidelines regarding which staplers to use and what positions to place them in a helix for better outcome [57, 58].

Similar staples do not exist for -sheet-containing motifs, but it has been shown that disulfide bonds can be used to stabilize these structures [59]. In addition to inter-strand staples, a series of turn mimetics have been developed to stabilize and assist -sheet nucleations. D-Pro…L-Pro turn is among the most commonly used motifs for -sheet nucleation [60]. -hairpin mimetics using this turn have been widely used as inhibitors of a wide variety of targets such as Viral RNA TAR element [61] and HIV-1 Rev-RRE [62], CXCR4 [63], and for general antimicrobial activity [64]. Aib…D-AA [65] or Aib…achiral AA [66] have also been used as type I’ turn mimetics and are shown to assist with beta sheet nucleation.

Not all protein-protein interactions occur through secondary structure motifs; many are mediated through the so-called hot loops, loops at the interface that majorly contribute to binding. Multiple groups are working on using developed or novel crosslinkers to stabilize these hot loops to obtain higher affinity binders for targets of interest. Gavenonis et al. used LoopFinder [67] to identify these hot loops at the protein-protein interfaces [68]. After identifying the loops, the computational ∆G contribution of each residue at the interface was calculated by computational Ala scanning using Rosetta software. Residues with >1 REU (Rosetta Energy Unit) were identified as hot spots. Hot loops were defined as regions with at least 3 hot spot residues, two of which must be consecutive, and average REU per residue of 1 or more. They further developed a binder with μM affinity against stonin2 and Eps15 interactions from a loop discovered in this analysis. In another study, inspired by the natural binder to LC3B, Cerulli et al. used unnatural amino acids and diversity-oriented stapling to obtain high affinity inhibitors of this protein [69].

In a similar approach, London et al. developed a protocol that searches protein-protein interactions and identifies hot segments by comparing the calculated binding energy (using Rosetta software) of the fragments to the binding energy of the entire protein to the target [70]. The peptides derived from these protocol were shown to be able to inhibit a number of protein targets including TLR4 [71], ubiquitin E3 ligase SCFFbx4 [52], and gankyrin-ATPase complex [72]. This protocol is available through the Peptiderive server [73] where starting from a protein-protein complex, a hot segment is suggested and if possible, is cyclized via addition of a disulfide bond or through N-to-C cyclization.

One of the most commonly occurring natural loop motifs are those involved in antibody-antigen interactions. These motifs can also be stabilized using crosslinkers to generate peptide binders. Kadam et al. showed that using a series of noncanonical amino acids, they could stabilize and close the binding loop of BnA antibody that binds at the stem region of influenza hemagglutinin, resulting in binders with affinities as low as 17 nM [74].

2.3. Grafting Motifs into New Scaffolds

An alternative to using crosslinkers for stabilizing naturally occurring motifs is to use peptide and protein-based scaffolds. In this method the entire structure of the motif or its functional groups is grafted into an already existing natural or designed scaffold (Table 4) [22, 53, 101]. These scaffolds are often highly stable and robust to mutations or insertions of new elements.

Starting pointTargetScaffoldsResults

Structure of Rev binding to HIV RNAl helicalRRE RNA-SheetMimicked by a -sheet. Kd as low as 0.1 μM [62]
RGD motifIntegrinKnottinCell binding affinities as low as  nM after grafting and affinity maturation [107]
Peptidic antagonist DALKBradykinin B1 receptorCyclotideFunctional stable peptides were obtained [110]
RRKRRR epitope that inhibits interaction of VEGF with KDR receptorVEGF-AKatalase B1 cyclotideMaintained biological activity with enhanced serum stability [111]
A number of known melanocortin 4 receptor activating sequencesMelanocortin 4 receptorKatalase B1 cyclotideKi as low as 29 nM after grafting [112]
CVX15 peptidechemokine receptor 4, CXCR4MCoTI-I cyclotideEC50 as low as  nM after trying multiple grafting strategies [113]
TCF4 strand bound to -cateninTCF4/ -catenin interaction-Hairpin mimicbest μM [116]
Structure of helical motif bound at the interfacebotulinum neurotoxin BMiniprotein of size ~45 residuesKd as low as 0.5 nM [115]
Structure of helical motif bound at the interfaceInfluenza haemagglutininMiniprotein of size ~45 residuesKd as low as 5.0 nM [115]
Structure of SARS-CoV2 spike protein with human ACE-2Human ACE-2Miniprotein of size 40-65 residuesKd of ~1 nM after affinity maturation [13]

One of the most successful classes of such scaffolds are disulfide-rich peptides [102, 103] such as knottins [104] and cyclotides [105, 106]. These highly rigid scaffolds allow for insertion of loops without losing their structure. Grafting different loops containing the integrin binding RGD motif into knottin scaffolds (Ecballium Elaterium trypsin inhibitor) followed by directed evolution resulted in integrin binders with low nM affinities [107]. Since then, it has been shown that engineered Knottins can be used as imaging, therapeutics, or delivery molecules [104, 108]. Engineering cyclotides by grafting different receptor binding loops into them resulted in peptide binders for a number of GPCR receptors [109] including bradykinin receptor [110], VEGF-A [111], melanocortin 4 receptor [112], and chemokine receptor 4 [113, 114].

In addition to interactions involving loops, secondary structures can also be grafted onto other scaffolds. Bigger peptides with folded structures called miniproteins have recently been used for stabilizing these structured motifs known to interact with protein targets of interset. This method has been successfully used to generate binders with very high affinity and selectivity against many targets including Influenza HA [115], and Botulinum neurotoxin B [115].

The secondary structure motifs have also been incorporated into smaller peptides. For example, Blosser et al. used -hairpin mimetics to stabilize an extended -strand that mediates TCF4/ -catenin interaction [116]. They started their work by first identifying the minimal interacting sequence that could compete with the native TCF4 motif (Ki of this fragment was μM). A rationally designed strand was then added to this strand with stabilizing hairpins in between. This new strand was designed to stabilize the structure, enhance solubility, and generate additional interactions. This process led to the generation of binders with 10-fold improvement in affinity (best μM).

3. De Novo Design of Functional Peptides

As its name suggests, de novo design of peptides refers to the process of designing peptides from scratch. The term de novo has been used for a number of different approaches. In this review, we will use de novo design for cases where the generated peptide does not stabilize a known naturally occurring motif at the interface. The de novo design methods to generate functional peptides can be divided into three categories (Figure 4): (a) methods that generate sequences followed by conformational modeling/sampling, (b) methods that generate structures first and make sequence modifications later, and (c) methods that sample sequence and structure simultaneously.

3.1. De Novo Generation of Sequences

The sequence-based methods (Figure 5) often start from a known binder sequence or a pool of known binder sequences and generate diversity in this pool to obtain the best binder. Due to the large space of potential sequences, often some constraints are applied to the generated sequences, and/or some optimization methods are used. The type of applied constraints used or optimization methods are often determined by the information at hand and the target of interest. The generated sequences are then ranked to find the best candidates. This ranking can be done in a number of different ways, but often requires considering the structure of the peptide-target complex or its model. While case-specific, methods for de novo sampling of sequences can be generalized to other similar cases or inspire the development of new general methods. Thus, in the following section, we describe examples of these methods for de novo design of peptides.

To design peptides that can selectively inhibit EZH2, Smadbeck et al. have developed a multi-stage framework [117]. In the first stage, a series of potential peptide sequences were generated starting from the native peptide binder to EZH2. The generated sequences were then ranked using an energy minimization method that ensured stability. This was followed by assessing the ability of the designed peptides to match the structure of the template peptide, determined by measuring the fold specificity of each sequence for the target structure. The top sequences were then subject to binding affinity calculation using a method previously reported by Lilien et al. [118]. The binding affinity for the final 45 candidates were then calculated by RosettaDock and the top designs were validated experimentally. Most of the predicted peptides were able to inhibit EZH2 with affinities in the low μM range. One of the designed peptides, SQ037, showed an IC50 of 13.57 μM, higher than that of the existing small molecule inhibitors of EZH2, which are in the nanomolar range.

Unal et al. [119] also addressed the de novo design of binders in a multi-step approach. In the first step, the binding site on the protein of interest was determined using a coarse-grained Gaussian Network Model [120122]. This step was followed by an arbitrary docking of a peptide (a known interacting partner or poly-Ala if the information on the partner didn’t exist) on this binding surface using Autodock. The next step involved calculating the binding energies of all possible amino acid pairs at the binding interface. The transition probabilities (probability of a particular amino acid occupying a position, given another amino acid is in the partner position) in the Markov based model was calculated and was used to identify the best peptide binder. They validated their method by testing binders against five different proteins, including HIV-1 protease and scytalidocarboxyl peptidase B protein. The experimental data showed improved binding affinity compared to the naturally occurring peptides. For instance, the tripeptide Trp-Tyr-Val, designed by their algorithm showed a binding affinity of −9.59 kcal/mol against HIV-1 protease. In reference, the binding affinity of a known inhibitor to HIV-1 protease, Glu-Asp-Leu, determined by the same method is -7.66 kcal/mol.

Li et al. [123] developed a method for the design of peptides that can inhibit VEGFR-3. First, they identified the binding pocket on the outer surface of the VEGFR-3 protein using FT Map [124]. They then constructed a library of 576 peptide sequences based on Mekler–Idlis sense-antisense amino acid pairing theory (Figure 6) [125]. In the next step, molecular docking simulations and conformational analyses were performed to rank these peptides with respect to their binding interactions with VEGFR-3. The best ranking peptides were then synthesized and further experimentally evaluated. Cell binding experiments identified several peptides, including CP-7, that could selectively target VEGFR-3. For example, designed peptide CP-7 (CVKTFDP) could bind to 94.6% of the A549 cells in the experimental setting compared to 57.8% bound by their positive control (LARGR).

3.2. De Novo Generation of Structure

In this method, researchers first generate peptides with predefined structures. These peptides will then be used as scaffolds to bind to a target surface of interest (Table 5). To guide the binding and increase the probability of success, these scaffolds often contain hotspot residues known to bind to the target of interest. The design process needs to ensure that these hotspot residues are represented in an orientation that will be competent for binding. Thus, these approaches are more hands-on than those described in sections 2.2 and 2.3 where an entire secondary structure or loop is crosslinked or grafted.

Designed scaffoldTargetResults

FoldamersMDM2IC50 as low as 24 nM [127]
FoldamersVitamin D receptorKd as low as 0.12 μM [127]
Terphenyl scaffoldsBcl-XLKi as low as 0.114 μM [128]
De novo coiled coilsMCL-1IC50 as low as 0.09±0.005 μM [130]
[3]-peptidegp41EC50 as low as 5.3±0.5 [134], one variant was later improved by near 7 folds through mutagenesis [135]
-peptidehDM2IC50 of 80.8 ± 3.2 nM [136], further improved to as low as 27.6 nM for MDM2 upon mutagenesis [137]
-hairpinHDM2IC50 values as low as 0.14±0.06 after optimization [129]
Designed peptide-peptoidTCF/-cateninBest binder had IC50 of 5.44 0.82 μM [133]
Miniproteins of ~65 residueSARS-COV25 pM after affinity maturation [13]

It has been shown that in cases where the interactions between the motif and the target are mainly through side-chains, presenting these side-chains in the right orientation is enough to achieve binding. These cases often involve secondary structure features such as -helicx and -sheets. A library of peptide-like foldamers developed that can fold into stable alpha helices and -sheet mimetic purely based on their backbone propensities, thus allowing open sites for presenting functional side-chains in the orientation similar to the motif.

The abundance of helical motifs at protein-protein interfaces (Figure 7) [126] created a large space for the use of -helix mimicking foldamers with many successful examples reported to date. One successful example of -helix mimicry are peptide-oligourea hybrids, used to inhibit a variety of proteins such as MDM2 and vitamin D receptors [127]. Another example is the use of Terphenyl scaffolds as helix mimetics to generate binders against Bcl-xL and Bak interactions [128].

Small -hairpins can also be used to mimic the placement of key residues presented by helical motifs. For example, Fasan et al. used a small -hairpin to stabilize two hot spot residues at the interface of p53 with HDM2. After a series of mutational studies, they obtained a final peptide with binding affinity in the order of 100 nM [129].

De novo designed alpha-helical coiled-coil structures are another class of scaffolds that can be used to stabilize interactions at the protein-protein interfaces. Fletcher et al. [130] have designed de novo parallel homodimeric and heterodimeric -helical coiled-coil scaffolds to target and disrupt the interface of the MCL-1/NOXA-B by stabilizing key interacting residues. The critical residues in NOXA-B that are involved for binding were first identified by computational alanine scanning, using Robetta [131], and classified into a short motif and an extended motif. These residues were then grafted on the sequence of the de novo peptide CC-Di [132] to get a series of possible hybrid designs. These designed peptides were synthesized in the lab and were evaluated to check for their purity and stability. The inhibitory capabilities of these hybrids were checked by fluorescence spectroscopy and it was concluded that their binding to MCL-1 stabilized these coiled-coiled assemblies.

Hotspot residues at the interface that guide the design of peptide binders can be generated computationally. For example, Cao et al. used miniproteins as their starting scaffolds to present computationally generated hotspot residues and developed inhibitors with IC50 as low as 5 pM against SARS-CoV-2 spike protein [13].

The de novo design of binders can also start in a hotspot-free manner. In this method, researchers start with searching a library of existing peptides for shape-complementarity to the surface of interest, and enhance binding through sequence design. For example, starting with the structure of TCF/-catenin interaction, Schneider et al. searched for the best scaffold from a series of previously designed high-resolution peptoid or peptoid-peptide hybrid scaffolds that can best mimic and inhibit the -hairpin motif at this interface [133]. After finding the best scaffold, they performed a series of computational mutagenesis and tested the best ranking binders experimentally. Their best binder had an IC50 of 5.440.82 μM and showed inhibition in cells as well.

3.3. Simultaneous Optimization of Sequence and Structure

Simultaneous optimization of sequence and structure, while intuitive, can be computationally expensive. Molecular dynamics (MD) simulation methods that take into account structural changes are often time-intensive, and including mutations in the process is not feasible. De novo methods to design sequences that stabilize a given structure can include slight backbone movements, but large-scale movements that can be observed in MD are often not possible as the sequence optimization problem for a single rigid backbone is in itself an expensive problem. However, for very small peptides, these two processes can be combined to design peptides with pre-defined structures or function.

While linear peptides are often too flexible to be accurately designed, constrained peptides such as cyclic peptides have fewer degrees of freedom and can be designed computationally to adopt predefined shapes. In 2018, Slough et al. designed rigidly structured peptides (pentamers) in an aquatic environment using MD [138]. They simulated a library of cyclic pentapeptides differing only in two residues (excluding the Pro) using BE-META MD. They used the result of these simulations to develop a score function for predicting structures of cyclic pentapeptides. This score function was developed based on the preference of 2-mer pairs to adopt a given conformation during the simulation. One of the peptides in the library was well-structured, and its structure was confirmed by NMR as well. However, most other sequences in the library were not well-structured. They observed that some turn combinations that build up these pentamers score poorly; thus, it is very unlikely to obtain a structured peptide containing these turn combinations using the substitutions they considered initially. Therefore, the authors proposed using ncAAs to stabilize these and other conformations. They also suggested that their method can be extended to larger peptides and other types of cyclic bonds.

In addition to MD-based methods, Rosetta software has been used to design structured cyclic peptides. The Rosetta-based design takes advantage of a method called generalized kinematic loop closure (GenKIC, Figure 8), which is inspired by a robotic approach used to calculate accessible conformations for a constrained object (e.g., a finger fixed at joints) [139]. Using this method, the authors sampled a large set of potential cyclic backbones starting from poly-Gly sequences. After clustering these sampled conformations, they performed sidechain assignment by restricting D- and L- amino acids to positions with phi > 0 and phi < 0, respectively. By preventing the use of Gly in the design due to its flexibility, the authors made sure that the assigned sidechains were optimized for stabilizing their backbone conformation. NMR studies confirmed that these peptides indeed take the conformations they are designed for, showing a success rate of >70%. An analysis of the computationally generated library of 7-10 residue cyclic peptides identified more than 200 stable structures using this method [140].

Compared to MD-based methods, the Rosetta-based method has the advantage of being very fast; thus, it can cover near-complete space of D- and L-amino acid combinations in peptides of size ranges as large as 11 amino acids [140]. Additionally, it can be used in line with other design modules to sample both conformations and sequences of cyclic peptides at the target protein surface to design peptide binders. To use this method for the design of target-specific cyclic peptides, a known binding molecule is used as the starting point or anchor on the target surface; then, the backbone is expanded and circularized at the pocket. The conformations of cyclic peptides are sampled using GenKIC, and for each conformation, sequences with maximum favorable interactions to the protein surface are designed. After sidechain and rotamer refinement, the computationally generated designs are filtered based on a number of interface metrics such as shape complementarity, and ∆G of binding, and the best designs will be evaluated using downstream experimental approaches [141].

Using this method, several groups have been able to successfully design peptides for therapeutic targets, such as programmed cell death 1 (PD-1) receptor [142], antibiotic resistance factor NDM1 [143], and histone deacetylase (HDAC-2 and HDAC-6) enzymes [144]. Although the overall design process of these studies is the same (described above), there are differences in each. Mulligan et al. introduced energy terms that could guide the design not just by the overall score but also by the interface quality. They also showed that more rigid peptides have better binding, as the flexible peptides have a higher enthalpy and will negatively affect ΔG of binding [143]. Hosseinzadeh et al. showed that high-affinity binding could be obtained with nonstructured peptides. They also added an exhaustive search near the binding pocket to find the best interacting residues before extending the peptide and showed it highly increased the success rate [144]. The best designed peptides for PD-1 had a Kd of 30 μM and 102 μM, which could outcompete the natural ligand. Six out of seven designed peptides chosen for experimental validation of NDM1 inhibition had lower IC50 values than D-captopril control and the top peptide had an IC50 value of 1.2 ± 0.1 μM, 50 times lower than the control. 17 of 39 peptides tested for HDAC-2 had IC50 values of 100 nM or lower. The best HDAC-6 and HDAC-2 binders showed IC50 values of around 4 and 9 nM, respectively, three orders of magnitude better than the starting anchor.

4. Machine Learning for Design of Functional Peptides

While the field of protein structure prediction has been revolutionized by the recent advances in deep learning [145, 146], predicting structure or function of peptides, or designing new functions is still a challenge. In this section, we first review the successful application of learning methods for the prediction of structure or function of peptides (Section 4.1). We will then describe recent advances in designing new peptides using machine learning (ML) in Section 4.2.

The learning methods that are applied for design and prediction heavily depend on the peptide representation used as well as the type and number of available data in the training set (Figure 9). Thus, our main effort in this section is to cover a wide variety of representations and methods to cover the wide scope of the field.

4.1. Predictive Methods
4.1.1. Prediction of Structure

Recent developments in computational algorithms have resulted in great progress towards designing cyclic peptides that stabilize a predefined structure and fold into one conformation (see section 3). However, many native peptides can adopt multiple conformations in solution. The ability to predict these conformations is helpful in the design of peptide fragments with desired functions and in predicting their function. But, predicting the various conformations that cyclic peptides can adopt in solution is still a challenge.

One advancement in this area is StrEAM that predicts peptide conformations through combining MD simulations with deep learning. Miao et al. [147] have tailored a metadynamics method using explicit solvent for modeling cyclic peptide conformations. In this method, they used a bias-exchange metadynamics process to capture the transitional motions of cyclic peptides. While simpler models that do not consider solvent can be used to design cyclic peptides, to accurately model the solvent-exposed backbone C=O and N–H bonds which are sometimes associated with caged water molecules, one needs to use an explicit solvent model. In this algorithm, they have described the structure of penta-peptides by assigning the (phi, psi) angle of each residue to one of the ten different bins in the (phi, psi) space, resulting in a string of torsion bins. They then used results from multiple MD runs on several peptides and interactions between residues i,i+1 or i,i+2 as inputs to a neural network predictor. This algorithm could efficiently predict structural ensembles of cyclic peptides with MD simulation. StrEAMM can also be extended to larger cyclic peptides due to inclusion of the above mentioned longer range interactions.

4.1.2. Prediction of Function

In addition to structure prediction, computational techniques such as probabilistic models, molecular modeling, and deep learning have been extensively used in screening novel biologically active peptides and predicting their functions. In the following sections, we discuss examples of these applications.

(1) Prediction of cell penetrating cyclic peptides. One function of interest for peptides is their cell penetration potency and extensive work has been done in developing predictors for cell penetrating peptides (CPPs). The methods developed for prediction of CPPs (Figure 10) can be adapted for prediction of other peptide functions. A summary of the examples discussed in this section is shown in Table 6.

MethodTraining SetFeaturesResultAccuracy

CellPPD [148]: SVM, MEME/MSAT708 experimentally validated, unique
CPPs from the CPPsite database
Amino acid composition, Dipeptide Composition, Physicochemical Properties of amino acidsDetailed 120 CPP motifs identified from their dataset97.40%
SkipCPP-Pred [150]: K-skip-n-gram model, RF classifier1855 experimentally validated CPPs from CPPsite 2.0Input sequencesHigh quality dataset CPP924 with reduced redundancy90.6%
CPPred-FL [152]: RF classifierBenchmark dataset CPP924 containing 924 peptide sequences6 features based on class information and 19 features based on probabilistic informationNew feature representation algorithm that includes all sequence based feature descriptors91.2%
BChemRF-CPPred [154]: MLP architecture, SVM, Gaussian process classifierCPPsite 2.0, C2Pred serverStructure based descriptors including molecular weight, number of rotatable bonds, topological polar surface area, number of hydrogen bond donors and acceptors, etc. Amino acid composition-based descriptors including dipeptide composition.
Pseudo-amino acid composition related to hydrophobicity, hydrophilicity, side chain mass, etc
Evaluated the influence of structural and physicochemical properties in the permeability of peptides89.62%

Gautam et al. [148] have developed a web-server, CellPPD, that can predict and design CPPs. This web-server uses a support vector machine (SVM) for prediction of CPPs. A dataset of 708 CPPs from CPPsite was used as the input along with features like amino acid composition, dipeptide composition, and physicochemical properties. In addition to SVM-based predictors, their web server also uses a hybrid motif-based predictor using MEME/MSAT program [149] which can achieve an accuracy of 97.40% for their dataset, greater than the SVM based or MEME/MSAT program alone. To understand the dominance of certain amino acids in CPPs, they calculated and compared the amino acid composition in the CPPs and concluded that residues like Arg, Lys and Trp were present abundantly in CPPs. The authors also analyzed the amino acid distribution in different positions and concluded that while basic residues like Arg and Lys were present throughout the sequence, the N- and C- terminals preferred different sets of residues. A list of 120 CPP motifs present in their dataset has also been provided. The program can also perform point-mutations on a user-specified sequence to suggest modifications with higher CPP properties.

Wei et al. [150] developed an improved predictor called SkipCPP-Pred. SkipCpp-Pred uses sequences collected for the CPPsite 2.0 as an input and converts them to a feature through application of the k-skip-n-gram model. [151] The k-skip-n-gram model is based on a language processing probability model that tries to predict the context or surrounding words in a string from a target word and thus can capture the correlation between residues. These features are then used to train a random forest classifier that performs the prediction. The accuracy of prediction by SkipCPP-Pred was 90.6% compared to 87% and 83.7% for CellPPD-DC and CellPPD-BP, respectively. In addition to developing this model, the authors also created a high-quality dataset with reduced sequence redundancy to decrease performance bias.

Qiang et al. [152] developed a tool called CPPred-FL. This tool uses random forest classifiers with composition and physicochemical properties of amino acids as descriptors. To address some of the challenges regarding proper feature representation for CPP prediction, the authors employed a new feature representation algorithm (including a learning and an optimization step) that transforms high dimensional feature space into a low dimensional space. With inclusion of their top 19 features, their model can achieve a performance with 92.1% accuracy.

In 2019, Fu et al. [153] developed a machine learning based method to predict CPPs. They considered four features based on amino acid composition, such as frequency of amino acid and composition of amino acid group pairs, and trained their SVM model to classify CPPs. Their predictor showed an accuracy of 92.3% in predicting CPP in the training dataset, better than existing state-of-the-art methods like CellPPD and SkipCPP-Pred.

Further improvement of existing algorithms has led to the design of better classifiers over the years. For instance, de Oliveira et al. [154] have developed an ML-based framework termed BChemRF-CPPred that is capable of differentiating CPPs from non-CPPs. This framework used different classifiers including an artificial neural multilayer perceptron or MLP architecture, SVM and a Gaussian process classifier. It then employed a voting classifier to make the final prediction based on the prediction of each one of these models. In contrast to other methods that use sequence features, this method uses structure-based descriptors that can be related to the permeability of biological membranes such as molecular weight, number of rotatable bonds, topological polar surface area, net charge, and number of negatively charged amino acids. The inclusion of structure-based features improved the accuracy of the model compared to the model based on sequence input only (from 86.5% to 90.66%). They also compared their method with existing methods like CPPred-RF and SkipCPP-Pred using an independent dataset of natural peptides (60 CPPs and 75 non-CPPs). BChemRF-CPPred obtained an accuracy of 89.62% in comparison to CPPred-RF and SkipCPP-Pred’s 68.88% and 62.58% respectively.

(2) Prediction of anti-cancer peptides. Peptide-based cancer treatment is one area of interest because of peptides’ high specificity, low toxicity, and membrane permeability. One of the hurdles to predict the biological function of anti-cancer peptides is the lack of sufficient experimental data. To overcome this limitation, Chen et al. [155] have designed a deep learning method termed xDeep-AcPEP that can predict anti-cancer activity of peptides against six different tumor cells: breast, colon, cervix, lung, skin, and prostate. In this method, multi-task learning (MTL) was explored where multiple neural networks were trained independently for individual prediction tasks. For training their model, the authors considered the experimentally verified anticancer peptides (ACP) from the CancerPPD database. Four descriptors AAINDEX, BLOSUM62, ZScale (which reflects the physicochemical properties of amino acids such as lipophilicity, steric properties and electronic properties of amino acids) [156] and one-hot-encoding were used in combination to encode a sequence at the residue level. The encoder model architecture was chosen as a combination of convolutional layers with ReLU activation functions, average pool layers, batch normalization layers and Max Pool. Six convolutional neural network (CNN) models were trained for six tissue-specific ACP dataset. Comparisons between single-task learning (STL) and MTL based models showed that MTL increases the generalizability of the model. The model was tested on 19 recently published ACPs to test its predictive power. The results showed that these models were capable of accurately predicting the anticancer properties of highly active peptides but were less accurate for inactive peptides or peptides with weak activities.

(3) Prediction of antimicrobial peptides. Antimicrobial peptides (AMP) are attractive potential substitutes for antibiotics to fight against antibiotic resistant infections. Extensive study on AMPs have led to considerable efforts to develop computational methods for accurate prediction of AMPs from conventional machine learning models to deep learning (see Table 7).

MethodTraining setResult/OutputAccuracy

14 ML algorithms including RF, CART, SVM, K-nearest neighbor, Logistic Regression (LOGREG), Adaptive Boosting (AB) [157]HemoPI-1, HemoPI-2 and HemoPI-3Predicts hemolytic nature with more accuracy than hemolytic activityHemoPI-1: 94-95%
HemoP2/3- 75-77%
AI4AMP [158]: PC6 encoded Deep Learning ModelAPD3, LAMP, CAMP3, DRAMPConsiders physicochemical properties of aminoacids as features90%
Deep-AmPEP30 [159]: CNN, Reduced AAC (RAAC)Subset of a training set used in AmPEP [160] (experimentally validated AMPs from APD3, LAMP, CAMPR3)Predicts AMP activity of short peptides from genomic DNA.77.13%
AntiBP [161]: ANN, Quantitative Matrices (QM), SVM486 peptides shorter than 61 residues from the APD databasePreference for certain residues at the N- and C-terminals that demarcate them from non antibacterial peptidesANN based: 88.17%, QM based: 90.37% SVM based: 92.11%
AMPA [162]: Web application to assess the antimicrobial domain in protein. Can serve as template for AMP designHigh throughput screening results from amino acid substitutions on bactenecin 2AQuick discovery of new AMP patterns in protein sequencesaccuracy of 85% and a sensitivity of 90% [163]
AVPpred [164]: SVM1245 experimentally validated antiviral peptidesPredicts highly effective AVPs86%
iAMP-2L [165]: Pseudo-amino acid composition, Fuzzy K-Nearest neighborCurated set of AMP and non-AMP peptides collected from UniprotPredicts if a given sequence is an AMP or not and if yes, the next classification is done on the type of AMP92.23%
AMP_Scanner [166]: Deep Neural Network with the Keras framework ( using a sequential model and a TensorFlow deep learning library back-endAPD vr.3 databaseThis model automatically extracts expert-free features and hence removes the reliance on domain experts for feature construction91.01%

One of the main problems with AMPs is their hemolytic nature. Hemolysis disrupts the erythrocyte membrane and decreases the lifespan of RBCs. Thus, it is very important to identify and predict the hemolytic activities of AMPs. Plisson et al. [157] have developed ML methods to discover AMPs with low hemolytic properties. The model was trained on three publicly available datasets (HemoPI-1, HemoPI-2, and HemoPI-3) and employed 14 ML algorithms, including CART, random forest, and SVM, for binary classification. The model was tested on 3081 natural antimicrobial peptides from the APD database and 317 known hemolytic antimicrobial peptides were considered for validation of the model. 56 physicochemical properties, including aromaticity, hydrophobic ratio, net charge, and bulkiness were used as features for the model. The algorithm predicted the hemolytic nature of the peptides with greater accuracy compared to the extent of the hemolytic activities.

Lin et al. [158] developed a web-based server called AI4AMP that can accurately predict the antimicrobial potential of a given sequence. In this work, they proposed a new protein encoding method called PC6 (physicochemical component 6) that included features like hydrophobicity, volume of side chains, and isoelectric point. The server has an precision of 90%.

Yan et al. [159] developed a deep learning based tool called Deep-AmPEP30 for prediction of AMP activity of short length peptides from genomic DNA. The prediction is based on a reduced dataset and uses a deep CNN. They validated their model by identifying a 20 amino acid long peptide from the genome of the fungus, Candida glabrata, and the activity of this peptide was verified through experiments.

(4) Prediction of peptide-protein interaction. Many peptides exert their functions through binding to a protein partner; therefore, predicting a peptide-protein interaction is of huge interest. Lei et al. [167] have designed a deep learning framework, called CAMP, that can predict whether a peptide and a protein pair with each other. CAMP uses peptide-protein complexes from the PDB and the known peptide-protein pairs from the DrugBank Database. It then uses PLIP [168] to extract peptide-protein pairs that interact via noncovalent interactions. The model then creates feature profiles for both the peptide and the protein partner using structure and physicochemical properties of residues. It also uses PepBDB [169] to extract labels for residues that are involved in binding. The features for the peptides and proteins are then fed into a CNN layer and a self-attention module to predict the interacting pairs as well as peptide residues that are involved in making the interaction. Their model had a mean of AUC value of 0.803 and mean of average MCC value of 0.489. They showed that their model can correctly predict known protein-peptide pairs such as GLP-1 receptor and Semaglutide.

4.2. Generative Methods

One of the key limitations in application of learning-based methods for the design of functional peptides has been the lack of data for training. To overcome this limitation, researchers have developed creative approaches to link data generation through (semi-)high throughput methods with learning to improve the process of design. It has been shown that learning from a representative subset, even if small compared to the entire space, can result in significant improvement in the success rate of the original methods [170]. In this section we will first discuss examples of these methods (Section 4.2.1) and then move on to fully computational methods that tackle design using available databases (Section 4.2.2). The inputs and learning methods highly depend on the function of interest and available data. Thus, our focus in this section is to review cases that cover a variety of inputs and models. A summary of the examples discussed in this section is shown in Table 8.

TaskInput dataModelOutput

Binders for Bcl-xL, Mcl-1, or Bfl-1Set of 10,000 previously generated bindersSORTCERY [171]: Support vector machine for prediction, applied to a larger sequence library36 high affinity binders for each target with Ki as low as 5.71.2 nM for Mcl-1, 1.320.18 nM for Bcl-xL, and 6.61.0 nM for Bfl-1
Finding peptide hits for Sfp and AcpSExperimental data of known hits and fragments that don’t bind, each round new data was addedPOOL [172]: Naive BayesianSelective binders for Sfp or AcpS after 4 rounds
Design of cell permeable peptides600 PMO-miniprotein conjugatesRNN for generation and CNN for prediction [173]Synthesized and characterized 12. Highly active for macromolecule delivery
Peptides with anticancer propertiesCurated peptides from CancerPPD that target breast or lung cancer, pepCAST descriptors were used to find featuresCPANN [174]: counter propagation artificial neural network for prediction and modLAMP for generation6/15 peptides predicted to have anticancer activity indeed showed anticancer activity
Design of chitin binding peptides21 million amino acid sequences taken from the Pfam database and 20000 sequences from the SCOPe databaseTwo Bi-LSTM each with 1024 hidden layers [175]Interactive residues in the two predicted peptides matched experimentally known ones
Designing nonhemolytic AMPsDBAASP databaseRNN for classifying active, inactive, hemolytic, and nonhemolytic sequences [176]12/28 were active

4.2.1. Design-Test-Learn Methods

Jenson et al. [171] used a combination approach to generate functional peptides with high affinity and selectivity to three different targets: Bcl-xL, Mcl-1, or Bfl-1. They used a high throughput method called amped SORTCERY to measure the binding affinity of a library of around 10,000 peptides selected from previously designed combinatorial libraries against their target proteins. They then employed support vector regression to fit this experimental dataset. Next, they applied this predictor to their entire library of nearly 27,600,000 peptides as well as a large input dataset of 1014 potential peptide sequences. To search new sequences with desired binding features (for example high affinity to one while selective against other members), the authors applied an integer linear programming solver. They were able to generate 36 peptides that could bind with high affinity and specificity to either of the three proteins, namely, Bcl-xL, Mcl-1, or Bfl-1. The advantage of such a sequence-based model is the faster process of searching the sequence space to design the structure of the peptide.

Tallorin et al. [172] have developed a hybrid computational and biochemical method that can rapidly optimize peptides for a given biochemical function. This method, named POOL (Peptide Optimization with Optimal Learning) iteratively considers experimental data and fits them into a mathematical model to predict new hit sequences. In its core, POOL is a Naive Bayes algorithm that learns the probability of a given sequence to be a hit based on the training data. The initial database given to POOL was a combination of known peptide and protein fragments that were substrates for protein targets of interests Sfp (surfactin phospho-pantetheinyl transferase from Bacillus subtilis) and AcpS (holo-acyl carrier protein synthase from Streptomyces coelicolor) as well as a set that were known to be inactive. After each experimental round, the new data was added to this dataset. At each round, POOL is asked to predict a certain number of potential peptide hits such that the total probability of obtaining at least one hit is maximized. One key component of POOL is the fact that after it identifies each hit, the process of finding the next hit assumes the previously recommended peptides are not hits. This assumption ensures that multiple models are considered for predicting new hits. Compared to traditional predict-then-optimize methods, POOL can more effectively sample the space of functional peptides due to a better balance between exploitation and exploration, as suggested by the authors. After four rounds of learning and testing, the authors obtained short peptides that could selectively label AcpS or Sfp.

Schissel et al. [173] have combined high throughput experiments and deep learning approach to design de-novo abiotic nuclear targeting miniproteins. Their model was trained on a dataset consisting of unnatural residues and structures. The authors synthesized and tested a library of 600 phosphorodiamidate morpholino oligomer (PMO)-miniprotein conjugates using a combination of abiotic peptide molecules. These sequences were labeled with their experimental activities and were used to train their CNN-based predictor. Their generator was a nested LSTM neural network architecture. Sequences generated using this generator were then fed into a cycle of prediction and optimization (using a genetic algorithm) to reach new sequences with high likelihood of delivery across cell membranes. The model predicted hundreds of miniproteins and they synthesized and characterized 12 of them. The predicted activities for nearly all the sequences matched the experimental values.

4.2.2. Fully Generative Models

De novo design of anticancer peptides (ACP) holds importance in cancer therapeutics. In this context, Grisoni et al. [174] have presented a counter propagation artificial neural network (CPANN) [177] to generate peptides that have potent anticancer properties. The counterpropagataion network [178] is a statistically optimal self-programming lookup table. Databases of peptides targeting breast and lung cancer were curated manually from the CancerPPD database and used to train the CPANN model. The pharmacophore features of the peptides were determined using the pepCATS descriptors. The CPANN model consists of an input hidden layer called the Kohonen layer and an output layer called the Grossberg layer. The Kohonen networks are used to cluster the dataset into previously unknown groups. This model was applied to 1000 alpha helical peptides generated by modLAMP [179] de novo design engine to predict their anticancer property. The modLAMP package is a Python based software package that is used to design, classify and visually represent peptide data. The top 15 peptides along with three random peptides and three negative control peptides from the library were synthesized and tested in vitro on MCF7 and A549 cell lines. Out of the 15 positive peptides, six showed anticancer activities. Five out of these six showed activities against both cell lines, implying the multi-target prediction ability of this method.

Caceres-Delpiano et al. [175] developed a deep learning model that can design peptide sequences that are analogous in structure to a representative peptide. In this model, each amino acid was considered as a token and represented the protein as a vector. Their network architecture was made up of two Bi-LSTM layers, each having 1024 hidden units. The model was trained on nearly 21 million amino acid sequences taken from the Pfam database and 20000 sequences from the SCOP database. The model could reconstruct the protein sequence and minimize cross-entropy loss to predict various information from the sequence, including secondary structure, structural similarity, and contact maps. This method achieved 91.2% accuracy for classifying structures from the SCOP database. The authors then tested the application of this method to generation of sequences that can adopt a known structure and thus can perform a desired function. They chose chitin binding as their proof of concept model. The generation cycle started with a poly-Ala sequence. A mutator would create a library of sequences starting from this input sequence and the structural similarity of these sequences to the desired structure were calculated using their predictor. The best peptides were then used as the new input and this cycle was continued until a structural similarity score greater than 3.5 to the starting structure was attained. These predicted structures were then subjected to molecular dynamics simulations in the presence of a chitin surface to test their anti-fungal properties and interactions. These de novo designed peptides were expressed in vivo and their various biochemical properties were assayed. The authors showed that these peptides were able to replicate the antifungal properties of the wild type peptide, AC2-WT.

Capechhi et al. [176] have used ML for designing non hemolytic AMPs. Their models were trained using a combination of predictive and generative RNNs on training sequences obtained from the DBAASP database, which contained annotated sequences with information about their activities and hemolytic properties, if known. The model was further fine-tuned via transfer learning using active and nonhemolytic peptides against specific strains: P. aeruginosa/A. Baumannii and S.aureus, respectively. These fine-tuned models were sampled to generate sequences of new AMPs. The generated sequences were further classified by the RNN AMP activity classifier and the hemolytic classifier sequentially. 28 short peptides (less than fifteen residues) were synthesized and tested for their minimum inhibitory concentrations (MIC) against bacteria. A MIC value ≤16mg/mL was considered as a threshold of the peptide’s antimicrobial activity. It was seen that 64% of these peptides were active below this threshold, and out of these, 50% were both active and nonhemolytic.

5. Future Outlook

Design of functional peptides and protein fragments that bind to protein interfaces offer an exciting opportunity for understanding and inhibiting protein-protein interactions and developing new therapeutics. The design methods aimed at mimicking interaction of one protein with another span a variety of techniques from stabilization of naturally occurring binding partners to designing new sequences at the interface to advanced deep learning methods.

Methods that start from an existing natural motif have resulted in generation of many high affinity peptide binders, some leading to companies based on peptide-based binders, as well as huge advances in the field of peptide and protein design such as ncAA parametrization, cyclic peptide conformational sampling and prediction, and machine learning methods that can learn from sparse datasets. However, mimicking naturally occurring motifs inherently limits the design to cases where a structure is known. Additionally, the obtained affinities are highly dependent on the starting motif affinity and downstream mutagenesis and optimizations are often required to obtain peptides that outcompete the native binder.

The paradigm in recent years has shifted more towards application of automated methods and computational approaches for designing peptides and predicting their functions. Computational methods such as Rosetta-based design of peptides and MD simulations, while successful, are still hampered by the limitations of their underlying score function. Computational interface metrics often fail to fully justify the experimental results and predicting the binding mode is still an extremely challenging task. Improvements in the score function as well as more efficient sampling methods will improve the power of these methods in designing new functional peptides.

Learning-based methods for prediction and design of peptides are still at their early phases. The major roadblock in the application of these methods is lack of experimental data. To empower these approaches, we need both technological advances in learning methods and increased training data. To this end, approaches that combine high-throughput testing of peptide libraries with learning methods are of high value as they generate much needed labeled datasets for peptide functions. An example can be integration of computationally designed peptides with library-based screening methods such as phage display or mRNA display to generate a focused library whose screening can generate more informative results.

The huge leap in our ability to predict protein structures has inspired many researchers to develop novel approaches for predicting peptide structures, especially in complex with their binding partners. Access to accurate predictions of protein-protein or protein-peptide complexes combined with a more accurate prediction of binding affinities from these structural data will be transformational for the design of peptide binders, particularly in designing selective binders, an area that still lacks a general and robust pipeline.

Most peptide binders are designed with the ultimate goal of being used as a therapeutic. Binding is the first step in the long process of therapeutic development. Many high affinity candidates fail in the downstream steps due to undesired pharmacokinetic properties, half-life in the body, toxicity, and more. Additionally, most binding assays fail to test for cell permeability for intracellular targets, resulting in high affinity binders that cannot access the intracellular targets they’re designed to bind to due to lack of permeability. Thus, predicting permeability and pharmacokinetic properties will be an important area of focus for downstream applications. Lack of large-scale data for these features necessitates the use of new features and methods for prediction from noisy datasets with high failure rates.

Conflicts of Interest

The authors declare no competing interests.

Authors’ Contributions

S. Gupta and P. Hosseinzadeh devised the outline of the review. All authors contributed to writing, editing, and making figures.


The authors acknowledge NIH grant DP2GM146249 for funding and support.


  1. V. Mikol, J. Kallen, G. Pflügl, and M. D. Walkinshaw, “X-ray structure of a monomeric cyclophilin A-cyclosporin A crystal complex at 2·1 Å resolution,” Journal of Molecular Biology, vol. 234, no. 4, pp. 1119–1130, 1993. View at: Publisher Site | Google Scholar
  2. M. J. Gunzburg, K. Kulkarni, G. M. Watson et al., “Unexpected involvement of staple leads to redesign of selective bicyclic peptide inhibitor of grb7,” Scientific Reports, vol. 6, no. 1, article 27060, 2016. View at: Publisher Site | Google Scholar
  3. J. Koehbach, M. O’Brien, M. Muttenthaler et al., “Oxytocic plant cyclotides as templates for peptide g protein-coupled receptor ligand design,” Proceedings of the National Academy of Sciences of the United States of America, vol. 110, no. 52, pp. 21183–21188, 2013. View at: Publisher Site | Google Scholar
  4. J. D. Hegemann, M. Zimmermann, X. Xie, and M. A. Marahiel, “Lasso peptides: an intriguing class of bacterial natural products,” Accounts of Chemical Research, vol. 48, no. 7, pp. 1909–1919, 2015. View at: Publisher Site | Google Scholar
  5. D. Frechet, J. D. Guitton, F. Herman et al., “Solution structure of RP 71955, a new 21 amino acid tricyclic peptide active against HIV-1 virus,” Biochemistry, vol. 33, no. 1, pp. 42–50, 1994. View at: Publisher Site | Google Scholar
  6. J. D. Hegemann, M. Zimmermann, X. Xie, and M. A. Marahiel, “Caulosegnins I–III: a highly diverse group of lasso peptides derived from a single biosynthetic gene cluster,” Journal of the American Chemical Society, vol. 135, no. 1, pp. 210–222, 2013. View at: Publisher Site | Google Scholar
  7. H. Nar, A. Schmid, C. Puder, and O. Potterat, “High-resolution crystal structure of a lasso peptide,” ChemMedChem, vol. 5, no. 10, pp. 1689–1692, 2010. View at: Publisher Site | Google Scholar
  8. C. Ulens, R. C. Hogg, P. H. Celie et al., “Structural determinants of selective α-conotoxin binding to a nicotinic acetylcholine receptor homolog AChBP,” Proceedings of the National Academy of Sciences of the United States of America, vol. 103, no. 10, pp. 3615–3620, 2006. View at: Publisher Site | Google Scholar
  9. S.-T. D. Hsu, E. Breukink, E. Tischenko et al., “The nisin–lipid II complex reveals a pyrophosphate cage that provides a blueprint for novel antibiotics,” Nature Structural & Molecular Biology, vol. 11, no. 10, pp. 963–967, 2004. View at: Publisher Site | Google Scholar
  10. P. H. Kussie, S. Gorina, V. Marechal et al., “Structure of the MDM2 oncoprotein bound to the P53 tumor suppressor transactivation domain,” Science, vol. 274, no. 5289, pp. 948–953, 1996. View at: Publisher Site | Google Scholar
  11. H. Dai, A. W. Case, T. V. Riera et al., “Crystallographic structure of a small molecule SIRT1 activator-enzyme complex,” Nature Communications, vol. 6, 2015. View at: Publisher Site | Google Scholar
  12. M. Wittekind, C. Mapelli, B. T. Farmer et al., “Orientation of peptide fragments from Sos proteins bound to the N-terminal SH3 domain of Grb2 determined by NMR spectroscopy,” Biochemistry, vol. 33, no. 46, pp. 13531–13539, 1994. View at: Publisher Site | Google Scholar
  13. L. Cao, I. Goreshnik, B. Coventry et al., “Novo design of picomolar SARS-CoV-2 miniprotein inhibitors,” Science, vol. 370, no. 6515, pp. 426–431, 2020. View at: Publisher Site | Google Scholar
  14. U. Dietrich, R. Dürr, and J. Koch, “Peptides as drugs: from screening to application,” Current Pharmaceutical Biotechnology, vol. 14, no. 5, pp. 501–512, 2013. View at: Publisher Site | Google Scholar
  15. J. L. Lau and M. K. Dunn, “Therapeutic peptides: historical perspectives, current development trends, and future directions,” Bioorganic & Medicinal Chemistry, vol. 26, no. 10, pp. 2700–2707, 2018. View at: Publisher Site | Google Scholar
  16. A. Henninot, J. C. Collins, and J. M. Nuss, “The current state of peptide drug discovery: back to the future?” Journal of Medicinal Chemistry, vol. 61, no. 4, pp. 1382–1414, 2018. View at: Publisher Site | Google Scholar
  17. M. Muttenthaler, G. F. King, D. J. Adams, and P. F. Alewood, “Trends in peptide drug discovery,” Nature Reviews Drug Discovery, vol. 20, no. 4, pp. 309–325, 2021. View at: Publisher Site | Google Scholar
  18. M. W. MacArthur and J. M. Thornton, “Deviations from planarity of the peptide bond in peptides and proteins,” Journal of Molecular Biology, vol. 264, no. 5, pp. 1180–1195, 1996. View at: Publisher Site | Google Scholar
  19. L. Chan, G. R. Hutchison, and G. M. Morris, “Understanding ring puckering in small molecules and cyclic peptides,” Journal of Chemical Information and Modeling, vol. 61, no. 2, pp. 743–755, 2021. View at: Publisher Site | Google Scholar
  20. K. Natarajan, J. Jiang, N. A. May et al., “The role of molecular flexibility in antigen presentation and T cell receptor-mediated signaling,” Frontiers in Immunology, vol. 9, p. 1657, 2018. View at: Publisher Site | Google Scholar
  21. G. P. Singh, M. Ganapathi, and D. Dash, “Role of intrinsic disorder in transient interactions of hub proteins,” Proteins: Structure, Function, and Bioinformatics, vol. 66, no. 4, pp. 761–765, 2006. View at: Publisher Site | Google Scholar
  22. P.-S. Huang, S. E. Boyken, and D. Baker, “The coming of age of de novo protein design,” Nature, vol. 537, no. 7620, pp. 320–327, 2016. View at: Publisher Site | Google Scholar
  23. A. K. Malde, T. A. Hill, A. Iyer, and D. P. Fairlie, “Crystal structures of protein-bound cyclic peptides,” Chemical Reviews, vol. 119, no. 17, pp. 9861–9914, 2019. View at: Publisher Site | Google Scholar
  24. R. E. Pavlovicz, H. Park, and F. DiMaio, “Efficient consideration of coordinated water molecules improves computational protein-protein and protein-ligand docking discrimination,” PLoS Computational Biology, vol. 16, no. 9, article e1008103, 2020. View at: Publisher Site | Google Scholar
  25. Y. Maeda, O. V. Makhlynets, H. Matsui, and I. V. Korendovych, “Design of catalytic peptides and proteins through rational and combinatorial approaches,” Annual Review of Biomedical Engineering, vol. 18, no. 1, pp. 311–328, 2016. View at: Publisher Site | Google Scholar
  26. J. Liu and X. Zhao, “Design of self-assembling peptides and their biomedical applications,” Nanomedicine, vol. 6, no. 9, pp. 1621–1643, 2011. View at: Publisher Site | Google Scholar
  27. S. Zhang, “Discovery and design of self-assembling peptides,” Interface Focus, vol. 7, no. 6, article 20170028, 2017. View at: Publisher Site | Google Scholar
  28. S. E. D’Souza, M. H. Ginsberg, and E. F. Plow, “Arginyl-glycyl-aspartic acid (RGD): a cell adhesion motif,” Trends in Biochemical Sciences, vol. 16, no. 7, pp. 246–250, 1991. View at: Publisher Site | Google Scholar
  29. C.-Y. Tsai, E. O. Salawu, H. Li et al., “Helical structure motifs made searchable for functional peptide design,” Nature Communications, vol. 13, no. 1, p. 102, 2022. View at: Publisher Site | Google Scholar
  30. N. Alam, L. Zimmerman, N. A. Wolfson, C. G. Joseph, C. A. Fierke, and O. Schueler-Furman, “Structure-based identification of HDAC8 non-histone substrates,” Structure, vol. 24, no. 3, pp. 458–468, 2016. View at: Publisher Site | Google Scholar
  31. I. M. A. Nooren and J. M. Thornton, “Structural characterisation and functional significance of transient protein–protein interactions,” Journal of Molecular Biology, vol. 325, no. 5, pp. 991–1018, 2003. View at: Publisher Site | Google Scholar
  32. I. M. A. Nooren, “New EMBO member’s review: diversity of protein-protein interactions,” The EMBO Journal, vol. 22, no. 14, pp. 3486–3492, 2003. View at: Publisher Site | Google Scholar
  33. B. W. Parker, E. J. Goncz, D. T. Krist, A. V. Statsyuk, A. I. Nesvizhskii, and E. L. Weiss, “Mapping low-affinity/high-specificity peptide–protein interactions using ligand-footprinting mass spectrometry,” Proceedings of the National Academy of Sciences of the United States of America, vol. 116, no. 42, pp. 21001–21011, 2019. View at: Publisher Site | Google Scholar
  34. S. Dutta, J. Ryan, T. S. Chen, C. Kougentakis, A. Letai, and A. E. Keating, “Potent and specific peptide inhibitors of human pro-survival protein Bcl-XL,” Journal of Molecular Biology, vol. 427, no. 6, pp. 1241–1253, 2015. View at: Publisher Site | Google Scholar
  35. J. M. Jenson, J. A. Ryan, R. A. Grant, A. Letai, and A. E. Keating, “Epistatic Mutations in PUMA BH3 Drive an Alternate Binding Mode to Potently and Selectively Inhibit Anti-Apoptotic Bfl-1,” eLife, vol. 6, article e25541, 2017. View at: Publisher Site | Google Scholar
  36. G. W. Foight, J. A. Ryan, S. V. Gullá, A. Letai, and A. E. Keating, “Designed BH3 peptides with high affinity and specificity for targeting Mcl-1 in cells,” ACS Chemical Biology, vol. 9, no. 9, pp. 1962–1968, 2014. View at: Publisher Site | Google Scholar
  37. C. O. Mackenzie, J. Zhou, and G. Grigoryan, “Tertiary alphabet for the observable protein structural universe,” Proceedings of the National Academy of Sciences, vol. 113, no. 47, pp. E7438–E7447, 2016. View at: Publisher Site | Google Scholar
  38. V. Frappier, J. M. Jenson, J. Zhou, G. Grigoryan, and A. E. Keating, “Tertiary structural motif sequence statistics enable facile prediction and design of peptides that bind anti-apoptotic Bfl-1 and Mcl-1,” Structure, vol. 27, no. 4, pp. 606–617.e5, 2019. View at: Publisher Site | Google Scholar
  39. M. J. I. Andrews, G. Kontopidis, C. McInnes et al., “REPLACE: a strategy for iterative design of cyclin-binding groove inhibitors,” Chembiochem, vol. 7, no. 12, pp. 1909–1915, 2006. View at: Publisher Site | Google Scholar
  40. R. Pearce, X. Huang, D. Setiawan, and Y. Zhang, “EvoDesign: designing protein–protein binding interactions using evolutionary interface profiles in conjunction with an optimized physical energy function,” Journal of Molecular Biology, vol. 431, no. 13, pp. 2467–2476, 2019. View at: Publisher Site | Google Scholar
  41. P. Xiong, C. Zhang, W. Zheng, and Y. Zhang, “BindProfX: assessing mutation-induced binding affinity change by protein interface profiles with pseudo-counts,” Journal of Molecular Biology, vol. 429, no. 3, pp. 426–434, 2017. View at: Publisher Site | Google Scholar
  42. M. Gao and J. Skolnick, “IAlign: a method for the structural comparison of protein–protein interfaces,” Bioinformatics, vol. 26, no. 18, pp. 2259–2265, 2010. View at: Publisher Site | Google Scholar
  43. X. Huang, R. Pearce, and Y. Zhang, “EvoEF2: accurate and fast energy function for computational protein design,” Bioinformatics, vol. 36, no. 4, pp. 1135–1142, 2020. View at: Publisher Site | Google Scholar
  44. X. Huang, R. Pearce, and Y. Zhang, “Novo design of protein peptides to block association of the SARS-CoV-2 spike protein with human ACE2,” Aging, vol. 12, no. 12, pp. 11263–11276, 2020. View at: Publisher Site | Google Scholar
  45. D. Rooklin, A. E. Modell, H. Li, V. Berdan, P. S. Arora, and Y. Zhang, “Targeting unoccupied surfaces on protein–protein interfaces,” Journal of the American Chemical Society, vol. 139, no. 44, pp. 15560–15563, 2017. View at: Publisher Site | Google Scholar
  46. J. M. Torner, Y. Yang, D. Rooklin, Y. Zhang, and P. S. Arora, “Identification of secondary binding sites on protein surfaces for rational elaboration of synthetic protein mimics,” ACS Chemical Biology, vol. 16, no. 7, pp. 1179–1183, 2021. View at: Publisher Site | Google Scholar
  47. I. Lamberto, B. C. Lechtenberg, E. J. Olson et al., “Development and structural analysis of a nanomolar cyclic peptide antagonist for the EphA4 receptor,” ACS Chemical Biology, vol. 9, no. 12, pp. 2787–2795, 2014. View at: Publisher Site | Google Scholar
  48. M. Kawai, D. A. Quincy, B. Lane et al., “Structure-function studies in a series of carboxyl-terminal octapeptide analogs of anaphylatoxin C5a,” Journal of Medicinal Chemistry, vol. 35, no. 2, pp. 220–223, 1992. View at: Publisher Site | Google Scholar
  49. W. S. Horne, L. M. Johnson, T. J. Ketas et al., “Structural and biological mimicry of protein surface recognition by α/β-peptide foldamers,” Proceedings of the National Academy of Sciences of the United States of America, vol. 106, no. 35, pp. 14751–14756, 2009. View at: Publisher Site | Google Scholar
  50. B. Hu, D. M. Gilkes, and J. Chen, “Efficient P53 activation and apoptosis by simultaneous disruption of binding to MDM2 and MDMX,” Cancer Research, vol. 67, no. 18, pp. 8810–8817, 2007. View at: Publisher Site | Google Scholar
  51. Y. S. Chang, B. Graves, V. Guerlavais et al., “Stapled α-helical peptide drug development: a potent dual inhibitor of MDM2 and MDMX for P53-dependent cancer therapy,” Proceedings of the National Academy of Sciences of the United States of America, vol. 110, no. 36, pp. E3445–E3454, 2013. View at: Publisher Site | Google Scholar
  52. J. Lee, D. W. Sammond, Z. Fiorini et al., “Computationally designed peptide inhibitors of the ubiquitin E3 ligase SCF Fbx4,” Chembiochem, vol. 14, no. 4, pp. 445–451, 2013. View at: Publisher Site | Google Scholar
  53. H. I. Merritt, N. Sawyer, and P. S. Arora, “Bent into shape: folded peptides to mimic protein structure and modulate protein function,” Peptide Science, vol. 112, no. 1, 2020. View at: Publisher Site | Google Scholar
  54. R. Rezaei Araghi and A. E. Keating, “Designing helical peptide inhibitors of protein–protein interactions,” Current Opinion in Structural Biology, vol. 39, pp. 27–38, 2016. View at: Publisher Site | Google Scholar
  55. M. T. J. Bluntzer, J. O’Connell, T. S. Baker, J. Michel, and A. N. Hulme, “Designing stapled peptides to inhibit protein-protein interactions: an analysis of successes in a rapidly changing field,” Peptide Science, vol. 113, no. 1, 2021. View at: Publisher Site | Google Scholar
  56. A. Jamieson and N. Robertson, “Regulation of protein–protein interactions using stapled peptides,” Reports in Organic Chemistry, vol. 5, pp. 65–74, 2015. View at: Publisher Site | Google Scholar
  57. Y. H. Lau, P. de Andrade, Y. Wu, and D. R. Spring, “Peptide stapling techniques based on different macrocyclisation chemistries,” Chemical Society Reviews, vol. 44, no. 1, pp. 91–102, 2015. View at: Publisher Site | Google Scholar
  58. M. Moiola, M. G. Memeo, and P. Quadrelli, “Stapled Peptides-A Useful Improvement for Peptide-Based Drugs,” Molecules, vol. 24, no. 20, article 3654, 2019. View at: Publisher Site | Google Scholar
  59. O. Khakshoor and J. S. Nowick, “Use of disulfide “Staples” to stabilize beta-sheet quaternary structure,” Organic Letters, vol. 11, no. 14, pp. 3000–3003, 2009. View at: Publisher Site | Google Scholar
  60. J. A. Robinson, “β-hairpin peptidomimetics: design, structures and biological activities,” Accounts of Chemical Research, vol. 41, no. 10, pp. 1278–1288, 2008. View at: Publisher Site | Google Scholar
  61. Z. Athanassiou, R. L. A. Dias, K. Moehle, N. Dobson, G. Varani, and J. A. Robinson, “Structural mimicry of retroviral tat proteins by constrained β-hairpin peptidomimetics: ligands with high affinity and selectivity for viral TAR RNA regulatory elements,” Journal of the American Chemical Society, vol. 126, no. 22, pp. 6906–6913, 2004. View at: Publisher Site | Google Scholar
  62. K. Moehle, Z. Athanassiou, K. Patora, A. Davidson, G. Varani, and J. A. Robinson, “Design of Beta-Hairpin Peptidomimetics that Inhibit Binding of alpha-helical HIV-1 rev protein to the rev response element RNA,” Angewandte Chemie (International Ed. in English), vol. 46, no. 47, pp. 9101–9104, 2007. View at: Publisher Site | Google Scholar
  63. S. J. DeMarco, H. Henze, A. Lederer et al., “Discovery of novel, highly potent and selective Beta-hairpin mimetic CXCR4 inhibitors with excellent anti-HIV activity and pharmacokinetic profiles,” Bioorganic & Medicinal Chemistry, vol. 14, no. 24, pp. 8396–8404, 2006. View at: Publisher Site | Google Scholar
  64. S. C. Shankaramma, Z. Athanassiou, O. Zerbe et al., “Macrocyclic hairpin mimetics of the cationic antimicrobial Peptide Protegrin I: A new family of broad-Spectrum antibiotics,” ChemBioChem, vol. 3, no. 11, pp. 1126–1133, 2002. View at: Publisher Site | Google Scholar
  65. S. Aravinda, N. Shamala, R. Rajkishore, H. N. Gopi, and P. Balaram, “A crystalline β-hairpin Peptide nucleated by a type I Aib-D-ala β-turn: evidence for cross-Strand aromatic interactions,” Angewandte Chemie, International Edition, vol. 41, no. 20, pp. 3863–3865, 2002. View at: Publisher Site | Google Scholar
  66. L. R. Masterson, M. A. Etienne, F. Porcelli, G. Barany, R. P. Hammer, and G. Veglia, “Nonstereogenic α-Aminoisobutyryl-Glycyl dipeptidyl unit nucleates type I β-turn in linear peptides in aqueous solution,” Biopolymers, vol. 88, no. 5, pp. 746–753, 2007. View at: Publisher Site | Google Scholar
  67. J. Gavenonis, B. A. Sheneman, T. R. Siegert, M. R. Eshelman, and J. A. Kritzer, “Comprehensive analysis of loops at protein-protein interfaces for macrocycle design,” Nature Chemical Biology, vol. 10, no. 9, pp. 716–722, 2014. View at: Publisher Site | Google Scholar
  68. T. R. Siegert, M. J. Bird, K. M. Makwana, and J. A. Kritzer, “Analysis of Loops That Mediate Protein–Protein Interactions and Translation into Submicromolar Inhibitors,” Journal of the American Chemical Society, vol. 138, no. 39, pp. 12876–12884, 2016. View at: Publisher Site | Google Scholar
  69. R. A. Cerulli, L. Shehaj, H. Brown, J. Pace, Y. Mei, and J. A. Kritzer, “Stapled Peptide inhibitors of autophagy adapter LC3B,” Chembiochem, vol. 21, no. 19, pp. 2777–2785, 2020. View at: Publisher Site | Google Scholar
  70. N. London, B. Raveh, D. Movshovitz-Attias, and O. Schueler-Furman, “Can self-inhibitory peptides be derived from the interfaces of globular protein-protein interactions?: deriving inhibitory peptides from protein-protein interactions,” Proteins: Structure, Function, and Bioinformatics, vol. 78, no. 15, pp. 3140–3149, 2010. View at: Publisher Site | Google Scholar
  71. M. Gao, N. London, K. Cheng et al., “Rationally designed macrocyclic peptides as synergistic agonists of LPS-induced inflammatory response,” Tetrahedron, vol. 70, no. 42, pp. 7664–7668, 2014. View at: Publisher Site | Google Scholar
  72. P. P. Nanaware, M. P. Ramteke, A. K. Somavarapu, and P. Venkatraman, “Discovery of multiple interacting Partners of Gankyrin, a proteasomal chaperone and an Oncoprotein-evidence for a common hot spot site at the Interface and its functional relevance: protein interaction network of a hub Oncoprotein,” Proteins: Structure, Function, and Bioinformatics, vol. 82, no. 7, pp. 1283–1300, 2014. View at: Publisher Site | Google Scholar
  73. Y. Sedan, O. Marcu, S. Lyskov, and O. Schueler-Furman, “Peptiderive Server: Derive Peptide Inhibitors from Protein–Protein Interactions,” Nucleic Acids Research, vol. 44, no. W1, pp. W536–W541, 2016. View at: Publisher Site | Google Scholar
  74. R. U. Kadam, J. Juraszek, B. Brandenburg et al., “Potent Peptidic fusion inhibitors of influenza virus,” Science, vol. 358, no. 6362, pp. 496–502, 2017. View at: Publisher Site | Google Scholar
  75. S. McGrath, M. Tortorici, L. Drouin et al., “Structure-enabled discovery of a stapled Peptide inhibitor to target the oncogenic transcriptional repressor TLE1,” Chemistry - A European Journal, vol. 23, no. 40, pp. 9577–9584, 2017. View at: Publisher Site | Google Scholar
  76. A. M. Finch, A. K. Wong, N. J. Paczkowski et al., “Low-molecular-weight Peptidic and cyclic antagonists of the receptor for the complement factor C5a,” Journal of Medicinal Chemistry, vol. 42, no. 11, pp. 1965–1974, 1999. View at: Publisher Site | Google Scholar
  77. B. Laudet, C. Barette, V. Dulery et al., “Structure-Based Design of Small Peptide Inhibitors of protein kinase CK2 subunit interaction,” The Biochemical Journal, vol. 408, no. 3, pp. 363–373, 2007. View at: Publisher Site | Google Scholar
  78. M. G. Wuo, A. B. Mahon, and P. S. Arora, “An effective strategy for stabilizing minimal coiled coil mimetics,” Journal of the American Chemical Society, vol. 137, no. 36, pp. 11618–11621, 2015. View at: Publisher Site | Google Scholar
  79. R. P. Morse, J. L. E. Willett, P. M. Johnson et al., “Diversification of β-augmentation interactions between CDI toxin/immunity proteins,” Journal of Molecular Biology, vol. 427, no. 23, pp. 3766–3784, 2015. View at: Publisher Site | Google Scholar
  80. A. Davidson, T. C. Leeper, Z. Athanassiou et al., “Simultaneous recognition of HIV-1 TAR RNA bulge and loop sequences by cyclic Peptide mimics of tat protein,” Proceedings of the National Academy of Sciences of the United States of America, vol. 106, no. 29, pp. 11931–11936, 2009. View at: Publisher Site | Google Scholar
  81. C. Phillips, L. R. Roberts, M. Schade et al., “Design and structure of stapled peptides binding to estrogen receptors,” Journal of the American Chemical Society, vol. 133, no. 25, pp. 9696–9699, 2011. View at: Publisher Site | Google Scholar
  82. A. Glas, D. Bier, G. Hahne, C. Rademacher, C. Ottmann, and T. N. Grossmann, “Constrained peptides with target-adapted cross-links as inhibitors of a pathogenic protein-protein interaction,” Angewandte Chemie, International Edition, vol. 53, no. 9, pp. 2489–2493, 2014. View at: Publisher Site | Google Scholar
  83. Y. Teng, A. Bahassan, D. Dong et al., “Targeting the WASF3–CYFIP1 complex using stapled peptides suppresses cancer cell invasion,” Cancer Research, vol. 76, no. 4, pp. 965–973, 2016. View at: Publisher Site | Google Scholar
  84. W. Kim, G. H. Bird, T. Neff et al., “Targeted disruption of the EZH2-EED complex inhibits EZH2-dependent cancer,” Nature Chemical Biology, vol. 9, no. 10, pp. 643–650, 2013. View at: Publisher Site | Google Scholar
  85. A. Patgiri, K. K. Yadav, P. S. Arora, and D. Bar-Sagi, “An Orthosteric inhibitor of the Ras-Sos interaction,” Nature Chemical Biology, vol. 7, no. 9, pp. 585–587, 2011. View at: Publisher Site | Google Scholar
  86. J. Spiegel, P. M. Cromm, A. Itzen, R. S. Goody, T. N. Grossmann, and H. Waldmann, “Direct targeting of Rab-GTPase-effector interactions,” Angewandte Chemie, International Edition, vol. 53, no. 9, pp. 2498–2503, 2014. View at: Publisher Site | Google Scholar
  87. P. M. Cromm, J. Spiegel, P. Küchler et al., “Protease-resistant and cell-permeable double-stapled peptides targeting the Rab8a GTPase,” ACS Chemical Biology, vol. 11, no. 8, pp. 2375–2382, 2016. View at: Publisher Site | Google Scholar
  88. S. Kushal, B. B. Lao, L. K. Henchey et al., “Protein domain mimetics as in vivo modulators of hypoxia-inducible factor signaling,” Proceedings of the National Academy of Sciences of the United States of America, vol. 110, no. 39, pp. 15602–15607, 2013. View at: Publisher Site | Google Scholar
  89. L. K. Henchey, S. Kushal, R. Dubey, R. N. Chapman, B. Z. Olenyuk, and P. S. Arora, “Inhibition of hypoxia inducible factor 1-transcription coactivator interaction by a hydrogen bond surrogate alpha-helix,” Journal of the American Chemical Society, vol. 132, no. 3, pp. 941–943, 2010. View at: Publisher Site | Google Scholar
  90. S. Kneissl, E. J. Loveridge, C. Williams, M. P. Crump, and R. K. Allemann, “Photocontrollable Peptide-Based switches target the anti-apoptotic protein Bcl-x L,” Chembiochem, vol. 9, no. 18, pp. 3046–3054, 2008. View at: Publisher Site | Google Scholar
  91. L. D. Walensky, A. L. Kung, I. Escher et al., “Activation of apoptosis in vivo by a hydrocarbon-stapled BH3 helix,” Science, vol. 305, no. 5689, pp. 1466–1470, 2004. View at: Publisher Site | Google Scholar
  92. J. W. Checco, E. F. Lee, M. Evangelista et al., “α/β-Peptide Foldamers Targeting Intracellular Protein–Protein Interactions with Activity in Living Cells,” Journal of the American Chemical Society, vol. 137, no. 35, pp. 11365–11375, 2015. View at: Publisher Site | Google Scholar
  93. S. A. Kawamoto, A. Coleska, X. Ran, H. Yi, C.-Y. Yang, and S. Wang, “Design of Triazole-Stapled BCL9 α-helical peptides to target the β-catenin/B-cell CLL/lymphoma 9 (BCL9) protein-protein interaction,” Journal of Medicinal Chemistry, vol. 55, no. 3, pp. 1137–1146, 2012. View at: Publisher Site | Google Scholar
  94. D. Wang, M. Lu, and P. S. Arora, “Inhibition of HIV-1 fusion by hydrogen-bond-surrogate-Based α helices,” Angewandte Chemie, International Edition, vol. 47, no. 10, pp. 1879–1882, 2008. View at: Publisher Site | Google Scholar
  95. G. H. Bird, N. Madani, A. F. Perry et al., “Hydrocarbon double-stapling remedies the proteolytic instability of a lengthy Peptide therapeutic,” Proceedings of the National Academy of Sciences of the United States of America, vol. 107, no. 32, pp. 14093–14098, 2010. View at: Publisher Site | Google Scholar
  96. F. M. Brunel, M. B. Zwick, R. M. F. Cardoso et al., “Structure-function analysis of the epitope for 4E10, a broadly neutralizing human immunodeficiency virus type 1 antibody,” Journal of Virology, vol. 80, no. 4, pp. 1680–1687, 2006. View at: Publisher Site | Google Scholar
  97. G. Philippe, Y. Huang, O. Cheneval et al., “Development of cell-penetrating Peptide-based drug leads to inhibit MDMX:P53 and MDM2:P53 interactions,” Peptide Science, vol. 106, no. 6, pp. 853–863, 2016. View at: Publisher Site | Google Scholar
  98. Y. H. Lau, P. de Andrade, N. Sköld et al., “Investigating Peptide sequence variations for ‘double-click’ stapled P53 peptides,” Organic & Biomolecular Chemistry, vol. 12, no. 24, pp. 4074–4077, 2014. View at: Publisher Site | Google Scholar
  99. M. M. Madden, A. Muppidi, Z. Li, X. Li, J. Chen, and Q. Lin, “Synthesis of cell-permeable stapled Peptide dual inhibitors of the P53-Mdm2/Mdmx interactions via Photoinduced cycloaddition,” Bioorganic & Medicinal Chemistry Letters, vol. 21, no. 5, pp. 1472–1475, 2011. View at: Publisher Site | Google Scholar
  100. L. K. Henchey, J. R. Porter, I. Ghosh, and P. S. Arora, “High specificity in protein recognition by hydrogen-bond-surrogate α-helices: selective inhibition of the P53/MDM2 complex,” Chembiochem, vol. 11, no. 15, pp. 2104–2107, 2010. View at: Publisher Site | Google Scholar
  101. W. S. Horne and T. N. Grossmann, “Proteomimetics as protein-inspired scaffolds with defined tertiary folding patterns,” Nature Chemistry, vol. 12, no. 4, pp. 331–337, 2020. View at: Publisher Site | Google Scholar
  102. C. K. Wang and D. J. Craik, “Designing macrocyclic disulfide-rich peptides for biotechnological applications,” Nature Chemical Biology, vol. 14, no. 5, pp. 417–427, 2018. View at: Publisher Site | Google Scholar
  103. A. C. Conibear, A. Bochen, K. J. Rosengren et al., “The cyclic Cystine ladder of theta-Defensins as a stable, bifunctional scaffold: A proof-of-concept study using the integrin-binding RGD motif,” Chembiochem, vol. 15, no. 3, pp. 451–459, 2014. View at: Publisher Site | Google Scholar
  104. R. H. Kimura, Z. Cheng, S. S. Gambhir, and J. R. Cochran, “Engineered Knottin peptides: A new class of agents for imaging integrin expression in living subjects,” Cancer Research, vol. 69, no. 6, pp. 2435–2442, 2009. View at: Publisher Site | Google Scholar
  105. A. Gould, Y. Ji, T. L. Aboye, and J. A. Camarero, “Cyclotides, a novel Ultrastable polypeptide scaffold for drug discovery,” Current Pharmaceutical Design, vol. 17, no. 38, pp. 4294–4307, 2011. View at: Publisher Site | Google Scholar
  106. K. Jagadish and J. A. Camarero, “Cyclotides, a promising molecular scaffold for Peptide-Based therapeutics,” Biopolymers, vol. 94, no. 5, pp. 611–616, 2010. View at: Publisher Site | Google Scholar
  107. R. H. Kimura, A. M. Levin, F. V. Cochran, and J. R. Cochran, “Engineered Cystine knot peptides that bind Alphavbeta3, Alphavbeta5, and Alpha5beta1 Integrins with low-Nanomolar affinity,” Proteins, vol. 77, no. 2, pp. 359–369, 2009. View at: Publisher Site | Google Scholar
  108. J. R. Kintzing and J. R. Cochran, “Engineered Knottin peptides as diagnostics, therapeutics, and drug delivery vehicles,” Current Opinion in Chemical Biology, vol. 34, pp. 143–150, 2016. View at: Publisher Site | Google Scholar
  109. E. Muratspahić, J. Koehbach, C. W. Gruber, and D. J. Craik, “Harnessing Cyclotides to design and develop novel Peptide GPCR ligands,” RSC Chemical Biology, vol. 1, no. 4, pp. 177–191, 2020. View at: Publisher Site | Google Scholar
  110. C. T. T. Wong, D. K. Rowlands, C.-H. Wong et al., “Orally active Peptidic bradykinin B1 receptor antagonists engineered from a Cyclotide scaffold for inflammatory pain treatment,” Angewandte Chemie, International Edition, vol. 51, no. 23, pp. 5620–5624, 2012. View at: Publisher Site | Google Scholar
  111. S. Gunasekera, F. M. Foley, R. J. Clark et al., “Engineering stabilized vascular endothelial growth factor-A antagonists: synthesis, structural characterization, and bioactivity of grafted analogues of Cyclotides,” Journal of Medicinal Chemistry, vol. 51, no. 24, pp. 7697–7704, 2008. View at: Publisher Site | Google Scholar
  112. R. Eliasen, N. L. Daly, B. S. Wulff, T. L. Andresen, K. W. Conde-Frieboes, and D. J. Craik, “Design, synthesis, structural and functional characterization of novel Melanocortin agonists Based on the Cyclotide Kalata B1,” The Journal of Biological Chemistry, vol. 287, no. 48, pp. 40493–40501, 2012. View at: Publisher Site | Google Scholar
  113. T. L. Aboye, H. Ha, S. Majumder et al., “Design of a Novel Cyclotide-Based CXCR4 antagonist with anti-human immunodeficiency virus (HIV)-1 activity,” Journal of Medicinal Chemistry, vol. 55, no. 23, pp. 10729–10734, 2012. View at: Publisher Site | Google Scholar
  114. W. G. Lesniak, T. Aboye, S. Chatterjee, J. A. Camarero, and S. Nimmagadda, “In vivo evaluation of an engineered Cyclotide as specific CXCR4 imaging reagent,” Chemistry - A European Journal, vol. 23, no. 58, pp. 14469–14475, 2017. View at: Publisher Site | Google Scholar
  115. A. Chevalier, D.-A. Silva, G. J. Rocklin et al., “Massively parallel de novo protein Design for Targeted Therapeutics,” Nature, vol. 550, no. 7674, pp. 74–79, 2017. View at: Publisher Site | Google Scholar
  116. S. L. Blosser, N. Sawyer, I. Maksimovic, B. Ghosh, and P. S. Arora, “Covalent and noncovalent targeting of the Tcf4/β-catenin Strand Interface with β-hairpin mimics,” ACS Chemical Biology, vol. 16, no. 8, pp. 1518–1525, 2021. View at: Publisher Site | Google Scholar
  117. J. Smadbeck, M. B. Peterson, B. M. Zee et al., “Novo Peptide design and experimental validation of histone methyltransferase inhibitors,” PLoS One, vol. 9, no. 2, article e90095, 2014. View at: Publisher Site | Google Scholar
  118. R. H. Lilien, B. W. Stevens, A. C. Anderson, and B. R. Donald, “A novel ensemble-Based scoring and search algorithm for protein redesign and its application to modify the substrate specificity of the gramicidin Synthetase a phenylalanine Adenylation enzyme,” Journal of Computational Biology, vol. 12, no. 6, pp. 740–761, 2005. View at: Publisher Site | Google Scholar
  119. E. B. Unal, A. Gursoy, and B. Erman, “VitAL: Viterbi Algorithm for de Novo Peptide Design,” PLoS One, vol. 5, no. 6, article e10926, 2010. View at: Publisher Site | Google Scholar
  120. T. Haliloglu and B. Erman, “Analysis of correlations between energy and residue fluctuations in native proteins and determination of specific sites for binding,” Physical Review Letters, vol. 102, no. 8, article 088103, 2009. View at: Publisher Site | Google Scholar
  121. T. Haliloglu, A. Gul, and B. Erman, “Predicting important residues and interaction pathways in proteins using Gaussian network model: binding and stability of HLA proteins,” PLoS Computational Biology, vol. 6, no. 7, article e1000845, 2010. View at: Publisher Site | Google Scholar
  122. T. Haliloglu, E. Seyrek, and B. Erman, “Prediction of binding sites in receptor-ligand complexes with the Gaussian network model,” Physical Review Letters, vol. 100, no. 22, article 228102, 2008. View at: Publisher Site | Google Scholar
  123. H. M. Li, Z. P. Dong, Q. Y. Wang et al., “Novo computational Design for Development of a Peptide ligand oriented to VEGFR-3 with high affinity and long circulation,” Molecular Pharmaceutics, vol. 14, no. 7, pp. 2236–2244, 2017. View at: Publisher Site | Google Scholar
  124. D. Kozakov, L. E. Grove, D. R. Hall et al., “The FTMap family of web servers for determining and characterizing ligand-binding hot spots of proteins,” Nature Protocols, vol. 10, no. 5, pp. 733–755, 2015. View at: Publisher Site | Google Scholar
  125. N. Stambuk, P. Konjevoda, A. Boban-Blagaić, and B. Pokrić, “Molecular recognition theory of the complementary (antisense) Peptide interactions,” Theory in Biosciences, vol. 123, no. 4, pp. 265–275, 2005. View at: Publisher Site | Google Scholar
  126. T. A. Edwards and A. J. Wilson, “Helix-Mediated Protein–Protein Interactions as Targets for Intervention Using Foldamers,” Amino Acids, vol. 41, no. 3, pp. 743–754, 2011. View at: Publisher Site | Google Scholar
  127. L. Cussol, L. Mauran-Ambrosino, J. Buratto et al., “Structural Basis for Α-Helix Mimicry and Inhibition of Protein–Protein Interactions with Oligourea Foldamers,” Angewandte Chemie, International Edition, vol. 60, no. 5, pp. 2296–2303, 2021. View at: Publisher Site | Google Scholar
  128. H. Yin, G. Lee, K. A. Sedey et al., “Terphenyl-Based Bak BH3 α-helical Proteomimetics as low-molecular-weight antagonists of Bcl-xL,” Journal of the American Chemical Society, vol. 127, no. 29, pp. 10191–10196, 2005. View at: Publisher Site | Google Scholar
  129. R. Fasan, R. L. A. Dias, K. Moehle et al., “Using a β-Hairpin To Mimic an α-Helix: Cyclic Peptidomimetic Inhibitors of the P53–HDM2 Protein–Protein Interaction,” Angewandte Chemie, International Edition, vol. 43, no. 16, pp. 2109–2112, 2004. View at: Publisher Site | Google Scholar
  130. J. M. Fletcher, K. A. Horner, G. J. Bartlett, G. G. Rhys, A. J. Wilson, and D. N. Woolfson, “Novo Coiled-Coil Peptides as Scaffolds for Disrupting Protein–Protein Interactions,” Chemical Science, vol. 9, no. 39, pp. 7656–7665, 2018. View at: Publisher Site | Google Scholar
  131. P. E. Czabotar, G. Lessene, A. Strasser, and J. M. Adams, “Control of apoptosis by the BCL-2 protein family: implications for physiology and therapy,” Nature Reviews Molecular Cell Biology, vol. 15, no. 1, pp. 49–63, 2014. View at: Publisher Site | Google Scholar
  132. F. Thomas, A. L. Boyle, A. J. Burton, and D. N. Woolfson, “A set of de novo designed parallel heterodimeric coiled coils with quantified dissociation constants in the micromolar to sub-Nanomolar regime,” Journal of the American Chemical Society, vol. 135, no. 13, pp. 5161–5166, 2013. View at: Publisher Site | Google Scholar
  133. J. A. Schneider, T. W. Craven, A. C. Kasper et al., “Design of Peptoid-Peptide Macrocycles to inhibit the β-catenin TCF interaction in prostate cancer,” Nature Communications, vol. 9, no. 1, p. 4396, 2018. View at: Publisher Site | Google Scholar
  134. O. M. Stephens, S. Kim, B. D. Welch, M. E. Hodsdon, M. S. Kay, and A. Schepartz, “Inhibiting HIV fusion with a β-Peptide Foldamer,” Journal of the American Chemical Society, vol. 127, no. 38, pp. 13126-13127, 2005. View at: Publisher Site | Google Scholar
  135. A. D. Bautista, O. M. Stephens, L. Wang, R. A. Domaoal, K. S. Anderson, and A. Schepartz, “Identification of a β3-Peptide HIV fusion inhibitor with improved potency in live cells,” Bioorganic & Medicinal Chemistry Letters, vol. 19, no. 14, pp. 3736–3738, 2009. View at: Publisher Site | Google Scholar
  136. J. A. Kritzer, J. D. Lear, M. E. Hodsdon, and A. Schepartz, “Helical β-Peptide inhibitors of the P53-HDM2 interaction,” Journal of the American Chemical Society, vol. 126, no. 31, pp. 9468-9469, 2004. View at: Publisher Site | Google Scholar
  137. J. Michel, E. A. Harker, J. Tirado-Rives, W. L. Jorgensen, and A. Schepartz, “In silico Improvement of β3-Peptide inhibitors of P53•hDM2 and P53•hDMX,” Journal of the American Chemical Society, vol. 131, no. 18, pp. 6356-6357, 2009. View at: Publisher Site | Google Scholar
  138. D. P. Slough, S. M. McHugh, A. E. Cummings et al., “Designing Well-Structured Cyclic Pentapeptides Based on Sequence–Structure Relationships,” The Journal of Physical Chemistry B, vol. 122, no. 14, pp. 3908–3919, 2018. View at: Publisher Site | Google Scholar
  139. D. J. Mandell, E. A. Coutsias, and T. Kortemme, “Sub-angstrom accuracy in protein loop reconstruction by robotics-inspired conformational sampling,” Nature Methods, vol. 6, no. 8, pp. 551-552, 2009. View at: Google Scholar
  140. P. Hosseinzadeh, G. Bhardwaj, V. K. Mulligan et al., “Comprehensive computational Design of Ordered Peptide Macrocycles,” Science, vol. 358, no. 6369, pp. 1461–1466, 2017. View at: Publisher Site | Google Scholar
  141. S. Berger and P. Hosseinzadeh, “Computational design of structured and functional peptide macrocycles,” in Peptide Macrocycles. Methods in Molecular Biology, M. B. Coppock and A. J. Winton, Eds., vol. 2371, pp. 63–100, Humana, New York, NY, USA, 2022. View at: Publisher Site | Google Scholar
  142. S. Guardiola, M. Varese, X. Roig, M. Sánchez-Navarro, J. García, and E. Giralt, “Target-templated de novo Design of Macrocyclic d -/l -peptides: discovery of drug-like inhibitors of PD-1,” Chemical Science, vol. 12, no. 14, pp. 5164–5170, 2021. View at: Publisher Site | Google Scholar
  143. V. K. Mulligan, S. Workman, T. Sun et al., “Computationally designed Peptide macrocycle inhibitors of New Delhi Metallo-β-lactamase 1,” Proceedings of the National Academy of Sciences of the United States of America, vol. 118, no. 12, article e2012800118, 2021. View at: Publisher Site | Google Scholar
  144. P. Hosseinzadeh, P. R. Watson, T. W. Craven et al., “Anchor extension: A structure-guided approach to design cyclic peptides targeting enzyme active sites,” Nature Communications, vol. 12, no. 1, p. 3384, 2021. View at: Publisher Site | Google Scholar
  145. A. W. Senior, R. Evans, J. Jumper et al., “Improved protein structure prediction using potentials from deep learning,” Nature, vol. 577, no. 7792, pp. 706–710, 2020. View at: Publisher Site | Google Scholar
  146. J. Jumper, R. Evans, A. Pritzel et al., “Highly accurate protein structure prediction with alpha fold,” Nature, vol. 596, no. 7873, pp. 583–589, 2021. View at: Publisher Site | Google Scholar
  147. J. Miao, M. L. Descoteaux, and Y.-S. Lin, “Structure prediction of cyclic peptides by molecular dynamics + machine learning,” Chemical Science, vol. 12, no. 44, pp. 14927–14936, 2021. View at: Publisher Site | Google Scholar
  148. A. Gautam, K. Chaudhary, R. Kumar et al., “In Silico Approaches for Designing Highly Effective Cell Penetrating Peptides,” Journal of Translational Medicine, vol. 11, p. 74, 2013. View at: Publisher Site | Google Scholar
  149. T. L. Bailey, M. Boden, F. A. Buske et al., “MEME SUITE: Tools for Motif Discovery and Searching,” Nucleic Acids Research, vol. 37, pp. W202–W208, 2009. View at: Publisher Site | Google Scholar
  150. L. Wei, J. Tang, and Q. Zou, “SkipCPP-Pred: an improved and promising sequence-Based predictor for predicting cell-penetrating peptides,” BMC Genomics, vol. 18, no. 7, p. 742, 2017. View at: Publisher Site | Google Scholar
  151. B. Liu, J. Xu, Q. Zou, R. Xu, X. Wang, and Q. Chen, “Using distances between top-n-gram and residue pairs for protein remote homology detection,” BMC Bioinformatics, vol. 15, no. 2, p. S3, 2014. View at: Publisher Site | Google Scholar
  152. X. Qiang, C. Zhou, X. Ye, P.-F. Du, R. Su, and L. Wei, “CPPred-FL: A sequence-Based predictor for large-scale identification of cell-penetrating peptides by feature representation learning,” Briefings in Bioinformatics, 2018. View at: Publisher Site | Google Scholar
  153. X. Fu, L. Ke, L. Cai, X. Chen, X. Ren, and M. Gao, “Improved prediction of cell-penetrating peptides via effective orchestrating amino acid composition feature representation,” IEEE Access, vol. 7, pp. 163547–163555, 2019. View at: Publisher Site | Google Scholar
  154. E. C. L. de Oliveira, K. Santana, L. Josino, A. H. Lima e Lima, and C. de Souza de Sales Júnior, “Predicting cell-penetrating peptides using machine learning algorithms and navigating in their chemical space,” Scientific Reports, vol. 11, no. 1, article 7628, 2021. View at: Publisher Site | Google Scholar
  155. J. Chen, H. H. Cheong, and S. W. I. Siu, “XDeep-AcPEP: deep learning method for anticancer Peptide activity prediction Based on convolutional neural network and multitask learning,” Journal of Chemical Information and Modeling, vol. 61, no. 8, pp. 3789–3803, 2021. View at: Publisher Site | Google Scholar
  156. S. Hellberg, M. Sjoestroem, B. Skagerberg, and S. Wold, “Peptide quantitative structure-activity relationships, a multivariate approach,” Journal of Medicinal Chemistry, vol. 30, no. 7, pp. 1126–1135, 1987. View at: Publisher Site | Google Scholar
  157. F. Plisson, O. Ramírez-Sánchez, and C. Martínez-Hernández, “Machine learning-guided discovery and Design of non-Hemolytic Peptides,” Scientific Reports, vol. 10, no. 1, article 16581, 2020. View at: Publisher Site | Google Scholar
  158. T.-T. Lin, L.-Y. Yang, I.-H. Lu et al., “AI4AMP: an antimicrobial Peptide predictor using physicochemical property-Based encoding method and deep learning,” mSystems, vol. 6, no. 6, article e0029921, 2021. View at: Publisher Site | Google Scholar
  159. J. Yan, P. Bhadra, A. Li et al., “Deep-AmPEP30: improve short antimicrobial peptides prediction with deep learning,” Molecular Therapy - Nucleic Acids, vol. 20, pp. 882–894, 2020. View at: Publisher Site | Google Scholar
  160. P. Bhadra, J. Yan, J. Li, S. Fong, and S. W. I. Siu, “AmPEP: sequence-Based prediction of antimicrobial peptides using distribution patterns of amino acid properties and random Forest,” Scientific Reports, vol. 8, no. 1, article 1697, 2018. View at: Publisher Site | Google Scholar
  161. S. Lata, B. Sharma, and G. Raghava, “Analysis and prediction of antibacterial peptides,” BMC Bioinformatics, vol. 8, no. 1, p. 263, 2007. View at: Publisher Site | Google Scholar
  162. M. Torrent, P. Di Tommaso, D. Pulido et al., “AMPA: an automated web server for prediction of protein antimicrobial regions,” Bioinformatics, vol. 28, no. 1, pp. 130-131, 2012. View at: Publisher Site | Google Scholar
  163. M. Torrent, V. M. Nogués, and E. Boix, “A theoretical approach to spot active regions in antimicrobial proteins,” BMC Bioinformatics, vol. 10, no. 1, p. 373, 2009. View at: Publisher Site | Google Scholar
  164. N. Thakur, A. Qureshi, and M. Kumar, “AVPpred: Collection and Prediction of Highly Effective Antiviral Peptides,” Nucleic Acids Research, vol. 40, pp. W199–W204, 2012. View at: Publisher Site | Google Scholar
  165. X. Xiao, P. Wang, W.-Z. Lin, J.-H. Jia, and K.-C. Chou, “IAMP-2L: A two-level multi-label classifier for identifying antimicrobial peptides and their functional types,” Analytical Biochemistry, vol. 436, no. 2, pp. 168–177, 2013. View at: Publisher Site | Google Scholar
  166. D. Veltri, U. Kamath, and A. Shehu, “Deep learning improves antimicrobial Peptide recognition,” Bioinformatics, vol. 34, no. 16, pp. 2740–2747, 2018. View at: Publisher Site | Google Scholar
  167. Y. Lei, S. Li, Z. Liu et al., “A Deep-Learning Framework for Multi-Level Peptide–Protein Interaction Prediction,” Nature Communications, vol. 12, no. 1, p. 5465, 2021. View at: Publisher Site | Google Scholar
  168. S. Salentin, S. Schreiber, V. J. Haupt, M. F. Adasme, and M. Schroeder, “PLIP: fully automated protein-ligand interaction profiler,” Nucleic Acids Research, vol. 43, no. W1, pp. W443–W447, 2015. View at: Publisher Site | Google Scholar
  169. Z. Wen, J. He, H. Tao, and S.-Y. Huang, “PepBDB: A comprehensive structural database of biological Peptide-protein interactions,” Bioinformatics, vol. 35, no. 1, pp. 175–177, 2019. View at: Publisher Site | Google Scholar
  170. D. M. Mason, S. Friedensohn, C. R. Weber et al., “Optimization of therapeutic antibodies by predicting antigen specificity from antibody sequence via deep learning,” Nature Biomedical Engineering, vol. 5, no. 6, pp. 600–612, 2021. View at: Publisher Site | Google Scholar
  171. J. M. Jenson, V. Xue, L. Stretz, T. Mandal, L. Reich, and A. E. Keating, “Peptide design by optimization on a data-parameterized protein interaction landscape,” Proceedings of the National Academy of Sciences of the United States of America, vol. 115, no. 44, pp. E10342–E10351, 2018. View at: Publisher Site | Google Scholar
  172. L. Tallorin, J. Wang, W. E. Kim et al., “Discovering de novo peptide substrates for enzymes using machine learning,” Nature Communications, vol. 9, no. 1, article 5253, 2018. View at: Publisher Site | Google Scholar
  173. C. K. Schissel, S. Mohapatra, J. M. Wolfe et al., “Deep learning to design nuclear-targeting abiotic miniproteins,” Nature Chemistry, 2021. View at: Publisher Site | Google Scholar
  174. F. Grisoni, C. S. Neuhaus, M. Hishinuma et al., “De novo design of anticancer peptides by ensemble artificial neural networks,” Journal of Molecular Modeling, vol. 25, no. 5, p. 112, 2019. View at: Publisher Site | Google Scholar
  175. J. Caceres-Delpiano, R. Ibañez, P. Alegre et al., “Deep learning enables the Design of Functional de novo antimicrobial proteins,” Bioengineering, 2020. View at: Publisher Site | Google Scholar
  176. A. Capecchi, X. Cai, H. Personne, T. Köhler, C. van Delden, and J.-L. Reymond, “Machine learning designs non-hemolytic antimicrobial peptides,” Chemical Science, vol. 12, no. 26, pp. 9221–9232, 2021. View at: Publisher Site | Google Scholar
  177. J. Zupan, M. Novič, and I. Ruisánchez, “Kohonen and Counterpropagation artificial neural networks in analytical chemistry,” Chemometrics and Intelligent Laboratory Systems, vol. 38, no. 1, pp. 1–23, 1997. View at: Publisher Site | Google Scholar
  178. R. Hecht-Nielsen, “Counterpropagation networks,” Applied Optics, vol. 26, no. 23, p. 4979, 1987. View at: Publisher Site | Google Scholar
  179. A. T. Müller, G. Gabernet, J. A. Hiss, and G. Schneider, “Modl AMP: python for antimicrobial peptides,” Bioinformatics, vol. 33, no. 17, pp. 2753–2755, 2017. View at: Publisher Site | Google Scholar

Copyright © 2022 Suchetana Gupta et al. Exclusive Licensee Nanjing Agricultural University. Distributed under a Creative Commons Attribution License (CC BY 4.0).

 PDF Download Citation Citation
Altmetric Score