Revolutionizing Peptide-Based Drug Discovery: Advances in the Post-AlphaFold Era

Abstract

Peptide-based drugs offer high specificity, potency, and selectivity. However, their inherent flexibility and differences in conformational preferences between their free and bound states create unique challenges that have hindered progress in effective drug discovery pipelines. The emergence of AlphaFold (AF) and Artificial Intelligence (AI) presents new opportunities for enhancing peptide-based drug discovery. We explore recent advancements that facilitate a successful peptide drug discovery pipeline, considering peptides' attractive therapeutic properties and strategies to enhance their stability and bioavailability. AF enables efficient and accurate prediction of peptide-protein structures, addressing a critical requirement in computational drug discovery pipelines. In the post-AF era, we are witnessing rapid progress with the potential to revolutionize peptide-based drug discovery such as the ability to rank peptide binders or classify them as binders/non-binders and the ability to design novel peptide sequences. However, AI-based methods are struggling due to the lack of well-curated datasets, for example to accommodate modified amino acids or unconventional cyclization. Thus, physics-based methods, such as docking or molecular dynamics simulations, continue to hold a complementary role in peptide drug discovery pipelines. Moreover, MD-based tools offer valuable insights into binding mechanisms, as well as the thermodynamic and kinetic properties of complexes. As we navigate this evolving landscape, a synergistic integration of AI and physics-based methods holds the promise of reshaping the landscape of peptide-based drug discovery.

1. INTRODUCTION

Computational pipelines have become an indispensable step in the drug discovery process^[1]. For small molecules, honing these pipelines over many years has produced efficient and accurate workflows for identifying lead compounds. However, small molecules cannot target all possible proteins involved in disease. Peptide based drugs are a promising alternative that are already first-in-class for many diseases – but they present several challenges to identify new lead compounds. Here we identify the advances needed for a successful peptide drug discovery pipeline. Peptide flexibility, diversity, and number of binding modes posed a bleak future to develop such pipelines. However, the AlphaFold (AF) and Artificial Intelligence (AI) revolution has drastically changed the outlook, and in the post-AF era we are closer to fully integrated pipelines that are faster and more accurate than ever before.

1.1. Peptides as therapeutics

Peptides have gained increasing recognition since the development of insulin for their role as endogenous hormones, growth factors, neurotransmitters, and signaling molecules. A primary reason for their growing importance as a distinct class of drugs is their ability to bind with high specificity to targets, resulting in remarkable potency and relatively low off-target side effects^[2].

Peptides exhibit high selectivity due to the presence of large surface area interfaces and significant chiral and protein-like structural features. These traits have been favored by evolution to ensure compatibility in shapes and sizes, facilitating interactions across a diverse range of structural and functional contexts^[2]. Often, the bound peptides display evolutionarily conserved structural motifs such as α-helices, β-sheets, β-turns, γ-turns, and in some cases disulfide bonds (e.g., insulin)^[3,4]. This structural complexity is often absent in small molecules, which may explain the higher tendency of small molecules for off-target binding^[5].

Peptides are of special interest to target proteins that small molecule drugs are incapable of binding due to the absence of suitable binding pockets^[6]. Additionally, they are well-suited for interacting with extracellular receptors like GPCRs (G protein-coupled receptors), with numerous approved drugs for oncology or metabolic diseases already available, such as somatostatin, ghrelin, angiostensin, vasopressin, oxytocin, calcitonin, and glucagon^[7]. However, targeting intracellular proteins remains a significant challenge^[8,9].

Early on, peptides faced challenges concerning their short half-lives and high renal clearance. However, over time, several strategies have emerged to enhance stability, improve pharmacodynamics, and refine pharmacokinetic properties. For instance, since renal clearance is often size-dependent, elevating the molecular weight of peptides can significantly prolong their presence in plasma circulation^[10]. Various approaches, including C-terminal amide, N-terminal acetylation, backbone and side chain modifications, disulfide mimetics (e.g., Desmopressin, Carbetocin)^[11], lipidation (e.g., Liraglutide)^[12], conjugation, cyclization^[13](e.g., Setmelanotide), and pegylation (e.g., Pegfilgrastim)^[14] play a protective role against proteolytic degradation. In some instances, low clearance is even desirable due to fast onset of action and rapid clearance, as observed with oxytocin.

A second set of challenges arises from their poor bioavailability and limitations to oral administration, primarily caused by the difficulty in crossing the gut lining and rapid degradation by enzymes like amylases. Recent advancements in delivery technologies, such as microtechnology-based implantable pumps^[15], micro-needles^[16], liquid jet injectors^[17], and nasal delivery^[18] seek to address these bioavailability issues. Nevertheless, these methods have their drawbacks, like discomfort, safety concerns, or impact on glands and organs, which limits their use. Ongoing efforts are dedicated to developing orally-administered peptides, further pushing the envelope in the field of peptide therapeutics^[19].

1.2. The peptide therapeutic market landscape

Bringing a drug to market entails a lengthy timeline (~15 years) and a substantial investment (~$2 billion), starting from target identification^[20]. Despite this considerable commitment, the journey from phase 1 trials to FDA approval is successful for only 10-20% of drugs. However, the drugs that make it to the market generate enough revenue to compensate for the failures. In 2022, the revenue generated from peptide-based drugs was approximately 40 billion USD (see "Further Reading" for links to market data) and the projection over the next decade (see Figure 1a) is that the peptide market will continue to increase, but at a slower pace than the small molecule or biologics market. Similarly, the number of FDA approved peptide drugs per year remains a small fraction of their small molecule counterpart (see Figure 1b). Although peptides hold potential as remarkably selective drugs, the current limitations and the absence of well-established peptide-based drug discovery pipelines impede their therapeutic development. Nonetheless, the allure of blockbuster peptide drugs (see Table 1) and the diversity of diseases they target continue to arise interest in the scientific community, spurring efforts to enhance peptide-discovery pipelines. Moreover, an analysis of FDA-approved drugs from 2015 to 2022 shows that roughly half (20 out of 38) of approved peptide drugs fall into the category of First-in-class drugs (see Figure 1c), emphasizing the pioneering nature of peptide drugs for treating medical conditions. Finally, analysis of DrugBank^[21] data reveals 174 approved peptide drugs, with nearly 212 currently in human clinical trials (investigational category) and 74 peptides at the preclinical or animal testing stage (experimental category). This data further illustrates the significant presence and potential of peptide drugs within the pharmaceutical landscape.

Figure 1.

Projection of peptide-based drugs (see "Further Reading" for links to market study data). A. Comparison in market size between small-molecules, peptides and biologics. B. Number of approved drugs in each category by year. C. List of peptide drugs by years highlighting their first-in-class nature for many diseases and their delivery method.

Table 1:

List of Blockbuster Peptide Drugs, sorted by retail sales (based on Retail sales in 2022, see "Further Reading" for links)^[22].

Peptide	Brand Name	Length	Year of Approval	Disease	Retail Sales (Billion USD)
Insulin	Humalog , NovoRapid, …	51	1982	diabetes	15.12
Semaglutide	Ozempic, Wegovy	31	2012	diabetes	11.287
Dulaglutide	Trulicity	31	2014	diabetes	7.439
Liraglutide	Saxenda	30	2010	nutritional deficiency	1.556
Carfilzomib	Kyprolis	4	2012	oncology	1.328
Lanreotide	Somatuline	8	2014	oncology	1.306
Octreotide	Sandostatin	8	1989	oncology	1.238
Fam-trastuzumab Deruxtecan-nxki	Enhertu	4	2019	oncology	1.096
Linaclotide	Linzess	14	2012	gastrointestinal diseases	1.088
Goserelin	Zoladex	10	1989	oncology	0.927
Leuprorelin Acetate	Leuplin	9	1995	oncology	0.843
Ixazomib	Ninlaro	2	2015	oncology	0.743
Glatiramer	Copaxone	4	2014	neurology	0.691
Cyclosporine	Restasis	11	1983	ophthalmology	0.666
Etelcalcetide	Parsabiv	8	2017	endocrinology	0.447

Orally

Subcutaneously

Intravenously

2. COMPUTATIONAL DRUG DISCOVERY PIPELINES

The hit identification stage in drug discovery is a critical and time-consuming process, accounting for approximately half of the total time needed to take a drug to market. The challenge lies in exploring the vast chemical drug-like space, estimated to consist of ~10⁶⁰ molecules. Historically, high throughput screening (HTS) methods, like biological assays, have been successful in screening smaller subsets of this chemical space – but even for these subsets, the cost can become prohibitive^[23] . Alternatively, computer-based drug discovery pipelines like virtual screening (VS) have been in use for over five decades and forego the experimental expense of synthesizing and acquiring compounds^[24]. These pipelines, refined over years of development, serve to expedite the identification of lead compounds that effectively inhibit specific targets, thereby reducing both time and costs. A well-constructed pipeline should encompass at least four crucial elements: 1) designing a library of possible candidates, 2) predicting structures of the complex to identify binders, 3) ranking candidate molecules to identify promising leads (e.g., by binding affinity), and 4) generating new candidates based on the current best molecules. For small molecules the initial library could be as large as the commercially available small molecules from labs like enamine (https://enamine.net). Docking techniques are a good compromise between speed and accuracy for the VS of large libraries of compounds allowing users to identify and rank binding hits.

However, when dealing with peptides, the natural amino acid building blocks allow the creation of vast libraries (20^N peptides, with N being the peptide length). Yet, the complex task of predicting and ranking peptide structures remains highly challenging due to the peptide’s diversity and intrinsic flexibility^[25]. Drawing parallels between small molecule and peptide pipelines, a noteworthy breakthrough has emerged with the integration of AlphaFold (AF) and other recent Artificial Intelligence (AI) based tools^[26-29] . For small molecules it could shorten the time needed in the hit identification stage from years to months^[1]. For the peptide field, this development lays the groundwork for forthcoming efficient peptide discovery pipelines. We thus categorize methods used for peptide studies as either being pre-AF or post-AF (see Figure 2).

Figure 2.

The peptide-based drug discovery pipeline: pre-AF and post-AF. Once the initial target has been identified computational pipelines are used in the three stages leading to identifying a candidate drug. Those compounds will enter the next stages of identifying the action mechanism and clinical trials before eventually being approved. The expectation is that the post-AF technologies will reduce the overall timescale from years to months in the drug discovery stage.

2.1. Pre-AF protein-peptide structure prediction methods have low success rates

The prediction and ranking of peptide-protein complexes are underdeveloped as compared to the small molecule. This can be attributed to the fact that peptides are complex, flexible molecules with large numbers of accessible conformations that each peptide can sample, as well as how sensitive their conformational preferences are to the chemical environment (e.g., disordered in free form and different defined structure when binding different protein targets). Repurposed traditional docking methods have limited ability to deal with conformational flexibility and accurately score the complexes – often being limited to short peptides (up to five residues)^[30]. Even the integration of MMGBSA or FEP (Free Energy Perturbation) techniques have yet to attain the precision required to be adopted as a standard for peptide-protein complexes^[31]. In cases involving larger peptides, methods such as AutoDock CrankPep (ADCP)^[32], Rosetta PIPER FlexPepDock (PFDP)^[33], InterPep2^[34], and PatchMAN^[31] have made notable strides, outperforming traditional tools within benchmark sets. However, their success in sampling diverse binding conformations often contrasts with a lower success rate in selecting the optimal docked models^[32]. Despite these challenges, the presence of virtual screening (VS) peptide pipelines in prominent toolkits like Schrodinger Modelling Suites and OpenEye Docking toolkits attests to the interest in the field^[33].

2.2. Post-AF opens new opportunities for developing peptide VS pipelines

Since its introduction, AF has revolutionized the structural biology community. The realization that what AF had learned was transferable to predicting protein-peptide complexes opened new and exciting possibilities for the field of peptide-based drug discovery^[37]. Concomitant to accuracy and ease of use, there was the need for speed: while the original AF takes about 10-12 hours to predict a protein-peptide complex, the ColabFold implementation uses a fast sequence search technique (MMseqs2^[34]) that reduces computation time by 40-60 fold – to within a few minutes depending on the size of the system^[35]. While users of the Google Colaboratory implementation of ColabFold might have limited access to free resources, a locally installable version, LocalColabFold, can be easily used to screen thousands of peptides in parallel in a supercomputer cluster. The success of these methods has led to new AI developments such as AF-multimer^[36] and OmegaFold (OF)^[28]. In a recent benchmark, AF-multimer was the most successful (53%) in identifying protein-peptide complexes^[37]. Interestingly, traditional tools such as ADCP can sometimes succeed in cases where AI fails, despite their lower overall performance, prompting studies combining different traditional methods with AF. However, such approaches are only marginally better than AF – and AF could itself improve its performance nearly 10% by changing its model parameters and enhancing sampling^[36].

A key advantage of AF is its transferability across different protein and peptide sequences. However, within the domain of drug discovery, certain proteins hold greater significance as potential drug targets, amassing a more substantial repository of data garnered from varied assays over time. This is the case of the Major Histocompatiblity Complex (MHC), which plays important roles in immune responses, T-cell recognitions, and is an active target for peptide-based vaccines. Not surprisingly, this wealth of data is ideal for machine learning and AI-driven methods. For example, one area of interest for peptide-MHC complex prediction is classifying peptides as either binders or non-binders. Such is the case of NetMHCpan, trained on a dataset of binding and non-binding peptide-MHC complexes to predict binding affinities^[38]. However, its applicability to broader biological contexts is hindered by its lack of structural insights. To address this limitation, a recent study fine-tuned the AF network by incorporating a simple classifier (binder vs non-binder)^[39]. This adaptation allowed the network to demonstrate transferability to SH3 and PDZ domains. This approach can be further generalized if sufficient experimental data on binding affinity is available. Another example of fine-tuning the AF network is AlphaFold-TCR, which incorporates T-cell Receptor (TCR) data into peptide-MHC complex prediction to account for immunoresponses^[40].

Recent strides have also been made in advancing the prediction of structures and binding affinities for MHC-peptide complexes, exemplified by the MHCfold tool^[41]. This method slightly outperforms AF multimer in accuracy while remarkably enhancing computational speed, enabling the calculation of 100,000 predictions in just four hours. Other noteworthy advancements include TFold, which extends the AF network's prowess by incorporating paired templates and multiple sequence alignments (MSA) derived from binding data, alongside a bespoke sequence-based network^[42]. Additionally, an independent study conducted by InstaDeep and BioNTech harnessed a simplified graph neural network (GNN) model with 98% fewer learnable parameters than AF, yet yielded comparable performance^[43].

In the post-AF era, a pertinent inquiry arises: Do traditional docking tools retain their relevance for peptide-based discovery? While AF has indeed revolutionized the field, it is not devoid of limitations. AF's success rate in predicting native-like protein-peptide complex structures starting from sequences hovers around ~50% -- similar to the recent PatchMAN methodology^[31,37]. Furthermore, AF remains unable to predict the impact of point mutations, modified amino acids, and non-standard residues^[44]. In such scenarios, traditional docking tools retain their value. Nevertheless, the thrust in new docking tool development has shifted toward hybrid approaches or the refinement of existing deep learning networks within this post-AF era.

2.3. Exploring the peptide sequence space

A brute force VS approach for all possible peptide sequences against a target is impractical. Instead, a pragmatic approach involves designing peptide libraries tailored to the binding site, leveraging knowledge of the target protein. Conventionally, researchers capitalize on the structural and biological attributes of the target Protein-Protein Interaction (PPI) interface to pinpoint a handful of "hotspot residues”. Subsequently, utilizing the complementary nature of this interface, researchers curate initial peptides and refine these preliminary candidates through an optimization process^[45]. Alternatively, a known experimental lead is used as a foundational reference point for peptide design^[46]. These methodologies either focus on direct optimization of structural attributes (e.g., Peptide Binding Design) or combine Multiple Sequence Alignments (MSA) with design algorithms like PinaColada and pepCrawler to enhance binding affinity (e.g., PepWhisperer)^[46-49]. Despite notable achievements, these methodologies are constrained to system-specific investigations and present limited flexibility in terms of varying peptide length. Furthermore, even if a designed peptide outperforms the wild type PPI interaction, that design could be far from the “best possible” design.

2.4. AI-enhanced peptide design

Peptide design shares similarities with protein design, particularly in generating mini-protein designs that encapsulate the binding epitope within a biologically active conformation. Pre-AF, there were a limited number of groups possessing the required expertise for successful design. Among the standout tools for protein design, Rosetta emerged as an emblematic choice^[50,51]. Rosetta based designs typically draw information from developer expertise and complementary experiments such as directed evolution. Yet, its practical utility remained constrained due to a relatively modest success rate, many designs are needed for one to be successful, and the substantial computational expenses required during sequence searches and energy minimization processes. Consequently, its application was largely confined to a select cohort of expert groups ^[52,53].

The development of AI-based methods accelerates the process by benefiting from the ever-growing dataset, model architecture innovation and hardware revolution. Notably, the user-friendly nature of these AI tools has facilitated their adoption in experimental laboratories, transcending the requirement for computational modeling experts^[35]. This has enabled non-experts to design candidates for experimental testing. Many AI tools for peptide design are commonly categorized as either supervised or unsupervised models learned from well-curated dataset. In practice, a combination of both is often essential for hands-on application.

2.4.1. Supervised learning

Supervised learning involves learning a mapping function y = f(x) that takes an input x and predicts a target label y using a training dataset. This trained model can make accurate predictions for y when provided with new, unseen data. In the context of sequence-based peptide prediction, a supervised model can be trained using labeled sequences denoting binders and non-binders. Subsequently, this model can recognize and predict the classification of new sequences. Peptides are known to play crucial roles in various biological activities, such as facilitating drug delivery into cells using cell-penetrating peptides and potentially combatting cancer through anti-cancer peptides. PreTP-EL was designed for predicting therapeutic peptides. It employs an ensemble learning approach based on support vector machine (SVM) and random forest classifiers to model nine distinct peptide features^[54]. This model outperformed other state-of-the-art methods, displaying consistent performance on multiple functional therapeutic peptide prediction tasks. Building on this success, the same group later developed TPpred_ATMV. This method improves performance over the former by applying multi-view learning and tensor learning frameworks to extract complementary information among different sequence-based features with both predicted class and probability information^[55].

Recent efforts in supervised learning aim at accurate prediction of anticancer peptides (ACPs). For instance, DeepACP combines a recurrent neural network (RNN) model with sequence level features showing its superiority over both convolutional neural networks (CNN) and long short-term memory (LSTM) based architectures^[56]. Unlike single-task learning, xDeep-AcPEP exploited multitask learning (MTL) by training a neural network with shared common layers across multiple related tasks. Furthermore, the Grad-CAM method is adapted to analyze the relative importance of each residue for effective ACP predictions^[57].

Structure-based design aims to solve the inverse of protein folding problem: it seeks to find a suitable protein sequence that can adopt a desired structure. In essence, a mapping needs to be established between sequence (s) and structure coordinates (x) to optimize the conditional probability p(s∣x). Conventional methods, such as Rosetta, use the Markov-chain Monte Carlo (MCMC) technique to generate sequences optimized against a force field⁵⁶. However, the accuracy of force fields and computational cost to sample the search space for long peptides limit the success rate. Recent deep learning models showed substantial progress given protein structure in a rapid and automatic fashion^[58-60]. ProteinMPNN^[61], for instance, was developed to learn the conditional distribution p(s∣x) by employing an encode-decoder style architecture. The approach builds on previous work that predicts protein sequences autoregressively. It decomposes the distribution into p(s ∣ x) = ∏_ip(s_i ∣ x, s_<i), where p(s_i ∣ x, s_<i) of amino acid s_i depends on both the overall structure and the preceding sequence. The sequence recovery rate was enhanced by including more structural features and taking advantage of the ever-expanding structure database.

Finally, in instances where the desired topology is absent, recent strides in deep learning-based de novo protein design have demonstrated their robustness and generalizability via deep network hallucination^[62]. For example, a novel “inpainting” method draws inspiration from a deep learning breakthrough in the computer vision field. The approach generates a protein scaffold, yielding both sequence and structure, given a starting functional site. The method uses a retraining of RoseTTAFold (RF) where a subset of the MSAs was masked^[27,63].

2.4.2. Unsupervised learning

Unsupervised learning involves training a model to learn the representation from a dataset without annotated examples. Among the recent breakthrough in unsupervised de novo protein design^[64,65], RFdiffusion^[66], a diffusion model developed for designing sequences that fold into a structure when no initial structure template is given. This approach can tackle a wide range of challenges in protein design including de novo peptide binder design^[67]. The model uses an RF representation of protein structures and is trained to reverse the noising process through minimizing the difference between a true structure and structures generated starting from noise. The model has already proven successful in designing flexible helical peptide binders with picomolar binding affinity. There are also several attempts to utilize generative model such as Generative Adversarial Nets (GANs) and their variants to generate functional peptides. For instance, PepGAN incorporates a mixed loss function and antimicrobial peptides (AMPs) label information to encourage the generation of AMPs that outperform conventional antibiotics^[68]. AMPGAN and its variant, AMPGAN-v2, adopt a bidirectional CGAN framework and an encoder neural network to generate peptides with specific targets, mechanisms, and minimum inhibitory concentration (MIC) values^[69]. These models can indeed generate novel and potent AMPs, as evidenced through experimental validation. Other learning strategies such as transfer learning have also been incorporated in generative models to tackle the data scarcity challenge. A more thorough overview of developing generative models for peptide design can be found elsewhere^[70]. Even though such models have been used to generate peptides with distinct properties, most still need wet-lab validation to support their effectiveness.

2.5. AI-based peptide ranking

While significant efforts have been invested in creating various AI-driven tools for tasks like designing new binder peptides and accurately predicting complexes with minimal computational costs, only a handful of methods have emerged to address the third step of the drug design pipeline: ranking the candidates within a library. For example, AF was able to correctly predict the structures of different peptides binding to a target, it did not have a specific measure to rank order the peptides by binding affinity^[71]. Nonetheless, two key insights from addressing the interpretability and transferability of what AF had learnt led to the realization that AF had learnt a scoring function^[72] and competing peptides for the same target protein allowed the scoring function to discern the peptides that bind with higher binding affinity^[73]. Achieving this initial capacity to predict peptide-protein structures and rank peptides based on their binding affinity addresses two key limitations in current peptide-based computational virtual screening pipelines. It is especially remarkable that these two requirements stem from an AI methodology that was developed with a different purpose in mind (protein structure prediction)^[26]. Thus, we expect that current and future AI method development efforts will further improve on the quality for peptide ranking, leading to more versatile and reliable peptide VS pipelines.

3. DO PHYSICS-BASED METHODS STILL PLAY A ROLE IN A POST-AF WORLD?

We have already mentioned the successes and limitations of physics-based tools such as docking or Rosetta. Now, our focus shifts to MD (molecular dynamics) based methods, a category characterized by generating molecular ensembles and utilizing statistical mechanics to gain insights into biological phenomena. Multiple MD-based approaches and protocols exist, each designed to address specific questions like structure prediction, stability, or relative binding affinities. Certain approaches, such as alchemical free energy calculations, have become standardized and routine in pharmaceutical research, with errors in the range of 0.5-1 kcal/mol that limit chemical accuracy. However, many approaches such as those needed for structure prediction demand a high level of user expertise and are not consistently successful across all systems. The progression is hindered by two key factors: the quality of force fields and the extent of sampling needed to observe events of interest (e.g., binding or folding) with statistical significance^[74]. As a result, these tools encounter challenges in scenarios involving flexible systems, multiple potential binding modes, and variations in binding affinities across chemically diverse systems. Despite their computational expense and the expertise required, there are some areas where currently MD-based approaches can provide unique advantages, complementary to AI (see Fig. 3). For instance, peptide systems that: 1) incorporate non-standard amino acids, 2) involve unconventional cyclizations, 3) respond sensitively to single point mutations, 4) exhibit multiple metastable states, 5) require an understanding of thermodynamic and kinetic relationships between states, 6) involve binding mechanisms of interest. Thus, we expect that peptide-based drug discovery pipelines will integrate methods from these three disciplines (see Fig. 4 for an example pipeline).

Figure 3.

Qualitative comparison of the strengths and weaknesses of AI, docking, and MD-based approaches in different aspects of modeling peptides and their complexes, highlighting their complementary nature.

Figure 4.

Proposed peptide-based discovery pipeline. (A) Prediction of the peptide-protein complex structure. (B) Given the receptor, AI designs many candidate sequences, which are then ranked by binding affinity by other AI tools. (C) MD-based tools are used in a later stage to capture finer aspects such as thermodynamic and kinetic properties (top), selectivity (middle, here the blue peptide only binds receptor A, while the orange peptide, non-selective, binds all three related receptors), and modifications to the canonical amino acids (bottom).

Below we provide more details on the different applications of MD-based approaches. The current literature provides limited application examples for peptides due to computational costs, and their folding upon binding natures. However, the continuous development of more efficient enhanced sampling methods^[75,76], the increase in computational power, and development of integrative approaches^[77] bring optimism to the field.

3.1. Predicting structures

An MD ensemble starting from a native-like state will typically explore conformations close to the initial state, allowing the user to understand the local dynamics, deformation modes, and alternate accessible states. On the other hand, when starting far from the native state, if there are no force field or sampling issues, the native state should arise as the one with the lowest free energy (e.g., the state with highest population). This principle has readily been used to study the structures of peptides as they bind their partners and to identify differences in binding behavior between related peptide sequences. In practice, to sample efficiently, different enhanced sampling approaches are used.

When both the initial and final state are known and one is interested in differences among proteins sequences, approaches such as umbrella sampling and even regular MD readily identify the best sequences in agreement with experiments^[78-80]. However, more often we do not know the end state and need to rely on generalized ensemble methods (e.g., REMD^[81])^[82], biasing potentials along certain collective variables (e.g., metadynamics^[83])^[84,85], or integrative approaches based on Bayesian inference (e.g., MELD^[86], Modeling Employing Limited Data)^[87,88] and other enhanced sampling approaches^[89].

3.2. Mechanism, pathways, and kinetics

A caveat of some of the enhanced sampled methods is that they might favor a particular pathway and kinetic information might be lost. In those instances, it would be preferrable to carry out simulations that are unbiased. Techniques such as Weighted Ensemble (WE) methods^[90-93], adaptive sampling^[94] and Markov State Models^[95,96], or milestoning^[97] are ideal for starting multiple independent simulations within a framework that allows to reconstruct the ensemble generated from the individual trajectories. These methods allow to obtain rates well beyond what would be possible to simulate directly – for example the use of MSMs to study of the PMI peptide to MDM2-PMI predicted association rates beyond the seconds timescale^[98].

A challenge of these methods is that they tend to require a large amount of sampling to assure convergence, especially for properties such as k_off. While the use of biases to accelerate convergence, such as in Multi-ensemble Markov Models (MEMM) from MD and Replica Exchange, can result in good agreement with experiments^[99] it can also lead to errors if biases are not chosen correctly. Similarly, seeding unbiased simulations from biased ensembles can introduce biases in the resulting MSMs^[100,101]. More recently, Gaussian accelerated MD (GaMD)^[102] has also been used to bias the system starting from the structure of the complex to allow rapid dissociation and reassociation. Unbiasing the ensembles yields kinetic (k_on and k_off) and thermodynamic properties for peptide-protein complexes^[103] at a comparatively low computational expense.

3.2.1. Induced fit vs conformational selection

Many peptides are intrinsically disordered on their own and only adopt a folded conformation upon binding. The balance between the amount of residual structural propensity (e.g., helical or hairpins) can shift the balance between a conformational selection and induced fit binding mechanism. In conformational selection, the peptide binds in one of the conformations accessible at equilibrium. In the induced fit mechanism, the peptide will first bind weakly and then fold into the bound conformation. Different systems will bind through a combination of both mechanisms. For example, Voelz and co-workers found that p53 binds MDM2 through an induced-fit mechanism, but increasing the helical propensity of the peptide leads to an increase of the conformational selection pathway ^[95]. Thayer et al.^[104] employed MSMs to explore allostery in the CRIB-PDZ system, concluding that the allosteric process was a combination of sequenced steps of induced fit for when the protein is allosterically activated, followed by the conformational selection upon peptide ligand binding.

3.2.2. Special cases: cyclic peptides and modified amino acids

Cyclic peptides remain attractive in the field as a way to block the peptide in a certain conformation^[105,106], decrease proteolytic degradation, and increase membrane permeability^[107]. Due to the multiple ways of cyclizing peptides, and the many possible modified amino acids these are not readily implemented in AI tools, but they can be readily parametrized and incorporated into MD pipelines^[108] such as the ones previously described.

Figure 5 summarizes the different methodologies introduced in this review, serving as a quick reference guide for their main application (structure prediction, binding affinity, kinetic rates, and in designing novel binders).

Figure 5.

Applicability of the tools mentioned in the review. Each method is cited on the corresponding boxes.

Conclusion

The blooming field of AI has overcome some of the traditional limitations associated to computational peptide drug discovery pipelines. Beyond the realm of peptides, AI has already catalyzed the development of small molecule drugs currently undergoing clinical trials, underscoring its potential to reshape conventional drug discovery paradigms. This transformative influence extends to peptide discovery, as AI-driven tools hold the promise of a substantial shift in the landscape in the near future. The transferability of AF from the protein structure prediction problem to the peptide-protein structure prediction problem opened the door to more targeted AI development for peptide-protein complexes. Moreover, protein language models are now successfully extending their capacity to tackle various challenges including protein-peptide binding interface prediction^[109] and protein sequence generation^[110-113]. We are seeing the first generation of such post-AF methodologies, which are addressing further shortcomings, and contributing to establish all the steps required for a computational peptide drug discovery pipeline – similar to the successful ones for small molecules.

While substantial progress has been made, certain challenges endure over the long term, such as sensitivity to single point mutations, effect of non-canonical amino acids or unconventional cyclization strategies that will improve peptides’ drug-like properties and stability. This is one area where traditional physics-based methods, empowered by increasing computer power, more accurate force fields and novel enhanced sampling methodologies can fill the gap. Furthermore, the combination of AI and physics-based tools can help focus sampling and identify mechanisms of action, binding pathways, and binding rates.