Accelerating the Discovery and Design of Antimicrobial Peptides with Artificial Intelligence

Abstract

Peptides modulate many processes of human physiology targeting ion channels, protein receptors, or enzymes. They represent valuable starting points for the development of new biologics against communicable and non-communicable disorders. However, turning native peptide ligands into druggable materials requires high selectivity and efficacy, predictable metabolism, and good safety profiles. Machine learning models have gradually emerged as cost-effective and time-saving solutions to predict and generate new proteins with optimal properties. In this chapter, we will discuss the evolution and applications of predictive modeling and generative modeling to discover and design safe and effective antimicrobial peptides. We will also present their current limitations and suggest future research directions, applicable to peptide drug design campaigns.

1. Introduction

Most living organisms produce antimicrobial peptides (AMPs) as defense mechanisms in response to infections. These peptides generally protect their host from pathogens by exhibiting broad-spectrum microbicidal activities against bacteria, fungi, eukaryotic parasites, and enveloped viruses [1]. In addition, AMPs have demonstrated the ability to stimulate the host’s immune responses, anti-inflammatory, and wound healing properties, as well as anti-biofilm activity [2–8]. Thus, these peptides are also referred to as host defense peptides or HDPs [9, 10]. Moreover, the evolution of these peptides should also be taken into consideration as a mechanism to structure microbial communities [11, 12]. AMPs have received special attention as potential alternatives to conventional antibiotics against drug-resistant infections since they possess several benefits, including potent efficacy, good safety profiles and tolerability, high selectivity, standardized synthesis protocols, predictable metabolism, and present opportunities such as the discovery of new peptides, the development of appropriate formulations, and the ability to synthesize multifunctional peptides and conjugates. Unlike conventional antibiotics, AMPs do not readily select for bacterial resistance mechanisms, which is thought to be due to their non-specific mechanisms of action on the bacterial membrane [13–15]. Designing AMPs less susceptible to evolutionary resistance mechanisms without compromising the host’s innate immune response is an exciting prospect [16, 17].

Research strategies aimed at searching for novel AMPs/HDPs include isolating peptides from biological organisms or tissues across the microbial, vegetal, and animal kingdoms [18]. More recently, computational methods have been developed enabling the identification of a myriad of small encrypted domains with the desired characteristics from genomic and proteomic sequences [19, 20] and from larger proteins [21–23]. These advances are beyond the scope of this review. Today, the Antimicrobial Peptide Database lists over 3500 natural peptides from six life kingdoms [18], and the Giant Repository of Antimicrobial Peptides Activities, or GRAMPA, regroups roughly 6200 natural and synthetic peptides [24]. However, very few of these molecules have reached the market. Most AMPs currently in clinical trials are modified from natural peptides, the majority of which are administered topically. Academic laboratories and biotechnological and pharmaceutical companies have devoted substantial effort to overcoming limitation such as short half-life and fast elimination (proteolytic degradation, low stability), along with toxicity toward host cells at therapeutic doses without compromising the antimicrobial potential of the peptides [25–29]. Identifying peptides that limit the development of bacterial resistance while remaining cost-effective in terms of pharmaceutical development has also been a focus.

Computational models have gradually emerged as cost-effective and time-saving solutions to predict and generate new proteins with optimal properties. In silico predictive modeling accelerates the discovery of new peptides with therapeutic potential by learning the relationships between the protein sequences or structures and their functions (e.g., antimicrobial activity, thermal stability, and aggregation). Generative models create peptides that share meaningful representations (e.g., amphipathic character, conserved domain) with the original sequences or scaffolds. The synergy between predictive and generative models provides an informative way to effectively explore the vast chemical peptide space—20^L solutions for each peptide of length L made with canonical residues. Several reviews have summarized the development of early probabilistic models applied to AMPs based on evolutionary information to those leveraging traditional machine learning (ML) and deep learning (DL) algorithms [30–33].

In this chapter, we will describe how artificial intelligence (AI), i.e., ML and DL algorithms, can be used to accelerate the discovery and development of AMPs. First, we will introduce the different AI models that have been developed to predict and identify novel AMPs. Second, we will discuss their evolution and current limitations. Third, we will describe the development of generative AMP models. Finally, we will address current multi-objective strategies, combining predictive and generative models to discover and design safe and effective AMPs.

2. Predicting Peptide Antimicrobial Activity

Predictive ML models typically learn the relationships between the peptide sequences and their biological functions, such as their antimicrobial activity, from medium to large datasets. The power of such models relies on several factors such as their input datasets, their independent features (descriptors) and the type of algorithms used (i.e., classical ML vs neural network architectures) as illustrated in Fig. 1. Choosing the right input dataset(s) is the first important step toward model development. Most input datasets are either from in-house experiments or publicly available databases such as APD3 [18] or DBAASP [34]. Over the years, numerous AMP databases have been created adding new information, e.g., sequences, taxonomy, structure predictions, biological measurements in different media; however, only a handful provide experimentally validated 3D structures [35, 36]. The selection of descriptors with independent features representative of peptide sequence and structural diversity greatly influences the performance of ML models—an active field of study known as representation learning [37, 38]. These descriptors can be calculated based on the peptide sequence (e.g., peptide net charge, hydrophobic moment) or can be determined empirically (e.g., solubility, chromatography retention time), and these representation learning features can be derived for both single amino acids and for entire peptides [39]. Finally, AMP predictive models have gradually evolved from evolutionary and genetic algorithms (i.e., hidden Markov and linguistic models) to leveraging classical ML algorithms (i.e., random forest, support vector classifier) and DL architectures such as recurrent neural networks (RNNs) [30–33]. Selecting the optimal computational method is critical to significantly accelerate the development of new peptide drug candidates.

Fig. 1.

Computational AMP discovery and design workflow. This is a schematic diagram illustrating the workflow for using AI to accelerate antibiotic discovery. In the first phase (Experimental Information), large amounts of data are collected on potential antimicrobial peptides, such as their primary sequences, secondary structures, and biological activities. The data is then converted into a digital format that can be read by ML algorithms (Molecular Representation). For this purpose, relevant features are extracted from the data and presented in a standardized format. In the third phase (Machine Learning Models), the use of different algorithms (ML or DL) to create a suitable model is validated using a range of techniques to ensure their accuracy and reliability. In the final phase (Predictions and Generations), the best candidates identified by artificial intelligence models are selected for further testing and validation. (The figure was designed taking inspiration from Figure 1 [33])

2.1. Evolution of AMP Predictors

The earliest predictive models for antimicrobial activity can be traced back to the early 2000s when researchers started to use ML algorithms, such as artificial neural networks (ANN) and support vector machines (SVM), to classify antimicrobial activity—see Table 1. Such models are predominantly built as binary classification problems [i.e., whether a peptide sequence is predicted to be antimicrobial (AMP) or not (non-AMP)]. However, these models were trained on relatively small peptide datasets, which limited their accuracy and overall generalization. One of the earliest and most widely used AMP predictors is the AntiBP server, which used an SVM model trained on 1060 peptides from the Antimicrobial Peptide Database (APD) [18] to predict their activity against 10 bacterial species [41]. These models are sometimes referred to as Quantitative Structure–Activity Relationship (QSAR) studies, a term mostly associated with small molecule drug discovery since the 1960s [64].

Table 1.

List of AMP predictors (ML algorithms)

Year	Name	Algorithm(s)	Training size	Feature(s)	References
2007	AntiBP	SVM, ANN, NNA	872	AAC, diAAC	[40]
2010	AntiBP2	SVM, QM, ANN	1998	AAC, PseAAC	[41]
2011	Biosino	NNA	12,766	AAC, PseAAC, PCP	[42]
2012	ANFIS	ANFIS	231	AAC, diAAC, PCP	[43]
2012	AMPA	SVM	100	AI	[44]
2012	CS-AMPPred	SVM poly.	620	PCP, AAC	[45]
2012	ClassAMP	RF, SVM	1362	AAC, PseAAC, PCP	[46]
2013	PeptideLocator	BRNN	4505	SF	[47]
2013	iAMP-2L	ML-FKNN	3175	PseAAC	[48]
2014	DBAASP	SVM	1153	AAC, PseAAC, PCP	[49]
2015	ADAM	SVM, HMM	7000	SF	[50]
2015	Ng et al.	SVM-LZ	12,766	PSSM	[51]
2016	CAMPR3	RF, SVM	7021	PCP	[52]
2016	MLAMP	RF-ML	3284	AAC, PCP	[53]
2017	iAMPPred	SVM	8913	PseAAC, NAAC, PCP, SF	[54]
2018	AmPEP	RF	170,059	CTD global	[55]
2018	Joker	“linguistic”	303	Patterns	[56]
2019	dbAMP	RF	19,424	AAC, PCP	[57]
2019	AMAP	SVM, XGBoost	11,267	AAC	[58]
2020	amPEPpy 1.0	RF	170,059	PseAAC, PCP	[59]
2020	IAMPE	KNN, SVM, RF, XGB	5739	NMR-based clustering, PCP	[60]
2020	AmpGram	RF	4926	Encoded N-Grams	[61]
2020	Meta-iAVP	RF, KNN, SVM	1088	AAC, diAAC, GDC, PseAAC	[62]
2021	Ensemble-AMPPred	Ensemble learning	1840	PseAAC, CTD, PCP	[63]

Algorithms— ANFIS adaptive neuro-fuzzy inference system, ANN artificial neural network, BRNN bidirectional recursive neural networks, (ML-F)KNN (multi-level fuzzy) K-nearest neighbor, NNA nearest neighbor algorithm, RF random forest, SVM support vector machine, LZ Lempel-Ziv complexity, QM quantitative matrix, XGB extra gradient boosting. Features—AA amino acid, (Am)PseAAC (amphiphilic) pseudo amino acid composition, (N)AAC (normalized) amino acid composition, AI antimicrobial indices, CTD composition-transition-distribution, diAAC dipeptide composition, GDC g-gap dipeptide composition, N-Grams amino-acid motifs, PCP physicochemical properties, PSSM position-specific scoring matrix, SF structural features

Due to the lack of extensive experimentally known AMP structures in databases, most ML models predict the antimicrobial activity solely based on sequence information and thus are referred to as sequence-first models [65]. These sequence-first approaches rely on the recognition of different sequence motifs and physicochemical properties of existing AMPs to predict new peptides. Sequence-first models are based on classical ML algorithms which require numerical features encoding different peptide characteristics (e.g., amino acid composition, k-mers, or sequence-derived physicochemical properties) on a vector of finite length [66]. Sequence-first AMP predictors based on classical ML algorithms include DBAASP [49], ADAM [50], and iAMPpred [54]—see Table 1. These predictive models integrate multiple features of AMPs in numerical forms, including sequence, structure, and physicochemical properties. The most widely used algorithms are SVM, Decision Tree (CART), Random Forest (RF), and Gradient Boosting (GB). Random Forest is commonly used for predictive modeling tasks due to its ability to handle high-dimensional data and capture complex interactions among features [67]. It works by generating an ensemble of decision trees, where each tree is trained on a random data subset and features. The final prediction is made by aggregating the predictions of all the trees. In addition to its ability to handle high-dimensional data, Random Forest is a powerful algorithm for dealing with imbalanced datasets, which is often the case in AMP prediction [55]. Thus, it is not surprising that one-third of the 25 models listed in Table 1 used this algorithm.

However, classical ML techniques such as SVMs find it challenging to predict complex interactions between AMPs and bacterial membranes, as well as potential synergistic interactions, when combining different peptides. Recent sequence-first AMP predictors using Deep Learning (DL) architectures—convolutional neural networks (CNNs), recurrent neural networks (RNNs), and generative adversarial networks (GANs)—have also been explored to improve AMP prediction using large data sets and complicated modeling structures. These architectures aim to capture the nonlinear interactions between peptide sequence, structure, and antimicrobial activity to increase the performance of AMP predictors. Unlike classical ML models, these nonlinear models predict the antimicrobial activity of AMPs directly from their sequence, without intermediate features. Algorithms leveraging DL approaches are particularly promising because they can learn complex features and relationships within peptide sequences without the need for manual feature engineering. Peptide sequences are converted into numerical vectors using multiple encoding techniques, such as one-hot encoding where each amino acid is represented by a binary vector of length 20. The vector encoding for the entire peptide sequence will vary with the peptide length. The input datasets representing the AMP dataset are vectors of variable lengths. In some cases, DL algorithms require vectors of fixed length constraining the modeler to pad the existing vectors based on the longest one. AMP predictors integrating DL architectures include AMP scanner [68], DeepAMP [69], and Pep-CNN [70]—see Table 2. An exhaustive list of predictive models applied to AMPs and other types of peptides can be consulted at https://biogenies.info/peptide-prediction-list/ [78].

Table 2.

List of AMP predictors (DL algorithms)

Year	Name	Algorithm(s)	Training size	Feature(s)	References
2018	PepCVAE	VAE	1.6 M	Learned representation with characters	[71]
2018	AmPEP	CNN	1400	N-Grams, Composition	[55]
2018	AMP Scanner	BiLSTM-RNN	3818	PSSM, Composition	[68]
2018	Nagarajan et al.	LSTM	1512	Character sequence	[72]
2020	PepGAN	GAN	22,231	Character sequence	[73]
2021	iAMP-CA2L	CNN-BiLSTM-SVM	–	Composition, PCP	[74]
2021	AMPlify	CNN	18,860	Genomic and biochemical features	[75]
2021	AI4AMP	BiLSTM	706	PCP	[37]
2021	sAMP-PFPDeep	DNN	–	PCP	[76]
2022	LMPred	CNN	3556	PCP	[77]

Algorithms—BiLSTM (bidirectional) long- to short-term memory, CNN convolutional neural networks, DNN deep neural networks, GAN generative adversarial networks, NLM natural language model, RNN recurrent neural networks, VAE variational autoencoder. Features—N-Grams amino-acid motifs, PCP physicochemical properties, PSSM position-specific scoring matrix

2.2. Limitations

The use of ML can help reduce the cost and time associated with peptide design and discovery. However, biases in input dataset (s) used to train predictive models may affect their performance and our understanding of how and what they learn. Biological datasets are riddled with biases, and a systematic auditing to debiasing before building predictive ML models is essential [79]. Current research efforts to debiasing AMP datasets during recollection include early detection of taxonomic bias and structural bias.

Taxonomic bias refers to the high prevalence of homologous peptide sequences while sampling data from a limited set of source organisms. The study conducted by Rádai and co-workers (2021) illustrated taxonomic bias in predicting AMPs of invertebrate origins, mainly arthropods [80]. The authors collected a dataset of 4575 invertebrate peptides and compared the performances of 20 AMP predictors—ADAM [50], AMPscanner [68], AmpGram [61], CAMP_R3 [52], ClassAMP [46], CS-AMPpred [45], iAMP-2L [48], IAMPE [60], iAMP [54], and StM [81]. The results indicated that the performance varied depending on the taxon. Indeed, some tools showed increased performance for certain taxa. Overall, ADAM showed the best performance across all 19 taxa. In addition, the predictive accuracy was found to be very low for many of these tools, suggesting room for improvement, perhaps through the development of generalized ML models. The authors also identified another bias, namely algorithmic bias, where the hidden Markov model provided the best results in most taxa. Two solutions have been proposed to mitigate taxonomic bias across model predictions: (1) reducing the number of homologous sequences in input dataset(s) with tools like CD-HIT [82] and (2) combining predictions from multiple AMP identifying tools [80].

Structural bias refers to the high prevalence of a specific tridimensional structure in datasets. Most predictive ML models rely exclusively on sequence data, and the 3D- structures of peptides are largely ignored. If there is a biased distribution of amino acids, dipeptides, or oligopeptides, which are often linked to the physicochemical properties used for model training, then particular secondary structures would be adopted. As a result, predictive ML models built from sequences favoring certain folds would predict these folds more accurately. With that concept in mind, Aldas-Bulos and Plisson reported a trustworthy approach to assess the many folds of sizable sequence sets [83]. The authors first illustrated its use for fold discovery and mapped the structural space of GRAMPA, a library of over 5900 AMPs. Helices and random coils represented most of the folds (~83%), the rest included β-stranded and mixed structures. They notably showed that all datasets targeting specific strains such as Escherichia coli, Staphylococcus aureus, and Pseudomonas aeruginosa presented skewed distributions toward α-helices. Consequently, ML models trained on these distributions would predict the antimicrobial activity of α-helical sequences more accurately. A handful of predictive ML models have taken structural information into consideration. Some researchers have excluded peptide sequences based on structural limitations, so they presumably fold into the same 3D structure(s), and the structural effects upon prediction are reduced. For example, Dean and co-workers have recently developed tree-based regression models capable of predicting the minimum inhibitory concentration (MIC) against the three strains E. coli, S. aureus, P. aeruginosa for peptides folding uniquely into α-helices [84]. Currently, no methods have been developed to address structural bias. One may encourage researchers to synthesize and test peptides folding into non-helical structures for antimicrobial activity via combinatorial libraries, high-throughput screening, and rational design. The structural information of these active and inactive peptides would be added to existing predictive models in order to generalize over antimicrobial activity regardless of structural diversity. Alternatively, generative algorithms and de novo protein design provide potential solutions to “hallucinate” AMPs from underrepresented folds.

Moreover, a potential bias might exist in the methods used to measure antimicrobial activity, as the biological assays used may not accurately reflect the activity of an AMP in vivo or may use laboratory bacterial strains that are not representative of clinical pathogens of interest [85]. Indeed, in vitro assays are usually performed with bacterial strains that are easy to grow and obtain, but studies involving clinical isolates are lacking. Furthermore, experimental variations between laboratories make it difficult to collect biological measurements without generating noisy data that often leads to poor implementation of regression models. The development of GRAMPA [24], the giant repository of Antimicrobial peptides (AMPs) made from multiple publicly available databases and their corresponding MIC values against different bacteria strains, provides an excellent opportunity to develop novel regression models. For example, Dean and co-workers have notably curated the original GRAMPA datasets to focus on α-helical AMPs [84].

Since AMPs are produced as part of the innate immune system, they interact with immune cells to eliminate pathogens and prevent infections. The assays used to measure the antimicrobial activity of AMPs do not take into account these complex interactions, so the efficacy of AMPs in vivo may additionally be influenced by complex host immune responses [86]. On the other hand, the use of positive and negative labels is common practice in prediction models for AMPs, but negative data sets may contain peptides with different activities, leading to a generalization of antimicrobial activity for each predictive model [87]. To develop accurate ML models, high-quality datasets with representative peptide samples and appropriate features (descriptors) must be developed and used for training and validation, keywords used for data collection should be included [18, 78] and the information about MIC values should be consistent.

2.3. Perspectives

Predictive models have only recently been applied to identify the mechanisms of action of AMPs. In 2016, Lee and co-workers reported a seminal support-vector classifier to predict the membrane activity of AMPs rather than their antimicrobial function [88, 89]. Two years later, Brand and co-workers implemented unsupervised ML methods to classify membrane-active peptides from differential scanning calorimetry and circular dichroism experiments [90]. Despite being limited to α-helical AMPs, these studies revealed that physicochemical properties such as amphiphilicity and helical propensity could help predict the peptides’ activity on the membrane.

3. Generating Novel AMPs

Generative models can produce artificial peptides and proteins with desired properties or functions. Such sequences may resemble the original peptides or proteins as the algorithm intends to learn critical representations (e.g., an interactive domain). The design of novel AMPs using in silico generators initiated in the early 2000s with the artificial substitutions of one or more amino acids across sequences, under selective conditions. These algorithms included random sequence generators, genetic algorithms, evolutionary algorithms, or the ant colony optimization algorithm [31]. In the last 5 years, deep neural network architectures have translated from image classification, object recognition, or language translation to support peptide design and protein engineering. The most common algorithms applied to AMPs include recurrent neural networks (RNNs), convolutional neural networks (CNNs), variational autoencoders (VAEs), or generative adversarial networks (GANs) [32, 33]. RNNs are popular for modeling sequences of data, such as sentences or protein sequences. These algorithms capture the long-term dependencies between the amino acids in the sequence, making them well suited for predicting the properties and functions of (mini-)proteins. Vastly applied in computer vision, CNNs learn the patterns by treating protein sequences as one-dimensional vectors or two-dimensional images. VAEs are neural networks that learn to map input data to a lower-dimensional latent space representation. They generate new protein sequences by sampling from the latent space distribution, subsequently decoding the samples into protein sequences. Finally, GANs consist of two neural networks, a generator and a discriminator, which work together to generate new protein sequences with desired properties. The generator learns to generate new protein sequences by sampling from a latent space distribution, while the discriminator learns to distinguish between real and generated protein sequences. The generator is trained to produce protein sequences that can fool the discriminator into thinking that they are real, while the discriminator is trained to identify the generated protein sequences.

3.1. Evolutionary-Based Generators

Conventional peptide design methods have had limited success in producing AMPs with high potency and low toxicity. Early models like de novo design, linguistic models, and pattern insertion methods designed AMPs through a grammar model which considered the primary sequences as a “vocabulary” and the pattern of repetition of amino acid residues as “rules,” whereas the genetic algorithms produced successive generations of mutation and deletion events on the target sequence to improve the peptide performance and identification of patterns associated with antimicrobial activity. This has led to the exploration of alternative approaches, including evolutionary-based generators. Evolutionary algorithms are a powerful optimization method inspired by natural selection to evolve a molecule toward a desired biological performance through successive generations of event mutations (i.e., mutation, selection, and recombination events). In the context of AMP design, evolutionary algorithms have been used to design new peptides with increased antibacterial activity, stability, and specificity [31, 91, 92]. One advantage of evolution-based generators is that they can search a much larger sequence space than conventional approaches. This is particularly advantageous when designing AMPs with complex biological functions such as broad-spectrum activity and/or high selectivity. Another advantage of evolution-based generators is that they can go beyond what is currently known in sequence space, leading to the identification of previously undiscovered motifs or combinations of motifs. Genetic algorithms can classify virtually any AMP sequence using fitness functions based on antimicrobial activity descriptors and structural information collected from databases [91, 92]. A study using a genetic algorithm [91] led to the design of guavanin 2, a synthetic peptide that displayed anti-infective efficacy in a preclinical mouse model, demonstrating that machines could be used to generate preclinical antimicrobial candidates.

3.2. Deep Learning Models

Several computational peptide sequence generators are based on deep learning algorithms that learn the underlying patterns and relationships between the amino acid sequences and their corresponding peptide/protein structures and biological functions. The algorithms include recurrent neural networks (RNNs), convolutional neural networks (CNNs), variational autoencoders (VAEs), and generative adversarial network (GANs), as shown in Table 3.

Table 3.

List of deep learning AMP generators

Year	Name	Training size	Feature(s)	Algorithm (s)	References
2018	Müller et al	29,142	One-hot encoding	LSTM-RNN	[93]
2018	Nagarajan et al.	1512	Character sequence	RNN	[72]
2018	PepCVAE	1.6 M	One-hot encoding	VAE	[71]
2020	Caceres-Delpiano et al.	2 M	Learned representation using structural and evolutionary data	VAE	[94]
2020	Tucs et al.	508,000	Character sequence	GAN	[73]
2020	AMPAGANv2	6238	Learned representation using character sequence	GAN	[95]
2021	Wang et al.	2851	One-hot encoding	LSTM-RNN	[96]
2021	Das et al.	41,566	Learned representation using character sequence	VAE	[97]
2021	PepGAN	22,231	Learned representation using character sequence	VAE	[73]
2021	PandoraGAN	1276	Character sequence	GAN	[98]
2023	HydrAMP	11,131	Learned representation using character sequence	cVAE	[99]

Algorithms— LSTM long-short term memory, RNN recurrent neural network, (c)VAE (conditional) variational auto-encoders, GAN generative adversarial networks

Due to their popularity in learning patterns from sequential data, RNNs have been applied to analyze peptide sequences and generate new peptides with similar characteristics. In 2018, Müller and co-workers used the long short-term memory (LSTM) network to create novel antibacterial peptides with low toxicity against Gram-negative and Gram-positive bacteria [93]. The authors used LSTM to overcome the vanishing gradient problem associated with training RNNs on long sequences. The authors proposed an RNN-based model called DeepEyes, which uses a generative approach to identify new AMPs. The model takes the target protein as input and generates a new peptide sequence that binds to the protein’s target site. They showed that their model can generate novel peptides with high binding affinity and antimicrobial activity [93].

Another study by García-Jacas and co-workers compared the performance of different ML models, including DL models, to identify AMPs [100]. They found that DL models, especially RNNs, performed better than other models. In 2019, Hamid and Friedberg proposed a new method that combines word embedding with deep RNNs to identify AMPs from amino acid sequences [101]. Their model, called DeepAMP+, uses a bidirectional LSTM network to capture both forward and reverse dependence of amino acid sequence. They compared the performance of their model with several other classification methods and showed that DeepAMP performed best at identifying AMPs. The same year, Bolatchiev and co-workers created DeepAMP+, an RNN-based model used to engineer new AMPs that can reduce mortality in experimental sepsis. The model was trained on a large database of known AMPs and used to design new peptides with potent antimicrobial activity and low toxicity [102]. Finally, Sharma et al. (2022) developed Deep-ABPpred; the model uses bidirectional LSTM and word2vec techniques to capture the complex dependencies and semantic relationships between amino acid residues in protein sequences for the identification of antibacterial peptides in protein sequences [103].

Other DL algorithms have been applied to the field of AMPs. For example, in 2019, Dean and Walper introduced Variational Autoencoder for Antimicrobial Peptides (VAE-AMP), a VAE trained on a large database of known AMPs to generate peptides with desired properties, such as increased antimicrobial activity or decreased toxicity [104]. In 2022, Ghorbani and co-workers developed a deep attention-based variational autoencoder (DAVAE) for antimicrobial peptide discovery [105]. The DAVAE model consists of an encoder that maps the amino acid sequence of a peptide into a latent space and a decoder that generates a new peptide sequence from a random point in the latent space. The attention mechanism in DAVAE enables the model to focus on specific parts of the input sequence that are more relevant to the prediction of antimicrobial activity. CNNs were trained to distinguish AMPs from non-AMPs by analyzing their amino acid sequences. In 2019, Su and co-workers utilized multi-scale CNNs for the identification of antimicrobial peptides [106]. The multi-scale CNN architecture allowed the model to effectively capture the spatial and structural features of peptides at different scales. In 2020, Yan and co-workers demonstrated that CNNs were effective in predicting the antimicrobial activity of novel peptides, suggesting their potential for developing new AMPs with improved properties [107]. Finally, GANs are another DL model used for generating unique AMPs, as they can produce new data that is comparable to a given dataset. In 2021, Van Oort and co-workers employed a generative adversarial network to engineer novel AMPs that showed stronger antimicrobial activity than the original training dataset [108]. Tucs and co-workers present a novel method for engineering high-activity AMPs using Activity-Aware Generative Adversarial Networks (AA-GAN) [73]. The model AA-GAN uses a combination of CNNs and GANs to generate new peptides. The authors showed that their method could generate peptides with activity comparable to that of the conventional antibiotic ampicillin.

3.3. Perspectives

Most generative models applied to AMPs and peptides are based on their primary sequences, lacking structural information. Alternatively, protein designers have devised robust computational strategies to challenge the sequence-structure-function problem upside down, namely by generating sequences that fold into a predefined tridimensional structure derived from native protein folds or generated de novo [109–111]. These approaches, merged under umbrella terms inverse protein folding [112], fixed-backbone, or structure-based protein design [113], sample amino acid distributions to find optimal sequences for a given fold. Computational biologists have notably generated new functional proteins with unprecedented structures [114], artificial immunoglobulins [115], or de novo TIM barrels [116, 117]. ProteinMPNN [118] and MutDock [119] are two recently reported tools democratizing fixed-backbone peptide/protein design practices. Ultimately, reinforcement learning, another branch of artificial intelligence, has moved from board games to design proteins [120].

4. Interplay

Peptide drug design is a multi-objective problem, as it requires optimizing for multiple criteria simultaneously, including efficacy, safety, bioavailability, and cost-effectiveness. Predictive and generative ML models interplay to provide a grounded sequence design to address these challenges. Predictive modeling enables efficient screening and selection of promising candidates. Generative modeling, such as molecular design algorithms and DL-based approaches, generates novel peptide/protein sequences that meet desired properties or functions and optimizes multiple objectives simultaneously. By combining these two modeling strategies, computational biologists and drug designers can explore the vast peptide sequence space in an efficient manner and identify candidates that present the desired properties—as summarized in Fig. 2. Here we listed a series of examples where the authors combined predictive and generative algorithms to develop AMPs.

Fig. 2.

Streamlining the DMTA cycle for computational peptide design. DMTA stands for design, make, test, and analyze. First, the computational protein designer “designs” the experiment and selects one or more seed peptides. These peptides will be encoded to train ML generators and predictors. Second, the generative model “makes” new peptide sequences based on certain conditions (e.g., maintain amphipathic character, keep 3D structure) similar to a medicinal chemistry campaign. Third, predictive models substitute biological assays and “test” all sequences for potential candidates. Finally, the last step “analyzes” if the generated sequences comply with the desired characteristics. (This figure was adapted from Figure 1 [121])

The first wave of computational strategies combined generative algorithms based on evolution with predictive modeling for antimicrobial activity. In 2013, Maccari and co-workers developed an evolutionary multi-objective optimization approach to design new AMPs with improved biological activity and selectivity. The authors used a genetic algorithm to explore the chemical space of peptides and optimized multiple objectives simultaneously, including antimicrobial activity, hemolytic activity, and physicochemical properties. The potential of their approach was demonstrated by generating new AMPs with potent antimicrobial activity and low toxicity [122]. Another study by Müller and co-workers introduced a sparse neural network model for predicting the antimicrobial activity of peptides from their sequence-derived physicochemical properties. The authors used an evolutionary algorithm to optimize the weights of the neural network [123]. In 2018, Beltran and co-workers reported an evolutionary feature weighting approach for selecting the most relevant molecular descriptors for AMP classification. The authors used a genetic algorithm to optimize the weights of the molecular descriptors and demonstrated the potential of their approach by achieving high accuracy in classifying a large dataset of AMPs [124]. The same year, Porto and co-workers developed a computational approach for designing AMPs using a genetic algorithm. The algorithm was based on a fitness function that incorporated both the antimicrobial activity and the physicochemical properties of guava peptides. The authors tested their approach on a large dataset of natural and synthetic AMPs and found that the algorithm was able to generate new sequences with higher activity than the original peptides [91]. Around the same time, Yoshida and co-workers reported the use of genetic algorithm and predictive modeling to develop analogues of 13-mer peptide Temporin-Ali and succeeded by producing highly potent antimicrobial peptides against E. coli [125]. In 2019, the platform ADAPTABLE, designed by Ramos-Martín and co-workers, integrates multiple computational approaches, including machine learning and evolutionary algorithms, to provide a comprehensive and user-friendly tool for AMP design [126]. Finally, in 2021, Plisson and co-workers designed ~50 new non-hemolytic AMPs by generating random AMPs sequences from measured amino acid frequencies and predicted their non-hemolytic nature or activity from GB algorithms [127].

In the last 5 years, researchers have combined predictive models with generative DL architectures. In 2021, Capecchi and Reymond generated and analyzed a large library of peptides, with the goal of understanding their chemical diversity and distribution in chemical space. This study highlighted multiple distinct regions in peptide chemical space that were enriched in sequences with specific properties, including α-helical structures, cationic charge, or hydrophobicity [128]. The same year, the authors trained RNNs with data from DBAASP (Database of Antimicrobial Activity and Structure of Peptides) to design small non-hemolytic α-helical antimicrobial peptides [129]. In 2022, Agüero-Chapin et al. (2022) proposed a computational approach to identify new AMPs combining evolutionary algorithms and machine learning techniques. First they generated a large library of peptide sequences and then trained a SVM model to predict their antimicrobial activity. The authors tested their approach on a dataset of AMPs with known activity and found that their method was able to identify new peptides with high activity against several bacterial pathogens [130]. In 2022, Röckendorf and co-workers reported a multi-objective optimization approach for designing new membrane-active peptides with improved antimicrobial activity and biocompatibility. The authors used a genetic algorithm to explore the chemical space of peptides and optimized multiple objectives simultaneously, including antimicrobial activity, hemolytic activity, and cell viability [131]. Finally, this year, Szymczack and co-workers reported HydrAMP, a variational autoencoder trained to generate peptide sequences conditioned on antimicrobial activities [99].

5. Conclusions

This chapter discusses the importance of predictive and generative machine learning (ML) models in peptide drug design. Predictive models rely on input datasets and descriptors with independent features to learn mathematical relationships between peptide sequences and their biological indices. Designing the right AI models can significantly accelerate the development of new peptide drug candidates. Generative models produce artificial peptides with desired properties or functions, using algorithms such as RNNs, CNNs, VAEs, or GANs. Peptide drug design is a multi-objective problem, and the interplay between predictive and generative ML models provides a viable design strategy to address these challenges. By combining these two modeling strategies, computational biologists and drug designers have the ability to search through the vast peptide sequence space and pinpoint candidates that fulfill the objectives of interest, ultimately leading to antimicrobial candidates.

However, these models are not exempt from biases and errors. The choice of peptide encoders and algorithms affects model performance. To obtain a robust model with experimental predictions, the coupling of in silico models tested with in vitro and in vivo validations needs to be improved. It will be essential to continue to enrich the databases with high-quality and readily available data. However, there are still some difficulties to overcome regarding the clinical use of synthetic AMP peptides such as production costs, proteolytic stability, and unknown toxicity during systemic administration. The entry of several synthetic AMPs into clinical trials, added to the many peptide mimetics as antimicrobial agents already in the pipeline, suggests an exciting future in this area.

AI could help circumvent all these hurdles by taking into account experimental data including D-amino acids or non-natural residues to reduce the lability to degradation by proteases and by reducing toxicity through the use of formulations. Reliable computational toxicology predictions are being developed to improve models that explicitly consider crucial preclinical toxicological end points for AMPs. It will also be essential to think about the development of models to assess the mechanisms of action of peptides. The combination of expertise, through a close collaboration between computational and experimental scientists, will be critical. Finally, federated learning across institutions could facilitate collaborations without the need to share confidential datasets.