1. Bibliographic Information

1.1. Title

The central topic of this paper is "Artificial intelligence-driven antimicrobial peptide discovery."

1.2. Authors

The authors are Paulina Szymczak and Ewa Szczureka. Both authors are affiliated with the Faculty of Mathematics, Informatics and Mechanics, University of Warsaw, Banacha 2, 02-097, Warsaw, Poland. Ewa Szczureka is the corresponding author (indicated by an asterisk).

1.3. Journal/Conference

This paper was published as a preprint on arXiv. arXiv is an open-access repository for preprints of scientific papers in fields like mathematics, physics, computer science, quantitative biology, quantitative finance, statistics, electrical engineering and systems science, and economics. Papers on arXiv are not peer-reviewed before posting, though many are later published in peer-reviewed journals. Its reputation is high as a platform for rapid dissemination of research, but its influence as a peer-reviewed venue is none, as it serves as a preprint server.

1.4. Publication Year

The paper was published on 2023-08-21T14:02:14.000Z, which corresponds to the year 2023.

1.5. Abstract

The paper reviews the advancements in artificial intelligence (AI) for discovering antimicrobial peptides (AMPs), which are crucial alternatives to conventional antibiotics due to rising antimicrobial resistance. The authors categorize AI methods into two main groups: discriminators and generators. Discriminators predict key peptide properties such as activity and toxicity, aiding in the identification of promising candidates. Generators, on the other hand, learn the underlying distribution of peptides and can sample novel AMP candidates either de novo (from scratch) or as analogues (modified versions) of existing peptides. The review also highlights advanced techniques for controlled generation, including discriminator-guided filtering, positive-only learning, latent space sampling, and conditional and optimized generation, which enable the creation of AMPs with desired properties. The paper concludes by summarizing recent achievements and pointing out exciting future directions in AI-driven AMP discovery.

1.6. Original Source Link

Original Source Link: https://arxiv.org/abs/2308.10921v1 PDF Link: https://arxiv.org/pdf/2308.10921v1.pdf This paper is published as a preprint on arXiv.

2. Executive Summary

2.1. Background & Motivation

The core problem the paper aims to address is the urgent global health hazard posed by antimicrobial resistance (AMR). AMR, primarily fueled by the overuse of antibiotics, has led to the emergence of drug-resistant bacterial strains, making infections increasingly difficult to treat. In 2019, AMR was ranked as the third leading cause of death globally, surpassing HIV and malaria. This situation has created a pressing need for the discovery of new antimicrobial pharmaceuticals, as no successful novel antibiotics have been developed for over 30 years.

Antimicrobial peptides (AMPs) are presented as a promising alternative to conventional antibiotics. These naturally occurring peptides are part of the innate immune system of many organisms and are effective against a broad spectrum of pathogens, including antibiotic-resistant ones, often with a slower rate of resistance development compared to traditional antibiotics. Despite their potential, only a limited number of AMPs have reached clinical trials or commercialization, mainly due to challenges like low activity, high toxicity, or instability in a clinical setting.

The paper's entry point is the recent, tremendous advancements in artificial intelligence (AI), particularly in generative models and large language models. These AI technologies have revolutionized various fields, including drug and protein design, and are now being extensively applied to AMP discovery. The paper aims to provide a comprehensive review of these AI-driven approaches in AMP discovery, focusing on developments over the last two years and identifying the most exciting future directions.

2.2. Main Contributions / Findings

The paper's primary contributions lie in its comprehensive review and structured categorization of AI-driven AMP discovery methods. It provides a detailed characterization of tasks that AI methods perform, outlines the diverse properties of AMPs, and discusses their model representations.

The key contributions and findings include:

• Categorization of AI Methods: The paper clearly distinguishes between discriminator and generator approaches in AI-driven AMP discovery. Discriminators are AI models designed to predict properties (e.g., activity, toxicity) of existing or proposed peptides, aiding in the identification of promising candidates. Generators are AI models that create novel peptide sequences, either de novo (from scratch) or as analogues of known peptides.

• Controlled Generation Techniques: It highlights advanced strategies for controlled generation of AMPs with desired properties. These include discriminator-guided filtering (using discriminators to select generated peptides), positive-only learning (training generators exclusively on desired examples), latent space sampling (navigating the latent space of generative models to find peptides with specific characteristics), conditional generation (explicitly guiding generation based on desired conditions), and optimized generation (modifying existing peptides to improve properties).

• AMP Characterization and Representation: The review details various properties of AMPs (activity, toxicity, stability, synergy) and their diverse representations for AI models, from simple amino acid sequences to complex structural details and pretrained language model embeddings.

• Evaluation Framework: The paper discusses the challenges and current approaches to evaluating AMP discovery methods, covering both methodological metrics (e.g., diversity for generators) and experimental validation (e.g., MIC, HC50, Molecular Dynamics simulations). It also proposes the AMP success rate curve for a more nuanced experimental validation.

• Identification of Challenges and Opportunities: The paper concludes by outlining significant challenges impeding AMP clinical translation, such as the lack of community-wide benchmarking data, scarcity of data for critical properties (cytotoxicity, stability), noise in existing data, the need for better ranking of generated candidates, and the development of faster AI-driven molecular dynamics simulations. These challenges are framed as opportunities for future advancements.

  In essence, the paper provides a structured map of the current landscape of AI-driven AMP discovery, clarifying existing methods, evaluating their utility, and charting a course for future research to overcome current bottlenecks in bringing AMPs to the clinic.

3. Prerequisite Knowledge & Related Work

3.1. Foundational Concepts

To understand the paper, a beginner should grasp several fundamental concepts related to antimicrobial peptides and artificial intelligence.

• Antimicrobial Peptides (AMPs):

  • Definition: AMPs are small proteins, typically 10-100 amino acids long, produced by various organisms (e.g., humans, animals, plants, bacteria) as a primary component of their innate immune system. They are characterized by an overall positive charge and a significant percentage ($\geq 30\%$) of hydrophobic (water-fearing) amino acids.
  • Mechanism of Action: Many AMPs work by disrupting microbial membranes (e.g., bacterial cell walls), leading to cell lysis (breakdown). Their positive charge helps them bind preferentially to the negatively charged bacterial membranes, distinguishing them from host mammalian cells which have neutral membranes.
  • Properties: Key properties include activity (effectiveness against microbes, often measured by Minimum Inhibitory Concentration (MIC)), toxicity (harmfulness to host cells, often measured by hemolytic activity or cytotoxicity), stability (resistance to degradation), and synergy (combined effect with other antimicrobials).
  • Antimicrobial Resistance (AMR): A global health crisis where microorganisms evolve to withstand the effects of antimicrobial drugs, making infections harder to treat. AMPs are seen as a potential solution due to their different mechanisms of action and slower resistance development.

• Artificial Intelligence (AI):

  • Definition: AI refers to the simulation of human intelligence processes by machines, especially computer systems. These processes include learning, reasoning, problem-solving, perception, and language understanding.
  • Machine Learning (ML): A subfield of AI that enables systems to learn from data without explicit programming. It focuses on developing algorithms that can parse data, learn from it, and make predictions or decisions.
  • Deep Learning (DL): A subfield of ML that uses artificial neural networks with multiple layers (hence "deep") to learn complex patterns in data. These networks can automatically learn hierarchical features from raw input.

• Neural Networks (NNs):

  • Basic Structure: Composed of interconnected nodes (neurons) organized in layers (input, hidden, output). Each connection has a weight, and each node has an activation function.
  • Training: NNs learn by adjusting weights based on input data to minimize the difference between predicted and actual outputs (loss function), often using backpropagation and gradient descent.

• Specific Deep Learning Architectures Mentioned in the Paper:

  • Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM):
    • RNNs: Neural networks designed to process sequential data (like peptide sequences). They have internal memory, allowing them to use information from previous steps in the sequence.
    • LSTMs: A specialized type of RNN that can learn long-term dependencies in sequences, overcoming the vanishing gradient problem of standard RNNs. They achieve this using gates (input, output, forget gates) that control the flow of information.
  • Convolutional Neural Networks (CNNs):
    • Primarily for images: CNNs are typically used for image processing but can be adapted for sequences. They use convolutional filters to detect local patterns (e.g., motifs in a peptide sequence).
  • Generative Adversarial Networks (GANs):
    • Definition: A framework consisting of two neural networks: a generator and a discriminator.
    • Generator: Tries to create new data samples (e.g., peptide sequences) that resemble the training data.
    • Discriminator: Tries to distinguish between real data samples from the training set and fake samples created by the generator.
    • Training: The two networks compete in a zero-sum game. The generator learns to produce increasingly realistic samples to fool the discriminator, while the discriminator learns to become better at identifying fakes. This adversarial process drives both networks to improve.
  • Variational Autoencoders (VAEs):
    • Definition: A type of generative model that learns a latent space (a lower-dimensional representation) of the input data.
    • Encoder: Maps input data (e.g., a peptide sequence) into a probabilistic distribution in the latent space (mean and variance).
    • Decoder: Samples from this latent distribution and reconstructs the original input data.
    • Generation: Novel samples can be generated by sampling points from the learned latent space and passing them through the decoder.
    • Latent Space: A compressed representation of the input data where similar data points are close to each other. It allows for smooth interpolation and generation of new, plausible data.
  • Transformers and BERT (Bidirectional Encoder Representations from Transformers):
    • Transformers: A neural network architecture that relies heavily on the self-attention mechanism to weigh the importance of different parts of the input sequence. They are highly effective for natural language processing tasks.
    • Self-Attention: A mechanism that allows a model to weigh the importance of different words (or amino acids in a sequence) when processing a particular word. For a query $Q$, keys $K$, and values $V$: $ \mathrm{Attention}(Q, K, V) = \mathrm{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V $ where $Q$ is the query matrix, $K$ is the key matrix, $V$ is the value matrix, $d_k$ is the dimension of the keys, and $\mathrm{softmax}$ is the softmax function. This formula calculates attention scores by taking the dot product of the query with all keys, scaling by $\sqrt{d_k}$, applying softmax to get probabilities, and then multiplying by the values.
    • BERT: A transformer-based language model designed for pre-training on large text corpora (or protein sequences) and then fine-tuning for specific downstream tasks. It uses a bidirectional approach, considering the context from both left and right sides of a word/amino acid.

3.2. Previous Works

The paper frames itself as a review, building upon a rich history of AI applications in AMP discovery. It acknowledges that prior reviews have either focused on specific AI methods (like language models [12]) or geometric deep learning [11]. This paper complements these by providing a broad overview of approaches over the last two years, highlighting recent innovations.

Key categories of previous work that form the foundation include:

• Early AMP Prediction (Discriminators): Initial AI efforts focused on classification (AMP vs. non-AMP) and regression (predicting MIC values). These often used hand-engineered features (e.g., amino acid composition, charge, hydrophobicity) combined with traditional machine learning models like Support Vector Machines (SVMs) or Random Forests. The paper notes that even now, descriptor-based methods can be competitive for basic classification [39].
• Deep Learning for Discrimination: More recent discriminators leverage deep learning architectures like LSTMs (e.g., AMPlify [17]) or pretrained BERT models (e.g., AMP-BERT [44]) to automatically learn features from peptide sequences. Some approaches even convert sequence/structure into images for image recognition methods (e.g., VGG16-AMP [43]).
• Generative Models for De Novo Design: The concept of generative AI for molecular design has been evolving. Early generative models might have used simpler statistical methods or rule-based systems. The recent surge in GANs and VAEs for generating novel molecules (e.g., drugs [7], proteins [8]) laid the groundwork for their application to AMPs.
• Database Development: The existence of comprehensive AMP databases (e.g., DBAASP [16]) is crucial for training and evaluating AI models. These databases record activity, toxicity, sequence, and structural information, though they present challenges regarding data consistency and unit standardization.

3.3. Technological Evolution

The evolution of technology in AMP discovery has mirrored the broader advancements in AI:

1. Rule-based & Feature Engineering Era: Initial attempts relied on expert knowledge to define physicochemical properties (e.g., charge, hydrophobicity, amphipathicity) and then used classical machine learning algorithms to build predictive models. This required significant domain expertise and feature extraction efforts.
2. Sequence-based Deep Learning Era: With the rise of deep learning, models shifted to learning directly from peptide sequences. RNNs and LSTMs became popular for their ability to handle sequential data, automatically extracting relevant features.
3. Language Model & Generative AI Era: The success of Transformers and large language models (like BERT) in natural language processing quickly extended to biological sequences. These models, pre-trained on vast protein datasets, can capture complex biological grammar and relationships, leading to more powerful embeddings (numerical representations of sequences). Simultaneously, generative models (GANs, VAEs) moved beyond prediction to de novo design, enabling the creation of entirely new peptide sequences with desired properties.
4. Multi-modal & Controlled Generation Era: The field is now moving towards integrating multiple data modalities (sequence, structure, physicochemical properties) and developing sophisticated controlled generation techniques. This allows for fine-tuning the generative process to produce AMPs with specific activity profiles, low toxicity, and improved stability, often guided by discriminators or by navigating learned latent spaces. The use of Molecular Dynamics (MD) simulations is also becoming more integrated for higher-fidelity evaluation.

3.4. Differentiation Analysis

This paper is a review, so its innovations are in its structure and focus rather than a new AI method. Compared to other reviews, its core differentiation lies in:

• Focus on Recent Advancements: It specifically "complements the review of approaches spanning the last two years," highlighting the most current and exciting directions in AI-driven AMP discovery.
• Structured Categorization: It provides a unique and detailed characterization of discriminators (categorized by predicted properties) and generators (categorized by unconstrained vs. analogue generation), further breaking down controlled generation into specific mechanisms (discriminator-guided filtering, positive-only learning, latent space sampling, conditional generation, and optimized generation). This structured approach offers a clear conceptual framework for understanding the field.
• Emphasis on Evaluation Challenges: It dedicates significant attention to the complexities of evaluating AMP discovery, both methodologically and experimentally, proposing the AMP success rate curve as a more robust metric.
• Highlighting Gaps and Opportunities: The paper goes beyond summarizing existing work by critically identifying unaddressed challenges and future opportunities, providing a roadmap for future research, particularly regarding data scarcity, noise, and the need for more sophisticated optimization methods.

4. Methodology

The paper primarily reviews the methodologies employed in AI-driven AMP discovery rather than proposing a new one. It systematically categorizes and explains the various AI approaches used, from data representation to model types and controlled generation strategies.

4.1. Principles

The core idea behind AI-driven AMP discovery is to leverage computational intelligence to overcome the limitations of traditional, laborious, and expensive experimental methods. The theoretical basis is that patterns within known AMP sequences, structures, and associated properties (like activity and toxicity) can be learned by AI models. Once learned, these models can either:

1. Discriminate: Predict properties of new or unknown peptide sequences, thus identifying potential AMP candidates.

2. Generate: Create entirely new peptide sequences that are likely to possess desired antimicrobial properties.

   The intuition is that the vast chemical space of possible peptides is too large to explore experimentally. AI provides a powerful tool to intelligently navigate this space, prioritize promising candidates, and even design novel ones, significantly accelerating the discovery process.

4.2. Core Methodology In-depth (Layer by Layer)

The paper organizes its discussion around AMP properties, AMP representation, and two main categories of AI methods: discriminators and generators.

4.2.1. AMP Properties and Their Measurement

AMPs are characterized by several properties crucial for their clinical utility. AI models are often designed to predict or optimize these properties.

• Activity: Measured as Minimum Inhibitory Concentration (MIC) or Minimal Bacterial Concentration (MBC).
  • MIC: The lowest concentration of an antimicrobial agent that prevents visible growth of a microorganism after an overnight incubation.
  • MBC: The lowest concentration of an antimicrobial agent that results in a $\geq 99.9\%$ kill of the initial bacterial inoculum.
  • Challenges: Units (µg/mL vs. µM) vary, and protocols (bacterial concentration, medium) can introduce bias (e.g., inoculum effect [14]).
• Toxicity: Measured as hemolytic activity or cytotoxicity.
  • Hemolytic Activity: The ability of a peptide to lyse (break open) red blood cells. Measured by HC50, the peptide concentration causing $50\%$ hemolysis.
  • Cytotoxicity: Toxicity to other cell types (fibroblast, colon, lung, cancer lines). Measured by IC50 (inhibitory concentration $50\%$) or EC50 (effective concentration $50\%$), or as $50\%$ cell death.
• Stability: Resistance to degradation by enzymes (proteases) in biological environments.
• Synergy: The enhanced effect when AMPs are combined with other antimicrobials. Expressed as Fractional Inhibition Concentration (FIC) index. FIC values $\leq 0.5$ denote significant synergy.

4.2.2. AMP Representation for AI Models

The way a peptide is represented for an AI model significantly impacts the model's ability to learn and perform.

• Sequence-based Representation:
  • Amino Acid Sequence: The most prevalent representation, where a peptide is a string of amino acids (e.g., 'KWKLFKKIGAVLKVL'). This is often encoded using one-hot encoding (each amino acid is a vector with one '1' and rest '0's) or embedding vectors.
  • N and C-terminal Modifications: While important for structure and charge, this information is less frequently incorporated [33, 25].
• Derived Properties:
  • Amino Acid Composition: Frequencies of each amino acid.
  • Physicochemical Attributes: Properties like charge, hydrophobicity, isoelectric point, molecular weight.
  • Sequence Similarity: Used to relate peptides.
• Structural Details:
  • Secondary Structure: Alpha-helices, beta-sheets, random coils.
  • Molecular Fingerprints: Numerical representations capturing structural features.
  • Atom-type Connectivity: Graph-based representations of atoms and bonds.
• Language Model-derived Embeddings:
  • Pretrained language models (like BERT trained on protein sequences) convert sequences into dense numerical vectors (embeddings) that capture semantic and functional information. These have proven highly effective [44, 45, 46, 47, 48, 49, 50, 51].
• Combinations: Many methods use combinations of these encodings [40, 33, 52].

4.2.3. Data Preparation and Challenges

Most AI-driven AMP discovery methods use a supervised learning setting, requiring positive and negative datasets.

• Defining Positive Examples: Often, these are collected from AMP databases. A major challenge is the heterogeneity of positive samples (tested against different species/strains) and bias (e.g., many bacterial AMPs tested only against E. coli).
• Defining Negative Examples:
  • Common Approach: Sampling from UniProt, excluding entries with keywords like 'antimicrobial'. However, this can be biased [53].
  • Alternative: Using shuffled AMP sequences as negatives [39].
• Activity/Toxicity Thresholds: A major challenge is the lack of consensus on what constitutes an "active" or "toxic" peptide, with different studies using different thresholds (see Table 1).

  • Unit Conflicts: MIC can be in µg/mL or µM, and conversions are not always straightforward due to counterion effects [54].

  • Conflicting Database Entries: A single peptide can have varied activity/toxicity reports. Solutions include averaging, decile discretization, selecting minimal values, discarding conflicts, or retaining all measurements.

    The following are the results from Table 1 of the original paper:

    Ref	Task	Property	Positive	Negative
    		Activity	Active	Inactive
    [35]	Discriminaton		MIC <25 µg/ml	MIC >100 µg/ml
    	Discrimination		MIC <25 µg/ml	MIC >100 µg/ml
    	Generation		MIC <25 µg/ml	MIC >100 µg/ml
    E	Generation		MIC <5 µM	MIC >5 µM
    	Generation		MIC <32 µg/mL	MIC >32 µg/mL
    [28]	Generation		MIC <32 µg/mL or 10 µM	MIC >32 µg/mL or 10 µM
    [24]	Generation		log MIC <4 µM	log MIC >4 µM
    	Generation	Toxicity	Non-toxic	Toxic
    	Discriminaton		Less than 20% hemolysis at a concentration of at least 50 µM	more than 20% hemolysis at any concentration
    AEE	Discriminaton		HC50 >100 µg/ml HC50 >100 µg/ml	HC50 <100 µg/ml
    	Generation		HC50 >100 µM	HC50 <100 µg/ml
    	Discrimination		MHC ≥ 50 µM	HC50 <100 µM MHC ≤ 50 µM
    	Generation		Hemolytic/cytotoxic activities >250 µg/ml
    				Hemolytic/cytotoxic activities <200 µg/ml

This table illustrates the wide range of thresholds used to define positive (active/non-toxic) and negative (inactive/toxic) examples across different studies and tasks (discrimination vs. generation). For instance, for activity, MIC thresholds for "Active" can range from $<25 µg/ml$ to $<5 µM$, and for toxicity, "Non-toxic" can be defined as Less than 20% hemolysis at 50 µM or $HC50 >100 µg/ml$. This variability underscores the difficulty in comparing different models directly.

4.2.4. AMP Discriminators

Discriminators are AI models that predict properties of peptides.

• Basic Classification (AMP vs. non-AMP): These models classify whether a peptide is an AMP or not.
  • AMPlify [17]: Uses a bidirectional long short-term memory (LSTM) model with a multihead attention mechanism. Bidirectional LSTMs process sequences in both forward and backward directions, capturing context from both ends. Multihead attention allows the model to jointly attend to information from different representation subspaces at different positions.
  • AMP-BERT [44]: A BERT model (pretrained on protein sequences) fine-tuned for AMP prediction. BERT's bidirectional context understanding is key here.
  • VGG16-AMP [43]: Converts sequence and structure into a 3-channel image-like representation, allowing VGG16 (a Convolutional Neural Network famous for image recognition) to classify AMPs.
• Subcategory Classification: Classifying AMPs into specific types (antibacterial, antiviral, antifungal, anticancer). Challenges arise from limited training data, often leading to multi-label classification problems [49, 40].
• MIC Prediction: Predicting the exact MIC value (regression) or classifying peptides as active/inactive based on a threshold (classification).
  • Microbial Strain-Specific (MSS) Prediction [35, 46]: Predicts activity for a given peptide-bacterial strain pair, leveraging genomic information of the strain.
  • Activity Comparison [25]: Uses Siamese neural networks to predict the difference in MIC between two AMPs. A Siamese network consists of two identical subnetworks that share weights, processing two inputs separately and comparing their outputs.
• Toxicity Prediction: Classifying peptides as toxic/non-toxic or hemolytic/non-hemolytic [47, 33, 52, 23].
  • Transfer Learning for Hemolytic Activity [47]: A large language model is first trained to recognize secretory peptides and then adapted to classify hemolytic activity.
• Solubility Prediction [52, 56, 57]: Predicting how well a peptide dissolves in a solvent. Note: existing methods often focus on longer sequences and may be less accurate for AMPs.
• Secondary Structure Prediction [48]: Classifying the local structural conformation (alpha-helix, beta-sheet) of a peptide. Some methods combine pretrained language models, hypergraph multihead attention networks, and bi-LSTM with conditional random fields (CRF) [48].
• Structure Prediction [58, 24]: Predicting the 3D atomic structure. Models like AlphaFold (originally for large proteins) are being applied, but with caveats for short peptides [31].

4.2.5. AMP Generators

Generative models create new AMP candidates.

• Generation Modes:
  • Unconstrained Generation (de novo): Peptides are sampled freely from the model, creating entirely new sequences [31, 29, 28, 26, 18, 24, 21, 41, 20, 19].
  • Analogue Generation: A given peptide serves as a prototype, and the model generates similar peptides (analogues) [31, 30, 22, 42, 32, 59, 51]. This can involve single or multiple steps of modification.
• Modeling Frameworks:
  • Generative Adversarial Networks (GANs): Popular for unconstrained generation [19, 21, 45, 29]. The adversarial process allows for learning complex data distributions.
  • Variational Autoencoders (VAEs): Used for both unconstrained [20, 41, 24, 31] and analogue generation [30, 2, 31, 59, 32, 60]. Their latent space is key for controlling generation and creating analogues.
  • Recurrent Neural Networks (RNNs): Used for unconstrained generation [28].
  • Graph Neural Networks (GNNs): Used for unconstrained generation, particularly for structural properties [26].

4.2.6. Controlled Generation of AMPs

To obtain peptides with desired properties, the generation process needs to be controlled. The following figure (Figure 1a from the original paper) illustrates the overall process of AI-driven AMP discovery, including properties of interest, AI methods, and evaluation:

该图像是示意图，展示了抗菌肽（AMP）发现过程中关键属性、发现方法及评估指标，重点说明了受控AMP生成的多种策略，包括判别器引导过滤、仅正样本学习、潜在空间采样、条件生成及优化生成等。

The figure, titled "Figure 1. AI-driven AMP discovery.", provides a comprehensive overview. Panel (a) shows the main AI methods (discriminators, generators) and the properties of AMPs (activity, toxicity, stability, synergy) that are typically targeted. Panel (b) further breaks down the discriminator and generator methods. Discriminators are grouped by the property they predict (AMP, MIC, Toxicity, Structure, Solubility, Synergy). Generators are categorized by their mode (Unconstrained, Analogue). Panel (c) details the various strategies for controlled generation: discriminator-guided filtering, positive-only learning, latent space sampling, conditional generation, and optimized generation. The evaluation stage includes methodological (diversity, reconstruction, interpretability) and experimental (MIC, HC50, MD) aspects.

Strategies for controlled generation include:

1. Discriminator-Guided Filtering:
   • Principle: Generative models produce many candidates, and separate discriminator models (trained on desired properties) are used to filter out those that do not meet the criteria.
   • Mechanism: The generator creates peptides, and then a discriminator (e.g., an activity predictor) scores each generated peptide. Only peptides scoring above a certain threshold for activity and below a threshold for toxicity are kept.
   • Examples: Used in [20, 24, 45, 28]. In CLaSS [20], discriminators are trained directly in the latent space of the generative model.
2. Positive-Only Learning:
   • Principle: The generative model is trained exclusively on examples that possess the desired properties. By learning the distribution of "good" peptides, the model is expected to generate more good peptides.
   • Mechanism: Typically implemented with GANs [29, 45], where the generator learns to mimic only positive examples. Transfer learning can also be combined, where a model is pre-trained on active peptides and then fine-tuned on non-hemolytic ones [28].
   • Example: PandoraGAN [29] was trained on experimentally validated peptides with high antiviral activity.
3. Latent Space Sampling:
   • Principle: Leverages the organized structure of the latent space in models like VAEs. Points in the latent space correspond to peptides, and moving through this space allows for generating peptides with varying properties.
   • Mechanism: Identify regions in the latent space corresponding to desired properties (e.g., active, non-toxic). Sampling from these regions or manipulating latent vectors allows for targeted generation.
   • Examples:
     • PepVAE [30] samples active peptides from latent space regions distant from inactive query peptides.
     • Renaud et al. [22] introduced PCA property aligned sampling, where Principal Component Analysis (PCA) is used to identify latent space dimensions correlated with properties (e.g., hydrophobicity), enabling controlled generation along these axes.
     • Wang et al. [41] encoded sequence and structure into a discrete latent space, allowing generation of peptides with desired secondary structures.
4. Conditional Generation:
   • Principle: The generative model is explicitly conditioned on desired properties during training. This means the model learns to generate outputs given specific input conditions (e.g., "generate an active peptide," "generate a non-toxic peptide").
   • Mechanism: The model takes additional input variables that encode the desired conditions (e.g., a one-hot vector representing 'active' or 'non-toxic'). The training objective encourages the generator to produce samples that match these conditions.
   • Examples: Conditional GANs (cGANs) [19, 21] and GNNs [26] are used. HydrAMP [31] is an extended conditional VAE that can perform both unconstrained and analogue generation in a temperature-controlled setting, generating highly active analogues even from inactive prototypes.
5. Optimized Generation:
   • Principle: Aims to modify an existing peptide (or initial generated candidates) to improve specific properties. This often involves combining generative models with optimization algorithms.
   • Mechanism: Iterative refinement of peptide sequences. This can involve Bayesian optimization [42], multi-objective evolutionary algorithms [51], or gradient-based optimization within the latent space of generative models.
   • Examples:
     • Hoffman et al. [59] used a VAE with gradient descent zeroth-order optimization to convert toxic peptides into non-toxic ones while retaining antimicrobial properties.
     • Tucs et al. [32] combined a binary VAE with quantum annealing and non-dominated sorting (for Pareto optimization) to design peptides with both high activity and low toxicity.
     • Jain et al. [60] proposed an active learning algorithm using GFlowNets for generating diverse batches of active peptide candidates, leveraging epistemic uncertainty estimation.

5. Experimental Setup

As this paper is a review, it does not present a novel experimental setup of its own. Instead, it discusses the general practices and challenges associated with experimental setups in the field of AI-driven AMP discovery. This section will summarize how other papers conduct their experiments and evaluations, as reviewed by Szymczak and Szczureka.

5.1. Datasets

The datasets used in AI-driven AMP discovery are primarily derived from publicly available AMP databases, augmented with negative examples.

• Sources: Numerous databases record AMP activity, toxicity, sequence, and structure. DBAASP [16] is highlighted as a comprehensive resource with an API, providing detailed information including medium and CFU data for activity entries, and synergistic effects.
• Characteristics and Challenges:
  • Positive Examples: Often constitute the entire collection of peptides from AMP databases. This leads to heterogeneity (diverse target species/strains) and bias (e.g., many bacterial AMPs tested only against E. coli).
  • Negative Examples: Frequently sampled from UniProt by excluding keywords like "antimicrobial." However, this selection can be biased [53]. Shuffled AMP sequences are an alternative [39].
  • Activity/Toxicity Measurement Conflicts: Databases often contain conflicting entries for the same peptide against the same strain/cell type, or report measurements in inconsistent units (µg/mL vs. µM).
    • Resolution Approaches: Researchers employ various strategies to resolve conflicts: averaging per strain/genera [21, 25, 46, 36], decile discretization [21], selecting the minimal reported value, discarding conflicting entries [34], or retaining all measurements [52, 23].
  • Lack of Consensus on Thresholds: As illustrated in Table 1 (transcribed in the Methodology section), different methods assume different thresholds to define "active" or "toxic" peptides. For example, MIC thresholds for activity can vary from $<25 µg/ml$ to $<5 µM$. Toxicity thresholds (e.g., HC50) also vary widely.
  • Data Scarcity for Specific Properties: Data on crucial properties like cytotoxicity, solubility, time-to-resistance, stability, and degradation are often rare or entirely absent in databases. Clinically important MIC measurements (e.g., for carbapenem-resistant Klebsiella pneumoniae) are also underreported.

5.2. Evaluation Metrics

The evaluation of AI models in AMP discovery involves both standard machine learning metrics for discriminators and specialized metrics for generative models, culminating in experimental validation.

5.2.1. Discriminator Evaluation Metrics

• Area Under the Receiver Operating Characteristic Curve (AUROC):
  • Conceptual Definition: AUROC is a performance metric for binary classification models. It measures the ability of a classifier to distinguish between classes across all possible classification thresholds. A higher AUROC indicates better discriminatory power, meaning the model can effectively separate positive from negative cases. An AUROC of 1.0 indicates perfect classification, while 0.5 indicates performance no better than random chance.
  • Mathematical Formula: AUROC is the area under the Receiver Operating Characteristic (ROC) curve. The ROC curve plots the True Positive Rate (TPR) against the False Positive Rate (FPR) at various threshold settings.
    • $TPR = \frac{TP}{TP + FN}$ (Sensitivity or Recall)
    • $FPR = \frac{FP}{FP + TN}$ (1 - Specificity) where TP is True Positives, FN is False Negatives, FP is False Positives, and TN is True Negatives.
  • Symbol Explanation:
    • TP: Number of positive instances correctly predicted as positive.
    • FN: Number of positive instances incorrectly predicted as negative.
    • FP: Number of negative instances incorrectly predicted as positive.
    • TN: Number of negative instances correctly predicted as negative.
• Root Mean Squared Error (RMSE):
  • Conceptual Definition: RMSE is a common metric used to measure the difference between values predicted by a model or estimator and the true values. It is frequently used for regression models. RMSE represents the square root of the average of the squared errors, giving higher weight to larger errors. A lower RMSE indicates better model accuracy.
  • Mathematical Formula: $ RMSE = \sqrt{\frac{1}{n} \sum_{i=1}^{n} (y_i - \hat{y}_i)^2} $
  • Symbol Explanation:
    • $n$: The number of data points.
    • $y_i$: The observed (true) value for the $i$-th data point.
    • $\hat{y}_i$: The predicted value for the $i$-th data point.

5.2.2. Generator Evaluation Metrics

The evaluation of generative models is more complex due to the subjective nature of "good" generation.

• Diversity Metrics: Measure the variety of generated sequences and their difference from training data.
  • Levehnstein Distance [41, 28, 31]:
    • Conceptual Definition: Also known as edit distance, it quantifies the minimum number of single-character edits (insertions, deletions, substitutions) required to change one word (or peptide sequence) into the other. A higher distance indicates greater dissimilarity.
    • Note: The paper criticizes Levehnstein (and Jaccard) for ignoring amino acid order [51], which is crucial for peptides.
  • BLEU (Bilingual Evaluation Understudy) Score [18]:
    • Conceptual Definition: Originally designed to evaluate the quality of machine-translated text by comparing it to reference translations. It measures the overlap of n-grams (contiguous sequences of n items) between generated and reference sequences. A higher score indicates greater similarity to reference.
  • Jaccard Similarity [22]:
    • Conceptual Definition: Measures the similarity between finite sample sets. For two sets, it is defined as the size of their intersection divided by the size of their union. For peptide sequences, it could involve comparing sets of k-mers or unique amino acids. A score of 1 means identical sets, 0 means no overlap.
  • Pairwise Sequence Similarity Score [20, 59, 22, 24]:
    • Conceptual Definition: A general term for metrics that quantify how similar two sequences are, often based on alignment scores (e.g., BLAST, Smith-Waterman) or other comparison methods.
• Latent Space Metrics (for VAEs):
  • Reconstruction Ability: How accurately the decoder can reconstruct the original input from its latent representation.
  • Interpretability: How well specific dimensions or regions of the latent space correlate with meaningful peptide properties.
  • Organization: How well similar peptides are grouped together in the latent space.
• Property Prediction using Discriminators: Generated peptides are often evaluated by applying existing discriminators (e.g., activity/toxicity predictors) to their sequences. The predictions from these discriminators serve as proxies for the true properties [20, 24, 45, 28]. The paper notes that limited accuracy of these discriminators can bias evaluation.
• Physicochemical Property Calculation: Properties like amphipathicity, hydrophobicity, and charge are easier to compute computationally but are less specific indicators of desired AMP features.
• Molecular Dynamics (MD) Simulations:
  • Conceptual Definition: A computational method that simulates the physical movements of atoms and molecules over time, providing insights into their dynamic behavior and interactions. In AMP discovery, MD can simulate peptide stability, interaction with bacterial or host membranes, and mechanism of action.
  • Application: Used to study secondary structure stability [45], mode of action [28], and peptide-membrane interactions [19, 31, 24].
  • Challenge: Computationally expensive, especially fully atomistic simulations. Coarse-grained simulations are faster but less specific [20].
• Experimental Validation: The ultimate evaluation involves laboratory experiments.
  • Activity: Measuring MIC/MBC on target bacteria.
  • Toxicity: Measuring HC50 on human erythrocytes or cytotoxicity on cell lines.
  • AMP Success Rate Curve [31]: A proposed method to summarize experimental activity validation, displaying the fraction of active peptides below various MIC thresholds. This provides a more comprehensive view than a single activity value.
  • Further Characterization: For promising candidates: antimicrobial assays against drug-resistant strains, membrane disruption studies, time-kill assays, resistance induction, biofilm efficacy.

5.3. Baselines

The paper, being a review, doesn't propose a new method with specific baselines. Instead, it discusses the general practice of comparing new AI methods against existing ones within the field.

• Discriminators: New discriminator models are typically compared against state-of-the-art (SOTA) methods for classification (e.g., other deep learning models like LSTMs, CNNs, or transformer-based models) or regression (e.g., traditional machine learning models or other deep learning architectures). The paper notes a challenge: "there is no SOTA method that each new method would be compared to." Due to diverse training datasets and thresholds, a fair comparison often requires re-training all models on a unified dataset.
• Generators: Generative models are often compared based on:
  • The diversity of generated peptides (against previous generative approaches).
  • The quality of generated peptides (e.g., predicted activity/toxicity using discriminators as proxies).
  • Their ability to achieve controlled generation of peptides with specific desired properties, often benchmarked against random generation or simpler analogue design strategies.
  • Ultimately, experimental validation is the most robust comparison.

6. Results & Analysis

As a review paper, this document doesn't present new experimental results from the authors. Instead, it synthesizes the achievements and trends reported in recent AI-driven AMP discovery research. The "results" discussed are the capabilities and performance of the various AI methods developed by others, as summarized and critically assessed by the authors.

6.1. Core Results Analysis

The paper highlights significant progress in using AI for AMP discovery, primarily in two areas: discrimination and generation.

• Achievements in Discriminators:

  • Broad Classification: Many methods effectively classify peptides as AMPs or non-AMPs, leveraging deep learning architectures like LSTMs (AMPlify [17]), BERT (AMP-BERT [44]), and even image recognition models (VGG16-AMP [43]).
  • Specific Predictions: Advancements include MIC prediction (regression or classification) for specific genera/species/strains, with some methods (MSS prediction [35, 46]) incorporating genomic information for bacterial strains.
  • Toxicity and Other Properties: Discriminators for toxicity (hemolytic activity, cytotoxicity) [47, 33, 52, 23], solubility [52, 56, 57], and secondary structure [48] have been developed.
  • Limitations: Despite these achievements, basic AMP/non-AMP classification often lacks specificity for practical utility. The paper notes that hand-engineered descriptor-based methods can still offer competitive performance for basic classification [39]. The accuracy of discriminators remains a limiting factor when using them as proxies for evaluating generative models.

• Achievements in Generators: The emergence of generative AI, particularly GANs and VAEs, has revolutionized the design process, moving beyond prediction to creation.

  • Unconstrained Generation: Models can generate novel AMP candidates de novo [31, 29, 28, 26, 18, 24, 21, 41, 20, 19].
  • Analogue Generation: Models can create modified versions of existing peptides (analogues) [31, 30, 22, 42, 32, 59, 51].
  • Controlled Generation: This is a major advancement, allowing researchers to guide the generation process towards desired properties (e.g., high activity, low toxicity, specific structure). Techniques like discriminator-guided filtering, positive-only learning, latent space sampling, conditional generation, and optimized generation have been successfully implemented.
    • HydrAMP [31] is highlighted as a cVAE-based model capable of both unconstrained and analogue generation, even generating highly active analogues from inactive prototypes, using a temperature-controlled creativity setting.
    • Optimized generation methods, such as those combining VAEs with gradient descent (Hoffman et al. [59]) or binary VAEs with quantum annealing (Tucs et al. [32]), demonstrate the ability to refine peptides for improved properties like reduced toxicity while maintaining activity.
  • Experimental Validation: While not all generated models undergo experimental validation, several prominent works have demonstrated success in the lab [24, 31, 30]. The paper notes that only a few models report high activity (e.g., MIC $\leq 2$ µg/mL or 1 µM) for their generated peptides, comparable to market antibiotics [31, 24, 30].

6.2. Data Presentation (Tables)

The paper summarizes key generative methods in a table. The following are the results from Table 2 of the original paper:

This table provides a concise overview of 15 key generative models in AMP discovery, highlighting their generation mode (unconstrained or analogue), underlying generation framework (GAN, VAE, RNN, GNN), controlled conditions (e.g., activity, toxicity, structure), whether they use external discriminators, their approach to controlled generation, and if they feature experimental validation or Molecular Dynamics (MD) simulations.

Key observations from the table:

• Dominant Frameworks: GANs and VAEs are the most prevalent generative frameworks, with VAEs being particularly versatile for both unconstrained and analogue generation.
• Controlled Conditions: A wide range of properties are targeted for controlled generation, with activity and toxicity being the most common, followed by AMP presence, secondary structure, and sequence length. More advanced controls include microbial target and mechanism of action.
• Role of Discriminators: Many generative models (10 out of 15 listed) integrate discriminators, primarily for discriminator-guided filtering or as part of optimized generation. This underscores the complementary nature of discriminators and generators.
• Controlled Generation Approaches: All five types of controlled generation (discriminator-guided filtering, positive-only learning, latent space sampling, conditional generation, optimized generation) are represented, showing the diverse strategies employed. Conditional generation and optimized generation appear to be gaining traction for analogue design.
• Experimental Validation and MD: While many models (9 out of 15) report experimental validation, a significant portion (6 out of 15) do not, indicating a gap between computational design and real-world testing. Molecular Dynamics (MD) simulations are included in 6 methods, suggesting their increasing importance for in-silico evaluation.

6.3. Ablation Studies / Parameter Analysis

The paper, as a review, does not conduct its own ablation studies or parameter analyses. However, it implicitly suggests the importance of such analyses through its discussion of evaluation challenges and opportunities. For instance, the discussion on temperature-controlled creativity in HydrAMP [31] implies a parameter that could be tuned and analyzed. The need for ranking generated AMPs and developing dedicated ensemble methods also points to the idea that different model components and parameters would need careful evaluation. The authors mention that optimization algorithms in optimized generation (e.g., Bayesian optimization, multi-objective evolutionary algorithms) are inherently about exploring parameter spaces to find optimal peptides.

7. Conclusion & Reflections

7.1. Conclusion Summary

This review paper provides a timely and comprehensive overview of the rapidly evolving field of AI-driven antimicrobial peptide (AMP) discovery. It effectively categorizes existing AI methodologies into discriminators (for prediction and filtering) and generators (for de novo or analogue design), highlighting their respective strengths and applications. A key focus is on controlled generation techniques, which represent a significant advancement in tailoring AMPs with desired properties like high activity and low toxicity. The paper acknowledges notable achievements, such as the use of large language model embeddings for better representation and the successful experimental validation of some AI-generated AMPs with potent activity. Ultimately, the paper concludes that AI has revolutionized AMP discovery, offering promising avenues to combat the global threat of antimicrobial resistance.

7.2. Limitations & Future Work

The authors meticulously outline several significant challenges and limitations that impede the clinical translation of AI-designed AMPs, which also serve as fertile grounds for future research:

• Data Quality and Quantity:
  • Benchmarking Data: A dire need for community-wide accepted benchmarking datasets for antibacterial and hemolytic activity, with standardized preprocessing, consensus activity thresholds, and detailed recording of experimental conditions (e.g., medium, pH, salt).
  • Knowledge Gaps: Scarcity of data for crucial AMP properties like cytotoxicity, solubility, time-to-resistance, stability, and degradation.
  • Clinical Relevance: Lack of MIC measurements against clinically important drug-resistant strains (e.g., carbapenem-resistant Klebsiella pneumoniae).
  • Novel Peptide Types: Minimal data for lipopeptides, glycopeptides, peptoids, or peptides with non-proteinogenic amino acids (NPAA), hindering AI development for these promising alternatives.
  • Noise and Scarcity Modeling: Current AI methods, especially generators, often don't explicitly account for data noise and scarcity, suggesting a need for model uncertainty measures.
• Methodological Limitations:
  • Exploiting Clustering: Current models often filter out similar peptides, overlooking opportunities to learn from matched pairs of original AMPs and their analogues that differ slightly but significantly in activity/toxicity. This could inform models for AMP comparison and local organization of latent spaces.
  • Ranking Generated AMPs: A critical need for robust methods to rank the thousands of AI-generated candidates, prioritizing those with the highest chance of experimental validation. Current discriminators often don't output suitable prediction scores, and ensemble methods are underdeveloped.
  • Accelerated MD Simulations: Molecular Dynamics (MD) simulations, while valuable for evaluation, are computationally expensive. AI could accelerate these simulations or predict peptide conformation more efficiently.
  • Optimized Generation: Current optimized generation methods haven't consistently designed AMPs significantly better than those in the training set. Future models need to be simultaneously trained in generation and optimization, embrace Pareto optimization for multiple conflicting properties (activity vs. toxicity), and balance idealism-realism tradeoff (generating ideal properties while ensuring biological meaningfulness and synthesizability).
• Automation and Integration:
  • AI-driven, Accelerated, Automated Lab and Design Process: The ultimate vision is a fully automated, iterative pipeline for AI-driven design, peptide synthesis, and multi-assay experimental validation, enabling adaptive on-line improvement of both robots and AI.

7.3. Personal Insights & Critique

This review paper by Szymczak and Szczureka offers a highly structured and insightful perspective on a crucial area of research. My personal insights and critique are as follows:

• Clarity and Structure: The paper's rigorous categorization of discriminators, generators, and controlled generation approaches is exceptionally clear and valuable for anyone trying to navigate this complex field. The figure summarizing AI-driven AMP discovery (Figure 1) and the table detailing generative methods (Table 2) are excellent aids for comprehension.
• Emphasis on Data Challenges: The candid discussion about data heterogeneity, conflicting entries, unit inconsistencies, and the sheer scarcity of clinically relevant data is a major strength. It highlights that even with advanced AI, the quality and quantity of experimental input remain fundamental bottlenecks. This resonates across many AI applications in biology, emphasizing the need for FAIR (Findable, Accessible, Interoperable, Reusable) data principles and concerted data generation efforts.
• The "Idealism-Realism Tradeoff": The concept of balancing idealism (optimizing for desired properties) with realism (ensuring biological meaningfulness and synthesizability) in optimized generation is a profound observation. It addresses a common pitfall in generative AI, where models can generate chemically plausible but experimentally intractable molecules. Incorporating metrics like Milton Coupling Efficiency [52] for synthesis feasibility is a practical step towards addressing this.
• Translatability to Other Domains: The structured approach to AI-driven discovery (discrimination, generation, controlled generation, and evaluation challenges) is highly transferable. For instance, similar frameworks could be applied to:
  • Enzyme Design: Generating novel enzymes with enhanced catalytic activity or specificity.
  • Drug Discovery (beyond AMPs): Designing small molecules with improved binding affinity, pharmacokinetics, or reduced off-target effects.
  • Material Science: Designing new polymers or crystals with desired physical properties. The latent space sampling and conditional generation techniques, in particular, are generalizable principles for controlled material or molecular design.
• Potential Issues/Areas for Improvement:

  • Lack of Mechanistic Interpretability: While the paper touches on structure prediction and MD simulations, a deeper dive into how AI models can contribute to understanding the mechanism of action of generated AMPs would be beneficial. Knowing why a peptide is active or toxic can significantly guide design.

  • Ethical Considerations: As with any powerful generative AI in drug discovery, ethical considerations (e.g., unintended generation of harmful compounds, misuse of technology) are becoming increasingly relevant. Although outside the technical scope, a brief mention of responsible AI principles could be valuable in future reviews.

  • Computational Cost: The paper mentions the computational cost of MD simulations. A more explicit discussion of the computational resources (GPUs, TPUs, cloud infrastructure) required for training and deploying the sophisticated deep learning models reviewed would be helpful for practitioners, especially given the scale of pretrained language models.

  • Accessibility of Tools: While the paper reviews methods, an assessment of the accessibility and user-friendliness of these AI tools for non-AI experts (e.g., experimental biologists) could highlight a different type of translational gap – from AI research to practical application in the lab.

    In conclusion, this review is an excellent resource for both novices and experts, effectively mapping the current state and future trajectory of AI in AMP discovery, while critically identifying the hurdles that must be overcome to bring these promising therapeutics to clinical reality.