3.1 AI in Designing Antimicrobial Peptides

AI-driven frameworks, such as explainable artificial intelligence (XAI), play a pivotal role in the rational design of antimicrobial peptides (AMPs) that combat biofilms. For example, research by Pikalyova et al. (2024) demonstrated a remarkable 100% success rate against methicillin-resistant Staphylococcus aureus (MRSA) using peptides generated through AI. Machine learning algorithms further optimize peptide properties, boosting their effectiveness while simultaneously reducing production costs (Pikalyova et al. 2024).

3.2 Advances in Biofilm Detection and Characterization

AI-powered methods, including machine learning and advanced image processing, significantly enhance the identification of biofilm-forming pathogens across diverse surfaces. These advancements, highlighted by Mishra et al. (2024), are instrumental in enabling prompt intervention. By providing faster diagnostics and facilitating tailored treatment approaches, AI addresses key challenges posed by biofilms in clinical environments (Mishra et al. 2024).

3.3 Emerging Therapeutic Innovations

Small-scale robotics offer a groundbreaking approach to treating biofilm infections by combining mechanochemical disruption with localized drug delivery systems. Additionally, AI supports the refinement of antimicrobial photo-sonodynamic therapy by predicting resistance patterns and customizing treatment regimens, as noted by Pourhajibagher et al (Pourhajibagher et al. 2024). These developments underscore the potential of AI to revolutionize treatment approaches and enhance patient outcomes.

Despite these advancements, the application of AI in biofilm-targeting therapies must be accompanied by careful consideration of its limitations and ethical implications. Ensuring that these technological innovations prioritize patient safety and adhere to clinical care standards is critical for their successful integration into healthcare practices.

Zoffman et al (Zoffmann et al. 2019). employed machine learning to prioritize compounds from the Roche library based on novelty, potency, and chemical structure for antibacterial activity testing against Gram-negative bacteria. They identified compound-induced phenotypic changes and mechanisms of action, noting similarities in phenotypic fingerprints among compounds with the same mechanism of action.

Stokes et al. proposed a deep learning approach to discover antimicrobial molecules from the Drug Repurposing Hub, identifying 99 potential candidates. They empirically tested these compounds, with halicin showing strong inhibitory activity against various pathogens. Halicin's mechanism of action involves disrupting bacterial metabolism by sequestering iron and inhibiting c-Jun N-terminal kinase (Stokes et al. 2020a). Parvaiz et al (Parvaiz et al. 2021). used machine learning to search for beta-lactamase inhibitors, identifying promising compounds from a large data set. They validated 74 compounds, with one showing potential as both a β-lactam enhancer and inhibitor. Machine learning facilitated the identification of structurally similar compounds with antibacterial activity. Liu et. al (Liu et al. 2023) used deep learning to identify a novel antibiotic, abaucin, targeting Acinetobacter baumannii, a drug-resistant bacterium often found in hospitals. Conventional methods struggled to find effective antibiotics, but by screening thousands of molecules and training a neural network with the data, they discovered Abaucin.

3.4 AI Techniques and Novel Antibiotics

One prominent example is the use of graph convolutional networks to leverage the structure of molecules to translate them into graphs. Neural networks are then used to learn from the chemical structure itself, similar to how they are used to study protein structures. This approach has been successful in creating new antimicrobial compounds with desired properties (Yang et al. 2019). Recurrent neural networks [RNNs] have also been adapted to process chemical information. RNNs have also been combined with reinforcement learning to create new drug representations. RNNs have also been applied to represent protein sequences. By training RNNs on amino acid sequences, researchers can create embedded representations of proteins that can be used to predict various properties, such as protein structure and function (Alley et al. 2019).

Researchers also have used a technique called multinomial logistic regression to create a collection of useful fragments that can be combined to form new antibiotics (Mansbach et al. 2020). Another approach involves reusing existing drugs as antibiotics. In this case, scientists use a combination of neural network models to create a new way of representing chemical compounds and then assess their potential to fight microbes (Stokes et al. 2020b). There are also methods that focus on antimicrobial peptides (AMPs), which are considered a promising source of new antibiotics. Traditional machine learning models like support vector machines [SVMs] have been used to analyze AMPs and understand how they work (Lee et al. 2018). Other techniques include extreme gradient boosting, which can predict how much of a drug is needed to inhibit bacterial growth (Yan et al. 2020), and recurrent neural networks (RNNs), which can be used to design new peptide sequences with antimicrobial activity (Nguyen et al. 2019). The focus on AMPs is due to their varied mechanisms of action (MOAs), which make it harder for bacteria to develop resistance (Nagarajan et al. 2018). Researchers are using techniques like Density-based spatial clustering of applications with noise (DBSCAN)1 to group AMPs with similar properties and then testing the most promising ones in the lab (Fjell et al. 2012).

Some researchers are even combining Machine learning (ML) with traditional lab techniques. By starting with a known AMP and its close relatives, they were able to train a model to create new AMPs that are much more effective (Vishnepolsky et al. 2019). Traditionally, researchers relied on concepts like drug-likeness, encompassing factors like absorption and toxicity (ADMET), to identify promising drug candidates (Yoshida et al. 2018). ML takes this a step further by predicting binding affinity, accelerating the selection of molecules with optimal interactions with their targets (Jia et al. 2020a). Choosing the right ML algorithm for drug-likeness prediction can be challenging, as various algorithms can perform similarly. A rigorous selection process comparing multiple algorithms like Gaussian processes and neural networks is crucial to ensure optimal performance for a specific antibiotic (D'Souza et al. 2020). Beyond drug-likeness, another major hurdle in antibiotic development is potential toxicity (De Oliveira et al. 2020). ML is being harnessed to predict these risks early on, including hemolytic activity, the ability to burst red blood cells, a major concern for drugs entering the bloodstream (Jia et al. 2020b).

3.5 AI Algorithms in Prioritizing Bacterial Targets for Drug Design

AI algorithms play a transformative role in drug discovery, particularly in prioritizing bacterial targets during drug design. These algorithms apply advanced computational techniques to analyze complex datasets and rank potential drug targets based on specific criteria. This systematic approach enhances both the efficiency and accuracy of identifying viable therapeutic targets, making it a powerful tool in addressing critical challenges such as antibiotic resistance and targeted pathogen control.

3.6 Target Prioritization Approaches

3.7 Implications for Drug Discovery

3.8 Challenges in AI-Driven Target Prioritization

While AI offers significant advantages, challenges persist in its adoption, particularly concerning reproducibility and transparency. Computational models often operate as “black boxes,” making it difficult to interpret their predictions. Ensuring the reproducibility of results across diverse datasets and enhancing the transparency of AI algorithms are critical for their acceptance in clinical applications. Addressing these issues will be key to fully integrating AI into the drug discovery pipeline.

By utilizing AI-driven methods, researchers can not only streamline the process of drug design but also develop more precise and effective antimicrobial therapies, paving the way for breakthroughs in combating antibiotic resistance and treating infectious diseases.

For peptide-based antibiotics, solubility and stability are critical factors for success. Fortunately, ML models can predict these properties from the amino acid sequence itself, allowing researchers to focus on designing more stable and effective antibiotics (Webel et al. 2020). Furthermore, ML can identify potential cleavage sites in these peptides, which can be detrimental to their stability. By predicting these sites, researchers can optimize the peptide sequence for increased stability (Hou et al. 2020).

While ML can identify drug-like molecules, the field is now incorporating new factors. One such factor is the potential damage to the gut microbiome, a crucial consideration given the link between microbiome disruption and Antimicrobial Resistance (AMR). Researchers are developing ML models to predict how candidate antibiotics might impact the microbiome, allowing for the selection of drugs with minimal collateral damage. Ideally, these drugs would be highly specific to target pathogens while leaving beneficial gut bacteria unharmed (Torres et al. 2019).

A particularly promising area lies in exploring protein space for antibiotic development. Antimicrobial host defense peptides (AMPs) have emerged as potential antibiotic templates due to their ability to target multiple cellular sites, potentially reducing the risk of AMR. Interestingly, there seems to be an inverse relationship between resistance to traditional antibiotics and susceptibility to AMPs, suggesting further potential for crossover between ML and protein informatics in AMR research (Chng et al. 2020). Current ML models for AMR prediction primarily rely on bacterial genetic and genomic data, overlooking the drug or target molecule itself. However, research is ongoing to develop more comprehensive models that incorporate both aspects (Lázár et al. 2018).

Interpretable Machine Learning (IML) also offers exciting possibilities. By helping to identify the factors that contribute to AMR at the organismal and population level, IML can guide the development of next-generation antibiotics with lower AMR risks (Deelder et al. 2019). Analyzing AMR at the molecular and population level using ML can shed light on overused mechanisms of action [MOAs] and identify promising new avenues for antibiotic development, even on a regional scale (Kavvas et al. 2020).

Generative adversarial networks (GANs) and variational autoencoders (VAEs) are prominent architectures used in this approach. GANs involve two models: a generator creating novel molecules and a discriminator attempting to distinguish real from synthetic molecules. VAEs, on the other hand, encode and reconstruct molecules, learning the underlying data properties (Kavvas et al. 2018). Generative deep learning (DL) has broader applications in chemical and protein engineering, including material design and protein folding prediction. It's increasingly being utilized in drug discovery for generating novel drug candidates within relevant chemical spaces. Various methods employing generative DL for drug design have been explored, including deep reinforcement learning, generative adversarial autoencoders, and combinations with Monte Carlo tree search (Kingma and Welling 2014). Additionally, generative recurrent neural networks [RNNs] and Simplified Molecular Input Line Entry System (SMILES) have shown promise in drug design due to their ability to handle sequential data (Segler et al. 2018).

Jukič and Bren (Jukič and Bren 2022) surveyed the latest machine learning approaches used in the discovery of new antibacterial agents and targets, covering both small molecules and antibacterial peptides. Besides, they summarized all applied machine learning approaches and available databases useful for the design of new antibacterial agents and address the current shortcomings. Muteeb et al (Muteeb et al. 2022a). screened a library of FDA-approved drugs to identify novel non-β-lactam ring-containing inhibitors of NDM-1 by applying computational as well as in vitro experimental approaches. They employed different steps of high-throughput virtual screening, molecular docking, molecular dynamics simulation, and enzyme kinetics to identify risedronate and methotrexate as the inhibitors with the most potential. Muteeb et al (Muteeb et al. 2022b). screened a library of more than 20 million compounds, available at the MCULE purchasable database. Virtual screening led to the identification of six potential inhibitors. More recently, Lluka and Stokes (Lluka and Stokes 2023) described how advancements in AI were reinvigorating the adoption of past antibiotic discovery models—namely natural product exploration and small molecule screening.

4.1 Data Mining and Analysis

AI can analyze large datasets of bacterial genomes, protein structures, and biochemical pathways to identify potential drug targets. By mining these data, AI algorithms can pinpoint vulnerabilities in pathogens that can be exploited by novel antibiotics.

4.2 Predictive Modeling

Machine learning algorithms can predict the effectiveness of potential antibiotics based on their chemical structure, pharmacological properties, and interactions with bacterial targets. This predictive modeling enables researchers to prioritize candidate compounds for further experimentation, saving time and resources.

4.3 Structure-Based Drug Design

AI algorithms can simulate the binding interactions between potential antibiotics and bacterial proteins at the atomic level. This approach, known as structure-based drug design, allows researchers to design molecules that specifically target essential bacterial proteins while minimizing off-target effects.

4.4 De Novo Drug Design

AI can generate novel antibiotic candidates from scratch using generative models and deep learning techniques. By learning from existing chemical libraries and structural databases, AI algorithms can propose entirely new molecular structures with desired antibacterial properties.

4.5 Optimization of Drug Combinations

AI can optimize combination therapy regimens by analyzing the synergistic interactions between multiple antibiotics. By considering factors such as drug pharmacokinetics, bacterial resistance mechanisms, and patient-specific variables, AI algorithms can tailor treatment regimens to maximize efficacy and minimize resistance development.

4.6 Antimicrobial Peptide Design

AI can design antimicrobial peptides (AMPs) with enhanced potency and specificity against drug-resistant bacteria. By analyzing the amino acid sequences of naturally occurring AMPs and their interactions with bacterial membranes, AI algorithms can generate novel peptide sequences with improved antibacterial activity.

4.7 Virtual Screening of Chemical Libraries

AI-driven virtual screening techniques can efficiently screen large chemical libraries to identify potential antibiotic candidates. By simulating the binding interactions between thousands of compounds and bacterial targets, AI algorithms can rapidly identify lead compounds for experimental validation.

4.8 Adaptive Learning and Evolutionary Algorithms

AI algorithms can adapt and evolve over time to keep pace with the rapidly evolving landscape of bacterial resistance. By continuously learning from new data and feedback from experimental studies, AI-driven antibiotic design platforms can iteratively improve their predictive accuracy and effectiveness.

By leveraging the capabilities of AI for antibiotic design, researchers can accelerate the discovery of novel antibiotics, overcome bacterial resistance mechanisms, and ultimately address the urgent need for effective treatments against Drug-Resistant Superbugs.