The discovery of penicillin marked a milestone in modern medicine, transforming the treatment of bacterial infections and establishing the golden era of antibiotic research [1]. However, widespread antibiotic use, both in human medicine and livestock production, has fueled the rise antibiotic resistance (AMR) [2, 3] globally. This poses a significant challenge in healthcare settings [4], particularly in Intensive Care Units (ICUs) [5], where multidrug resistant (MDR) pathogens are a major concern. Despite efforts in drug discovery, new antibiotic development has declined [6, 7], with current antibiotics insufficient to tackle AMR effectively [8], leading to high morbidity and mortality rates [4, 9].

Infectious disease management requires precise and rapid bacterial identification and antibiotic susceptibility testing for optimal patient outcomes [4, 5, 10, 11]. Recent advancements in artificial intelligence (AI) [12] have revolutionized various sectors, including healthcare, by analyzing extensive datasets, identify patterns and making predictions, eventually improving diagnostic accuracy [13, 14]. AI algorithms, particularly with machine learning (ML) capabilities, enables faster and more accurate diagnoses than conventional methods [15]. Integrating AI/ML into healthcare settings extends beyond predictive modeling to real-time monitoring, decision support systems, and drug discovery [16], facilitating proactive interventions and targeted antimicrobial stewardship [17, 18].

AI/ML is transforming drug discovery, particularly in antimicrobial peptides (AMPs) [12, 19], with potent antimicrobial properties. Computational modeling and predictive analytics accelerate AMP discovery and optimization, offering novel therapeutics against drug-resistant infections [20]. Continued advancements in AI/ML, combined with clinical expertise, hold promise in mitigating the impact of AMR and improving patient outcomes [17, 18].

In the fight against antimicrobial resistance, ML offers a variety of techniques (Fig. 1) and applications (Table 1). Through computational modeling, virtual screening, and structure-based design, researchers can pinpoint potential drug targets, screen chemical compound libraries, and optimize lead candidates for antimicrobial activity. ML algorithms, trained on vast datasets of known antimicrobial agents, predict bioactivity, pharmacokinetic properties, and safety profiles of novel drug candidates, accelerating drug development and reducing time and costs associated with traditional approaches [21,22,23].

ML methods and applications in the fight against antimicrobial resistance

Supervised learning, a common ML method, predicts antibiotic sensitivity or treatment responses by training models on labeled datasets, such as microbial genomes or patient records (Fig. 1). As an example, it was used to predict the susceptibility of Streptococcus pneumoniae to β-lactam antibiotics by correlating penicillin-binding protein (PBP) sequences with minimal inhibitory concentration (MIC) values as labeled data. Additionally, sequences from the NCBI database that lacked MIC values were used as unlabeled data. This approach helped uncover the correlation between S. pneumoniae resistance phenotypes, serotypes, and sequence types [24]. Similarly, supervised ML identified genetic traits associated with antibiotic susceptibility in Escherichia coli across different sequence types (ST). These genetic markers help understand the development and spread of STs within clonal complexes that have a high transmission probabilities [26]. In a study on the prediction of virulence factors in Streptococcus pyogenes, López-Kleine et al. were able to narrow down the list of 1,507 genes to just 12 candidates without applying subjective filters or focusing on specific biological processes. These genes represent interesting targets for further biological validation and possible drug development [27]. Recently, a novel workflow utilizing ML to test genotype–phenotype associations has been proposed to improve the collection of high quality data on the virulence phenotype of S. pyogenes in conjunction with clinical outcomes [36].

Unsupervised learning analyzes unlabeled data to uncover hidden patterns or clusters among microbial populations, aiding the understanding of resistance mechanisms [37, 38] or discovery of new resistance genes (Fig. 1). Clustering algorithms, like K-means Clustering, group bacterial isolates based on resistance profiles or genetic characteristics [39]. A recent study classified β-lactamases as resistant or wild type, revealing distinct clusters with various strain characteristics [29]. Additionally, k-means clustering can assist in identifying novel resistance clusters or outbreaks, enabling timely interventions to mitigate the spread of antimicrobial-resistant infections. It has recently been used successfully to investigate the simultaneous presence of metal resistance and antibiotic resistance in Salmonella enterica [30].

Deep learning models, including convolutional neural networks (CNNs) and recurrent neural networks (RNNs) [40], extract complex features from genetic sequences to identify resistant strains or predict resistance mechanisms (Fig. 1). For example, CNNs accurately classify microbial strains as resistant or susceptible based on genomic sequences, while RNNs predict antimicrobial susceptibility from sequential data like treatment history or microbial evolution over time. Some CNN models have made it possible to identify functional genetic variations, provide physiologically relevant explanations, and have practical applications in clinical settings [40]. For example, a CNN was used to predict the resistance of Mycobacterium tuberculosis to 13 drugs by analyzing 18 specific sites in the genome previously not associated with resistance [31]. In addition, deep learning techniques such as generative adversarial networks (GANs) have shown promising results when used in the field of antimicrobial peptides [32]. GANs can develop new antibacterial peptides by changing the probability distribution of the generated sequences. Tucs et al. [33] generated six peptide variations with one of these peptides showing potent antibacterial activity against Escherichia coli.

Reinforcement learning (RL) trains algorithms to make decisions based on trial-and-error feedback, optimizing antibiotic treatment strategies or drug combinations against resistance [41] (Fig. 1). RL approaches optimize tasks with limited knowledge about system dynamics such as evolutionary simulations of bacterial populations. In a study using E. coli as a model, each genotype in the population was associated with a particular fitness landscape of the simulated evolution. The authors demonstrated that the reduction in population fitness due to the use of drug cycles was not constrained by an increase in genome size [34]. Recently, the RL approach has been shown to be useful in providing reasonable recommendations for antibiotic treatment in sepsis that are consistent with clinical practice [35].

Each ML methods offers unique advantages in tackling AMR, from identifying genetic markers and predicting resistance patterns to optimizing treatment strategies in real-time. The choice of methods depends on the specific problem and variables involved, ultimately improving efficiency and precision in developing new antimicrobial agents [42, 43].

Antibiotic stewardship involves coordinated efforts to optimize antibiotic use, reduce unnecessary prescribing, and minimize the development of antibiotic resistance. It encompasses strategies at the institutional or healthcare system level to promote judicious antibiotic prescribing, optimize antibiotic selection and dosing, and prevent spread of multidrug-resistant organisms. While CDSS support individual clinical decisions, antibiotic stewardship programs address broader antibiotic use practices, aiming to maintain drug efficacy and combating antibiotic resistance [44]. AI systems enhance antibiotic stewardship in hospitals by monitoring antibiotic use and detecting overuse or inappropriate use. They analyze prescribing patterns, identifying anomalies [45], allowing targeted education or intervention to improve practices. They can be used to highlight situations where administered antibiotics do not match the approved first-line treatment or where broad-spectrum antibiotics are used without an explicit indication [46]. For example, a decision algorithm focused on urinary tract infections could reduce second-line antibiotics use by 67% compared to decisions made by clinicians, while decreasing inappropriate therapies by 18% [47].

AI analyzes microbiological data to detect evolving antibiotic resistance patterns, advising on antibiotic selection, dosage, duration, and de-escalation tactics based on individual patient data and antibiotic resistance trends. ML models, such as Extreme Gradient Boosting (XGB) and Light Gradient Boosting Machine (LGBM), discriminate scenarios requiring discontinuation of medication, transition of drug administration, and early or late reduction of antibiotic use, align well with clinical intuitions, resulting in improved efficiency [48]. This information helps to adapt antibiotic formularies or revise guidelines according to local resistance profiles. The integration of antibiotic stewardship into telemedicine with AI/ML technologies, transform outpatient healthcare. Telemedicine platforms, leveraging AI/ML algorithms, optimize antibiotic prescribing practices, improve clinical decision making, and ensure the judicious use of antibiotics [49]. A recent study compared telemedicine and in-person visits for acute respiratory tract infections found slightly higher guideline-concordant antibiotic management in telemedicine visits (92.5%) compared to in-person visits (90.7%). The findings suggest that with active antibiotic stewardship, telemedicine integrated into primary care can consistently deliver guideline-concordant care [50]. This integration addresses remote healthcare complexities and promotes effective antibiotic stewardship, leading to improved patient outcomes and sustainable antibiotic practices.

ML/AI approaches have emerged to address the ever-growing problem of antibiotic resistance. These technologies enable systems to analyze bacterial genomes, predict resistance, monitor epidemic patterns, and discover new antibacterial drugs or vaccines [51,52,53,54,55,56,57]. We refer to the extensive reviews by Anahtar et al. [55] or Wong et al. [23] that provide comprehensive overviews of ML application in the antimicrobial space.

Access to genome sequences and global surveillance data facilitates the prediction of antibiotic resistance based on genomic content, patient history, and infection characteristics [55]. ML excels at identifying factors contributing to resistance, such as resistance-associated genes [58], resistance-associated alleles [59], and treatment conditions [60], critical for optimizing therapies. Many studies now utilize ML algorithms to predict antibiotic resistance based on gene mutations, presence/absence of genes, and antibiotic sensitivities as data for training algorithms (Table 2). Many of these studies use rule-based or ML to predict the antibiotic resistance status of bacteria based on their treatment history, patient demographics, and genomic content. Promisingly, high prediction accuracies, often exceeding 90%, have been reported in various studies [61,62,63]. For example, Khaledi et al. [59] achieved high sensitivity (> 90%) in predicting resistance in Pseudomonas aeruginosa clinical isolates by using whole-genome sequencing (WGS) coupled with transcriptomics to identify a panel of biomarkers to make accurate predictions. Wang et al. [64] reported over 90% accuracy in predicting antibiotic resistance in Staphylococcus aureus bloodstream infections and obtained resistance predictions up to 6 h faster than traditional bacterial identification and antibiotic resistance testing methods.

Emerging technologies like matrix assisted laser desorption ionization-time of flight mass spectrometry (MALDI-TOF) mass spectrometry have further reduced the time to predict antibiotic resistance, leading to potential improvements in antibiotic stewardship and patient outcomes [80]. They also demonstrated in a retrospective analysis that by accelerating the prediction of antibiotic resistance using MALDI-TOF, 89% of their patients’ antibiotic regimens would have been changed, directly improving antibiotic stewardship, and potentially leading to measurable improvements in patient outcomes.

Moreover, ML is crucial in identifying epistatic interactions that can lead to resistance. For example, infectious diseases caused by Mycobacteria, Pseudomonas, and Staphylococcus all share a common problem: rifampicin resistance caused by missense mutations in the rpoB gene. These alterations reduce the nucleic acid affinity of the RNA polymerase complex by lowering the affinity of its constituent proteins. Portelli et al. have created a computational model that can predict whether or not a particular bacterial strain will develop resistance to the antibiotic rifampicin [81]. The discovery of these epistatic combinations or even rare mutations conferring resistance may be overlooked in current assessments, likely due to the need for prior knowledge of the genetics of resistance to be well understood for sequenced isolates to make accurate predictions. ML has been applied more frequently to predict antibiotic resistance caused by known resistance genes, or to identify genes whose role in resistance has been well characterized [59, 64, 82]. This requires a sufficiently large dataset of genomes and associated antibiotic susceptibility test results from both resistant and sensitive isolates to accurately train a de novo resistance prediction algorithm. The purpose of susceptibility testing is to determine whether antimicrobials suppress the growth of bacteria or fungi responsible for a particular disease. Combining WGS with routine antimicrobial susceptibility testing allows unprecedented mapping of genotype to phenotype. This information can then be used to train ML algorithms that can then be applied to new infections caused by the same organism to determine the most appropriate treatment. Clinically obtained isolates provide the most comprehensive examples to identify the spectrum of mutations that enable successful infection and have been used to inform ML algorithms [53, 58, 59, 62, 64, 73].

A potential disadvantage of using only clinical isolates to determine resistance is that a complex spectrum of mutations is found during infection [83,84,85,