Microorganisms/microbes are the oldest known life forms on Earth, dating back to approximately 3.42 billion years (Schopf et al., 2018). As the support system of the biosphere, these ubiquitous organisms are paramount for the survival of more complex organisms. They are involved in various intricate interactions including breakdown of biological components, food spoilage, climate change, and operation of basic metabolic cycles in plants (Omkar Khade, 2024). In addition to exercising these functions, several microorganisms have been reported as potential candidates for causing detrimental effects on other life forms. Such microbes that cause harm to the host form the class of pathogenic microorganisms. Salmonellosis, listeriosis, campylobacteriosis, yersiniosis, tuberculosis, gonorrhea, and syphilis are some of the life-threatening infections in humans caused by pathogenic microorganisms. In addition to the number of increasing infections by these microbes, another threat known as antibacterial resistance (ABR) has now taken a global turnover exhibiting the possibility of a future pandemic. A number of bacterial species have been identified as resistant to the available antibiotics that pose a threat to humanity in the near future (Ventola, 2015).

Several ecologists have come up with a broad spectrum of molecular techniques to investigate microbial communities (Davey and O’toole, 2000; Douterelo et al., 2014; Agrawal et al., 2015; Braga et al., 2016). These techniques aided not only in understanding the diversity among microbes but also in the characterization and selection of treatment strategies to overcome diseases caused by the pathogenic forms. With the extensive diversity and considerable number of known strains, characterization based on experimental techniques makes it expensive, labor-intensive, and time-consuming (Nemati et al., 2016; Qu et al., 2019). This reduces the potential for meta-analysis. Owing to the enormous amounts of data collected, microbiology has now emerged into a field with big data competencies (Falony et al., 2015; Kyrpides et al., 2016; Goodswen et al., 2021). Utilization of machine learning (ML) techniques for analysis of data has become a proven strategy in acquiring insights about microorganisms (Aida et al., 2022; Jiang et al., 2022; Munjal et al., 2022; Wu and Gadsden, 2023). Comprehensive studies on drug target prediction, drug resistance against antimicrobial drugs, prediction of disease outbreaks, and exploration of microbial–host interactions are now being carried out using ML techniques (Cazer et al., 2021; Kim and Ahn, 2021; Salim et al., 2021; Sudhakar et al., 2021; Kuang et al., 2022; Joshi et al., 2024). K-mer analysis and deep learning have been previously carried out to identify 16S short-read sequences from amplicon and shotgun data (Fiannaca et al., 2018). The tool MARVEL based on Random Forest algorithm aided in the prediction of dsDNA bacteriophage sequences from metagenomic studies (Amgarten et al., 2018).

Genetic programming, Random Forest, and logistic regression were previously used for the classification of microbes associated with bacterial vaginosis (Beck and Foster, 2014). Recently, another study showcased a new approach to analyze microbial–disease association through integration of multiple data sources (Fan et al., 2019). Similarly, a method for the diagnosis of malarial parasite(s) and for gastrointestinal parasite diagnosis was developed through binary image classification using convolutional neural network (Rajaraman et al., 2018; Mathison et al., 2020). Utilization of an support vector machine-based model for the prediction of secretory proteins from malarial parasites using amino acid compositions was another study that introduced ML in microbiology (Verma et al., 2008). Prediction of parasite load in the absence of quantitative polymerase chain reaction trained on clinical records of Leishmania infantum-infected dogs also indicated the application of ML in microbiology (Torrecilha et al., 2017). Certain studies have also investigated antimicrobial resistance (AMR) using ML-based approaches and have developed methods to classify genomes into resistant and susceptible against specific antibiotics (Drouin et al., 2016; Naidenov et al., 2019; Hyun et al., 2020). A study also presented the mapping of Acinetobacter baumannii, Streptococcus pneumoniae, Staphylococcus aureus, and Mycobacterium tuberculosis into three classes: susceptible, intermediate, and resistant (Davis et al., 2016). Similar studies by numerous research groups led to the development of methods for prediction of minimum inhibitory concentration (MIC) (Nguyen et al., 2018; Mujeeb et al., 2020; Umar et al., 2020; Valizadehaslani et al., 2020; Khan et al., 2021).

The current study employed ML-based algorithms and nucleotide sequences of pathogenic bacteria with humans as host and resistant to known antibiotics for clustering into groups of microorganisms sharing similar genomic features. Thereafter, a prediction model was developed to predict the cluster that is closest to the organism in question. Furthermore, the study includes the development of a multi-label classifier capable of predicting the antibiotic that the organism is resistant to, based on the cluster information. The clustering model was evaluated using the Silhouette coefficient, the Calinski–Harabasz index, and the Davies–Bouldin index (Caliński and Harabasz, 1974; David and Davies, 1979; Rousseeuw, 1987). The prediction models were evaluated on the basis of sensitivity, specificity, and 5-fold cross-validation (CV) accuracy. Although microbial infections involve the interplay of several molecular features, the pathogenic features corresponding to a certain pathogen remain unique to that pathogen (Voter et al., 2020; Liu et al., 2021; Parthasarathi et al., 2021). Here, the genes corresponding to the features selected in the prediction model were also identified that shed light on the biological importance of the features in distinguishing one strain from another. This study would aid in coming up with improved strategies for the segregation of pathogenic bacteria. Furthermore, based on genomic similarities and differences with other well-studied microorganisms, it may aid in assessing the severity of the disease produced by the bacterium. Furthermore, incorporating ML-based algorithms into clinical practice not only is viable, reproducible, and resilient, but also aids in the production of clinician-friendly outcomes. Overall, the computational prediction analyses directed the benefit of ML in clustering pathogenic bacterial forms, which may aid in the development of better strategies to improve treatment options.

The workflow of the study is depicted in Figure 1 .

Advances in the processing capacity, improvements in the classical data processing algorithms, and the availability of bacterial whole genome sequences (WGS) in public databases allow for a retrospective population study of many bacterial populations. Identifying patterns in the genomic sequences resulting in mosaic structures poses challenges in comprehending and visualizing the diversity and similarities within and across various bacterial strains. However, increasing interest in the quantitative techniques to predict phenotypes from genotypes beginning with bacterial WGS are becoming popular. The pathogenicity and ABR could be the key phenotypes for predicting clinical outcomes and estimating possible treatment options.

The present study emphasizes on the utilization of ML-based techniques to examine the relatedness in the different bacterial strains. Clustering analysis was performed to segregate the pathogenic forms based on their genomic similarities. Various species undergo horizontal gene transfer in the evolution process to increase their chances of survival (Burmeister, 2015). The most evident advantage of horizontal gene transfer is that a cell can acquire a beneficial gene that originated in another cell. The emergence of new beneficial genes is likely extremely rare; therefore, stealing a gene from a neighbor should be considerably faster than waiting for it to evolve independently (Vogan and Higgs, 2011). Moreover, it would also allow a cell to reclaim a gene that had been lost by another member of the population (Vogan and Higgs, 2011). Horizontal gene transfer can also acquire beneficial features that aid adaptation to new environments, such as metabolic and antibiotic resistance genes (Hall et al., 2020). This enables organisms to become interdependent, ensuring cooperation in preserving their relationship (Hall et al., 2020). This phenomenon could be visualized in the present study as the strains from Escherichia, Salmonella, Mycobacterium, and Pseudomonas did not form single clusters. The strains belonging to E. coli and S. enterica clustered together, indicating the strains within these species to share similarities. Among the 10-mers selected, 854 10-mers mapped to Salmonella strains while 876 k-mers mapped to Escherichia strains. All the 10-mers belonging to Salmonella overlapped with the k-mers from Escherichia. A total of 342 genes identified based on the selected 10-mers were common among the two organisms ( Supplementary Table 2 ). The two species are a part of the same family—Enterobacteriaceae. According to evolutionary rate estimates derived from 5S and 16S rRNA sequence analysis, Escherichia and Salmonella species diverged from a common ancestor (Bisi-Johnson et al., 2011). They are estimated to have separated from the common ancestor approximately 140 million years ago (Ochman and Wilson, 1987; Hu et al., 2010). Despite their contrasting lifestyles, there has been no significant rewiring at the level of local regulons involved (Peyman Zarrineh et al., 2014). There is notable conservation in signaling pathways and stress sensing across these phylogenetically similar species (Peyman Zarrineh et al., 2014). Moreover, a similarity of 76% to 100% between their housekeeping genes makes them evolutionarily closely related species (Sharp, 1991; Samuel et al., 2004; Hu et al., 2010). They are foodborne pathogens and create complex biofilms that contribute to their virulence, antibiotic resistance, and surface survival (Milho et al., 2019). Interspecies interactions occur in mixed biofilms, resulting in diverse consequences for each species (Milho et al., 2019).

Furthermore, two clusters, cluster 3 and cluster 6, had a mixture of Mycobacterium and Pseudomonas pathogenic strains. These two clusters comprised a small fraction of strains joining the opponent cluster. Species collaborate when they are mutually advantageous, when their interests are aligned, and when each individual improves the fitness of the other, thus encouraging the advancement of diverse, unique phenotypes or interactions (Hall et al., 2020). One of the most common forms of prokaryote cooperation is the secretion of products required to build biofilms, digest complex chemicals, and modulate a host’s immune response, among other important functions (Hall et al., 2020). Genes involved in such benefit production can be transmitted between organisms, opening up new possibilities for collaboration and adaptation (Hall et al., 2020).

Currently, there are no studies indicating an evolutionary relationship between Mycobacterium and Pseudomonas bacteria. However, there is evidence of the two pathogens interacting to co-colonize the same infection niches and create a mixed-species biofilm that enhances both their immune system and antibiotic resistance (Camus et al., 2022). Further studies are needed to understand their evolutionary and clinical phenotype implications. Another interesting finding from this study was that strains belonging to S. aureus segregated to two separate clusters, indicating the within-species diversity. In the course of evolution, it undergoes both horizontal and vertical gene transfer events that have resulted in the genetically diversified bacterial population (Furqan Awan et al., 2021). Their diversity makes them resistant towards almost all the antimicrobial drugs used (Mlynarczyk-Bonikowska et al., 2022).

The obtained clusters paved the way for the introduction of a strategy based on the Random Forest algorithm to segregate the strains into organisms sharing similar genomic features. The proposed model attempts to integrate genomic sequences of the disease-causing microbes and further cluster them into groups of pathogens sharing similar characteristics. Along with developing the model, the study also identified key 10-mers that were capable of differentiating the strains into clusters. This could significantly accelerate the processing time required to deliver the output in terms of cluster identification. The majority of 10-mers mapping to genes in each organism varied across organisms.

Of the 876 10-mers, majority of them mapped to gltB gene in Acinetobacter baumannii. It codes for glutamate synthase subunit alpha. Glutamate is one of the carbon sources that can support growth of Acinetobacter species, making glutamate synthetases an important protein in these organisms (Ren and Palmer, 2023).

Majority of 10-mers corresponding to E. coli mapped to toxB gene. toxB is a virulence gene present in the virulence plasmid of E. coli species. It functions in enhancing bacterial adhesion and in inhibiting host lymphocyte activation (Tozzoli et al., 2005).

The fas gene stood out within 10-mers in M. tuberculosis strains. Biosynthesis of fatty acids regulated by FAS-I polypeptide is crucial in the formation of mycobacterial cell wall components, specifically mycolic acids that form a protective lipid layer on the cell wall. This is required for the survival of the bacterium in the host environment (Kinsella et al., 2003; Apoorva Bhatt et al., 2007). rpoB gene was the most common gene in the 10-mers mapped to P. aeruginosa and S. aureus. The list of 10-mers mapped were different in both the species, although there were some overlaps. These two species form one of the most commonly observed clinical polymicrobial communities that lead to the emergence of antibiotic-resistant strains (Pajon et al., 2023). rpoB gene is a DNA-directed RNA polymerase, and studies have reported that mutations in this gene lead to resistance against rifampin, an antibiotic used against multidrug-resistant bacterial strains (Yee et al., 1996; Guo et al., 2021). In S. enterica-related 10-mers, majority of them mapped to ftsK. It is involved in cell division and peptidoglycan biosynthesis. Mutations in ftsK could result in increased susceptibility against β-lactams and ciprofloxacin-related tolerance (Curiao et al., 2016). The gene spr0328 was identified as the topmost gene in S. pneumoniae. It encodes for a conserved hypothetical protein with a role in cell wall surface anchorage. The protein was one of the selected candidates for a study related to vaccine testing due to its ability to raise immune response in infected patients (Olaya-Abril et al., 2013).

The mapping of 10-mers to ABR genes led to the exploration of developing an MLP-based model to predict the antibiotics that a specific strain could be resistant to due to its genomic properties. This could thus aid in tracking ABR strains in a time-efficient manner. The 10-mers identified in the present study could open up new avenues in the field of drug designing-based studies. However, the present study is based on a limited number of sequences, although the same model could be implemented to a larger bacterial cohort based on sequence availability.

Amid growing advances in whole genome sequencing and applications of ML-based techniques, the characterization of pathogenic microbial communities could become a rapid process in the near future. The current study demonstrates one such strategy in identifying bacterial strains based on the presence and absence of 10-mers in their genomes. A subset of 10-mer sequences across the strains in the present study could also act as signatures to explore the diversity through understanding their biological significance. Furthermore, the MLP model enabled the classification of strains to ABR and non-ABR strains against various antibiotics. Overall, the computational prediction analyses demonstrated the advantage of ML to uncover the ABR determinants that might facilitate the exploration of better treatment options. However, the study is a data-driven approach, and thus, outcomes of the study may appear in the form of overfitting or underfitting.

The original contributions presented in the study are included in the article/ Supplementary Material . Further inquiries can be directed to the corresponding author. Customized Python scripts used for developing the model are available through the GitHub repository via the following URL: https://github.com/js-iob/Bacterial_clustering_AMR.

KTSP: Data curation, Formal analysis, Investigation, Methodology, Resources, Software, Validation, Visualization, Writing – original draft, Writing – review & editing. KG: Data curation, Methodology, Resources, Software, Writing – review & editing. SRaj: Data curation, Methodology, Writing – review & editing. SRan: Writing – review & editing. AP: Funding acquisition, Supervision, Writing – review & editing. HS: Writing – review & editing. JS: Conceptualization, Funding acquisition, Investigation, Methodology, Supervision, Writing – review & editing.