Role of Deep Learning Applications in Antibiotic Resistance: History
View Latest Version
Please note this is an old version of this entry, which may differ significantly from the current revision.
Contributor:
Stefan Lucian Popa
,
Cristina Pop
,
Miruna Oana Dita
,
Vlad Dumitru Brata
,
Roxana Bolchis
,
Zoltan Czako
,
Mohamed Mehdi Saadani
,
Abdulrahman Ismaiel
,
Dinu Iuliu Dumitrascu
,
Simona Grad
,
Liliana David
,
Gabriel Cismaru
,
Alexandru Marius Padureanu

Antibiotic resistance (AR) is a naturally occurring phenomenon consisting mostly of acquired adaptative mechanisms. Antibiotic development faces economical and technical difficulties, with increased costs of research and marketing new antibiotics being some of the economic hardships. Technical difficulties include lack of equipment, infrastructure, and trained personnel in some regions. To overcome AR, international programs should be created and enforced globally. Such programs are currently developed with the aid of deep learning (DL)-constructed computer systems. DL is a subtype of machine learning (ML) that utilizes artificial neural networks (ANN). Every ANN consists of multitudinous interconnected neurons, grouped into layers, having similar functionality to biological neurons. They receive an input, process the data, and generate an output signal. In the beginning, the program’s generated predictions will be highly inaccurate. The parameters linking the neural networks (NN), namely the weights and biases, are thus far unrefined to form an appropriate processing pathway. However, the neural network architecture constantly evolves to form the correct pattern recognition.

  • antibiotic resistance
  • deep learning
  • drug repurposing

1. Present

Antibiotic resistance (AR) is a pressing issue in current times, with a concerning lack of awareness from the health practitioners and the general public. This frequently determines an unjustified self-prescribed usage of antibiotics, as well as an over-prescription of antibiotics by medical practitioners, with studies showing that half of the prescribed antibiotics are unnecessary [1].
Infection prevention control strategies are deficient in a number of countries, especially in developing ones. The main factors to be considered are the governing political parties, infrastructure, lack of knowledge on proper prophylactic measures, urbanization, and the general low level of education in relation to antibiotic intake [2].
Modern technologies could come to aid in impeding AR. As infrastructure develops, so do the quality and stability of the newer antibiotic formulas. Moreover, bioinformatics has become crucial in the development of in silico methods that aid researchers in studying antibiotic interactions.

1.1. Antimicrobial Peptides Testing

Ruiz Puentes et al., proposed a new deep learning (DL) model, called AMPs-Net, that successfully classified a set of 23,967 peptides from known databases as antimicrobial peptides (AMPs) or non-AMPs, assigning each AMP to its class: antibacterial, antiviral, antifungal, or antiparasitic. To test this DL model and discover new AMPs, they used a restriction enzyme to cut an Escherichia coli (E. coli) genome and generate a new peptide library. With the help of a neural network generated algorithm, they were able to identify 252 peptides out of 423,697 sequences created by the restriction enzyme by filtering these 252 peptides with antimicrobial activity they found four AMPs which were further analyzed. Molecular dynamics assays were used to describe the capability of these AMPs to penetrate membranes. In vitro evaluation of these AMPs concluded that three out of four peptides presented antimicrobial activity. RD-10, a peptide consisting of ten amino acids, showed promising bacteriostatic activity and it was considered a potential candidate for further preclinical examinations [3].
Li et al., used a DL-derived tool that predicted new AMPs derived from the Rana catesbeiana (bullfrog) genome. They tested the new acquired peptides against the World Health Organization’s (WHO) priority list. They concluded that 4 of the 16 new discovered AMPs were highly active against multiple species of bacteria from the organization’s priority list, including multi-drug-resistant (MDR) carbapenemase-producing (CPO) E. coli. The algorithm works by partitioning its training set, and the authors concluded that this novel approach is superior to other deep-learning models. It was speculated that the AMPs, with no effect on the four bacterial isolates used in this research, may be active on other microbes such as fungi or viruses [4].
Lin et al., created a DL-derived AMP predictor that used six physicochemical properties (PC6) to encode peptides based on their chemical structure. Furthermore, the team implemented this algorithm in a web application, AI4AMP, by training the DL model with more than 13,000 peptide sequences. This application was able to calculate a score, the AI4AMP score, which is corelated with the antibacterial activity; higher score represent possible candidates for more preclinical testing. The authors concluded that an ideal encoding DL-based model should consider protein features, such as the physicochemical properties of peptides. Good analysis and choice of parameters lead to a better performance of the prediction software, thus reducing the loss of information when encoding peptides for further processing [5].

1.2. Detection of Antibiotic Resistance Genes

Yu Li et al., developed a multi-task method based on DL to ease detection of AR genes (ARGs). The technology provides details about the antibiotic class for which resistance is granted, the resistance mechanism, and gene mobility (if the gene is innate or plasmid-acquired). For the beta-lactamase antibiotic family, the algorithm can tell which subclass the ARG accounts for. The model presented in this study was trained using a dataset that was compiled manually out of 7 pre-existing ones for the technology to capture the most relevant features associated with ARGs. The DL technology proposed in this study performed better than other known models in characterizing ARGs presents in the human intestinal microbiota [6].
The study conducted by Arango-Argoty et al., proposes two DL models which focus on the similarity distribution of sequences in the ARGs database. This technology addresses the limitations of the traditional “best-hit” methodologies which can have a significantly higher rate of false negative results (they classify many ARGs as not being ARGs.) The DL models were trained separately using a collection of three distinct databases (CARD, ARDB, and UNIPROT). The first DL model was focused on predicting ARGs directly by reading short DNA sequences, while the other DL model could read complete ARGs sequences and was therefore used to discover novel ARGs. This technology’s limitations are that it cannot predict AR granted by single nucleotide polymorphisms (SNPs) and it only recognizes ARGs if they are part of one of the groups used in the training of the algorithm [7].
Steiner et al., studied the evolution of antiretroviral therapies (ART) resistance by using the human immunodeficiency viruses (HIV) genome. The study uses the genome of HIV-1 and drug resistance assay results for 18 known ARTs to test three different artificial neural networks (ANN) architectures. The genotype and phenotype data were obtained from Stanford University’s HIV Drug Resistance database. All three DL methods were trained and tested in all 18 data sets for ART resistance. The study compares the performance of three ML algorithms in testing drug resistance and shows a higher classification performance of the NN algorithms [8].

1.3. Other Measures to Decrease Antibiotic Resistance

Nevertheless, prophylaxis is desired instead of treatment. Therefore, knowledge on proper hygiene is mandatory for preventing bacterial infections. Appropriate hand hygiene alongside food safety measures can reduce the number of infections and lead, in time, to a lower antibiotic usage [9]. Hospital-acquired multi-resistant bacteria pose a threat to patients and require extensive antibiotic treatments. Thus, proper disinfection and sterilization procedures represent helpful measures against nosocomial infections [10]. In addition, resistant bacteria hotspots should be located and eliminated to prevent further development of bacterial reservoirs.
The rise in urbanization levels leads to increased pollution and overcrowded areas and, together with other factors such as the individual’s susceptibility and pathogen’s virulence, these facilitate chronic respiratory tract infections. Recurrent infectious episodes lead, in time, to the evolution of multi-resistant bacteria [11].
Usage of antibiotics in the fodder fed to livestock affects humans as well. Animal excrements containing traces of antibiotics determine an environmental selection pressure. The consequence is the appearance of drug-resistant bacteria, which cause serious human infections. In addition, animals can be responsible for AR through mobile genetic elements (MGE), such as transposons. For example, the mcr-1 transposon was responsible for the development of AR in swine farms in Northeastern China. When mcr-1-positive E. coli was frequently isolated from meat in China, its prevalence in animal food attracted negative public attention [12].
However, the antibiotic-producing industry is more focused on generic drugs that resemble established, original molecules, instead of trying to develop entirely new formulas. Instead of creating innovative antibiotics with unique action mechanisms, the trend is to produce virtually identical chemical drugs [13].
Tourism is another possible factor for the emergence of AR bacteria. With expanding tourism, there is also a higher chance of people becoming vectors for bacteria. Through this extensive spread, bacteria are responsible for creating new reservoirs that can transform into endemic outbreaks [14].

2. Future Perspectives

2.1. Critical Findings Concerning Artificial Intelligence in Antibiotics Development

ML is a technology that analyzes models and builds up algorithms depending on the data which is initially introduced in the database. These computational methods can adapt and become better based on previous experience. The predictive power of an algorithm grows proportionally to the quantity and quality of the initial dataset and allows researchers to explore a vast amount of information beyond the reach of traditional approaches.
NNs are also a subset of ML which are rapidly emerging. NNs are inspired by the human brain, in which neurons obtain input information, process it, and transform it to a specific output signal. Likewise, NN can recognize patterns and can learn from examples before processing incoming data and generating predictions. Weights are assigned for each input variable, and by constantly adjusting in the processing pathway, NN are able to improve their accuracy over time, which makes them a powerful asset to AI [15]. DL algorithms are basically NN which have more than three hidden layers.

2.2. Drug Repurposing Testing

An important aspect in which AI technologies are valuable is drug repurposing testing. By combining the above-mentioned methods, already-known approved or unapproved substances can be screened for antimicrobial effects. With the help of computer-generated simulations, interactions between these substances and bacterial structures can be studied. A good example of drug repurposing is aspirin, currently being used as an antiaggregant in addition to its anti-inflammatory effects [16]. The beta-blocker propranolol is also used nowadays in the treatment of infantile hemangioma, proving that drug repurposing can also be an efficient method in antibiotic development programs [17]. Given that many new substances never reach clinical testing, drug repurposing can be very helpful and AI techniques can be beneficial by helping with scanning and filtering the large number of known drugs as well as drugs that undergo preclinical and clinical testing for other indications.

2.3. Discovery of Antibiotic Peptides

ML is also used in the discovery of antibiotic peptides [18]. Peptides with antimicrobial activity are also widely distributed in different lifeforms where they play an essential role as part of the innate immune system. The most known mammalian AMPs are cathelicidins and defensins, acting as human host defense peptides (HDPs). They are secreted in different parts of the organism such as the skin, eyes, respiratory tract, lung, and intestine. Their main role is to act fast by being part of the innate immune system and provide a broad-spectrum protection against invading pathogens [19]. Microorganisms like bacteria and fungi also produce AMPs which help them fight against each other. AMPs are simple peptides without a complex 3D structure like large proteins, however their cost of production is expensive, at about USD 100–1000 for 1 mg of AMP, and they are only produced in lab conditions for experiments [20].
The most widely known mechanism through which AMPs kill microbes is osmotic shock, which occurs either through the formation of pores, or paving as carpet on the membrane surface to weaken membrane integrity [18][21].
Unlike most conventional antibiotics, which have specific functional or structural targets, AMPs act directly on the microorganisms, often causing cell lysis, or modulate the host immunity to enhance defense against microorganisms [22]. Moreover, they act faster than conventional antibiotics [3], have a narrower active concentration window for killing [23], and do not typically damage the DNA of their targets [4][24], though there have been studies describing how AMPs inhibit critical intracellular functions by binding to DNA, RNA, or intracellular proteins [25]. As a result, they do not induce resistance to the extent that is observed with conventional antibiotics [5]. Nevertheless, if bacteria are exposed to AMPs for extended periods of time, they can and do develop resistance even to peptide-based drugs, including the last-resort and life-saving drug, colistin [4][5].

2.4. Other Applications Combined with Artificial Intelligence for Antibiotic Discovery

Lu et al., used the Raman spectroscopy combined with ML in order to depict AR status of K. pneumoniae. Raman spectroscopy is a powerful tool that requires a sample to generate a fingerprint with the chemical constituents of the sample. After training the neural network with the initial dataset consisting of 71 K. pneumoniae isolates, the algorithm was programmed to output a probability distribution across seven ARGs, two virulence genes, and the drug-resistant phenotypes (sensitive or non-sensitive), among 15 commonly used antimicrobial agents. This convolutional neural network (CNN) has proved to be better at predicting ARGs, virulence factors, and drug-resistant phenotypes compared to other past ML algorithms. It is also much faster than conventional antimicrobial-susceptibility testing (AST) methods which imply growing of bacterial colonies on agar plates in the presence of antibiotics and the study group. When compared to other culture free methods, the Raman spectroscopy combined with CNN offers a faster and easier interpretation of the samples, thus reducing the risk of antimicrobial resistance development that comes with empirical treatment [23].
Brincat and Hofmann created a text-mining system that incorporates DL algorithms to help speed up the curation process. This kind of algorithm can replace the manual curation process by screening a large amount of literature to find genes related to AR. The new text-mining algorithm was used to identify ARGs of Helicobacter pylori (H. pylori), a bacterium that infects the stomach of more than half of the world’s population. The algorithm was then tested by using the nitroimidazole antibiotic group as a case study. Results showed that out of 28 identified genes, 23 should be included in knowledge databases because they could serve as new candidates in studies regarding H. pylori resistance to antibiotics [24].
Ciloglu et al., used, for the first time, a new bacterial resistance identification technique to distinguish between MRSA and methicillin-susceptible Staphylococcus aureus (MSSA). This new method combined surface-enhanced Raman spectroscopy (SERS) with DL algorithms. Spectral data was acquired by illuminating the bacteria coated by silver nanoparticles (AgNPs) and authors anticipated spectral differences between MRSA and MSSA cell wall. The data was then processed by an autoencoder to train the DL model, represented by a CNN. Authors stated that SERS successfully identified variations between the MRSA and MSSA surface that could indicate the presence of structures related to AR, such as modified penicillin binding proteins (PBPs). In conclusion, SERS technique combined with deep-learning processes could be a valuable tool in the future that will guide clinicians’ decision making by rapid bacterial diagnoses [21].
Wang et al., developed a new DL-derived ensemble method that was able to predict the ARGs to their class. After training the program with ARGs data acquired from the COALA database (collection of all antibiotic resistance gene databases), the algorithm could identify known sequences and align them before predicting their class (e.g., SULFONAMIDE class; TETRACYCLINE class; BETA-LACTAM class). This method will save the time required for classic antimicrobial susceptibility testing (AST) and reduce the failure rate that comes with the empirical treatment of infections [26].
Jang et al., tested how accurate predictions about ARGs occurrence could be by using three different DL models. These methods were compared to determine which one predicts ARGs occurrence linked with certain changes in the environment (e.g., rainfall, tide, salinity etc.). The focus was on determining the abundance of four ARGs found in bacteria that contaminated the water in a recreational beach area in South Korea. Each model was developed based on both meteorological and aquatic variables as input data. All three models’ predictions were compared to the observed emergence of ARGs and thus “loss” values were calculated. One NN technology had an enhanced performance in detecting single ARGs, whereas another model showed superior performance in predicting multi-ARGs and allowed identification of the importance of the input variables [27].

This entry is adapted from the peer-reviewed paper 10.3390/antibiotics11111674

References

  1. CDC Newsroom. CDC. 2016. Available online: https://www.cdc.gov/media/releases/2016/p0503-unnecessary-prescriptions.html (accessed on 20 September 2022).
  2. Sengupta, S.; Barman, P.; Lo, J. Opportunities to Overcome Implementation Challenges of Infection Prevention and Control in Low-Middle Income Countries. Curr. Treat. Options Infect. Dis. 2019, 11, 267–280.
  3. Ruiz Puentes, P.; Henao, M.C.; Cifuentes, J.; Muñoz-Camargo, C.; Reyes, L.H.; Cruz, J.C.; Arbeláez, P. Rational Discovery of Antimicrobial Peptides by Means of Artificial Intelligence. Membranes 2022, 12, 708.
  4. Li, C.; Sutherland, D.; Hammond, S.A.; Yang, C.; Taho, F.; Bergman, L.; Houston, S.; Warren, R.L.; Wong, T.; Hoang, L.M.N.; et al. AMPlify: Attentive deep learning model for discovery of novel antimicrobial peptides effective against WHO priority pathogens. BMC Genom. 2022, 23, 77.
  5. Lin, T.-T.; Yang, L.-Y.; Lu, I.-H.; Cheng, W.-C.; Hsu, Z.-R.; Chen, S.-H.; Lin, C.-Y. AI4AMP: An Antimicrobial Peptide Predictor Using Physicochemical Property-Based Encoding Method and Deep Learning. mSystems 2021, 6, e00299-21.
  6. Li, Y.; Xu, Z.; Han, W.; Cao, H.; Umarov, R.; Yan, A.; Fan, M.; Chen, H.; Duarte, C.M.; Li, L.; et al. HMD-ARG: Hierarchical multi-task deep learning for annotating antibiotic resistance genes. Microbiome 2021, 9, 40.
  7. Arango-Argoty, G.; Garner, E.; Pruden, A.; Heath, L.S.; Vikesland, P.; Zhang, L. DeepARG: A deep learning approach for predicting antibiotic resistance genes from metagenomic data. Microbiome 2018, 6, 23.
  8. Steiner, M.C.; Gibson, K.M.; Crandall, K.A. Drug Resistance Prediction Using Deep Learning Techniques on HIV-1 Sequence Data. Viruses 2020, 12, 560.
  9. Mathur, P. Hand hygiene: Back to the basics of infection control. Indian J. Med. Res. 2011, 134, 611–620.
  10. Curtis, L.T. Prevention of hospital-acquired infections: Review of non-pharmacological interventions. J. Hosp. Infect. 2008, 69, 204–219.
  11. Neiderud, C.J. How urbanization affects the epidemiology of emerging infectious diseases. Infect. Ecol. Epidemiol. 2015, 5, 27060.
  12. Cheng, P.; Yang, Y.; Cao, S.; Liu, H.; Li, X.; Sun, J.; Li, F.; Ishfaq, M.; Zhang, X. Prevalence and Characteristic of Swine-Origin mcr-1-Positive Escherichia coli in Northeastern China. Front. Microbiol. 2021, 12, 712707.
  13. Iskandar, K.; Murugaiyan, J.; Halat, D.H.; El Hage, S.; Chibabhai, V.; Adukkadukkam, S.; Roques, C.; Molinier, L.; Salameh, P.; Van Dongen, M. Antibiotic Discovery and Resistance: The Chase and the Race. Antibiotics 2022, 11, 182.
  14. Chroboczek, T.; Boisset, S.; Rasigade, J.P.; Meugnier, H.; Akpaka, P.E.; Nicholson, A.; Nicolas, M.; Olive, C.; Bes, M.; Vandenesch, F.; et al. Major West Indies MRSA clones in human beings: Do they travel with their hosts? J. Travel Med. 2013, 20, 283–288.
  15. Kriegeskorte, N.; Golan, T. Neural network models and deep learning. Curr. Biol. 2019, 29, R231–R236.
  16. Corsello, S.M.; Bittker, J.A.; Liu, Z.; Gould, J.; McCarren, P.; Hirschman, J.E.; Johnston, S.E.; Vrcic, A.; Wong, B.; Khan, M.; et al. The Drug Repurposing Hub: A next-generation drug library and information resource. Nat. Med. 2017, 23, 405–408.
  17. Pandey, V.; Tiwari, P.; Gangopadhyay, A.N.; Gupta, D.K.; Sharma, S.P.; Kumar, V. Propranolol for Infantile Haemangiomas: Experience from a Tertiary Center. J. Cutan. Aesthetic Surg. 2014, 7, 37–41.
  18. Rabaan, A.A.; Alhumaid, S.; Al Mutair, A.; Garout, M.; Abulhamayel, Y.; Halwani, M.A.; Alestad, J.H.; Al Bshabshe, A.; Sulaiman, T.; AlFonaisan, M.K.; et al. Application of Artificial Intelligence in Combating High Antimicrobial Resistance Rates. Antibiotics 2022, 11, 784.
  19. Huan, Y.; Kong, Q.; Mou, H.; Yi, H. Antimicrobial Peptides: Classification, Design, Application and Research Progress in Multiple Fields. Front. Microbiol. 2020, 11, 582779.
  20. Cao, J.; de la Fuente-Nunez, C.; Ou, W.R.; Torres, T.M.D.; Pande, G.S.; Sinskey, J.A.; Lu, T.K. Yeast-Based Synthetic Biology Platform for Antimicrobial Peptide Production. ACS Synth. Biol. 2018, 7, 896–902.
  21. Ciloglu, F.U.; Caliskan, A.; Saridag, A.M.; Kilic, I.H.; Tokmakci, M.; Kahraman, M.; Aydin, O. Drug-resistant Staphylococcus aureus bacteria detection by combining surface-enhanced Raman spectroscopy (SERS) and deep learning techniques. Sci. Rep. 2021, 11, 18444.
  22. Kwon, S.; Bae, H.; Jo, J.; Yoon, S. Comprehensive ensemble in QSAR prediction for drug discovery. BMC Bioinform. 2019, 20, 521.
  23. Lu, J.; Chen, J.; Liu, C.; Zeng, Y.; Sun, Q.; Li, J.; Shen, Z.; Chen, S.; Zhang, R. Identification of antibiotic resistance and virulence-encoding factors in Klebsiella pneumoniae by Raman spectroscopy and deep learning. Microb. Biotechnol. 2022, 15, 1270–1280.
  24. Brincat, A.; Hofmann, M. Automated extraction of genes associated with antibiotic resistance from the biomedical literature. Database 2022, 2022, baab077.
  25. Singh, T.; Choudhary, P.; Singh, S. Antimicrobial Peptides: Mechanism of Action. In Insights on Antimicrobial Peptides; IntechOpen: London, UK, 2022.
  26. Wang, Z.; Li, S.; You, R.; Zhu, S.; Zhou, X.J.; Sun, F. ARG-SHINE: Improve antibiotic resistance class prediction by integrating sequence homology, functional information and deep convolutional neural network. NAR Genom. Bioinform. 2021, 3, lqab066.
  27. Jang, J.; Abbas, A.; Kim, M.; Shin, J.; Kim, Y.M.; Cho, K.H. Prediction of antibiotic-resistance genes occurrence at a recreational beach with deep learning models. Water Res. 2021, 196, 117001.
More
© Text is available under the terms and conditions of the Creative Commons Attribution (CC BY) license; additional terms may apply. By using this site, you agree to the Terms and Conditions and Privacy Policy.
This entry is offline, you can click here to edit this entry!
Video Production Service
Feedback