Review: Machine Learning Methods in Antibiotic Discovery

Background: The World Health Organization considers antibiotic resistance as one of the greatest economic and public health challenges of our time, which is why there is an urgent need for the discovery of new antibiotics. To improve the research process regarding the time and economic resources, machine learning techniques are used. These can be applied at various points in the development process and can be realized with different concepts. An overview of the most current and promising approaches is therefore of significant importance.
Objective: We aimed to provide an overview of the use of machine learning techniques in antibiotic discovery. The objective was to identify the most important methods that are useful and promising in antibiotic research. We sought to classify existing approaches by area of application.
Methods: The research paper was based on a systematic literature search including the databases ACM Digital Library, AIS EBSCOhost, IEEE Xplore Digital Library and PubMed. After defining central core terms and a resulting search string, 489 results were obtained. These were filtered and grouped by temporal currency (from december 2019 to June 2020) and relevance. ... mehrMarginal topics and under-researched methods were omitted, while promising approaches were researched in more detail through an additional backwards search. Finally, the work is based on 30 currently relevant research publications.
Results: The review identified two main areas in which machine learning techniques can be usefully applied in antibiotic discovery: virtual screening and end-to-end approaches. In the area of virtual screening, the quantitative structure-activity relationship (QSAR) methods, successful extensions of classical methods and virtual screening of molecular fragments emerge as promising. QSAR methods use classification methods (e.g., decision trees). For the extensions of classical screening methods, random forest or direct-message passing deep neural network (D-MPNN) provide useful support. In the fragment-based drug discovery approach the hunting FOX algorithm is applied. In terms of end-to-end concepts, research is currently being conducted on the two holistic concepts for prototype-based drug discovery and drug interaction, which are built on conditional diversity networks or convolutional neuronal networks.
Conclusion: The actual success of the novel models presented can only be assessed in a few years. Furthermore, the discovery of antimicrobial peptides, which offer an alternati