**7. Future work: machine learning tools**

Phage characterisation based on host range analysis, studying phage host interactions, phage infection kinetics and designing phage cocktails is resourceintensive. Machine learning (ML) based tools can be developed to predict these interactions, and the application of computational biology, artificial intelligence (AI) and modelling in phage research is rapidly developing [75]. The combination of these techniques with high-throughput Next Generation Sequencing promises greater insights into phage biology alongside the development of new tools to address previously intractable problems in phage therapy [76]. Computational tools applied to phage research are based on:


A key aim of these approaches, as applied to phage therapy, is to facilitate or automate the matching of phages to target bacterial pathogens. This would revolutionize the field as it would reduce or eliminate the need for extensive host range profiling in the laboratory and would allow the rapid countering of resistance.

Homology-based methods have been used more extensively than ML so far, but more for the identification and annotation of phage DNA from metagenomic data than for phage host matching. Homology-based approaches have used genomic similarity (e.g. HostPhinder [77]), oligonucleotide frequency (e.g. VirHostMatcher [77]), and phage abundance profiling [82]. However, the success of these methods varies widely, with correct identification of the host to genus or species level only occurring between 9.5% and 75% of the time.

Phage host matching using ML has also met with varied success. Approaches include using chemical parameters of all phage and host proteins [80], or focusing on a subset of these, such as receptor binding proteins [75], which have accurately predicted phage hosts 30 to 90% of the time. Relatively few studies have used DL methods. As with the homology-based methods above, some studies have focused on the use of DL to identify and separate phage sequences from metagenomic data. DL was used by Li et al. [79] to accurately to match phage and host species 81% of the time using 27 features of phage and host proteins.

A disadvantage of ML and DL is that large datasets are required, and these are often skewed heavily towards phage which infect a small number of well-studied bacteria. For example, in one study approximately 86% of phage used in the ML model infected a single species (*M. smegmatis*). Moreover, DL methods are not

*Potential Roles for Bacteriophages in Reducing* Salmonella *from Poultry and Swine DOI: http://dx.doi.org/10.5772/intechopen.96984*

readily interpretable and regarded as 'black boxes' due to the lack of human involvement in feature selection and application. Additionally, even the best performing ML and DL models are currently unable to predict phage hosts at the strain level, which will a necessary step in real-world therapeutic applications.

Phage host matching is likely to be more useful when using phage therapy for highly diverse pathogens, such as *Salmonella* and *E. coli*, than for more homogenous bacteria such as *Staphylococcus aureus*. ML and DL have the potential to automate the process of phage selection of their predictions are shown to be reliable, and potentially in the future could help design personalised phage therapeutics for human and agricultural use.

#### **8. Conclusions**

Phages could provide a natural alternative to traditional antimicrobial therapies in pig and poultry production. Multiple intervention points exist from farm-to-fork allowing for the development of targeted phage therapeutic strategies. The promising results obtained from diverse experimental approaches demonstrate the potential of phages to reduce *Salmonella* in live animals, as well as in finished retail products. With correct stewardship, phages may well become an integrated solution in livestock production especially within the remit of controlling significant pathogens such as *Salmonella*. While some products have made it to market, current legislation needs further development prior to widespread acceptance of phage therapeutics in animals and on retail products. The next generation of phage research is set to take advantage of developments in the fields of machine-based learning and other computationally oriented approaches. Such exciting techniques may offer a more refined approach towards the application of phages for elimination of *Salmonella* from pig and poultry production.

#### **Acknowledgements**

Dr. Steve Hooton and Dr. Adriano M. Gigante are funded by the Biotechnology and Biological Sciences Research Council (BBSRC – BB/T006196/1) and we thank BBSRC for providing the funding for this work.

## **Conflict of interest**

The authors declare no conflict of interest.

Salmonella *spp. - A Global Challenge*
