Prediction of bacterial hosts of bacteriophages

Abstract

The growth of metagenomics has accelerated the discovery of new organisms. One such class of organism is viruses that infect bacteria called bacteriophages (or simply phages). The therapeutic use of phages to cure bacterial infections is termed phage therapy. With increasing drug resistance in bacterial strains, antibiotics can no longer be used for treatments. Thus phage-based therapies serve as viable alternatives to antibiotics in treating those, as mentioned earlier, drug-resistant bacterial infections. In the last few decades number of bacteria and phages have been sequenced due to advances in sequencing technology. Despite the rapid discovery of phages, their interactions with host bacteria remain unknown. Another major challenge is identifying the most appropriate phage to treat bacterial infections. Computational methods for predicting bacterial hosts of these bacteriophages thus help bridge this gap. This thesis presents a review of recent developments in clinical trials, approved medications, tools, and databases in phage therapy. We also survey the recent works relating to host prediction for bacterio- phages. Next, we present a novel Ensemble model for predicting bacterial hosts of phages. The dataset used in this study comprises 2288 unique phage-host interactions. First, we develop four individual models - BLASTP hage, BLASTHost, CRISP RP red (alignment-based) and Hybrid Model (hybrid of alignment-free and alignment-based). The alignment-based models employ BLAST-based predictions, while the alignment-free model utilizes a combination of predictions from a machine learning model and alignment scores. The individual models (BLASTP hage, BLASTHost, CRISP RP red and Hybrid Model) achieve performance of (45.2%-71.2%), (34.8%- 62.8%), (42.4%-79.5%) and (49.7%-90.6%) across five taxonomic levels from Genus to Phylum respectively on the test dataset. The final prediction model uses the four predictive methods by stacking them sequentially to create an Ensemble model. The Ensemble model outperformed all other methods with an accuracy of 61.6%, 74.4%, 80.5%, 85.7%, 91.2% across all five taxonomic levels. A disproportionate representation of bacterial hosts among train and test splits is also found. Hence, we benchmark our approach on a modified test set formed after filtering out test phage-host interactions. All hosts in the modified dataset are present in the training dataset at least at the Genus level. We find that the performance for all individual models and the Ensemble model improves on the modified set. The final Ensemble model achieves an accuracy of 67.9%, 80.6%, 85.5%, 90%, 93.5% across the five levels on the modified dataset. In order to serve the scientific community, a webserver was developed named “PhageTB,” which can predict the host of bacteriophage, phage-host interactions, and lytic phage for a bacteria. The webserver is freely accessible at https://webs.iiitd.edu.in/raghava/phagetb/index.html, and GitHub is available at https://github.com/raghavagps/phagetb. A case study was also conducted a case study using the developed tool to find phages corresponding to the six known drug-resistant bacteria belonging to the ESKAPE group.