3-4
[00:05.30] We trained SpikeHunter with a carefully curated dataset of phage proteins from INPHARED.
[00:08.58] SpikeHunter takes in a protein sequence,
[00:13.58] generates a sequence representation using using the state-of-the-art ESM-2 protein language model,
[00:15.33] and then sends it through a three-layer neural network.
[00:21.67] This network is trained to tell us if the sequence is a tailspike protein or not.
[00:28.01] To measure how well SpikeHunter performs, we tested it using a separate set of proteins.
[00:37.11] The results showed high precision and recall values, indicating that our model is effective in distinguishing tailspike proteins from other phage proteins.
[00:46.89] By utilizing the protein language model, we can identify tailspike proteins without the need for meticulous feature selection or setting similarity thresholds.
[00:52.87] Next, we are interested in understanding the specificity of tailspike protein.
[01:00.04] To do that in silico, we needed a large dataset to match tailspike proteins with their bacterial hosts.
[01:04.43] Where did we get it? From prophages.
[01:10.83] Prophages can integrate themselves into bacterial chromosomes or exist as plasmids.
[01:20.06] By looking for prophages within sequenced bacterial genomes, we get a clear picture of which phages with tailspike proteins are linked to which bacteria.
[01:24.95] We can then check what surface glycans are encoded by these bacteria.
[01:30.06] Bacteria are often categorized using a system called serotype.
[01:43.30] The K antigen and O antigen are surface components of bacteria that help determine the serotype. K antigen is related to the bacterial capsule and O antigen is part of the lipopolysaccharide.
[01:55.37] By mapping bacterial serotypes and prophage-encoded tailspike proteins across a large set of bacterial genomes, we can estimate the likelihood of a tailspike protein recognizing specific glycans.
[02:03.84] In this example, we found that a particular cluster of tailspike proteins is only present in genomes with the O81 serotype.
[02:08.59] These proteins are distributed among phylogenetically distant clades.
[02:20.49] It is unlikely that this happened due to random chance. It is likely that this cluster of tailspike proteins specifically recognizes the glycan of the O81 serotype.
[02:28.02] With a large-scale collection of such associations, we would have more power to infer the tailspike protein's specificity.
[02:34.55] We started with nearly 800,000 bacterial genomes from the NCBI pathogen Detection database.
[02:46.35] We focused on the genomes of E. coli, P. aeruginosa, Klebsiella, Acinetobacter, and Salmonella, as they are important pathogen and we have tools to predict their serotypes.
[02:54.67] Using SpikeHunter, we identified over 230,000 tailspike proteins on prophages within our dataset.
[03:05.08] Although only 2.7% of the prophages have a predicted tailspike protein, more than one forth of the bacterial genomes have at least one prophage with a tailspike protein.
[03:11.88] We then paired the predicted serotypes with the found tailspike proteins, creating an association network.
[03:20.05] This extensive dataset covers a diverse range of host species and provides us with a wealth of tailspike proteins to work with.
[03:24.34] Already, we have come across some fascinating cases.
[03:33.16] In the example on the left, we found multiple distinct prophages in the same bacterial genome with highly similar tailspike proteins, shown in red.
[03:40.60] This shows that the tailspike protein is the only shared factor among phages infecting the same strain.
[03:51.19] In the second case, five Klebsiella prophages had distinct tailspike proteins despite having overall similar genomes, and they were found in genomes with different serotypes.
[03:58.20] This shows that the tailspike proteins could be the varying factor in these phages that infect different serotypes.
[04:05.50] Both cases demonstrate a strong correlation between the tailspike protein and specific bacterial serotypes.
