PhD Projects » ESR 10: Computational prediction of virus-host interactions in the microbiome

Dan Liu

My name is Dan Liu and I come from China, I studied software engineering during my bachelor’s in Wuhan, and during my master’s degree in computer science, I joined a computational biology group at CCNU and started to focus on bioinformatics and machine learning. My research interests mainly regarding developing computational algorithms to solve biology problems, and my master’s dissertation was based on logistic matrix factorization with neighborhood regularization and the heterogeneous network to predict virus-host interactions. I am so excited to join CVR to continue my studies in virus-host interactions.

During my Ph.D., I will study virology and work with Prof. David Robertson to figure out virus-host interactions at species level. Our aim is to develop computational approaches to predict virus and host interactions in microbiomes. In addition, I will be working on quantifying the co-phylogenetic signal for testing the preferential host-switching model in bacteria and determining the extent of co-speciation versus host-switching.

In my free time, I am into sports like yoga and badminton, and I also like hiking, traveling and photography.

Field:
Bioinformatics
Host institution:
University of Glasgow (UG), United Kingdom
Local supervisor:
Prof. Dr. David Robertson (UG, Centre for Virus Research)
Local co-supervisor:
Prof. Dr. José R Penadés (UG)
Project partner:
Xue Peng (ESR 5)
Work packages:
WP 1.2 Host prediction
WP 1.3 Virus-host interactions

Project description

Virus-host interactions are highly specific as viruses (including bacteriophages/phages) have co-evolved with the systems they infect and depend on for replication. As a consequence, phylogenetic relatedness and shared genomic properties of host species are a strong predictor of host infectivity. However, our knowledge of host relationships remains sparse and is unknown for many phages, particularly for metagenomically assembled data. In this project, existing resources and genomic signals will be leveraged to infer missing/unobserved and probable hosts, e.g., where infection has not been observed due to host immunity. Machine learning models will be trained on this data to predict phage-bacteria interactions at the species and strain levels.

Specific objectives are:

• Construction of bipartite networks of observed virus-host species links based on databases such as Virus-Host DB and MVP, text-mined literature data from company partner BioRelate, and data from ESR 13. Combined this will comprise many 10,000s of interactions from observed phage-bacteria interactions, and inferred from prophage and CRISPR- Cas associated viral sequences embedded in host genomes.
• Quantification of the co-phylogenetic signal using the software Jane to test the preferential host-switching model in bacteria and to determine the extent of co-speciation versus host-switching. This will give information on how stable phage-host interactions are at the species level.
• Application of combined SVM models to virus-host prediction by using different representations of the virus genome information: nucleotide sequence, amino acid sequence, amino acid properties and protein domains.
• Analysis of virus-host links at the strain level. This objective will use the bacteria Staphylococcus aureus (the area of speciality of José Penades) as a model system to study strain level phage-host interaction prediction in the context of CRISPR-Cas immunity and host pan-genomes.

@article{nokey,

title = {Prediction of virus-host association using protein language models and multiple instance learning},

author = {Dan Liu and Francesca Young and David L. Robertson and Ke Yuan},

doi = {10.1101/2023.04.07.536023},

year  = {2023},

date = {2023-04-08},

urldate = {2023-04-08},

journal = {bioRxiv},

abstract = {Predicting virus-host association is essential to understand how viruses interact with host species, and discovering new therapeutics for viral diseases across humans and animals. Currently, the host of the majority of viruses is unknown. Here, we introduce EvoMIL, a deep learning method that predicts virus-host association at the species level from viral sequence only. The method combines a pre-trained large protein language model and attention-based multiple instance learning to allow protein-orientated predictions. Our results show that protein embeddings capture stronger predictive signals than traditional handcrafted features, including amino acids and DNA k-mers, and physio-chemical properties. EvoMIL binary classifiers achieve AUC values of over 0.95 for all prokaryotic and nearly 0.8 for almost all eukaryotic hosts. In multi-host prediction tasks, EvoMIL achieved median performance improvements of 8.6% in prokaryotic hosts and 1.8% in eukaryotic hosts. Furthermore, EvoMIL estimates the importance of single proteins in the prediction and maps them to an embedding landscape of all viral proteins, where proteins with similar functions are distinctly clustered together.},

keywords = {},

pubstate = {published},

tppubtype = {article}

}

@misc{nokey,

title = {Computational prediction of virus-host interactions by pre-trained transformer and Multiple Instance Learning},

author = {Dan Liu},

url = {https://cbc.dcs.gla.ac.uk/},

year  = {2022},

date = {2022-09-01},

urldate = {2022-09-01},

abstract = {Viruses exist in the environment extensively and live in different hosts including plants, animals, eukaryotes, and bacteria and generate offspring depending on the replication mechanisms of the hosts. Detecting virus-host interactions is essential for us to understand interaction mechanisms between virus and host and explore clues regarding diseases caused by viruses. The virus mimics the host's patterns for escaping the immune response, so viral contain predictive signals for host prediction. Here, we proposed a computational strategy to predict hosts for viruses, which extracted representative vectors of viral protein sequences through a pre-trained transformer language model and then passed protein features to attention-based Multiple instance learning for host prediction and obtaining each weighted protein for viruses. Protein weight is helpful for understanding essential proteins contributing to virus-host interactions, and it is explainable for prediction signals of host prediction at the species level. Our results demonstrated that transformer embedding is potential for virus-host interaction prediction and is able to explain essential virus proteins leading to host infection.

},

howpublished = {Oral presentation at Computational Biology Conference 2022, Glasgow, UK},

keywords = {},

pubstate = {published},

tppubtype = {presentation}

}