Table 1.

Comparison of models employed for phage identification by Seeker, VirFinder, VirSorter, DeepVirFinder, PPR-Meta and VIBRANT

Name	Model	Parameters Count	Description
Seeker	Single LSTM trained with Python or MATLAB	157 (Python model), or 212 (MATLAB model)	Segments the genome to 1K fragments, and assigns the average score assigned by the LSTM to segments. Scores above 0.5 were considered as phage prediction.
VirFinder (23)	Three logistic regression models	Each model has 10890 parameters, totaling 32,670 parameters	Searches for multiple K-mer signatures that were frequently observed in known viral sequences. Scores are between 0 and 1, and scores above 0.5 were considered as phage prediction.
DeepVirFinder (25)	Four convolutional neural network models	Each model uses 1043001 parameters, totaling 4172004 parameters	Convolutional neural networks that extract motif intensities in sequences and then used them as features for prediction.
PPR-Meta (24)	Three convolutional neural network models used for phage, plasmid and chromosome	Each model uses 564632 parameters, totaling 1693896 parameters	Convolutional neural networks for different sequence lengths, for long sequences segments the genome into 1.2 kb fragments and reports average.
VirSorter (22)	Protein similarity	Not applicable.	Predicts proteins in sequences and detects similarity to known viral proteins. Predicted phages assigned with category scores 1 or 2 were considered as phage prediction.
VIBRANT (37)	Hybrid protein similarity and multi-layer perceptron approach	The multi-layer perceptron uses 63363 parameters	First extracts protein signatures based on HMM hits and then applies multi-layer perceptron to those signatures.