This repository contains simulated Illumina read datasets for novel human virus prediction and associated metadata extracted from the Virus Host Database (https://www.genome.jp/virushostdb/). The reads are 250bp long and were simulated with Mason (https://www.seqan.de/apps/mason/) from genomes downloaded from NCBI. The training-validation-test split was done on whole viral sequences to ensure "novelty" of validation and test viruses. The training sets contain 10 million reads per class, validation sets - 1.25 million reads per class, and test sets - 1.25 million paired reads per class. The negative class sets contain reads simulated from chordate-infecting ("cho"), metazoan-infecting ("met"), eukariote-infecting ("euk") and all-nonhuman viruses. The positive class contains human-infecting viruses. The stratified dataset ("strat") contains an equal number of reads from "cho", "met but not cho", "euk but not met" and "all but not euk". Species-level datasets ("humspec", "allspec" and "chospec", with the corresponding fasta and *_species.rds files) are constructed analogously, but ensuring that all viruses of a given species were assigned to either training, val or test set. This is a stricter setting modelling a "novel viral species" scenario while reflecting within-species phenotype diversity. blast_hits.gz contains blast hits of human virome reads form Moustafa et al., 2017 (https://doi.org/10.1371/journal.ppat.1006292) blasted against our training database (see paper for details). In the second column you can find the matched label and the accession number of the matched reference. blast_labels_complete.gz contains extracted labels for all virome reads, including those without any matches. Note: one of the read headers (>3c8ac47039d32b11c8fe23f588e444e9) from Moustafa et al. is slightly corrupted with null characters. You can remove them with sed 's/\x0//g' or equivalent.