BACKGROUND: Identifying viral sequences in mixed metagenomes containing both viral and host contigs is a critical first step in analyzing the viral component of samples. Current tools for distinguishing prokaryotic virus and host contigs primarily use gene-based similarity approaches. Such approaches can significantly limit results especially for short contigs that have few predicted proteins or lack proteins with similarity to previously known viruses. METHODS: We have developed VirFinder, the first k-mer frequency based, machine learning method for virus contig identification that entirely avoids gene-based similarity searches. VirFinder instead identifies viral sequences based on our empirical observation that viruses and hosts have discernibly different k-mer signatures. VirFinder's performance in correctly identifying viral sequences was tested by training its machine learning model on sequences from host and viral genomes sequenced before 1 January 2014 and evaluating on sequences obtained after 1 January 2014. RESULTS: VirFinder had significantly better rates of identifying true viral contigs (true positive rates (TPRs)) than VirSorter, the current state-of-the-art gene-based virus classification tool, when evaluated with either contigs subsampled from complete genomes or assembled from a simulated human gut metagenome. For example, for contigs subsampled from complete genomes, VirFinder had 78-, 2.4-, and 1.8-fold higher TPRs than VirSorter for 1, 3, and 5 kb contigs, respectively, at the same false positive rates as VirSorter (0, 0.003, and 0.006, respectively), thus VirFinder works considerably better for small contigs than VirSorter. VirFinder furthermore identified several recently sequenced virus genomes (after 1 January 2014) that VirSorter did not and that have no nucleotide similarity to previously sequenced viruses, demonstrating VirFinder's potential advantage in identifying novel viral sequences. Application of VirFinder to a set of human gut metagenomes from healthy and liver cirrhosis patients reveals higher viral diversity in healthy individuals than cirrhosis patients. We also identified contig bins containing crAssphage-like contigs with higher abundance in healthy patients and a putative Veillonella genus prophage associated with cirrhosis patients. CONCLUSIONS: This innovative k-mer based tool complements gene-based approaches and will significantly improve prokaryotic viral sequence identification, especially for metagenomic-based studies of viral ecology.