Metagenomes can be considered as mixtures of viral, bacterial, and other eukaryotic DNA sequences. Mining viral sequences from metagenomes could shed insight into virus-host relationships and expand viral databases. Current alignment-based methods are unsuitable for identifying viral sequences from metagenome sequences because most assembled metagenomic contigs are short and possess few or no predicted genes, and most metagenomic viral genes are dissimilar to known viral genes. In this study, I developed a Markov model-based method, VirMC, to identify viral sequences from metagenomic data. VirMC uses Markov chains to model sequence signatures and construct a scoring model using a likelihood test to distinguish viral and bacterial sequences. Compared with the other two state-of-the-art viral sequence-prediction methods, VirFinder and PPR-Meta, my proposed method outperformed VirFinder and had similar performance with PPR-Meta for short contigs with length less than 400 bp. VirMC outperformed VirFinder and PPR-Meta for identifying viral sequences in contaminated metagenomic samples with eukaryotic sequences. VirMC showed better performance in assembling viral-genome sequences from metagenomic data (based on filtering potential bacterial reads). Applying VirMC to human gut metagenomes from healthy subjects and patients with type-2 diabetes (T2D) revealed that viral contigs could help classify healthy and diseased statuses. This alignment-free method complements gene-based alignment approaches and will significantly improve the precision of viral sequence identification.