Background: Transposable elements (TE) are mobile genetic entities present in nearly all genomes. Previous work has shown that TEs tend to have a different nucleotide composition than the host genes, either considering codon usage bias or dinucleotide frequencies. We show here how these compositional differences can be used as a tool for detection and analysis of TE sequences. Results: We compared the composition of TE sequences and host gene sequences using probabilistic models of nucleotide sequences. We used hidden Markov models (HMM), which take into account the base composition of the sequences (occurrences of words n nucleotides long, with n ranging here from 1 to 4) and the heterogeneity between coding and non-coding parts of sequences. We analyzed three sets of sequences containing class I TEs, class II TEs and genes respectively in three species: Drosophila melanogaster, Cænorhabditis elegans and Arabidopsis thaliana. Each of these sets had a distinct, homogeneous composition, enabling us to distinguish between the two classes of TE and the genes. However the particular base composition of the TEs differed in the three species studied. Conclusions: This approach can be used to detect and annotate TEs in genomic sequences and complements the current homology-based TE detection methods. Furthermore, the HMM method is able to identify the parts of a sequence in which the nucleotide composition resembles that of a coding region of a TE. This is useful for the detailed annotation of TE sequences, which may contain an ancient, highly diverged coding region that is no longer fully functional.