Propensity for subsequent distant metastasis in head and neck squamous-cell carcinoma (HNSCC) was analysed using 186 primary tumours from patients initially treated by surgery that developed (M) or did not develop (NM) metastases as the first recurrent event. Transcriptome (Affymetrix HGU133_Plus2, QRT–PCR) and array-comparative genomic hybridization data were collected. Non-supervised hierarchical clustering based on Affymetrix data distinguished tumours differing in pathological differentiation, and identified associated functional changes. Propensity for metastasis was not associated with these subgroups. Using QRT–PCR data we identified a four-gene model (PSMD10, HSD17B12, FLOT2 and KRT17) that predicts M/NM status with 77% success in a separate 79-sample validation group of HNSCC samples. This prediction is independent of clinical criteria (age, lymph node status, stage, differentiation and localization). The most significantly altered transcripts in M versus NM were significantly associated to metastasis-related functions, including adhesion, mobility and cell survival. Several genomic modifications were significantly associated with M/NM status (most notably gains at 4q11–22 and Xq12–28; losses at 11q14–24 and 17q11 losses) and partly linked to transcription modifications. This work yields a basis for the development of prognostic molecular signatures, markers and therapeutic targets for HNSCC metastasis.