BACKGROUND: The prognosis of patients with metastatic melanomas is extremely heterogeneous. Therefore, identifying high-risk subgroups by using innovative prediction models would help to improve selection of appropriate management options. METHODS: In this study, two datasets (GSE7929 and GSE7956) of mRNA expression microarray in an animal melanoma model were normalized by frozen Robust Multi-Array Analysis and then combined by the distance-weighted discrimination method to identify time course-dependent metastasis-related gene signatures by Biometric Research Branch-ArrayTools (BRB)-ArrayTools. Then two datasets (GSE8401 and GSE19234) of clinical melanoma samples with relevant clinical and survival data were used to validate the prognosis signature. RESULTS: A novel 192-gene set that varies significantly in parallel with the increasing of metastatic potentials was identified in the animal melanoma model. Further, this gene signature was validated to correlate with poor prognosis of human metastatic melanomas but not of primary melanomas in two independent datasets. Furthermore, multivariate Cox proportional hazards regression analyses demonstrated that the prognostic value of the 192-gene set is independent of the TNM stage and has higher areas under the receiver operating characteristic curve than stage information in both validation datasets. CONCLUSION: Our findings suggest that a time course-dependent metastasis-related gene expression signature is useful in predicting survival of malignant melanomas and might be useful in informing treatment decisions for these patients.