Metastatic　propagation　is　the　leading　cause　of　death　for　most　cancers.　Prediction　and　elucidation　of　metastatic　process　is　crucial　for　the　treatment　of　cancer.　Even　though　somatic　mutations　have　been　linked　to　tumorigenesis　and　metastasis,　it　is　less　explored　whether　metastatic　events　can　be　identified　through　genomic　mutational　signatures,　which　are　concise　descriptions　of　the　mutational　processes.　Here,　we　developed　MetaWise,　a　Deep　Neural　Network　(DNN)　model,　by　applying　mutational　signatures　as　input　features　calculated　from　Whole-Exome　Sequencing　(WES)　data　of　TCGA　and　other　metastatic　cohorts.　This　model　can　accurately　classify　metastatic　tumors　from　primary　tumors　and　outperform　traditional　machine　learning　(ML)　models　and　a　deep　learning　(DL)　model,　DiaDeL.　Signatures　of　non-coding　mutations　also　have　a　major　impact　on　the　model＇s　performance.　SHapley　Additive　exPlanations　(SHAP)　and　Local　Surrogate　(LIME)　analyses　identify　several　mutational　signatures　which　are　directly　correlated　to　metastatic　spread　in　cancers,　including　APOBEC-mutagenesis,　UV-induced　signatures,　and　DNA　damage　response　deficiency　signatures.