Background: Despite of routine medical history inquiry, laboratory analysis, histopathological analyses, endoscopy and imaging methods, 3-5% of metastatic cancers are diagnosed with unknown primary (CUP). It poses a considerable challenge as modern therapeutic options for patients with CUP are mainly limited to empirical chemotherapy (paclitaxel, carboplatin, etc.), leading to an overall poor prognosis. In current clinical trials (CUPISCO trial, PACET-CUP trial) targeting CUP for precision therapy, genomic data suggest that only 31% of CUP patients could be matched for at least one targeted or immune agent. The determination of primary tumor origin and genetic mutations could probably provide more therapeutic choices and help to significantly improve the outcome in CUP patients. Methods: To identify the primary lesion of CUP, we built a weighted ensemble model by Automated machine learning, including 13 models (XGBoost, RandomForestGini, ExtraTreesGini, etc.), using genetic and clinical datasets containing 22,018 patients of 28 tumor types, enrolled between 2016-2021. Our genetic and clinical datasets contained 10 features including age, tumor site, gene alteration, TMB, etc. Among them, 813 sites related to tumor were covered by gene alteration. The 5-fold Cross-Validation method was used to estimate the accuracy of our diagnosis models. Results: The models were evaluated using top-k differential diagnosis accuracy, which measured how frequently the top-k predictions capture any of the primary diagnoses in the reference standard. Our diagnostic model was then performed on 8 CUP patients. All of them were successfully confirmed their diagnosis of primary origin tumor. Among these patients, 2/8 CUP patients have begun relevant systemic therapy based on the confirmed diagnosis: 1 patient with diagnosis of esophageal squamous cell carcinoma receiving the immune combination TP Regimen (Albumin paclitaxel plus cisplatin) and achieved a 9-month PFS; another patient diagnosed as colorectal cancer achieved a 2-month PFS after treatment with Oxaliplatin-based chemotherapy regimen. Preliminary results of our study demonstrated the efficiency of this Artificial Intelligence model for the diagnosis in CUP patients. Conclusions: We developed a diagnostic model on CUP using deep learning based on genomic and clinical data and preliminarily applied in clinic. Genetic characterization of CUP patients contributes to the selection of targeted therapies under unclear diagnosis, whereas our diagnosis model could equally adjust more proper precision treatment for CUP patients. In CUP patients, the combined strategy of deep learning-predicted diagnosis and genetic mutations-based therapy could help to maximize the clinical benefit in a clinical setting and optimize the clinical development of innovative therapies.