Accurate detection of driver gene mutations is crucial for treatment planning and prognosis prediction for patients with non-small cell lung cancer. Conventional genomic testing requires high-quality tissue samples and is time- and resource-consuming, and as a result, is not available for most patients especially those in low-resource settings. Herein, we introduce DeepGEM, an advanced annotation-agnostic Deep multiple instance learning-enabled artificial intelligence method that predicts Gene Mutations from routinely acquired histology slides. We evaluated DeepGEM on a collected largest multi-center dataset in China (16 centers, 3658 patients) to date and public TCGA datasets (473 patients). DeepGEM achieved a median area under the curve (AUC) of 0.939 (excisional biopsy) and 0.883 (aspiration biopsy) in the internal set and median AUC of 0.859 (excisional biopsy) and 0.853 (aspiration biopsy) on the multi-center external set. The DeepGEM model demonstrated high performance on the TCGA dataset, indicating its robustness across diverse racial backgrounds. Moreover, the DeepGEM model trained on primary region biopsies can be generalized to biopsies from lymph node metastases and shows potential for prognostic prediction of targeted therapy. DeepGEM is interpretable and can generate single-cell level spatial gene mutation maps for input slides, which were validated to be consistent with the Immunohistochemistry (IHC) results. Collectively, DeepGEM has shown tremendous potential to be an economical, timely, and accurate gene mutation and distribution prediction tool for guiding clinical drug recommendations for lung cancer patients.