Amplification of HER2 can drive the proliferation of cancer cells, and several inhibitors of HER2 have been successfully developed. Recent advances in next-generation sequencing now reveal that HER2 is subject to mutation, with over 2000 unique variants observed in human cancers. Several examples of oncogenic HER2 mutations have been described, and these primarily occur at allosteric sites outside the ATP-binding site. To identify the full spectrum of oncogenic HER2 driver mutations aside from a few well-studied mutations, we developed mutation-allostery-pharmacology (MAP), an in silico prediction algorithm based on machine learning. By applying this computational approach to 820 single-nucleotide variants, a list of 222 known and potential driver mutations was produced. Of these 222 mutations, 111 were screened by Ba/F3-retrovirus proliferation assays; 37 HER2 mutations were experimentally determined to be driver mutations, comprising 15 previously characterized and 22 newly identified oncogenic mutations. These oncogenic mutations mostly affected allosteric sites in the extracellular domain (ECD), transmembrane domain, and kinase domain of HER2, with only a single mutation in the HER2 orthosteric ATP site. Covalent homodimerization was established as a common mechanism of activation among HER2 ECD allosteric mutations, including the most prevalent HER2 mutation, S310F. Furthermore, HER2 allosteric mutants with enhanced covalent homodimerization were characterized by altered pharmacology that reduces the activity of existing anti-HER2 agents, including the monoclonal antibody trastuzumab and the tyrosine kinase inhibitor lapatinib. Overall, the MAP-scoring and functional validation analyses provided new insights into the oncogenic activity and therapeutic targeting of HER2 mutations in cancer.