Carbapenems are a class of last-resort antibiotics; thus, the increase in bacterial carbapenem-resistance is a serious public health threat. Acinetobacter baumannii is one of the microorganisms that can acquire carbapenem-resistance; it causes severe nosocomial infection, and is notoriously difficult to control in hospitals. Recently, a machine-learning approach was first used to analyze the genome sequences of hundreds of susceptible and resistant A. baumannii strains, including those carrying commonly acquired resistant mechanisms, to build a classifier that can predict strain resistance. A complementary approach is to explore novel genetic elements that could be associated with the antimicrobial resistance of strains, independent of known mechanisms. Therefore, we carefully selected A. baumannii strains, spanning various genotypes, from public genome databases, and conducted the first genome-wide association study (GWAS) of carbapenem resistance. We employed a recently developed method, capable of identifying any kind of genetic variation and accounting for bacterial population structure, and evaluated its effectiveness. Our study identified a surface adhesin gene that had been horizontally transferred to an ancestral branch of A. baumannii, as well as a specific region of that gene that appeared to accumulate multiple individual variations across the different branches of carbapenem-resistant A. baumannii strains.