As whole genome sequencing is becoming more accessible and affordable for clinical microbiological diagnostics, the reliability of genotypic antimicrobial resistance (AMR) prediction from sequencing data is an important issue to address. Computational AMR prediction can be performed at multiple levels. The first-level approach, such as simple AMR search relies heavily on the quality of the information fed into the database. However, AMR due to mutations are often undetected, since this is not included in the database or poorly documented. Using co-trimoxazole (trimethoprim-sulfamethoxazole) resistance in Staphylococcus aureus, we compared single-level and multi-level analysis to investigate the strengths and weaknesses of both approaches. The results revealed that a single mutation in the AMR gene on the nucleotide level may produce false positive results, which could have been detected if protein sequence analysis would have been performed. For AMR predictions based on chromosomal mutations, such as the folP gene of S. aureus, natural genetic variations should be taken into account to differentiate between variants linked to genetic lineage (MLST) and not over-estimate the potential resistant variants. Our study showed that careful analysis of the whole genome data and additional criterion such as lineage-independent mutations may be useful for identification of mutations leading to phenotypic resistance. Furthermore, the creation of reliable database for point mutations is needed to fully automatized AMR prediction.