Background: Nearly 10% to 20% of myasthenia gravis (MG) patients are refractory to conventional treatment for unclear reasons. The study aimed to explore the relationship between drug metabolism gene polymorphisms and refractory MG. Methods: One hundred and thirty-one MG patients (33 in the refractory group; 98 in the non-refractory group) admitted to Tongji Hospital were included in this retrospective study. Improved multiplex ligation detection reaction (iMLDR) was used to genotype 13 polymorphisms (NR3C1 rs17209237, rs9324921; FKBP5 rs1360780, rs4713904, rs9296158; HSP90AA1 rs10873531, rs2298877, rs7160651; MDR1 rs1045642, rs1128503, rs2032582; CYP3A4 rs2242480; and CYP3A5 rs776746). We applied multivariable logistic regression to investigate the association between refractory MG and nucleotide polymorphisms. Generalized multifactor dimensionality reduction (GMDR) was used to examine gene-gene interactions. Results: CC genotype of HSP90AA1 rs7160651 was associated with the increased risk of refractory MG than CT genotype [odds ratio (OR) =0.26; P=0.041] and CT + TT genotype (dominant model, OR =0.24; P=0.022). For CYP3A5 rs776746, AA genotype was associated with refractory MG compared with AG genotype (OR =0.11; P=0.017), GG genotype (OR =0.18; P=0.033), and AG + GG genotype (dominant model, OR =0.16; P=0.020). The frequency of CAT haplotype of HSP90AA1 rs10873531, rs2298877, rs7160651 was less common in refractory patients (OR =0.33; P=0.044). No significant gene-gene interactions were observed. Conclusions: HSP90AA1 rs7160651 and CYP3A5 rs776746 were significantly associated with refractory MG. Further studies are warranted to confirm the results and investigate the use of polymorphisms for treatment individualization.