Cytochrome P450 2D6 (CYP2D6) is crucial for the metabolism and bioactivation of over 20% of clinically used drugs. Its highly polymorphic nature results in varying enzyme activity levels among individuals, impacting the safety and efficacy of drugs metabolized by CYP2D6. Here, we show the use of machine learning to better understand the impact of missense mutations on assigned haplotype functional status. We collected information on missense mutations associated with several CYP2D6 alleles from the PharmVar database. We then modeled the mutants’ 3D structure using ColabFold. Finally, we employed graph-based algorithms to extract structural signatures and developed a machine-learning model to assess two classes of haplotype functional status: normal and decreased. Furthermore, we compared the accuracy of 3D structure-based analysis with sequence-based models. The best structure-based model was built using Gradient Boosting and achieved an accuracy of 92.9%, superior to the best model built using sequences (accuracy of 78.6%). Additionally, we used our model to identify the most impactful mutations for each class. Our results indicate that the use of graph-based signatures can be efficient in detecting structural differences, which can help in explaining mutations related to each haplotype’s functional status. Machine learning models can be useful for identifying individuals at risk of being poor metabolizers, helping to personalize medical treatment, and improving drug safety and efficacy.