Application of Association Rule Mining Methods to Genetic Mutations

This thesis applies association rule mining methods to genetic mutation data from the 1000 Genomes Project. The aim is to compare two distinct approaches, the FP-Growth algorithm, which operates on transactional data, and the AMIE 3 algorithm, which mines rules from knowledge graphs in RDF format, in terms of processing speed, number of discovered rules, rule quality, and interpretability. To ensure a methodologically sound comparison, the same biological data were preprocessed into two different formats corresponding to the requirements of each algorithm. The analysis was performed on a region of chromosome 21 across three tasks focused on associations between mutations themselves, their functional properties, and membership in human populations. The results show that both algorithms are comparable in terms of speed, although FP-Growth allows running the analysis for a specific task, while AMIE 3 generates all rules in a single run. The number and type of discovered rules are difficult to compare directly, as each algorithm generates rules of a different type. Direct comparison of rule quality is limited by the non-equivalence of the confidence and PCA confidence metrics. In terms of interpretability, the suitability of the algorithm depends on the type of task. FP-Growth is better suited for analysing direct variant co-occurrence, while AMIE 3 is better suited for capturing named relationships between different entity types. Both algorithms agreed in their main biological findings.