Methylation risk scores are derived from epigenome-wide association studies and can be used to assess the propensity for complex diseases such as schizophrenia in test subjects, and to measure the overall influence of methylation on the schizophrenia phenotype. Methods for creating methylation risk scores are lagging behind those for polygenic risk scores, and advances have been made using machine learning for polygenic risk scores that have so far not been tested for methylation risk scores. This study aimed to adapt the polygenic risk score creation R package LDpred2 for methylation data by incorporating topologically associating domains as a prior. The resulting method, named “TADpred”, was trained on summary statistics derived from public data (N=2015) and tested on individual-level data from schizophrenia cases and healthy controls (N=1227). TADpred outperformed the previous best-performing pruning and thresholding method, with the pseudo-R2 values being 0.036 and 0.021 for TADpred and pruning and thresholding, respectively, when regressing the scores against the known schizophrenia case-control phenotypes. Pathway and enrichment analysis of the 2269 most important biological features in the TADpred model did not reveal any known connection to schizophrenia. TADpred shows that it is feasible to adapt methods from polygenic risk scores for methylation, despite differences between genotype and methylation data, and that topologically associating domains are suitable as a prior in statistical methods dealing with methylation.