Follicular lymphoma (FL) accounts for approximately 20% of all new lymphoma cases. Increases in cytological grade are a feature of the clinical progression of this malignancy, and eventual histologic transformation (HT) to the aggressive diffuse large B-cell lymphoma (DLBCL) occurs in up to 15% of patients. Clinical or genetic features to predict the risk and timing of HT have not been comprehensively described. In this study, we analyzed whole genome sequencing data from 423 patients to compare the protein coding and non-coding mutation landscapes of untransformed FL, transformed FL and de novo DLBCL. This revealed two genetically distinct subgroups of FL which we have named DLBCL-like (dFL) and constrained FL (cFL). Each subgroup has distinguishing mutational patterns, aberrant somatic hypermutation rates, biological, and clinical characteristics. We implemented a machine-learning-derived classification approach to stratify FL patients into cFL and dFL subgroups based on their genomic features. Using separate validation cohorts, we demonstrate that cFL status, whether assigned with this full classifier or a single-gene approximation, is associated with a reduced rate of HT. This implies distinct biological features of cFL that constrain its evolution and we highlight the potential for this classification to predict HT from genetic features present at diagnosis.