Multiple sclerosis (MS) is subdivided into four phenotypes on the basis of medical history and clinical symptoms. These phenotypes are defined retrospectively and lack clear pathobiological underpinning. Since Magnetic Resonance Imaging (MRI) better reflects disease pathology than clinical symptoms, we aimed to explore MRI-driven subtypes of MS based on pathological changes visible on MRI using unsupervised machine learning. In separate train and external validation sets we looked at a total of 21,170 patient-years of data from 15 randomised controlled trials and three observational cohorts to explore MRI-driven subtypes and test whether these subtypes had differential clinical outcomes. We processed MRI data to obtain measures of brain volumes, lesion volumes, and normal appearing white matter T1/T2. We identified three MRI-driven subtypes who were similar in how they accumulated MRI abnormality. Based on the earliest abnormalities suggested by our model they were called: cortex-led, normal appearing white matter-led, and lesion-led subtypes. In the external validation datasets, the lesion-led subtype showed a faster disability progression and higher disease activity than the cortex-led subtype. In all datasets, MRI-driven subtypes were associated with disability progression (βSubtype=0.04, p=0.02; βStage=-0.06, p<0.001), whilst clinical phenotypes and baseline disability were not. Only the lesion-led subtype showed a significant treatment response in three progressive multiple sclerosis randomised controlled trials (-66%, p=0.009) and in three relapsing remitting multiple sclerosis trials (-89%, p=0.04). Our results show that MRI-driven subtyping using machine learning can prospectively enrich clinical trials with patients who are most likely to respond to treatments.