The need for training data can impede the adoption of novel imaging\nmodalities for learning-based medical image analysis. Domain adaptation methods\npartially mitigate this problem by translating training data from a related\nsource domain to a novel target domain, but typically assume that a one-to-one\ntranslation is possible. Our work addresses the challenge of adapting to a more\ninformative target domain where multiple target samples can emerge from a\nsingle source sample. In particular we consider translating from mp-MRI to\nVERDICT, a richer MRI modality involving an optimized acquisition protocol for\ncancer characterization. We explicitly account for the inherent uncertainty of\nthis mapping and exploit it to generate multiple outputs conditioned on a\nsingle input. Our results show that this allows us to extract systematically\nbetter image representations for the target domain, when used in tandem with\nboth simple, CycleGAN-based baselines, as well as more powerful approaches that\nintegrate discriminative segmentation losses and/or residual adapters. When\ncompared to its deterministic counterparts, our approach yields substantial\nimprovements across a broad range of dataset sizes, increasingly strong\nbaselines, and evaluation measures.\n