Cancers evolve obeying Darwinian laws and therefore the evolutionary paradigm lays the ground for predictive oncology. However, the predictive power of evolutionary metrics in cancer has been seldom tested. There is a need for quantitative measurements in controlled clinical trials with long term follow-up information. This is particularly true in locally-advanced prostate cancer, which can recur more than a decade after diagnosis. Here we mapped genomic intra-tumour heterogeneity in 642 samples from 114 patients who took part in the prostate radiotherapy trials at The Royal Marsden Hospital, for which full clinical information and 12y median follow-up was available. We concomitantly assessed phenotypic (morphological) heterogeneity using deep learning in 1,923 histological sections from 250 IMRT patients (fully overlapping with the genetic set). We found that evolvability, measured as genetic divergence as well as morphological diversity, was a strong independent predictor of recurrence (respectively HR=72.06, 95% CI 2.97-1748.5, p=0.009 and HR=6.2, 95% CI 1.86-20.72, p=0.003). Combined, these two measurements together also identified a group of patients with half the median time to recurrence compared to the rest of the cohort (5.6 vs 11.5 years). We also found a small subset of MYC/FGFR1 amplified cases (4.4%) with particularly poor prognosis. The overall burden of chromosomal alterations correlated with higher Gleason score. We identified associations between 24 chromosomal arm copy number changes and Gleason score (e.g. -22q, +5p, +8q, +16p, +7p), and show that loss of chromosome 6p (encompassing the HLA locus) was correlated with markedly reduced immune cell infiltration. This study shows that combining genomics with AI-aided histopathology in clinical trials leads to the identification of novel clinical biomarkers.