Biologically-informed machine learning identifies a new clinically-actionable bladder cancer subgroup characterised by NRF2 overactivity

Abstract

Muscle-invasive bladder cancer is a diverse disease where subtyping is ambiguous. Gene expression profiling followed by unsupervised machine learning (ML) has broadened our understanding of tumour biology, but has failed to provide high-confidence clinically-actionable subgroups. To focus on tissue-specific urothelial biology, we generated co-expression networks from histologically normal bladder, including multiple differentiation states and prioritising transcription factors (TFs). This strategy revealed an emergent set of 98 TFs which we used to stratify The Cancer Genome Atlas bladder cancer cohort, revealing a subdivision of basal tumours characterised by the detoxification and glutaminolysis activity of NRF2, rendering them resistant to standard bladder cancer interventions. These 20 tumours (4.9%) expressed squamous markers, were highly aggressive (15% 2-year survival), and had signatures of active PI3K, MTOR and retinoic acid signalling. Intriguingly, only half of the subgroup had activating mutations in the NRF2/KEAP1 pathway, whilst half of putative driver NFE2L2 mutations were excluded. This highlighted the importance of expression-based classification, particularly as re-analysis of NFE2L2-mutated lung cancer trial data showed only mutations consistent with our classification strategy responded to NRF2 inhibition. Our approach provides the first direct evidence that unsupervised ML can be biologically-informative in identifying clinically-actionable subgroups.