Deep network modelling in neuro-oncology

Brain tumours—the ninth most common cancer in the UK—are a major cause of death and disability. The commonest type of brain tumour, glioblastoma, remains stubbornly resistant to treatment, with survival essentially unchanged for thirty years. Only one in twenty patients with glioblastoma survive beyond five years, and those that do have severely impaired quality of life. Treatment outcomes in other cancers, such as breast, lung, and blood, have seen striking improvements in recent decades: why is the brain different? One possible answer is the marked complexity of both brain tumours and the biology of the brain itself. The disease mechanisms that cause brain tumours are many and diverse, varying greatly from one patient to another, even within the same patient at different times. Treatment here must be closely personalised to be both effective and equitable across different people, yet the diversity makes determining what treatment works for any individual patient very difficult. To understand this diversity, we must first describe it in enough detail for each individual’s experience to be distinct and recognisable. Such a detailed description or rich phenotype can be considered a “fingerprint” pattern that is distinctive to each individual yet enables comparison with others. Machines able to make sense of complex patterns—such as the artificial intelligence systems now revolutionising the world—can then be deployed to identify the right treatment for the right patient, helping deliver personalised care founded on robust evidence. Crucially, there is already a wealth of detailed information collected during routine NHS brain cancer care—brain scans, tissue examinations, genetic analyses—from which such patterns may be extracted without disturbing patient care. Moreover, such clinical information can be combined with a deep understanding of the brain’s biology, drawing on what neuroscientists have cumulatively learned over the last 100 years. Though no framework to deliver it across the NHS currently exists, we will build its foundations, establish its feasibility, and pilot its application at minimal cost to the NHS. At the heart of our proposed approach is making sense of biological patterns as networks. As an analogy, the path of a brain cell from normal to cancerous may be seen as a journey through the London Underground, where the final destination of tumour growth is reached via a series of stops defining a characteristic path. Capturing the set of all possible paths—a map of the “tumour underground”—allows us to understand what is going on, how the disease process varies from one patient to another, and how this links to the underlying normal networks of the human brain. Tailoring treatments to each patient may be the only way to identify treatments that stop this tumour network from developing altogether. Success will provide a framework that enables better and more equitable individual prediction of patient outcomes, more accurate, personalised treatment selection, and a powerful means of illuminating disease mechanisms on which future treatment innovation depends.