Personally tailored cancer management based on knowledge banks of genomic and clinical data

Abstract

Sequencing of cancer genomes, or parts thereof, has become widespread and will soon be implemented as part of routine clinical diagnostics. However the clinical ramifications of this have not been fully assessed. Here we assess the utility of sequencing large and clinically well-annotated cancer cohorts to derive personalized predictions about treatment outcome. To this end we study a cohort of 1,540 patients with AML (acute myeloid leukemia) with genetic profiles from 111 cancer genes, cytogenetic data and diagnostic blood counts. We test existing and develop new models to compute the probability of six different clinical outcomes based on more than 100 genetic and clinical variables. The predictions derived from our knowledge bank are more detailed and outperform strata currently used in clinical practice (concordance C=72% v C=64%), and are validated on three cohorts and data from TCGA (C=70%). Our prognostic algorithm is available as an online tool (http://cancer.sanger.ac.uk/aml-multistage). A simulation of different treatment scenarios indicates that a refined risk stratification could reduce the number of bonemarrow transplants by up to 25%, while achieving the same survival. Power calculation show that the inclusion of further genes most likely has small effects on the prognostic accuracy; increasing the number of cases will further reduce the error of personalized predictions.