Personally Tailored Risk Prediction of AML Based on Comprehensive Genomic and Clinical Data

Over the past years it has emerged that acute myeloid leukemia (AML) is a disease often driven by multiple co-occurring genomic lesions. It is a great challenge to understand the logic of these mutational patterns and how the particular constellation of genomic risk factors affects a patient's outcome in conjunction with common clinical variables such as blood counts.

Here we present a novel prognostic framework based genomic sequencing data of 111 cancer genes matched with detailed diagnostic, treatment and survival data from 1,540 patients with AML enrolled in three different trials run by the German-Austrian AML Study Group (AML-HD 98A, AML-HD 98B, and AMLSG 07-04). A systematic evaluation of risk modeling strategies reveals that much of the risk determining overall survival is captured in our comprehensive panel of genomic and prognostic clinical variables. Cox proportional hazards models with random effects achieved the highest cross-validated prognostic accuracy (Harrel's concordance C=0.72), better than models with variable selection (C=0.70 for AIC and BIC), and clearly superior to the ELN risk classification (C=0.63).

It emerges that patient risk is the aggregate of many small and few large factors, such as previously established mutations in NPM1, CEBPA^-/-, FLT3^ITD and TP53; fusion genes generated by t(15;17), inv(16), and inv(3) rearrangements; and complex karyotype, del(5q) and trisomy 21. Multiple risk factors act mostly additively, with the exception of gene-gene interaction terms, including NPM1:FLT3^ITD:DNMT3A (n=93; HR=1.50; P<0.03; Wald test, Benjamini-Yekutieli adjusted) that indicate the presence of epistatic effects on outcome. We found substantial heterogeneity in the presence of risk factors with almost unique constellations for each patient. We observed that approximately 2/3 of the predicted inter-patient risk variation was related to genomic factors (balanced rearrangements, copy number changes and point mutations), the remainder being mostly attributed to diagnostic blood counts, demographic data and treatment. Hence a large share, but not all, prognostic information seems to be determined by genomic factors.

Using multistage models with random effects we have assessed differential effects of prognostic variables at different stages of therapy. These models yield detailed predictions about the probability of being alive in induction, first complete remission and after relapse, as well as the mortality during each of the three stages. Importantly, our model computes how these probabilities change depending on a patient's constellation of risk factors. The resulting personalized predictions provide a quantitative risk assessment and allow evaluating the effect of treatment decisions such as allogeneic stem cell transplant versus standard chemotherapy in first complete remission.

Our analysis shows that detailed and accurate predictions can be made based on knowledge banks of genomic and clinical data. As a proof of principle we have implemented our prediction framework into a web portal to explore risk predictions. Our method is able to impute missing variables and quantify the uncertainty due to missingness and finite training data. Power calculations show that cohorts of 10,000 patients will be needed for precise clinical decision support.