Treatment of AML is guided by cytogenetics, age, patients’ medical condition and increasingly, new molecular markers. However, no definitive algorithms and biomarkers predicting response to induction chemotherapy exist. We recently showed that distinct protein profiles of AML correlate with cytogenetics, FAB classes and treatment outcome for a class/group of patients as a whole (Tibes, ASH 2005; Kornblau, ASH 2006 and manuscript in submission). To further individualize therapy, we applied a set-based decision modeling methodology based on the Rough Set Theory (RST) to our initial proteomic dataset of 34 (phospho)-proteins for 96 samples from 73 patients (Tibes, ASH 2005). The first model was built based on partitioning discretized protein expression levels of proteins for the 96 samples into 2 outcome categories: complete remission (CR) vs. refractory to treatment; we generated reducts for all proteins and computed pseudo-cores. Several proteins (incl. phospho-NPM1, PTEN, phospho-AKT, p53, cyclin D1 etc.) were able to distinguish outcomes in terms of CR and refractoriness. A second model based on discretized clinical condition attributes (e.g., blast %, FAB, cytogenetics, age, sex, blood counts etc.) yielded the most important clinical reducts. In a third model, protein expression and clinical attributes were combined to generate decision rules to predict outcome to induction therapy for each patient. For proof-of-concept purposes, the data set was divided into a training set and a test set. Prediction accuracy for CR vs. refractory state was in the 75% range for individual patients based on their pre-treatment clinical and proteomic attributes. In conclusion, predictive models obtained with RST yield reducts corresponding to a small but significant number of proteins (6–8) and clinical parameters which can further be used to derive association rules, capable of accurately predicting a patient’s response to induction therapy. The availability of several dozens of distinct decisions rules within each group (CR vs. refractory) based on protein expression and clinical characteristics, allows an individualized approach for each patient. This preliminary study is now being expanded on an independent proteomic dataset of 52 proteins for 258 AML samples (Kornblau, ASH 2006) to be presented at the Annual Meeting. The clinical utility lies in the fact that measuring expression levels of our protein biomarker profiles can be adopted to routine clinical testing (e.g., immunohistochemistry on diagnosis marrows) and potentially help to direct patients towards standard chemotherapy, upfront treatment on a clinical trial or evaluation for early transplant depending on the chance of response and relapse. Lastly, this model can be applied to predict other clinical outcomes (relapse vs. continuous CR), as well as to derive predictive signatures of gene expression datasets.
Disclosure:Research Funding: Protein expression data was funded in part under a Leukemia Lymphoma Society Translational Research Grant # 6089.