Melanoma recurrence risk stratification using Bayesian systems biology modeling

9089 Background: Estimating the risk of recurrence in patients with melanoma is extremely challenging. Standard of care is AJCC staging system, but the accuracy and robustness of this method is still under development. We conducted a proof of concept study exploring the use of machine-learned Bayesian Belief Networks (ml-BBNs) using a miRNA profiled cohort of melanoma patients with extended follow up to create Bayesian Biological Systems Models (BBSMs). We sought to determine if ml-BBNs could describe the biological system and if we could use the model to identify new cases with higher risk of recurrence. Methods: Our study cohort consisted of 89 patients (42 of which recurred) with a median follow up time of 118 months, that were examined for 869 miRNAs. Prior to modeling we segmented the data into training data (72 cases/80%) and testing data (17 cases/20%) at random. We recursively trained ml-BBNs on the training set, using all miRNAs. We used the directed graph structure of the ml-BBNs to identify miRNAs that consistently had more connectivity and goodness of fit as determined by Bayesian Information Criteria (BIC) scoring. MiRNAs that were in the top 50 BIC-scoring nodes across all models were selected for use to train the recurrence BBSMs. To compensate for a small number, bootstrapping was used to increase the sample to 100 records. We then compared our test set cases to our recurrence model, and used a similarity scoring algorithm to evaluate the similarity of values in each test instance to our biological models. For comparison we also trained an ml-BBN using clinical data from the same cohort. We then evaluated the scores against known recurrence outcome using Receiver Operating Characteristic (ROC) curve analysis. Results: BIC-scoring analysis selected 35 miRNAs for use in BBSM modeling. Area Under the Curve (AUC) for detection of recurrence is 0.76 in the training set and 0.62 in the testing set, while the clinical data yielded an AUC of 0.5 in this cohort. Conclusions: Our data suggest that ml-BBNs can be used to describe a biological model of melanoma recurrence. Additional data and refinement need to be made using independent datasets to be of a level that is clinically useful.