Abstract:
|
Sophisticated machine-learning algorithms, coupled with an explosion of high-throughput assays, have inspired the idea of using big data for precision personalized medicine. Yet despite considerable advances in both algorithms and assay technologies, the promise of precision genomic medicine remains largely unrealized. A key barrier is the lack of generalizability: predictors developed using data from one platform often lose accuracy in other settings or on other platforms, severely limiting translatability to the clinic. In this talk, I will describe new calibration approaches that can be applied on a single-subject basis to generate unitless measures of gene expression that are comparable across different assay technologies. I will illustrate the effectiveness of this approach for developing highly generalizable circadian biomarkers. Finally, I will discuss our error propagation approach that enables us to quantify how (im)precision in the calibration will impact the resulting prediction, which can be applied on a single-subject basis to obtain personalized prediction intervals.
|