Figure 1. Summary of the DREAM6 gene expression challenge.

( A) Training data consisted of DNA sequences for 90 yeast RP promoters whose activities were experimentally determined ^{30, 34}. DNA sequences for blinded test set of 53 promoters whose activity was hidden also experimentally determined but withheld from the challenge participants was also provided. ( B) Outline for strategy of modeling promoter activity. Each promoter was segmented into 100 bp non-overlapping windows with the full promoter regarded as a separate window. For each window, DNA sequence features were extracted and feature selection using a linear regression wrapper performed prior to machine learning. Performance of machine learning models trained on each window was determined in 5- and 10-fold cross-validations using Pearson correlation.