A Quantitative Model of Transcriptional Regulation Reveals the Influence of Binding Location on Expression
We compared three models relating regulator binding to relative transcription levels in liver and 3T3-L1 cells. The first model (A) learns a binding site's expression influence based on its location relative to the TSS. The second model (B) treats all binding events equally regardless of position. The third model (C) learns a binding site's expression influence based on its level of sequence conservation. We tested each model using several different distance cutoffs to identify bound genes. The bar graph shows how the number of genes included in the analysis increases as this cutoff is increased. Each model's performance, as measured by mean-squared prediction error on held out test data, is shown as a function of distance cutoff. Error bars indicate +/− s.e.m. The model that learns influence from position significantly outperforms the other approaches.