Binding/Perturbation Comparisons

This page asks how well binding predicts response: for each pair of an active binding dataset and an active perturbation dataset, how many of a regulator’s bound targets respond when the regulator is perturbed. It also compares, for the same experiment, the four promoter definitions and the two ways of scoring binding (promoter enrichment and peak calling).

Structure

The sidebar has:

  • Metric:
    • Top-N % responsive: the median, across regulators, of the percent of each regulator’s top-N bound targets that are responsive;
    • DTO % significant: from dual threshold optimization (DTO), the percent of regulators shared by the two datasets whose bound and responsive target sets overlap more than chance (empirical p < 0.01).
  • Controls for the selected metric. For Top-N: Top N (10, 25, 50, 75 or 100; default 25), Require full overlap (on by default: keep only regulator/sample pairs whose top-N list is complete after ties are resolved) and Responsiveness (Relaxed, default: |effect| > 0 and p <= 0.05; Stringent: |effect| > 0.77 and p <= 0.05, effect only for datasets with no p-value). For DTO: Perturbation Ranking (log2fc or pvalue).
  • Controls for the active tab (below).

The workspace has four tabs:

  • Compare Datasets: a binding-by-perturbation matrix of the metric, on a white (0%) to green (100%) scale. Its controls choose the Binding Method (Promoter Enrichment or Peaks) and the Promoter Set (default Kang). Clicking a row header shows the distribution across regulators for that binding dataset against every perturbation dataset, and a column header the reverse (Top-N only; DTO gives one value per pair).
  • Compare Promoter Definitions: one table per active perturbation dataset, binding datasets as rows and the checked Promoter Sets as columns.
  • Compare Analysis Methods: for one Binding Dataset with a peak-calling arm (Rossi ChIP-exo or Mahendrawada ChEC-seq), one table per perturbation dataset with promoter enrichment and peak calling as rows and the checked promoter sets as columns. The two methods are compared over the same regulators: a regulator missing from either method in a column is left out of both. Common regulators only further restricts each table to regulators present in every cell.
  • Method × Promoter Model: the pooled OLS of percent responsive on method, promoter set and assay, with regulator fixed effects and regulator-clustered standard errors, fitted when the database is built. One panel per perturbation dataset, for the selected Top N and Responsiveness.

Collapsible sections above the tabs explain the binding methods and the four promoter definitions.

All values are read from tables computed when the database is built (topn_results, dto, method_promoter_model_*).

Behaviour

Tables update as soon as a control or the dataset selection changes; a busy indicator shows while a fetch runs. If no binding or no perturbation dataset is active, the page shows “Select at least one binding and one perturbation dataset.” A combination with no data shows “—” in its cell; a tab with nothing to show says “No data for the selected combination.” (or “No data for the selected datasets.” for the Compare Datasets distribution).

Effect on other pages

None.