Régine2026

How it works

Régine (REGINE, Resampling-based Engine for Gene-set Inference and Network Enrichment) was first described and used in Beckers et al. (2026) to find candidate regulators of the PUCHI-dependent lipid genes of lateral roots.

Data. Two networks. Binding: the DAP-seq cistrome of O'Malley et al. (2016): for 387 Arabidopsis transcription factors, the genes with a binding peak in vitro (2.85 million TF–gene links once duplicates are removed; a median TF binds 22% of the genes). Regulation: the 65 TARGET experiments of the ConnecTF database (Brooks et al. 2019, 2021; Varala et al. 2018; Alvarez et al. 2020), 57 TFs: each TF, fused to the glucocorticoid receptor, is moved into the nucleus with dexamethasone in root (59 experiments) or shoot (6) protoplasts, with cycloheximide so that only direct targets respond; the genes induced or repressed after 3 h (FDR < 5%) are its targets. 30 of these TFs are not in the DAP-seq set. Annotations of the TFs and of your genes come from TAIR / Araport11 and UniProtKB; the family is inferred from them.

Kernels. Because DAP-seq is permissive, two random genes already share about 28 TFs. Régine therefore looks at kernels: groups of k genes of your list, and asks which TFs bind all of them. A real regulator of your list binds many of its kernels; a promiscuous TF also binds many random kernels. For a TF that binds k_t of the n genes of your list and K_t of the N background genes:

These are the exact values of what Régine 2022 estimated by drawing 1 000 random kernels (columns Real and Noise, here as fractions instead of counts), so the result no longer depends on chance. The enrichment P is the hypergeometric probability of finding at least k_t targets of the TF in a random list of n background genes, corrected over the 387 TFs (Benjamini–Hochberg). The table is ranked by FDR by default; the score favours TFs that bind almost all of your genes (often broad binders), the P favours TFs whose targets are rare in the genome; look at both.

Automatic kernel size. For each k from 2 to 10, random kernels are drawn from your list and from the background, and the number of TFs they share is counted (the boxplots, as in the Régine 2022 PDF). The difference between the two raw counts always peaks at k = 2, where everything shares many TFs, so Régine measures the separation of the two distributions instead: the AUC, the probability that a kernel of your list shares more TFs than a random kernel. The kernel size with the highest AUC is used. You can click any other size.

TARGET: direction and combination. Each TARGET experiment is tested three times: on all its regulated genes (the kernels and the score), on its induced genes and on its repressed genes; a TF whose induced (repressed) genes are enriched in your list probably activates (represses) them. A TF tested several times keeps one row per experiment, because repeated experiments overlap little. The TARGET files list the significant genes only, so the TARGET background is the 18 727 genes regulated in at least one experiment. The DAP-seq + TARGET view puts the two networks side by side, one row per TF: the best TARGET experiment of the TF (its P multiplied by the number of experiments), the genes both bound in DAP-seq and regulated in TARGET, and a combined P (Fisher's method) ranked by its FDR over all TFs.

Background. The random genes should be genes that could have been in your list: when your list comes from an RNA-seq or array experiment, paste the expressed (or measured) genes as your own background (it is then used for both networks).