AI discovery of sequence rules for RNA polymerase II pausing revises the pause-release model of gene activation
Synopsis
Using DEFT, an LLM-guided interpretable decision-tree framework, the authors found that pause sites are discriminated by a G at the pause position, C or T at +1 and a minimum G content in the 49 upstream bases (held-out AUROC 0.92, accuracy 0.87); inserting a 377 bp G-less cassette at the 5' ends of NDRG1 and HSP90AA1 abolished the promoter-proximal pause in an orientation-dependent way without impairing hypoxic or heat-shock activation (NDRG1 was even super-activated), while ChIP 5' RNAPII, Ser5P and NELF peaks persisted without NET-seq-detectable pausing and CTD deletion weakened the pause and shifted its peak from ~+80 to ~+40, supporting a model in which the pause is initiated by G-dependent arrest and stabilized by factors including CTD-mediated tethering.
Interpretation
DEFT identified the core sequence feature separating pauses from non-pauses: G at the pause position, C or T at +1, plus a minimum G content in the 49 upstream bases; the high-G branch contained 66,897 pause sites but only 6,680 non-pause sites. Previously NET-seq showed only that pause sites are enriched at G,T/C dinucleotides, and no sequence element required for establishing the pause had been identified, so the pause could not be treated as a variable to be eliminated and tested. Trained on roughly 95,000 pause sites and matched non-pause sites from HCT116 eNET-seq, and evaluated on a held-out set of 48,357 samples (24,178 pauses, 24,179 non-pauses) with AUROC 0.92 and accuracy 0.87.
Inserting a 377 bp G-less cassette at +8 of NDRG1 and +9 of HSP90AA1 abolished the promoter-proximal pause in an orientation-dependent way, while activation by hypoxia (DMOG, CoCl2) and heat shock (42 °C, 30 min) still occurred, with mRNA super-activation of NDRG1. It provides the first gene-specific elimination of the promoter-proximal pause, allowing the pause-release model to be tested causally rather than by correlation. Multiple independent G-less clones in HAP1 and HCT116, an inverted C-less control, RNA-seq confirmation that transcription starts within the insertion, a synthetic-read control for mapping bias over the G-less region, and NuPoP V2.5 nucleosome-positioning predictions.
The ChIP 5' RNAPII peak, together with Ser5P and NELF peaks, persisted after eNET-seq-detectable pausing was abolished, showing that the 5' peak does not necessarily correspond to paused elongation complexes. It separates the widely used ChIP 5' RNAPII peak from NET-seq-measured pausing; the authors propose the peak includes initiation or abortive-initiation complexes whose transcripts are too short to be captured by NET-seq. Parallel eNET-seq, RNAPII ChIP-seq, Ser5P ChIP-seq and NELF ChIP-seq on the same genes, with consistent patterns at both NDRG1 and HSP90AA1.
Deleting the CTD (Rpb1 residues 1593-1961, all 52 heptad repeats) suppressed the promoter-proximal pause and lowered the pausing index at thousands of genes without reducing recognition of G,T/C pause sites, and shifted the pause peak from ~+80 to ~+40. It gives direct experimental support for the CTD-as-promoter-tethering-arm model and shows that G-dependent pausing is necessary but not sufficient for the promoter-proximal pause. Anti-Avitag eNET-seq in doxycycline-inducible Avitag Rpb1 WT versus ΔCTD HEK293 Flp-in cells, n=4265 with p<1.88e-260, illustrated at genes such as FOS, ARC, DNAJA1, HSPA1B and LDHA and in genome-wide metaplots.
Perspective
In the two TATA-box, focused-start, highly regulated genes NDRG1 and HSP90AA1, G-less insertion is validated as a gene-specific means of suppressing the promoter-proximal pause, which can be used to test pause function at other genes; the DEFT decision-tree framework can be transferred to feature discovery for other classes of DNA function; and the genome-wide low-G analysis, using HEK293 CAGE-defined TSSs, links the sequence rule to pausing indices of natural genes. It speaks mainly to researchers who want causal rather than correlative tests of transcription pausing, and to method-oriented readers interested in promoter-proximal regulation and sequence-to-function mapping.
The molecular basis of G-dependent pausing is unknown, and it is not known whether the signal lies in the nascent RNA or on the non-template DNA strand; the minimal number and distribution of G residues required for pause formation, and whether the same rule holds at CpG island promoters and promoters with diffuse start sites, remain to be tested; whether other highly regulated genes depend on pause release is undetermined; and the identity of the non-paused RNAPII complexes that give rise to large ChIP peaks (the authors speculate abortive initiation complexes in vivo) is currently hard to assay. The work is a preprint that has not been certified by peer review, and part of the evidence is presented through supplementary figures (S1-S4 and others) and their legends, which limits the detail a text-only reader can verify.