Biologists have known about the initiator since 1989. What they have never had is a reliable way to find one. The initiator — the short stretch of DNA marking the exact base where a gene starts being read — spent thirty-seven years defined by consensus strings so loose they matched half the genome by accident. A team at UC San Diego has replaced the string with a model: a library of roughly 500,000 initiator variants, each measured for how strongly it fired, support vector regression models trained on 200,000 of them, and the result used to score every one of the 1,048,576 possible ten-nucleotide sequences. About 60 percent of natural human promoters, they report, carry a working initiator.
The study appeared in Genes & Development after posting online July 31 and was publicized by UC San Diego on August 21. It is led by graduate student researcher Torrey E. Rhyne-Carrigg with Long Vo ngoc, Claudia Medrano, Kassidy E. Gillespie and James T. Kadonaga.
"These AI models were found to provide, for the first time, strong predictions of the presence or absence of the initiator in human genes, and were thus able to decode the DNA base sequence pattern of the initiator," Kadonaga said.
Why a four-letter switch was so hard to read
A core promoter is the roughly 80-base window around a transcription start site where RNA polymerase II assembles. Within it sit a few recognizable motifs: the TATA box, in an estimated 15 to 23 percent of human promoters; the downstream promoter region, or DPR, at 25 to 34 percent; and the initiator, previously pegged between 40 and 56 percent. Roughly 30 percent of promoters contain none of them.
The initiator's problem was always that it is short, degenerate and defined by eye. A 1980 survey of eukaryotic promoters produced a YYCA+1YYYYY consensus; Smale and Baltimore established in 1989 that this was a functional element and named it. Every refinement since has been an alignment exercise — count the letters at each position across natural promoters, write down the ambiguity codes, hope the string separates real initiators from noise. It does not, particularly. As the authors put it, a consensus sequence "does not provide any quantitative information on the transcriptional activity of any Inr that matches this motif."
So the team stopped aligning and started measuring. Using a technique the lab calls HARPE — high-throughput analysis of randomized promoter elements — they randomized a 10-nucleotide window from position -4 to +6, and separately a 14-nucleotide window from -5 to +9, inside an otherwise fixed synthetic promoter backbone, each variant carrying its own 14-base barcode. They ran the libraries three ways: driven by a DPR, driven by a TATA box, and with neither, then repeated the experiment in HeLa cells.
The machine learning is deliberately unglamorous — support vector regression in scikit-learn 1.3.0, grid-searched hyperparameters, two independent test sets, run on the Expanse cluster at the San Diego Supercomputer Center. No transformer, no foundation model. The power comes from the data: several hundred thousand controlled variants of one element beat 30,000 natural promoters that differ from each other in every other respect at once.
The motif that fell out was a surprise. The most active initiators converge on TCA+1KTY — essentially identical to the fly consensus, unchanged across roughly 700 million years of divergence. In natural human promoters the minimal signature is looser, CA+1NW. Applied across 30,642 human and 4,489 Drosophila promoters, the models also surfaced a previously unrecognized variant: an initiator with its core A shifted to position +3, inert with a DPR but functional with a TATA box. It appears in 8.3 percent of TATA-containing human promoters (368 of 4,448) and 0.1 percent of DPR-containing ones (10 of 7,630). In Drosophila the split is starker — 4.6 percent versus zero out of 3,070.
Why It Matters
The immediate payoff is diagnostic. Variant interpretation in noncoding DNA remains largely guesswork; a mutation two bases upstream of a start site has no obvious protein consequence to reason about. A quantitative initiator model turns that into a score — compare predicted activity for reference and variant, and you get a defensible estimate of whether the mutation cripples gene activation. Run in reverse, the same models give synthetic biologists a dial for tuning promoter strength in gene therapy vectors.
The broader finding is structural. Initiator and DPR work in strict synergy — with a DPR-driven promoter, the best initiators are about 100-fold more active than a median dead sequence, versus only 10-fold with TATA — while TATA and DPR behave as alternatives, and TATA-only promoters skew toward signaling, immune and stress-response genes. That is a picture of at least two distinct grammars for switching genes on, not one universal switch.
One caveat on the headline number: the 60 percent figure is measured in focused promoters, those where transcription starts at one site or a tight cluster. The paper generalizes it to human promoters; press coverage generalized it further to human genes.
Kadonaga frames the work as a down payment. "Ultimately, within the six billion bases of DNA in each of our cells, there is a gene expression code that specifies when, where and to what extent each of our genes should be turned on or off," he said. "The new AI model for the initiator is a small but important part of this gene expression code, and I am optimistic that we will expand our AI models of the human gene expression code in the not-too-distant future."
What to watch is whether the HARPE-plus-regression recipe scales past single elements. The lab now has calibrated models for TATA, initiator and DPR independently; the untouched target is the roughly 30 percent of human promoters that match none of them, which the authors name as their next objective. If those promoters run on motifs nobody has written down, the same half-million-variant approach is the obvious way to find them — and the large sequence models being trained on natural DNA will need exactly this kind of ground truth to check themselves against.
“These AI models were found to provide, for the first time, strong predictions of the presence or absence of the initiator in human genes, and were thus able to decode the DNA base sequence pattern of the initiator.”— James T. Kadonaga, Professor of Molecular Biology, UC San Diego