Drilling Into the Transcriptional Core

Machine learning helps to tease out the patterns of DNA at promoters that initiate transcription.

One of the holy grails of molecular biology is the study of transcriptional initiation. While there are many levels of regulation in cells, the initiation of transcription is perhaps, of all of them, the most powerful. An organism's ability to keep the transcription of most genes off, and turn on genes that are needed to build particular tissues, and regulate others in response to other urgent needs, is the very soul of how multicellular organisms operate. The decision to transcribe a gene into its RNA message (mRNA) represents a large investment, as that transcript can last hours or more and during that time be translated into a great many protein copies. Additionally, this process identifies where, in the otherwise featureless landscape of genomic DNA, genes are located, which is another significant process, one that it took molecular biologists a long time to figure out.

Control over transcription is generally divided into two conceptual and physical regions- enhancers and promoters. Enhancers are typically far from the start site of transcription, and are modules of DNA sequences that bind innumerable regulatory proteins which collectively tune, in fine and rough ways, initiation. Promoters, in contrast, are at the core and straddle the start site of transcription (TSS, for short). They feature a much more limited set of motifs in the DNA sequence. The promoter is the site where the proteins bound to the various enhancers converge and encourage the formation of a "preinitiation complex", which includes the RNA polymerase that actually carries out transcription, plus a lot of ancillary proteins. The RNA polymerase can not initiate on its own or find a promoter on its own. It requires direction by the regulatory proteins and their promoter targets before finding its proper landing place. So the study of promoter initiation and regulation has a very long history, as a critical part of the central flow of information in molecular biology, from DNA to protein.

A schematic of a promoter, where initiation of transcription of Gene A, happens, with the start site (+1) right at the boundary of the orange and green colors. At this location, the RNA polymerase will melt the DNA strands, and start synthesizing an RNA strand using the (bottom) template strand of the DNA. Regulatory proteins bound to enhancers far away in the genomic DNA bend through space to activate proteins bound at the core promoter to load the polymerase and initiate this process.

A recent paper provided a novel analysis of promoter sequences, using machine learning to derive a relatively comprehensive account of the relevant sequences. Heretofore, many promoters had been dissected in detail and several key features found. But many human promoters had none of them, showing that our knowledge was incomplete. This new approach started strictly from empirical data- the genome sequence, plus large experimental compilations of nascent RNAs, as they are expressed in various cells, and mapped to the precise base where they initiated from- that is, their respective TSS. These were all loaded into a machine learning model that was supplemented with explanatory capabilities. That is, it was not just a black box, but gave interpretable results useful to science, in the form of small sequence signatures that it found are needed to make particular promoters work. These signatures presumably bind particular proteins that are the operational engines of regulatory integration and promoter function.

The TATA motif, found about 30 base pairs upstream of the transcription start site in many promoters. This is a motif view, where the statistical prevalence of the base is reflected in the height of the letter (top, in color) and its converse is reflected below in gray. Regular patterns like this found in DNA usually mean that some protein typically binds to this site, in this case TFIID.

For example, the grand-daddy of them all is the TATA box, which dates back to bacteria / archaea and was easily dug up by this machine learning system. The composition of the TATA box is shown above in a graphical form, where the probability of occurrence (of a base in the DNA) is reflected in height of the base over the axis line. A few G/C bases surround a central motif of T/A, and the TSS is typically 30 base pairs downstream. What happens here is that one of the central proteins of the RNA polymerase positioning complex, TFIID, binds strongly to this sequence, and bends the DNA here by ninety degrees, forming a launchpad of sorts for the polymerase, which later finds and opens DNA at the transcription start site. TFIID and the TATA box are well known, so it certainly is reassuring that this algorithmic method recovered it. TATA boxes are common at regulated promoters, being highly receptive to regulation by enhancer protein complexes. This is in contrast to more uniformly expressed (housekeeping) genes which typically use other promoter DNA motifs, and incidentally tend to have much less precise TSS positions. They might have start sites that range over a hundred base pairs, more or less stochastically.

The main advance of this paper was to find more DNA sites, and new types of sites, which collectively account for the positioning and activation of all promoters in humans. Instead of the previously known three or four factors, they found nine major DNA sequences, and a smattering of weaker patterns, which they combine into a predictive model that matches empirical data. Most of these DNA sequences were previously known, but not as part of core promoters. For example, one is called YY1, because it binds the YY1 protein, which has long been appreciated to be a transcriptional repressor, from enhancer positions. But now it turns out to also be core promoter participant, identifying and turning on a class of promoters that, as for most of the new-found sequence elements, tend to operate genes that are not heavily regulated, but rather universally expressed and with delocalized start sites. 

Motifs and initiator elements found by the current work. Each motif, presumably matched by a protein that binds it, gets its own graph of relation of the motif location (at 0 on the X axis) vs the start site of transcription that it directs, which for TATA is about 30 base pairs downstream. Most of the newly discovered motifs are bi-directional, directing start sites and transcription both upstream and downstream. This wastes a lot of effort, as the upstream transcripts are typically quickly discarded. The NFY motif has an interesting pattern of 10.5 bp periodicity of its directed start sites, which suggests that the protein complex that binds this site hugs one side of the DNA quite closely, setting up start sites on that side of the helix.

Secondly, these authors find that most of the new sequences they identify have bidirectional effects. That is, they set up promoters to fire in both directions, both typically about forty base pairs downstream and also upstream from their binding site. This explains a great deal of transcription data derived from new sequencing technologies, which shows that many promoters fire in both directions, even though the "upstream" or non-gene side transcript tends to be short-lived.

Overview of the new results, summarized by type of DNA sequence pattern. The total machine learning prediction was composed of predictions for larger motifs, which were the dominant pattern, plus a small contribution from "initiators", which comprise a few patterns right at the start site, plus a large but diffuse contribution from tiny trinucleotide patterns, such as the CG pattern known to mark active genes and carry activating DNA methylation marks.

A third finding was the set of trinucleotide motifs that serve as the sort of fudge factor for their machine learning model, filling in details to make the match to empirical data come out better. The length was set more or less arbitrarily, but they play a big part in the model fit. They note that one common example is the CG pattern, which is one of the stronger trinucleotide motifs. This pattern is known as CpG, and is the target of chemical methylation of DNA by regulatory enzymes, which helps to mark and regulate genes. The current work suggests that there may be more systems of this kind yet to be discovered, which play a modulating role in gene/promoter selection and activation.

The accuracy of this new learning and modeling system exemplifies some of the strengths of AI, of which machine learning is a sub-discipline. When there is a lot of data available, and a problem that is well defined and on the verge of solution (like the protein folding problem), then AI, or these machine learning methods, can push the field over the edge to a solution. AI / ML are powerful ways to explore a defined solution space for optimal results. They are not "intelligent" in the normal sense of the word, (at least not yet), which would imply having generalized world models that would allow them to range over large areas of knowledge, solve undefined problems, and exercise common sense.

• A lecture on promoter complex assembly, in bacteria and other organisms.
• A dose of honesty.
• Starfish are all head, no body.
• The meaning of financialization.