Systematic engineering and machine learning analysis of intrinsic terminators reveal crucial nucleotides directly upstream of the terminator hairpin.

Adenosine nucleotides upstream of terminator hairpins boost bacterial protein production

In the toolkit of synthetic biology, controlling how much protein a cell produces is a fundamental challenge. Researchers often focus on the "engine"—the coding sequence that dictates amino acid assembly—but a critical regulatory component is frequently overlooked: the 3' untranslated region (3'UTR), the segment of mRNA that follows the stop codon. Within this region lie intrinsic terminators, DNA sequences that signal the RNA polymerase (RNAP, the enzyme that builds RNA from DNA) to stop transcription and release the newly formed mRNA.

Current engineering practices treat these terminators as modular "parts." They assume strength is determined solely by two structural features: a GC-rich hairpin (a loop-like structure formed by base-pairing) and a downstream poly(U)-tail (a stretch of uracil nucleotides). However, this modular view is incomplete. Until now, the specific nucleotide identity of the "spacer" sequence—the stretch of RNA sitting between the stop codon and the terminator hairpin—has been treated as a mere buffer. This study reveals that this spacer is not a passive bystander. Instead, specific nucleotides within it can act as a precision dial to tune gene expression.

The hidden variable in transcriptional termination

The standard model of intrinsic termination relies on a mechanical tug-of-war. As the RNAP transcribes the poly(U)-tail, the weak interaction between the RNA and the DNA template causes the enzyme to stall. This pause provides the window necessary for the upstream GC-rich hairpin to fold within the RNAP exit tunnel. The formation of this hairpin exerts physical strain on the unstable RNA-DNA hybrid. This eventually forces the RNAP to dissociate from the template.

While the strength of the hairpin and the length of the poly(U)-tail are known drivers of efficiency, the intervening spacer sequence has remained a black box. Engineers have long struggled with "read-through"—a failure mode where the RNAP ignores the stop signal and continues transcribing downstream. This wastes cellular resources and can interfere with the expression of neighboring genes in multi-gene circuits. Previous research suggested that the distance between the stop codon and the terminator matters to avoid steric hindrance (physical crowding) from ribosomes. However, the actual chemical identity of those intervening bases has never been systematically mapped.

Decoding the spacer through randomization and machine learning

To isolate the effect of the spacer, the authors employed a dual strategy of massive combinatorial screening and predictive modeling. First, they generated a library of completely randomized 30-base pair (bp) 3'UTR spacer sequences. They inserted these between the stop codon and a standard terminator .

By screening these variants in E. coli, the researchers discovered that the spacer identity alone could drive up to a 5.4-fold difference in protein production for RFP (Red Fluorescent Protein) and a 2.7-fold difference for GFP (Green Fluorescent Protein) [Figure 1B]. Crucially, the authors demonstrated that this effect was largely independent of the upstream coding sequence. This means the spacer acts as a universal regulator of the transcript's fate [Figure 2A].

The researchers then moved from observation to mechanism by performing systematic substitutions. They identified that spacers ending in adenosine (A) or guanine (G) consistently yielded higher protein levels. The proposed mechanism is a structural "extension." These bases are predicted to base-pair with the free uracil residues of the downstream poly(U)-tail [Figure 2B]. This interaction effectively elongates the terminator hairpin stem. This stabilizes the structure and facilitates more efficient termination.

To validate this across a broader landscape, the team developed a machine learning approach using a Random Forest regressor (a model that uses many decision trees to make predictions). They trained the model on two major datasets: the "Chen dataset" of synthetic terminators and the "TERMITe dataset" comprising natural terminators from B. subtilis and E. coli. They encoded the terminators based on their predicted RNA secondary structures. They separated them into the A-tract, stem, loop, and U-tract. This allowed them to ask the model specific questions, such as "Is the $n$-th base of the spacer an Adenosine?"

Identifying the adenosine "sweet spot"

The machine learning analysis provided definitive proof that the spacer is a key regulatory feature. When the models were forced to generalize across different species and terminator types, the identity of the bases directly upstream of the hairpin emerged as a dominant predictive feature .

Specifically, the model identified that having an adenosine at the position immediately preceding the stem was highly predictive of high termination efficiency [Figure 5E, 5F].

The authors confirmed this through rational design. They found that inserting even a single adenosine at the end of the spacer could double the level of protein production. It also increased mRNA copy numbers by up to six-fold [Figure 3E, 3F]. This increase in mRNA is a direct consequence of improved termination efficiency. By reducing transcriptional read-through, the cell maintains a more abundant pool of the target transcript [Figure 4A]. Interestingly, the study showed that simply extending the hairpin stem with traditional G-C pairs did not provide the same boost. Nor did adding more adenines to a long "A-tract" [Figure 3E]. The specific base-pairing interaction between the spacer's terminal adenosine and the poly(U)-tail was the key driver.

Limits of the in vitro model

Despite the strength of the findings, the study notes several boundaries to its conclusions. First, the in vitro (outside the living cell) transcription assays did not perfectly replicate the in vivo (inside the living cell) results. These assays used purified RNAP and accessory proteins like NusA and NusG. While the in vivo data showed a clear benefit from the adenosine insertion, the in vitro assays could not definitively distinguish between two possibilities. They could not tell if the changes were due to improved termination efficiency or a change in the overall transcription rate .

Furthermore, the machine learning models struggled with natural terminators compared to synthetic ones. The model achieved a high Pearson correlation of 0.90 for synthetic terminators. However, its correlation for natural sequences was significantly lower, ranging from 0.34 to 0.48 [Figure 5A]. This discrepancy arises because natural terminators are evolved to operate within a complex regulatory landscape. In contrast, synthetic terminators often occupy extreme, easily classified states of "on" or "off." Finally, the study did not explore whether these spacer effects extend to mRNA stability (how quickly mRNA breaks down). Preliminary tests in mRNA-turnover mutant strains suggested the boost in protein levels was driven by transcription rather than reduced degradation.

A new rule for synthetic gene design

The verdict is clear: the 3'UTR spacer is a high-leverage target for genetic engineering. For practitioners designing synthetic circuits or heterologous expression systems (expressing foreign genes in a host), the takeaway is actionable. Include at least one adenosine nucleotide directly upstream of the terminator hairpin. This simple modification offers a robust, sequence-independent method to reduce read-through and maximize protein yields.

The research moves the field away from treating terminators as monolithic blocks. It moves toward a more granular, nucleotide-level understanding. Future work will likely need to integrate these spacer-specific features into existing terminator-design algorithms. This will help move from empirical "trial and error" to truly predictive synthetic genomics. The code and models used in this study are available via the MEWTWO repository on Zenodo and GitHub.

Figures from the paper