One of the earliest and most popular applications of synthetic promoters has been their use in functional genomics to drive constitutive tissue-specific or condition-specific expression of target or reporter genes (Ulmasov et al., 1997; Rushton et al., 2002). Since natural promoters can behave in unexpected ways, synthetic sequences composed of modular CREs upstream of minimal promoters have proven advantageous in generating more precise patterns of expression (Brooks et al., 2023). However, some studies have had success with modifying natural promoters to create synthetic ones, either through mutations, deletions, or chimeric fusions of multiple natural promoter sequences or elements (Efremova et al., 2020; Kumari et al., 2024; Sherpa and Dey, 2025). Two recent papers reported the creation of copper-inducible systems in Nicotiana benthamiana by utilizing four copies of the copper-binding site (CBS) motif fused upstream of various minimum promoters to create synthetic promoters that a copper responsive TF, CUP2, can bind to only in the presence of copper (Table 1) (Garcia-Perez et al., 2022; Chiang et al., 2024). In one of these studies, the copper-inducible system outperformed other inducible systems that were tested, specifically systems that rely on β-estradiol, dexamethasone, or doxycycline (Chiang et al., 2024). Another study leveraged promoters from chickweed ANTIMICROBIAL PEPTIDE1 (AMP1) and ANTIMICROBIAL PEPTIDE2 (AMP2) genes to create a constitutive chimeric promoter that expresses higher than pro-SmAMP2 and similarly to pro-SmAMP1 in transient assays, exhibiting a diversified promoter sequence that can achieve levels of activity similar to a natural promoter (Table 1) (Efremova et al., 2020). Additionally, the study was able to identify proline-inducible motifs from these natural promoters that when excluded from pro-SmAMP2, led to high constitutive expression profiles. Though this may not seem like a very significant finding, it is important to remember that generating promoters with diverse sequences that are able to constitutively express target genes is one of the goals of synthetic biology research, as these diversified promoter libraries are essential to downstream generation of genetic logic gates with non-overlapping inputs. To this end, natural promoters from plant pararetroviruses can also be leveraged to create synthetic promoters that constitutively express in plants, as seen with the creation of 35S promoter variants (Ali and Waliullah, 2021; Amack et al., 2022). MSD3 is an example of a synthetic promoter that is an amalgamation of motifs from various monocot and dicot pararetroviral-based promoters that is able to drive expression at similar levels to the 35S gold-standard constitutive promoter in rice, pearl millet, and tobacco plants (Table 1) (Kumari et al., 2024). Another synthetic promoter that was recently developed in the same manner, MFH17, contains fused elements from Mirabilis mosaic virus (MMV), Figwort mosaic virus (FMV), and Horseradish latent virus (HRLV) derived promoters and is also able to drive high expression in both dicots and monocots (Table 1) (Sherpa and Dey, 2025). These promoters represent valuable additions to the growing library of constitutive promoters available for use in plant biotechnology. Their effectiveness highlights the significance of a combinatorial approach, specifically by exploring and integrating promoter elements from diverse species to develop highly efficient, strongly expressing synthetic promoters.

In addition to constitutive promoters, researchers have also explored creating tissue-specific promoters for various purposes. Four synthetic promoters created through informed fusion of regulatory sequences derived from natural promoters, BiGSSP2, BiGSSP3, BiGSSP6, and BiGSSP7, were shown to induce bi-directional (i.e. can initiate transcription in either orientation) expression of reporters in rice (Table 1) (Bai et al., 2020). These promoters function specifically in green tissues, including leaf, sheath, panicle, and stem, and represent a promising example of how synthetic promoter design can provide agronomically relevant tissue specificity with direct potential for crop improvement (Bai et al., 2020). The capacity to restrict gene expression to photosynthetically active tissues is particularly valuable for metabolic engineering applications, where minimizing ectopic expression reduces unintended fitness costs while still achieving strong expression in target organs.

Identifying tissue-specific promoters with inducible functionality can provide a mechanism for tunable spatiotemporal control of gene expression. Inducible promoter designs are highly advantageous when driving expression of genes with potentially detrimental effects, as constitutive expression could cause cytotoxicity and compromise plant health. Tissue-specific inducible promoters ensure that expression is activated only when necessary in a restricted tissue domain. A study by Yang et al. (2024) exemplifies this approach by designing and testing a drought-inducible, green-tissue-specific promoter in poplar (Table 1) (Yang et al., 2024). Their methodology involved identifying relevant motifs from genes upregulated in leaf palisade and vascular tissues under water-deficit stress. This resulted in the creation of two synthetic promoters: a constitutive synthetic promoter Syn3, created by concatenating four repeats of a conserved 20 bp motif derived from promoters of green-tissue-specific genes differentially expressed under drought conditions, that preferentially expresses in green tissue, and a second promoter, Syn3-10b-1, created by concatenating the 5’ region of the 20 bp motif utilized in Syn3, that specifically drives gene expression only in green tissues under drought conditions in transgenic poplar (Table 1). Another study, through experimentation with copy number, orientation, and spacing of motifs, generated and characterized a suite of potential root-specific and drought-inducible synthetic promoters (Jameel et al., 2020). Root-specific promoters are of particular interest because root tissues are the primary site of water and nutrient uptake, and reprogramming their behavior under stress is critical for agricultural applications, such as enhancing crop resilience or adaptability. In the study by Jameel et al., one synthetic promoter in particular containing eleven concatenated in-silico identified CREs, SynP16, stood out in its specificity to root tissue in response to drought conditions in soybean and Arabidopsis (Table 1). The particular combination and arrangement of root-specific and drought-inducible elements in SynP16, though arising fortuitously in this case, illustrates how methodical researchers must be when constructing synthetic promoters if we hope to achieve highly controlled and context-dependent expression patterns.

The emergence of rational inducible synthetic promoter designs enabled researchers to monitor specific hormones, stresses, or metabolites in direct proportion to signal intensity through the use of synthetic promoter-driven reporters, commonly referred to as transcriptional biosensors. Auxin is arguably the most intensively studied plant hormone in terms of how it is spatially distribution and developmental patterning, and synthetic promoters were pivotal to that progress (Table 1) (Ulmasov et al., 1997; Liao et al., 2015). The canonical DR5 promoter, constructed from multimerized AUXIN RESPONSE ELEMENT (AuxRE) motifs (i.e. TGTCTC) recognized by Auxin Response Factors (ARFs), has been used for decades to visualize auxin dynamics in roots, embryos, and shoots (Ulmasov et al., 1997). Switching to TGTCGG motifs and optimizing repeat motif orientation gave rise to the DR5v2 promoter, resulting in increased sensitivity and refined cellular resolution (Table 1) (Liao et al., 2015). In more recent years, auxin responsive synthetic promoters have been leveraged to deconvolute auxin regulated transcription (Martin-Arevalillo et al., 2025). Preferential binding sites for auxin responsive factors (ARFs) in the form of AuxRE pair configurations were utilized in synthetic promoters to identify auxin-induced transcriptional responses at a single-cell level. The study found that different clades of ARFs prefer distinct binding sites, leading to spatial specificity of the AuxRE pairs in planta. This finding is yet another example illustrating how combinations of motifs can drive divergent expression patterns in plants.

Another useful hormone to create biosensors for is ethylene, due to its prominent role in plant growth and development (Dubois et al., 2018). The ethylene signaling pathway has been leveraged repeatedly to construct ethylene biosensors by generating synthetic promoters containing EIN3-binding sites (EBS) to drive fluorescent or histochemical reporters. In early efforts, Stepanova et al. developed a synthetic ethylene-responsive promoter using 5 copies of multimerized EBS motifs (Table 1) (Stepanova, 2001; Stepanova et al., 2007). This promoter was activated in the presence of ethylene but was limited in its sensitivity and expression uniformity across tissues. To address these restrictions, a new generation of ethylene biosensors was constructed by optimizing the sequence and arrangement of EBS motifs. One example is the 10×2EBS-S10 promoter, which contains ten tandem homotypic copies of a dual, everted EBS motif placed upstream of the minimal 35S(-46) promoter (Table 1). This promoter displays moderate ethylene responsiveness in Arabidopsis seedlings, showing different expression domains than the original 5xEBS promoter (Fernandez-Moreno et al., 2024). Building on this, an even more recent study reported the EBSn promoter, which is composed of ten heterotypic, natural, dual, everted EIN3-binding sites placed in tandem upstream of the same minimal 35S (-46) promoter (Table 1) (Fernandez-Moreno et al., 2025). EBSn was found to outperform earlier ethylene reporters in terms of sensitivity and breadth of response, detecting changes in ethylene production in both seedlings and adult plants of Arabidopsis. In tomato fruits, EBSn reporters displayed ripening-related ethylene accumulation, showcasing the promoter’s utility for crop applications (Fernandez-Moreno et al., 2025).

These ethylene biosensors suggest two important design principles: 1) multiplicity of binding sites within a promoter amplifies signal but may suffer from sequence-stability and synthesis issues, and 2) promoters that concatenate heterotypic motifs associated with natural binding-site variation can outperform promoters with simple homotypic repeats by better matching the native TF recognition landscape. Both biosensors, driven by either 10x2EBS-S10 or EBSn, are state-of-the-art examples of how informed synthetic promoter design can result in robust tools for monitoring hormone signaling to aid fundamental plant research.

Motifs associated with other plant hormones have also been utilized to construct synthetic promoters for hormone biosensors. The 6xABRE synthetic promoter harboring six tandem ABRE motifs upstream of a minimal promoter is a widely used ABA-responsive promoter that reveals spatiotemporal dynamics of ABA signaling (Table 1) (Wu et al., 2018). TCS and TCSn are the classic cytokinin-responsive synthetic promoters, built from concatenated type-B Arabidopsis response regulators (ARR) binding sites (Table 1) (Müller and Sheen, 2008). TCSn in particular offers improved sensitivity and has been leveraged as a template for the creation of other cytokinin-responsive promoters, such as TCSv2, which alters the orientation of the type-B ARR binding sites utilized in TCSn (Table 1) (Steiner et al., 2020; Zürcher et al., 2013). In more recent years, dual hormone-responsive promoters responsive to either auxin or cytokinin were created by combining CREs from both pathways (AuxRE for auxin and 2xCKRE for cytokinin) into one promoter, enabling the study of hormone crosstalk and the design of conditional control circuits (Li et al., 2023). Synthetic promoters responsive to salicylic acid (SA) and jasmonic acid (JA) were created in a similar manner, given that these two hormones play a large role in plant defense against pathogens and thus could be extremely useful in both sensing pathogen presence and hormone levels (Li et al., 2023). The most promising synthetic promoter, SJ-609, was constructed by inserting a JA-responsive CRE into an SA-responsive promoter and conferred higher expression in response to either hormone, as well as pathogen presence, in tobacco, Arabidopsis, tomato, and cotton than either of the native JA or SA responsive promoters tested (Table 1) (Li et al., 2023). This study not only directly adds to the repertoire of inducible hormone biosensors but also showcases a novel design strategy for synthetic promoters that can respond to multiple alternative inputs simultaneously.

Another study developed pathogen-sensing constructs in potato using synthetic promoters built from S-box elements, motifs that are responsive to fungal infection, coupled to a transcription-factor system (Q-system) (Persad-Russell et al., 2022). Utilizing the 4×S-Box synthetic promoter to drive a Q-system transcriptional activator variant QF2, which then binds to its TFBS upstream of an mEmerald GFP reporter gene, yielded approximately a six-fold signal amplification upon infection with Clavibacter michiganensis subsp. nebraskensis (CMN), demonstrating enhanced sensitivity for pathogen detection in potato (Table 1). This promoter, and others like it, would allow for early detection of plant pathogen infection in real-world applications, and possibly even trigger targeted defense responses in the plant upon pathogen encounter.

Ability to respond to salt and osmotic stress has also been explored through the design of synthetic promoters. Bhadouriya et al. reported a synthetic salt-inducible promoter, named PS, composed of motifs extracted from native salt-responsive promoters and rationally arranged based on their original configuration within their respective promoters (Table 1) (Bhadouriya et al., 2024). The PS promoter was used to drive reporter expression in transient assays within Nicotiana tabacum and in stably transformed Arabidopsis lines, showing stronger and more sustained expression under salt/abiotic stress conditions than CaMV 35S. The results demonstrate that synthetic promoter design can outperform common constitutive promoters in a stimulus-dependent manner, validating the cis-engineering approach for abiotic stress biosensors (Bhadouriya et al., 2024).

In recent years, synthetic promoters have begun to be utilized in the design of genetic logic gates in plants. Genetic logic gates borrow concepts from mathematics and Boolean logic to achieve conditional control of gene expression based on combinations of biological inputs such as transcription factors, signaling molecules, or environmental cues, allowing for complex and programmable transcriptional responses (Bradley et al., 2016; Andres et al., 2019). A logic gate takes in one or more binary inputs (i.e., that resolve to true (one) or false (zero)) and produces one binary output depending on the type of logic gate. In this framework, synthetic promoters allow for the integration of multiple CREs that respond to distinct upstream regulators. The resulting promoter output, either transcription activation or repression, functions analogously to Boolean logic operations such as AND, OR, and NOT (Miyamoto et al., 2013) (Figure 4A). At the transcriptional level, an AND gate may require two different TFs to bind simultaneously for activation, ensuring that expression only occurs when both signals are present (Figure 4B). Of course, given the nature of promoter activation by TFs, AND gates built in this manner are commonly ‘leaky’, as the presence of a single TF’s TFBS can often be sufficient for some level of expression. The idea is the TFs act cooperatively, significantly increasing expression when both TFBS are present (AkhavanAghdam et al., 2016). Conversely, OR gates can be designed by embedding multiple independent activator-binding sites, enabling expression when any one of several signals is detected (Figure 4B). Most current synthetic promoters are representative of the OR gate logic, as it is more difficult to identify motifs that can be combined to produce the AND logic described above. Similarly, NOT gates can be implemented through integrating repressor-binding motifs in the promoter of the output gene and expressing synthetic transcriptional repressors that block transcription in the presence of a specific input (Figure 4B). Such promoter-based logic designs enable precise, context-dependent expression of genes, allowing plants to make finely tuned transcriptional “decisions” in response to complex environmental or developmental conditions.

A landmark study of synthetic promoter utility in plants, provided by Brophy et al., demonstrated how logic circuits could be constructed from synthetic promoters to reprogram root architecture in Arabidopsis (Brophy et al., 2022) (Figure 5A). In this study, the authors developed a library of synthetic TFs capable of activating or repressing expression in a controlled, tissue-specific manner. By generating and leveraging synthetic promoters containing TFBSs recognized by synthetic regulators and placing them upstream of a 35S minimal promoter, Brophy et al. constructed AND, OR, NOT, NAND, NOR, IMPLY, and NIMPLY logic gates to enable combinatorial control of gene expression. These gates were first validated in N. benthamiana leaves using GFP reporters before a subset of them were deployed in Arabidopsis stable transgenic lines. While showcasing the application of these gates, the authors were able to successfully modulate expression of the dominant solitary root (slr) mutant gene in Arabidopsis specifically within lateral root stem cells, resulting in precise alterations to lateral root density without affecting other root phenotypes, such as primary root growth or root hair density (Brophy et al., 2022). The authors highlighted that some of the logic gates required iterative optimization to achieve the desired spatial and temporal expression, underscoring both the challenges and the modularity inherent in synthetic promoter design. Two years later, another study by Khan et al. deployed a CRISPR interference (CRISPRi)-based gene circuit platform to construct reversible Boolean logic gates, particularly NOR, in Arabidopsis protoplasts and stably transformed plants (Figure 5B). These gates were then utilized to build complex logic circuits to confer NIMPLY and AND gate logic (Figure 5B). This work further demonstrated the utility of logic gates in programming desired, controlled plant responses by showing that combining gates, NOR gates in this case, to create circuits that impose logic functions that are generally more difficult to implement biologically (e.g., AND gates), can enable the rise of tighter, more specific, and non-leaky expression profiles (Khan et al., 2025). Both studies provide practical examples of genetic logic gates, the basic building blocks for the construction of complex synthetic circuits, functioning as intended in vivo to alter specific plant traits.

Building on this foundational work, Kong et al. established a predictive framework in plants to reliably generate and scale the development of promoter-driven genetic circuits in as little as 10 days (Kong et al., 2025) (Figure 5C). In this study, a synthetic cytokinin-inducible promoter TCSn and a native auxin-inducible GH3.3 promoter were employed alongside orthogonal sensors and NOT logic gates to create a variety of genetic circuits capable of integrating two inputs (i.e., cytokinin trans-zeatin (tZ) and the synthetic auxin naphthaleneacetic acid (NAA)) to control the transcriptional output of a firefly luciferase (LUC) reporter. Synthetic promoters responsive to either auxin or cytokinin were utilized to drive expression of a transcriptional repressor designed to regulate LUC. The initial gate was designed to express the reporter gene only if neither NAA nor tZ were present by leveraging a transcriptional repressor (LmrA) to control LUC expression, creating a NOR gate (Figure 5C). This gate was validated in transient expression assays in Arabidopsis protoplasts. Simultaneously, a computational model was created to predict expression of this gate, and later other logic gates, under different input conditions (Kong et al., 2025). Various genetic circuits were tested transiently in Arabidopsis protoplasts in the presence and absence of repressor expression, where they behaved as predicted by the model, demonstrating the model’s high prediction accuracy. A subset of gates were adapted and implemented in stable Arabidopsis transgenic lines to test whether these gates could be utilized to reprogram Arabidopsis roots, with the NAND gate in particular showing strong suppression of reporter expression only when both inputs were present (Figure 5C). The authors then tested an OR gate in vivo to control cell death in Nicotiana benthamiana. They leveraged a NUCLEOTIDE-BINDING LEUCINE-RICH REPEAT RECEPTOR (NLR) gene driven by an inducible promoter to induce a plant hypersensitive response after exposure to either NAA or tZ (Figure 5C). Cell death was significantly higher when either or both inducers were present as compared to samples that had no added inducer, indicating that the OR gate was implemented successfully (Kong et al., 2025). This study serves as an indicator that promoter engineering is beginning to move beyond simple expression tuning and towards predictable phenotypic reprogramming.

A particularly exciting example of employing logic gates for controlled expression is a recent study that created a biosensor for phenylpropanoids, a class of plant metabolites (Ferreira and Antunes, 2024) (Figure 5D). Ferreira & Antunes created Boolean operators in plants that sense phenylpropanoid-related metabolite levels, implementing transcription-based AND, NAND, IMPLY and NIMPLY gates to control downstream expression in a biosensor system. This study used bacterial repressor allosteric TFs (aTFs) coupled with synthetic plant promoters that incorporate TFBS for aTFs and transcriptional activators to modulate gene activity in a ligand-specific manner, where levels of phenylpropanoid metabolites were leveraged as input to the genetic circuit (Ferreira and Antunes, 2024). In the default state of the biosensors created in this study, a constitutively expressed aTF binds to a response element (RE) in a promoter driving a firefly LUC gene, preventing reporter expression (Figure 5D). When phenylpropanoid metabolites are added, those metabolites bind to the aTF and prevent the TF from binding to its respective RE, therefore enabling LUC expression. Multiple aTF-metabolite pairs were integrated in various combinations and structures to create the Boolean operators, such as the NIMPLY circuit under which repression can be switched off with the addition of the phenylpropanoid metabolite naringenin, which were tested and validated in Arabidopsis protoplasts and N. benthamiana leaves (Figure 5D). By integrating logic gates with biosensors, researchers could program metabolic or signaling pathways to activate expression only under specific, defined conditions (Ferreira and Antunes, 2024).

In plant biotechnology contexts, synthetic promoters have emerged as powerful tools for engineering crops and enhancing agronomic traits. Over the past five years, several notable examples have demonstrated how promoter design enables precise tuning of complex traits, from metabolic rewiring for yield improvement to inducible systems that provide targeted resistance against pests and pathogens. The following studies illustrate how synthetic promoters can be tailored to the unique challenges of crop engineering, offering fine-grained control over gene expression that traditional promoters cannot achieve.

One of the longest-standing applications of promoter engineering in crops is fruit-specific expression. Ripening is an ideal target for trait engineering in crops because it is highly consequential for crop quality. The tomato E8 and E4 promoters are classical examples and both are strongly induced during ripening in response to ethylene (Deikman et al., 1992; Xu et al., 1996). Researchers have leveraged these promoters to create a hybrid, synthetic promoter combination to drive ripening-stage expression of transgenes. According to a United States patent published in the year 2000, an E8–E4 hybrid promoter was designed and seemingly leveraged by commercial companies to reduce ethylene biosynthesis in specific fruit tissues (Bestwick and Kellogg, 2000). Though this patent has since expired, this hybrid promoter is a prime early example of commercial companies utilizing synthetic promoters for a direct application in crops.

The use of synthetic promoters also enables inducible metabolic production, in which plants can generate compounds of interest in a controlled manner. Forestier et al. demonstrated this in Nicotiana benthamiana using a heat-inducible promoter system to drive biosynthesis of casbene, a compound that is utilized in the pharmaceutical industry (Forestier et al., 2023). In plants containing a cassette of four casbene biosynthesis genes driven by a heat-inducible synthetic promoter, casbene accumulated substantially when the plant was exposed to elevated temperatures (40 °C for 2 hours) while expression was low in control conditions (22 °C) (Forestier et al., 2023). Although demonstrated in a model host, this system provides a proof-of-concept for the application of synthetic promoters to induce compound production or engineer metabolites in plants, showing how researchers could leverage this system to engineer crops to produce specialty metabolites on demand, thereby avoiding unintended consequences that may decrease crop viability.

Pathogen resistance traits offer another application for synthetic promoters in crop engineering. Sultana et al. identified CREs responsive to soybean cyst nematode (SCN) and constructed synthetic promoters, containing four copies of SCN-responsive CREs (i.e., 4×M1.1 and 4×M2.3 fused to a minimal 35S core), that were active only during nematode infection (Sultana et al., 2022). In transgenic soybean assays, these promoters showed negligible background expression but strong induction at infection sites, particularly in root tissues, while showing no induction under abiotic stress (e.g., high salinity, drought, and cold) (Sultana et al., 2022). Similarly, another study engineered a synthetic pathogen-inducible promoter for potato, 2xS-4xD-NpCABE_core, incorporating tandem cis-elements fused to a core promoter from Nicotiana plumbaginifolia (Kauder et al., 2025). When used to drive expression of a modified version of the pathogen-inducible a virulence gene Avr3a in potato lines also containing the potato resistance gene R3a (as R3a is required for triggering cell death in cells that detect the Avr3a gene), the promoter conferred resistance to Phytophthora infestans, a causal agent of late blight, through programmed local cell death. The optimized version of the synthetic promoter provided robust inducibility while maintaining low basal activity under non-infected conditions, thereby avoiding mass cell death that may be associated with constitutive expression of Avr3a (Kauder et al., 2025). Aside from providing an application of synthetic promoters in the engineering of plant defense responses in crops, this study reinforces the design principles discussed earlier, that promoter performance depends on the choice and arrangement of CREs, as well as the minimal core promoter used.

In future endeavors, the scalability and effectiveness of crop engineering through the use of synthetic promoters will depend on the expansion of promoter libraries, systematically characterizing synthetic promoter performance in crops rather than model systems, and developing design rules for synthetic promoters in order to achieve predictable function across species. For a more extensive overview of synthetic promoter applications prior to 2023, see Yasmeen et al., which surveys promoter construction strategies and applications across various plant systems (Yasmeen et al., 2023).

As previously mentioned, the necessary first step in designing promoters is building a library of the short DNA motifs (TFBSs) that TFs recognize. Computational discovery of CREs has a long history and has benefited from both motif-discovery algorithms that utilize information gained from high-throughput plant TF-TFBS binding assays. Since their inception, motif discovery tools have established themselves as an integral part of CRE extraction. Early in silico prediction of CREs was limited due to a lack of known motifs across plant genomes (Rombauts et al., 2003). This limitation arose from the absence of sequencing data and centralized annotation databases of motifs rather than a lack of computational power and strategies, as algorithms developed in the early 2000s are still being leveraged decades later (Ho and Geisler, 2019). The classic Multiple Expectation maximizations for Motif Elicitation (MEME) Suite contains software tools like Find Individual Motif Occurrences (FIMO) that enable de novo motif discovery and motif scanning across promoters (Bailey et al., 2009; Grant et al., 2011). These tools are routinely used in plant research to identify which CREs are overrepresented in the promoters of co-regulated genes. For a more detailed review on using computational tools to identify CREs in plants, see the review by Zemlyanskaya et al., 2021 summarizing bioinformatics approaches that can be utilized to study transcriptional gene regulation (Zemlyanskaya et al., 2021).

Building on these foundational motif-discovery approaches, newer computational frameworks have begun to leverage deep learning (DL) to extract regulatory information directly from sequence data. iCREPCP is a DL-based platform that enables the identification of CREs within plant core promoters, trained on plant promoter activity datasets from Arabidopsis, maize, and tobacco (Deng et al., 2022) (Table 2). This program uses convolutional neural networks (CNNs) to identify individual bases or base pair windows that best inform promoter strength. CNNs can also aid in identifying proximal positional binding preferences of TFs, which can further inform where CREs should be placed within a promoter to promote gene activation (Rockenbach et al., 2025). By identifying the key sequence features that drive promoter activity, DL algorithms like iCREPCP can aid in guiding the design of synthetic promoters by streamlining CRE discovery in plants.

One recent study has extended the use of AI-based computational approaches beyond individual motif identification to the broader concept of creating compact enhancers for crop engineering (Yao et al., 2026). Yao et al. (2026) curated a set of almost 7000 natural short transcriptional enhancers (STEs) from maize, wheat, tomato, and soybean using a high-throughput screening platform, STEM-seq, which evaluated tens of thousands of candidate genomic elements and revealed thousands of functional enhancers with a wide range of activation levels. Building on this large data source, the authors developed BaseSearch, an AI-driven computational design framework that generated 5000 synthetic STE candidates, 455 of which were functional (1.0- to >2.0-fold activation), a success rate 11.4x higher than genome-wide screening alone could provide (Yao et al., 2026). Ten of those STEs exhibited very high activation, with the most active STE reaching 64.5-fold activation. These STEs were all, however, constitutive enhancers, with the authors noting that compact enhancers with spatiotemporal specificity may require more complex validation platforms, highlighting an existing limitation that can be explored further in future studies (Yao et al., 2026). By integrating high-throughput experimental data with AI-guided sequence design, this study exemplifies how computational modeling can accelerate the identification and creation of novel compact CREs for crop engineering.

In contrast to CRE-focused methods, sequence-based approaches aim to establish predictive relationships between entire promoter sequences and total expression output. For example, Siwo and coauthors developed an algorithm to predict gene activity in yeast when given a particular DNA promoter sequence (Siwo et al., 2016) (Table 2). This algorithm was developed in response to an open community challenge by the Dialogue for Reverse Engineering Assessments and Methods (DREAM) consortium aiming to decipher the relationship between promoter sequence and gene expression. Particularly for this challenge, promoter sequences of yeast ribosomal protein (RP) genes were leveraged to understand the features in the promoter’s DNA sequence associated with expression patterns of a target gene. Since the promoter sequence was defined as the 1200bp upstream of a gene, the promoter region was segmented into 100bp non-overlapping windows to predict activity. The authors found success in utilizing support vector machine (SVM) regression to predict promoter activity along with sequential minimal optimization (SMO) to train the model, achieving a correlation coefficient of 0.65, higher than the other 20 teams that participated in the challenge. Though this model was originally developed and trained on yeast RP gene promoters, in the future, it could also be utilized to predict plant-specific promoter activity as long as there is a robust training set. Interestingly, Siwo et al.’s model focused solely on the 100bp region upstream of the gene rather than individual TFBSs in the promoter region. As such, the authors indicated that the incorporation of TF and nucleosome binding data into their model boosted the model’s predictive power, highlighting the importance of TFBS-independent feature identification in addition to traditional methods that focus on TFBS-based promoter activity prediction.

In the past few years, frameworks like Extended Vision Mutant Priority (EVMP) have been developed to strengthen machine learning (ML) models in prediction performance of promoter strength (Yang et al., 2023). These types of tools do not generally transfer well into plant applications, as the model requires high-throughput generation and characterization of synthetic promoter variants for its training dataset, a laborious process when conducted in vivo in plants. This bottleneck highlights the need for computational approaches that are plant-specific or more easily applicable to plant systems. Recent studies have begun addressing this gap by developing plant-specific DL tools. Peleke et al. developed a CNN to accurately predict gene expression profiles from gene flanking regions in Arabidopsis, tomato, sorghum, and maize (Table 2) (Peleke et al., 2024). The resulting DL model, which leveraged publicly available large-scale genomic data as its training dataset, achieved prediction accuracy exceeding 80%, relying only on upstream and downstream proximal cis-regulatory regions for prediction of gene expression profiles. Another study developed two DL models to predict expression of either long (120kb) or short (3kb) genomic sequences in maize, rice, tomato, or Arabidopsis in order to guide gene editing of promoters in plants (Table 2) (Qiu et al., 2025). These models allow for the identification of important CREs and prediction of their functional roles through in silico mutagenesis, which enables the prediction of expression changes associated with the point mutations made. Although these models are more geared towards regulating transcription through promoter editing in a genomic context, these algorithms can also possibly be utilized to gain more information on how to design a synthetic promoter to have desired patterns and levels of gene expression.

A model that is especially applicable to plants was developed by Washburn et al. in 2019 that sought to leverage DL to estimate relative gene transcript abundance by analyzing the associated DNA sequence (Washburn et al., 2019). Two models were created with both leveraging promoter and/or terminator sequences as input, one for single species applications to classify whether the gene expresses or not (called the pseudogene model) and the second to compare relative expression levels of two orthologous genes in multi-species applications (called the ortholog contrast model). These models were tested utilizing Zea mays and Sorghum bicolor data, with the pseudogene model reaching a total of 86.6% accuracy when both promoters and terminators from Z. mays were considered. The second model revealed various Pearson correlation coefficients depending on the source of ortholog pairs (i.e., comparing two Z. mays sub-genomes (0.59) or comparing Z. mays with S. bicolor (0.78)) (Washburn et al., 2019). These findings suggest that DL models leveraging promoter and/or terminator DNA sequences can capture informative features relevant to gene transcript abundance in plants, while also noting that species-specific factors and context may influence the model’s predictive performance. Applications of promoter strength prediction outside of plant research, particularly in mammalian studies, can provide some insight into the direction the plant researchers could take. Fu et al. developed a DL model to predict the transcriptional activity of B-cell specific promoters in human and mice (Table 2) (Fu et al., 2023). To train this CNN-based model, these authors computationally designed over 23,000 promoters by incorporating and shuffling around motifs from natural promoters driving Immunoglobulin genes into computationally generated background sequences (i.e., natural promoter sequences that were shuffled around). The promoters were designed by utilizing the MEME suite to identify motifs upstream of target genes and subsequently mapped to motif databases (e.g. JASPAR) (Bailey et al., 2009; Rauluseviciute et al., 2024). These synthetic promoters were then tested by massively parallel reporter assays to identify the features from promoter sequences that play a role in transcriptional regulation (Fu et al., 2023). The model was later used to predict the effect that polymorphisms in immunoglobulin V gene promoters would have on transcription of the gene across the global human population, leading to a better understanding of B-cell specific regulation mechanisms. The success of this pipeline highlights potential paths that researchers in the plant community can take to advance our understanding of plant-specific transcriptional mechanisms. By designing a targeted library of synthetic promoters, testing those promoters using massively parallel reporter assays, and feeding this large dataset into a DL model, we can accomplish multiple objectives at once: generate a diverse set of synthetic promoters with a targeted function for use in various applications; create a prediction model to accurately identify expression levels of a given promoter sequence; and deepen our understanding of how to engineer desired spatiotemporal patterns of expression accurately in plants.

Another study developed a generic framework based on DL methods that included a de novo synthetic promoter generation model alongside a transcriptional activity prediction model (Table 2) (Seo et al., 2023). A set of almost 7,000 cyanobacteria-derived natural promoters was utilized as a training set for the variational autoencoder (VAE)-based generative model to design 10,000 synthetic promoters for Synechocystis sp. PCC 6803, a model unicellular cyanobacterium (Seo et al., 2023). The promoters were then inputted into a CNN-based model to predict promoter strength and experimentally validated using a cell-free transcription assay. After refinement, the predictive model was able to reach a Pearson correlation coefficient of 0.57 in predicted versus tested transcriptional activity levels. Both the generative and predictive models were designed such that they could be implemented in other species, particularly in non-model organisms, by virtue of using a cell-free system to validate transcriptional activity (Figure 6). However, VAEs may be limited in their ability to generate long, diverse promoter sequences with accurately predicted levels of activity due to limitations in latent space, making VAEs more suitable for generating short promoter sequences (Wang et al., 2025). Similar workflows leveraging VAEs for promoter generation and CNNs for activity prediction could, in principle, be deployed directly for plant synthetic promoter engineering with minimal architectural changes. Thus, while not plant-trained, the study demonstrates an additional conceptual pathway for translating computational models developed in non-plant species into plant-specific synthetic promoter generation and validation.

To enhance generation of promoters with predictable expression and refine synthetic promoter engineering methods, we can turn our attention to computational approaches that focus on promoter classification and discrimination. Researchers have developed ML and DL models that classify given sequences into promoter or non-promoter categories, like DeePromoter and iProm-Zea (Table 2) (Kim et al., 2022; Oubounyt et al., 2019). One study trained computational models on either yeast, Arabidopsis, or human data, then evaluated them in a within-species and a cross-species manner to determine model performance (Bhandari et al., 2021). The observed decline in performance underscores the variation in promoter architectures across species, suggesting that these models should ideally be trained on data from the target species to ensure reliable predictions. Models like iPromoter-ET take this classification a step further, utilizing a two-layer predictor to first identify whether a sequence is a promoter or not, then determine the predicted strength of the promoter (Table 2) (Liang et al., 2021). Synthetic promoter design in plants can benefit from these types of models by feeding into them proposed synthetic promoter sequences and observing whether the model identifies the sequence as a promoter, as well as predicted promoter strength.

Combining both promoter generation with discrimination methods allows for the emergence of a framework that supports iterative model optimization. A recent study reported a novel promoter design method called PromoDGDE developed to address the need for intentional design of promoters with varying transcriptional strengths (Gu et al., 2025). This study combined an altered generative adversarial network (GAN) based method with an evolutionary algorithm and reinforcement learning to create optimized constitutively expressing synthetic promoters for E. coli and Saccharomyces cerevisiae (Table 2). Essentially, the GAN model is made up of two neural networks: a generator (G), that generates fake promoter sequences, and a discriminator (D), that discriminates between real promoters and the ones generated by the generator, in order to learn the true distribution of features in real promoter data. The synthetic promoters generated by this module are then fed into an optimization module leveraging evolutionary algorithm with reinforcement learning. Their transcriptional strength is predicted using a module that incorporates both a CNN and a long short-term memory (LSTM) network to help extract both local and long-range features that affect transcriptional activity, effectively increasing prediction accuracy. The prediction module allows for dynamic optimization of synthetic promoter design, as the optimization module can learn from the results of the prediction module through RL and, in turn, refine synthetic promoters to produce desired levels of expression. When tested in vivo, over 60% of the promoter sequences behaved as predicted, with a Pearson’s correlation coefficient of 0.79 and 0.9 for E. coli and S. cerevisiae respectively. However, this model does not take into consideration prior information and thus may fail to capture weak features during generation that may affect overall model performance. Despite this, the adaptive and optimization-oriented framework of PromoDGDE offers a compelling model for future efforts in plant synthetic promoter design. Theoretically, the plant research community could begin moving in this direction by tweaking optimization parameters and tailoring the training datasets to plant-specific constitutive promoters, then possibly narrow the training dataset to a particular promoter type for more specialized applications, as long as there is a large enough dataset to train the model without overfitting the data.

The integration of these diverse computational strategies with traditional experimental validation illustrates a shift in the advancement of promoter engineering (Figure 6). Studies in yeast have particularly highlighted the importance of this integration, with one DL model called CRMnet utilizing data from millions of experimentally validated promoters to pre-train the model before fine tuning (i.e. transfer learning) (Ding et al., 2023a; Dhaval Vaishnav et al., 2022; de Boer et al., 2020). Some recent studies in E. coli have even begun to move past activity prediction and towards generating target strength promoter sequences, showcasing the precision of DL models trained on large-scale data (Zhou et al., 2025). The successes achieved in non-plant systems can provide a blueprint for generative and predictive algorithms that accelerate the development of synthetic promoters with desired patterns and levels of expression. As plant-specific datasets grow, and model architectures become more transferable across species, these computational approaches are positioned to enable rational, guided, and efficient design of synthetic promoters in plants.

In addition to promoter sequence-based activity prediction, considering the effects of epigenetic information can improve prediction accuracy of transcriptional output. In the modern day, chromatin accessibility assays such as ATAC-seq and DNase-seq are widely used in plants to annotate candidate regulatory regions and to differentiate motifs that fall within accessible versus not accessible sites in the tissue or conditions of interest. ATAC-seq has been utilized across various organs and cell types in Arabidopsis, maize, rice and other species to reveal both conserved (across species) and species-specific regulatory landscapes. Since promoter activity is correlated with local chromatin accessibility and histone modifications (e.g., H3K4me3), integrating ATAC-seq, ChIP-seq, and DAP-seq data into ML models that predict promoter design would greatly increase applicability of synthetic promoters that leverage native TFs. Models like CharPlant and Predmoter predict open chromatin or histone modifications directly from sequence, offering insights into where motifs are most likely to be functional (Table 2) (Kindel et al., 2024; Shen et al., 2021). These models are trained on DNase-seq, ChIP-seq, and/or ATAC-seq cross-species data and aim to narrow the search space of condition and tissue-relevant TF-TFBS pairs. Other types of models enable the integration of chromatin accessibility with CRE prediction by taking both a DNA sequence and a vector of epigenomic features as input and outputting expected promoter strength (Zhao et al., 2021). Motif enrichment in differential elements of accessibility (MEDEA) is a computational tool that evaluates high-throughput chromatin accessibility data to identify TFBS enrichment within accessible regions of the genome (Mariani et al., 2020). This helps identify biologically relevant TFBS that can be leveraged in synthetic promoters. Models that do not take into consideration epigenetic information will likely exhibit reduced predictive accuracy and produce promoters with limited generalizability across tissues, developmental stages, and environmental conditions. Tools incorporating chromatin accessibility are a useful resource for plant synthetic promoter design, as they can identify elements that look promising in sequence space but will be inaccessible due to blockage by heterochromatin in the tissue of a target species.

In the context of promoter design, DBTL aims to develop potential promoters using computational modeling, build synthetic promoters or promoter libraries, test their activity experimentally, and learn from the data gathered by feeding it back to the model to improve predictions (Figure 7) (Vollen et al., 2024). This creates a recursive cycle of creating and testing synthetic promoters in an informed manner. To continue advancing the field, DBTL pipelines should be the next focus in plant promoter engineering (Wang and Demirer, 2023). Innovations such as high-throughput transient assays (e.g., protoplasts, leaf infiltration), massively parallel reporter assays, and improved DNA part assembly methods allow for rapid experimental testing of promoters, while computational approaches can be utilized to reduce the number of wet-lab experiments needed to identify synthetic promoters that function as expected. As previously mentioned, there has already been some work done to quantify the expression of synthetic part libraries using transient protoplast assays and dual-luciferase reporters, with transient characterization accurately predicting behavior in stable transformants (Schaumberg et al., 2016). These quantitative datasets make it possible to build trained models to inform promoter design based on predicted promoter activity, closing the DBTL loop by guiding experimental studies (i.e., representing the ‘Learn’ in DBTL). Over time, this iterative cycle continuously improves model performance, making it easier for researchers to construct synthetic promoters that show desired behavior in subsequent iterations.

One web-tool called PromoterCAD, created in 2013 but no longer available, utilized experimental data to inform data-mining tools that could identify CREs in Arabidopsis and provided a user interface (UI) for the design of synthetic promoters (Cox et al., 2013). PromoterCAD was a direct example of utilizing computational approaches to guide promoter design, representing the ‘design’ and ‘build’ parts of the DBTL cycle. In more recent years, GPro, a Generative Artificial Intelligence (AI) toolkit for promoter design, has been developed and made accessible to the public (Table 2) (Wang et al., 2024). This tool has been trained on both prokaryotic and eukaryotic sequences to generate synthetic promoter sequences and their predicted expression output. Additionally, the developers have documented an extensive wiki to enable researchers to customize and train the model for their specific needs, thus making it potentially useful in plant promoter design. This tool directly supports the DBTL framework, as the ability to customize the model allows researchers to revise the model based on experimental studies, representing the ‘learn’ part of the cycle.

Reddy et al. developed a workflow using conservative objective models (COMs) for model-based optimization (MBO) to discriminate between promoters specific to three different cell lines, highlighting a path to designing cell-type specific promoters (Reddy et al., 2024a). The models utilized were pretrained on large promoter-driven expression datasets from massively parallel reporter assays and then fine-tuned with a much smaller cell-specific promoter-driven expression dataset. The authors tested their model experimentally and found that their workflow was able to help redesign the promoters in the smaller promoter-driven expression dataset in a way that improved cell-specificity for more than 70% of sequences for two of the cell lines. The most important aspect of this workflow is that it can be implemented iteratively, following the DBTL cycle to improve cell-specificity of promoters with every iteration. As this workflow utilizes data from massively parallel reporter assays and was implemented using cell lines to represent different cell types, this may be difficult to translate into plants. There are no stable, fully differentiated plant cell lines that can be leveraged to refine cell-specificity of promoters to the degree that other fields may be able to given the availability of well-characterized animal cell lines. This is a challenge to keep in mind when translating this method to plant-specific applications, but it does not preclude the plant community from attempting to use the workflow to improve specificity of relevant promoters with currently existing plant cell-type specific data or utilizing fluorescence-sorted protoplasts from well-characterized cell-type-specific marker lines.

By utilizing iterative DBTL cycles for promoter engineering, researchers can progressively refine computational model performance and expand experimental libraries, as well as increase the accuracy of in-silico prediction with in-planta performance. Models like the one created by Seo and coauthors (Seo et al., 2023) showcase how computational methods can quickly generate a large library of novel synthetic promoters, but validation of activity predication is limited to current experimental approaches. On the other hand, the ability of computational approaches to accurately predict expression of a given promoter sequence is often limited by the availability of large training datasets. This is where large-scale experimental studies come into play, providing the bulk of data fed to computational models that enable the rational design of promoters (Qi et al., 2022; Jores et al., 2021; de Boer et al., 2020). These promoters, in turn, get built and tested experimentally (via transient or stable transformation) (Figure 6). The resulting validation data provide a feedback platform for computational models to refine predictions and improve future promoter designs, effectively closing the DBTL loop between computational and experimental approaches (Figure 6). Thus, DBTL approaches enable synthetic promoter design to go from a trial-and-error endeavor to a rational, predictive, and rapid workflow.