Altogether 42 U6 promoters and 24 U3 promoters from maize and six other monocotyledonous species were mined from public and Bayer proprietary databases as a training set for a proprietary generative machine learning model. An arbitrarily selected subset of endogenous U6 promoters were directly tested ( Supplementary Material S1 ). Three endogenous maize U6 promoters, one each from chromosomes 1, 2 and 3 (referenced below as Chr01, Chr02 and Chr03, respectively), were tested in different length variants (160 bp and 400 bp) for genome editing. Based on inspection for TATA, USE, and MSP elements, 160 bp was chosen for minimal length variants for these promoters. The 400 bp variants were included to test for other potential upstream element(s) that could contribute to the promoter activity. A chimeric derivative between Chr08 and Chr01 U6 promoters (160 bp) was used as a positive control ( Figure 1A ) across several experiments in this investigation.

Novel Pol III promoters were computationally derived as previously described for Pol II promoters (To et al., 2021). Briefly, from the training set promoters, we extracted genomic sequence from -500 to +50 bases relative to the transcription start site (TSS). We identified statistically overrepresented sequence motifs that are likely to contribute to promoter function using the POWRS v1.2 algorithm, using “—bins 5” and otherwise default settings (Davis et al., 2012). As expected, the primary motifs were the TATA box, the MSP, and the USE. Multiple sequence alignment using MUSCLE v3.8 (Edgar, 2004) confirmed that lengths of intervening sequences between USE and TATA and between TATA and TSS were extremely consistent. The endogenous monocot promoters and the putative sequence motifs were used to train a proprietary generative machine learning model. The sequence design algorithm has been described in detail (Davis and Elich, 2025). In brief, the algorithm uses a scoring function to estimate how promoter-like a sequence is. The function has terms for position-specific 6-mer frequency in the training set, position-specific 6-mer enrichment relative to non-training-set promoters, and dinucleotide entropy. Starting from a random sequence, we randomly placed overrepresented motifs, then iteratively improved the sequence according to the scoring function, via random mutation with simulated annealing (Kirkpatrick et al., 1983). The model pre-dates modern deep learning, and is similar to Naïve Bayes models. As such, the frequency and enrichment terms are calculated from the training sequences, but there is no test/train split or cross-validation metrics per se. Two models were produced, using either U6 promoters alone, or U6 and U3 promoters together. Novel expression elements generated by these models were screened for allergenicity and toxicity risks with bioinformatic tools before use (Codex_Alimentarius, 2009; To et al., 2021).

The CRISPR/Cas12a system from Lachnospiraceae bacterium ND2006 (LbCas12a; Zetsche et al., 2015) was used throughout this study. The LbCas12a coding region was codon optimized for maize ( Figure 1B ; Supplementary Material S2 ; Kouranov et al., 2022) and was flanked by two nuclear localization signals from the heat stress transcription factor 1 (HSFA1) gene of tomato (Solanum lycopersicum L.; Doring et al., 2000). For both protein-coding and non-coding transgenes, expression was conferred by a combination of gene expression elements that are collectively called a gene expression cassette. The Cas12a gene cassette was driven by a maize ubiquitin promoter (Zm.UbqM1; Christensen et al., 1992) and terminated by the 3’ untranslated region (UTR) of the rice lipid transfer protein (Os.LTP; Vignols et al., 1994). The crRNA cassettes for protoplast studies included the variable promoter sequences followed by a standard LbCas12a repeat (GAATTTCTACTAAGTGTAGAT), where the initial ‘G’ was added for proper expression by U6 promoters, a spacer corresponding to the genomic target sites, and a poly-T terminator ( Figure 1B ). In all constructs for for the multiplex editing of 30 targets, and for the constructs pM199, pM200, pM223, pM225, pM226, and pM230, all 23 bp spacer sequences were flanked with LbCas12a repeats on both sides. Sequences were either synthesized (BioBasic, Amherst, NY, USA; biobasic.com) or PCR amplified with Pol III promoters using Ultramer primers (IDT, Coralville, IA, USA; idtdna.com).

Protoplasts were isolated from leaves of etiolated seedlings of Zea mays cultivar 01DKD2. PEG-mediated protoplast transformation was previously described (Sheen, 2001). In each well, 0.8 pmol LbCas12a and 1.6 pmol crRNA plasmid DNA were co-delivered into approximately 320,000 protoplast cells. Two additional plasmids carrying Renilla and Firefly luciferase cassettes optimized for maize gene expression were included as transformation quality controls. Each transformation was replicated in four wells. The transformed protoplasts were incubated at room temperature in the dark for 48 h. Luminescence, indicative of protoplast transformation, was quantified in each treatment using Dual-Glo Luciferase Assay System (Promega, WI, USA; https://www.promega.com). Luminescence values were checked against control wells lacking transformed DNA within each experiment. For a treatment to pass this quality control step, the ratio of luciferase in the treatment compared to the no-DNA control needed to be at least 100 fold. Total genomic DNA was isolated from the entire cell suspension.

Mature seed embryo explants from the cultivar 01DKD2 were transformed via Agrobacterium as described previously (Ye et al., 2022). Briefly, binary vectors carrying the epsps-cp4 selectable marker, LbCas12a, and various crRNA expression cassettes were transformed into a VirG constitutive Agrobacterium strain using gentamicin 30 mg/L and kanamycin 50 mg/L for selection. After co-culture of mature seed embryo explants with Agrobacterium, the explants were transferred to consecutive multiple bud induction media. Primary transformant (R0) plants were selected and regenerated on a hormone-free medium containing 25 µM glyphosate. Transgene copy numbers were determined by TaqMan-based qPCR using the Os.LTP terminator (Vignols et al., 1994) of the LbCas12a cassette as an assay template. Leaf tissues were collected from regenerated seedlings at the V1 growth stage. DNA isolation and copy number assay were performed as described previously (Kouranov et al., 2022).

All genomic target sites used in this study are listed in Supplementary Material S3 . Valuable edits for biotechnology products may require targeting low efficiency sites. Therefore, high- and low-efficiency target sites in genic and intergenic regions across a range of chromosomes were used to test a variety of editing conditions. Non-essential, non-linked sites across chromosomes were selected to test multiplexed editing of many independent locations simultaneously. Regions around the target sites were amplified using either protoplast or plant genomic DNA samples as templates. Amplicon sizes were kept below 200 bp to ensure that a paired-end workflow generating 2x150 bp sequences, would cover the entirety of the target site from at least one direction. The amplicons were sequenced by the Illumina HiSeq platform (San Diego, CA, USA; illumina.com). Reads were trimmed and mapped to the reference sequences as previously described (Rymarquis et al., 2024). The reads carrying insertions or deletions (indels) in the target sites were counted, which was used as an indirect measure for chromosome cleavage. For both protoplast suspensions and transgenic plants, the indel rates were determined by using the following formula: Indel% = 100 × reads carrying indels/total reads. In protoplasts, these individual edit percentages were averaged among the four replicates for each treatment. In plants, the above-calculated indel percentages per plant were used as an input to further calculate editing rates at the population level using two different methods. In the first approach, referred to as “average indel rate in sequencing reads” below, individual editing rates from leaf samples were averaged over the population of plants. In the second approach, we estimated the proportion of the population likely to transmit the edits to the next generation, which we termed “advanceable.” This estimate was based on the proportion of plants with an indel rate over a 10% threshold. Maize plants with edits over this threshold are more likely to have heritable edits (data not shown).

Leaf tissues were collected from regenerated seedlings at the V1 growth stage. Direct-zol RNA MiniPrep Kits (Zymo Research, Irvine, CA; https://www.zymoresearch.com) were used to extract total RNA according to manufacturer’s instructions. A subset of RNA samples was run on 5300 Fragment Analyzer System (Agilent, Santa Clara, CA; https://www.agilent.com) to confirm RNA quality (RNA Integrity Number > 7). Expression levels from computationally derived promoters were measured in two approaches, as detailed below.

In the multiplex editing experiment, the pre-crRNA species, each carrying multiple tandem spacers targeting different loci, were quantified. Total RNA was converted to first strand cDNA using a random primer mix in the High-Capacity cDNA Reverse Transcription Kit (Applied Biosystems, Foster City, CA; https://www.thermofisher.com). Real-time PCR reactions were done using TaqMan (TM) assays. The TM primers and probes assayed three consecutive spacers on each pre-crRNA species. A dilution series of pooled cDNA was used to confirm amplification efficiency and a no template control was used to confirm specificity. Expression values for the RNA of interest from each sample were normalized to the geometric means of the expression values of the housekeeping genes eIF1A and EF1a from that same sample.

When testing for increased crRNA expression targeting a single locus, first strand cDNAs were generated from processed, mature crRNAs using specific reverse transcription primers and Custom TaqMan Small RNA Assay kits (Applied Biosystems, Foster City, CA; https://www.thermofisher.com). Real-time PCR reactions were done using TM assays. The TM primers and probes assayed mature crRNAs. A dilution series of synthetic RNA was used to confirm amplification efficiency and a no template control was used to confirm specificity. Expression values for the RNA of interest from each sample were normalized to the geometric means of the expression values of the housekeeping genes eIF1A and EF1a from that same sample.

Altogether 42 U6 promoters and 24 U3 promoters from maize and seven other monocotyledonous species were mined from public and Bayer proprietary databases as a training set for a proprietary generative machine learning model. A subset of 15 endogenous U6 promoters were directly tested ( Supplementary Material S1 ). Three endogenous maize U6 promoters, one each from chromosomes 1, 2 and 3 (referenced below as Chr01, Chr02 and Chr03, respectively), were tested in different length variants (160 bp and 400 bp) for genome editing. A chimeric derivative between Chr08 and Chr01 U6 promoters (160 bp) was used as a positive control ( Figure 1A ) across several experiments in this investigation.

To establish a baseline for promoter performance, LbCas12a indel editing rates enabled by maize U6 promoters were tested in maize protoplasts. The promoters were mined from U6 genes on chromosomes 1, 2, and 3, along with a chimeric U6 promoter from chromosomes 1 and 8 ( Figure 1 ; Supplementary Materials S1 ). Editing rates were evaluated at an intergenic (Zm.7.1b) and a genic (Zm.Bmr3_2691) site ( Supplementary Material S3 ). The chimeric Chr08:Chr01 promoter was used as an internal positive control for subsequent experiments. The three endogenous promoters were tested in two length variants ( Supplementary Material S4 ). Above-background chromosome cutting was detected with most promoter species. The Chr02 U6 promoter, however, resulted in low editing efficiency. For example, the Chr02 U6 160 bp variant resulted in 0.018 ± 0.013% indel (mean ± standard deviation) at the Zm.Bmr3_2691 site. Both length variants included all required U6 promoter elements (TATA box, USE, and MSP) for all four promoter species. While the two length variants performed comparably for both Chr01 and Chr02, the longer version (400 bp) of Chr03 performed significantly better than its shorter counterpart (160 bp) for the Zm.7.1b target site (2.01 ± 0.60% indel compared to 0.93 ± 0.23% indel; Kruskal-Wallis rank sum test followed by Wilcoxon rank sum exact test, p < 0.05).

Twelve U6 promoters from six monocots other than maize were tested in the same protoplast system. Editing rates were evaluated at an intergenic (Zm.7.1c) and a genic (Zm.Bmr3_2691) target site ( Supplementary Material S5 ). Most promoters performed comparably with maize Chr08:Chr01. However, Et_3478 from Eragrostis tef showed low performance for both Zm.7.1c (2.54 ± 1.32% indel) and Zm.Bmr3_2691 (0.051 ± 0.102% indel) target sites. Additionally, So_1047 from Saccharum officinarum showed low performance for Zm.7.1c (8.15 ± 1.25% indel).

Thirty-seven novel 500 bp Pol III promoters were computationally derived using a proprietary generative machine learning model (To et al., 2021; Davis and Elich, 2025; Supplementary Material S6 ). Editing rates were evaluated at an intergenic (Zm.7.1c), and a genic (Zm.Bmr3_2691) target site in protoplast assays ( Figure 2 ). In this experiment, most promoters were trimmed to between 280 and 300 bp in length. Trimmed promoters generally showed the same activity as the original 500 bp designs ( Supplementary Material S7 ). Distributions of deletions along target sites were compared among a subset of the novel Pol III promoters ( Figure 2A ). Eight to ten base pairs around the upstream nick site on the non-target strand were deleted most frequently for all promoters tested. Editing rates were compared to a Chr08:Chr01 U6 positive control, a negative control lacking a crRNA cassette, and two 300 bp randomized sequences. A promoter was counted as active if it performed at or above the lower bound of standard deviation for the Chr08:Chr01 U6 ( Figures 2B, C ). In total, 27 promoters were counted as active for both the genic and intergenic sites. Thus, the hit rate for derivation of novel crRNA promoters was 73%. The distributions of editing rates were similar between promoters derived from a training set of U6 promoters and those derived from both U6 and U3 promoters ( Supplementary Material S8 ).

Nine promoters, representing a broad range of activities in the protoplast results ( Figure 2 ), along with the chimeric Chr08:Chr01 control, were selected for in planta genome editing at the Zm.7.1c target site of maize ( Table 1 ). Leaf samples from R0 plants with novel crRNA promoters showed averages of 41.8-76.1% indels, as measured by sequencing reads. A plant was counted as “advanceable” to the next generation if more than 10% of the sequencing reads from a leaf sample showed insertion or deletion. Between 65.5 and 100% of plants with novel crRNA promoters were advanceable.

The editing rates determined in protoplast testing ( Figure 2 ) were compared to the in planta editing rates ( Table 1 ) to test correlation ( Supplementary Material S9 ). Of the two in planta editing calculation methods, the “average indel rate in sequencing reads” approach is more similar to the one used in the protoplast system. Therefore, we used this method and found a significant correlation between protoplast and in planta rates (Spearman correlation = 0.68, p < 0.05).

Five computationally derived Pol III promoters were used to drive the expression of five crRNA cassettes, each containing six unique spacer sequences, to target a total of 30 distinct genomic regions interspersed across all 10 maize chromosomes ( Supplementary Material S3 , Figures 3A–C ). Editing was observed at all 30 target sites among the 333 R0 plants that were transformed with this construct. Within a single plant, up to 27 sites were edited at an advanceable rate (>10% indel reads) when multiple copies of the transgene were present and up to 18 sites were edited at an advanceable rate when only a single copy of the transgene was present ( Figure 3C ). The expression of each unprocessed crRNA transcript was measured using RNA Taqman assays ( Figure 3D ). The editing rate for individual crRNAs ranged from 3.6% to 83.8% ( Figure 3E ). Expression was compared to a control construct containing LbCas12a and no crRNA cassettes. All five of the computationally derived Pol III promoters showed expression, although the levels varied.

To improve low editing rates, the same spacer was tested in various multiplex crRNA configurations. The spacer for the low-efficiency Zm.Bmr3_2691 target site was repeated multiple times in the same crRNA array driven by one promoter. This was compared to a configuration in which the same spacer was placed in separate cassettes driven by diverse Pol III promoters ( Figure 4A ). Multiple spacers in a single array did not improve editing rates in plants as compared to their singleton counterparts ( Table 2 ). However, stacking multiple crRNA cassettes resulted in a significant increase in mature crRNA accumulation (Kruskal-Wallis rank sum test followed by Wilcoxon rank sum exact tests, p < 0.01, as detailed in Figure 4B ) and increased the rate of advanceable edited plants from 0% to 10.4% ( Table 2 ). Selectable marker gene RNA expression levels, used as a control, were constant across constructs ( Figure 4B ). A construct with multiple spacers in a single array (pM226) did show significantly higher LbCas12a RNA expression but did not result in an increased editing rate or crRNA accumulation (Kruskal-Wallis rank sum test followed by Wilcoxon rank sum exact tests, p < 0.05).

In similar experiments conducted in protoplasts ( Figure 5 ) and plants ( Table 3 ), single-spacer cassettes driven by five computationally derived promoters were placed as singletons or in multiplex in the same constructs along with a single LbCas12a cassette as depicted in Figure 5A . The same trend was observed: using multiple Pol III cassettes increased editing rates ( Figure 5 , Table 3 ). Specifically, the double, triple and quadruple constructs all enabled significantly higher editing than their singleton counterparts (Kruskal-Wallis rank sum test followed by Wilcoxon rank sum exact tests, p < 0.05 or 0.06, as detailed in Figure 5B ). For the Zm.Bmr3_2691 target site, the double construct (pMON677) enabled 1.73 ± 0.08% indel, the triple construct (pMON678) enabled 2.05 ± 0.39% indel, and the quadruple construct enabled 2.12 ± 0.72% indel. In contrast, the most efficient singleton construct (pMON434), enabled 0.61 ± 0.16% indel.

Results from protoplasts showed that the addition of the second crRNA cassette contributed the most to editing rates. The GSP2239-driven cassette plus GSP2244-driven cassette was more than 20 fold more efficient than GSP2239-driven cassette alone, and more than 3 fold more efficient than GSP2244-driven cassette alone. The improvements diminished with the third and then the fourth added cassettes. For example, the triple was about 1.2 fold more efficient than the double. In contrast, the efficiency gains from these same constructs in generating advanceable plants were more steady. The GSP2239-driven cassette plus GSP2244-driven cassette was about 2 fold more efficient than GSP2239-driven cassette or the GSP2244-driven cassette alone. The triple was about 1.4 fold more efficient than the double.