I enjoyed reading your paper. It had me thinking about the range of applications that duckweed offers as a model system.

A few questions: 1. I am curious how you settled on the 3h post-treatment sampling window? Was this based on a pilot time-course, or derived from prior literature on plant stress kinetics? 2. The histone kinase enrichment specific to TEMP+NH4 (Suppl. Fig. 6) is striking but undiscussed. Any hypothesis for the NH4-specificity? If there's any accessibility signal near these genes in your ATAC data, it could be a nice bridge between the transcriptional and chromatin sides of the story.

Gene expression optimization is an important aspect of protein engineering in plants. I really like the core idea here, particularly the emphasis on balancing multiple objectives rather than simply maximizing CAI. I had a few questions and comments while reading the paper:

1. What is the intended use case for PDD? Is it primarily for replacing endogenous genes with codon-optimized versions in the same host, or for transferring trait genes across species? The benchmarking suggests the former, since all six genes are rice genes optimized for rice. However, codon optimization should theoretically be most useful when expressing genes across divergent species.

2. Would you consider a different term than “effector protein”? “Effector protein” has a fairly specific meaning in plant pathology, so something like “trait gene” might be clearer here.

3. How do you envision optimizing expression level? PDD uses complete cassettes designed for strong constitutive expression, but some of the highlighted traits have known growth or yield tradeoffs when constitutively overexpressed. Is expression strength itself intended to be an optimization objective, or is strong constitutive expression simply the current implementation?

4. Could the benchmarking be extended to a eudicot? Since PDD is presented as applicable to 18 crops spanning both clades, a benchmark on at least one dicot would strengthen the case that the multi-objective advantage generalizes beyond the current rice/monocot examples.

5. I was confused about the CRISPR component of PDD. The guide design appears aimed at knocking out traits, which seems somewhat at odds with the “add a trait” framing elsewhere in the paper. Is the goal to remove these traits from their native species? I’d also appreciate clarification on the off-target scoring: are guides evaluated against reference genomes for each species, or only against the input CDS? If the latter, I’m not sure how useful this metric is within the PDD pipeline.

This looks like a nice demonstration of how flipping the genotype to phenotype approach can be applied to an agricultural system, and the null-distribution validation is a nice rigorous touch. A few questions:

1. Fig 6/Table 1: Did you run the same hybrid-vs-baseline architecture comparison for the other three known SNPs (ant.loc9, pale.loc4, serr.loc5)? The hybrid model's speed/performance tradeoff is a nice result, but currently rests on a single locus. Does it hold up across all four?

2. In regards to the 50-SNP screen: two SNPs clearing the 95th-percentile null threshold out of 50 is close to what you'd expect by chance alone (~2-3 false positives expected at α=0.05). Thoughts on applying an FDR correction? This would be especially important if the strategy was expanded beyond 50 to (potentially) 1000+ SNPs.

3. As a follow-up: any guesses on what those two SNPs actually are? Do their genomic neighborhoods contain genes with functions plausibly connected to the patterns you're seeing that could explain the "two plausible novel associations"? I understand this may be restricted given the consortium data-sharing agreement, but curious if you can share even a high-level idea.

4. One limitation of the current design, if I understand correctly, is the GPU-hour bottleneck that seems to come mostly from retraining the full CNN from scratch per SNP. I am curious about your thoughts on overcoming this problem in order to make this tractable at scale?

Thank you for sharing this preprint. I think this work highlights an often overlooked design consideration with broad relevance to protein engineering and construct design. I had a few questions/comments:

1. The six-frame ORF analysis in Figure 1 is the conceptual centerpiece of the paper, but I could not determine which specific eGFP sequence was used as the input. Because the presence of hidden ORFs is highly sequence-dependent (as your avGFP comparison nicely demonstrates), including the exact sequence in the supplement would help readers better contextualize the findings and evaluate their generalizability.
2. I would be interested in seeing the size distribution of hidden ORFs across the ~6,000 plasmids analyzed. A histogram of hidden ORF lengths, perhaps separated by CF+2 and RF categories, could help readers better understand the realistic range and prevalence of unintended translation products.
3. Do you have any additional insight into the lack of an RF signal on the immunoblot? For example, do you think this is primarily due to a lack of transcription, inefficient translation, rapid proteasomal degradation, or some combination of these factors?

This is one of the most compelling aspects of the manuscript to me. Gaining broad adoption really depends on lowering costs, reducing failure points, reducing the need for specialized equipment, and, importantly, speed.

Are these efficiencies the results of a single experiment or averages of multiple experiments? How much experiment to experiment variation would you expect to see?

I think it is interesting that this method is incompatible with cell-wall-less strains. Do you know why that is?

You have listed 6 strains as being tested, but only have data for 4A-. Is data for the remaining strains going to be included as supplemental data in the final pub? Are there any differences in transformation efficiency between strains?

Enhanced mating efficiency at 28 °C is an intriguing observation; however, the current experimental design doesn’t distinguish between two plausible explanations: (1) increased secretion of gametolysins, MMPs, and related factors at 28 °C directly enhances mating competence, or (2) broader physiological changes associated with acclimation to 28 °C (e.g., altered membrane properties, flagellar remodeling) are the primary drivers, with changes in the secretome being correlative rather than causal.

To disentangle these possibilities, have you considered a reciprocal autolysin transfer experiment, similar to the approach described in the Bio-protocol publication by Findinier 2023 (DOI:10.21769/BioProtoc.4705)? In this design, autolysin preparations from cells grown at 18 °C and 28 °C would be cross-applied to gametes conditioned at each temperature, generating four conditions: (i) 18 °C autolysin + 18 °C gametes; (ii) 28 °C autolysin + 28 °C gametes; (iii) 18 °C autolysin + 28 °C gametes; and (iv) 28 °C autolysin + 18 °C gametes.

If the secreted proteome is the primary determinant of enhanced mating efficiency, then 28 °C-derived autolysin should increase mating efficiency regardless of the temperature at which the recipient gametes were produced. In contrast, if physiological acclimation is dominant, mating efficiency should track with the growth temperature of the gametes rather than the source of the autolysin. This framework would also allow assessment of potential synergy between these factors, with the strongest increase in mating efficiency observed in the matched 28 °C condition relative to either of the reciprocal treatments.

I see the appeal of identifying evolutionarily agnostic regulatory activity and appreciate the effort to tackle this goal. I was hoping you could speak toward the magnitude of the effects observed in the initial screens.

In the plant and human screens, the largest log fold changes reported appear relatively modest. Do you think this reflects limitations of the screening platforms themselves, the size or composition of the candidate CR library (e.g., that stronger regulators may not have been captured), or might it suggest that CRs lose some regulatory potency when moved outside of their native evolutionary or chromatin context?

The seed-coat is T0 maternal tissue, so how would this be helpful to determine segregation of the transgene in the T1 seeds?

Earlier in the paragraph you suggest that T-DNA molecules are undergoing concatemerization prior to chromosomal integration, rather than multiple independent partial T-DNA insertions. However, if this was the case, then the mutlple transgene copies would be linked and segregation would occur as if there was a single insertion event.

Based on your segregation data, is it possible to determine if the RUBY observed in the seeds is being expressed in the T1 seeds, or if it is being expressed in T0 maternal floral tissue and being transferred/leaked into the seed?

Since editing continues through T0 generation development, which developmental stage did you use for checking T0 CRISPR editing efficiency? How soon is too soon, and is there a recommended tissue and/or developmental time point to perform the T0 screening?

It's exciting to see that Cas9 can be linked to RUBY using P2A, and I can see multiple benefits. Not just for selection, but it also helps reduce the size of the binary vector for plant transformation. With that in mind, since ZmUbi1 promoter is used for GIF1 and Cas9-RUBY, is there a reason for the current design over ZmUbi1::GRF-GIF1-P2A-CAS9-P2A-RUBY?

I’m excited to see the development of a commercial V. natriegens for the community to try out. I had a few questions arise when reading your paper.

First, in the introduction section you mention “This can be scaled from small culture tubes in the laboratory (to generate microgram quantities of DNA), to large-scale industrial bioreactors (to generate kilograms of DNA for vaccine production).” I am wondering where you envision the biggest benefit occurring? Would it be mostly for commercial DNA manufacturing? For general cloning in a research lab, I’m not sure how much I’ve ever worried about E.coli’s slower doubling time for my typical cloning workflow, since I often start cultures in the evening and let them grow overnight. With this system, I’m envisioning you would more likely to be starting cultures in the morning and harvesting in the afternoon. It would be nice to see a figure demonstrating what some hypothetical cloning workflows (with timelines) would look like to have a better understanding of how much time could be saved.

Second, I am wondering if you tested transformation of any larger plasmids. Most plasmids you tested ranged from ~3kb to 6kb. However, some of the larger plasmids seem like they may have a growth penalty (Figure 6) and a decrease in transformation efficiency (Fig 9). Would these possible phenotypes be more pronounced if you are transforming a plasmid that is 20-25kb, which is about the max that E.coli can support?