Jennifer Doudna and colleagues used an AI model system to make better CRISPR/Cas gene-editing tools, but the results are underwhelming. Report: Claire Robinson*
For many years GMWatch has made the science-based case that the process of CRISPR/Cas gene editing is neither precise in operation nor predictable in outcome. In doing so, we have gone against the much-hyped narrative that this technology is precise, efficient, and predictable.
Now, in what has become a typically cynical pivot by the pro-GMO lobby and its attendant media, it’s being admitted that CRISPR/Cas gene-editing tools are not precise after all and produce unintended effects both at the intended edit site (on-target) and elsewhere in the genome (off-target).
However, this admission has come about only in the context of gene editing scientists supposedly coming up with a “solution” to the problem – in the shape of CRISPR pioneer Jennifer Doudna and her team using AI with the aim of designing more specific and efficient CRISPR/Cas gene-editing tools.
Based at the Innovative Genomics Institute at UC Berkeley, USA, which Doudna founded, the researchers wanted to see if their AI model system approach could build custom gene-editing tools that operate more efficiently and with less off-target alterations.
How gene editing works
Gene-editing tools use a type of protein called a nuclease as “molecular scissors” to cut the DNA across its double strand at a pre-determined sequence. This nuclease activity is provided by the “Cas” component of the CRISPR/Cas gene editing tool. Components of the Cas protein, working together with an RNA guide sequence, steer the nuclease activity to the DNA sequence that is intended to be cut. After the cut is made, the DNA is repaired by the cell’s own DNA repair mechanism. The gene editing scientist aims to have the repair incorporate either a DNA deletion, insertion of additional DNA base units, or some other type of modification, which will confer a new or altered trait in the organism.
However, both the targeting and the repair process are not fully under the control of the gene editing scientist and are prone to errors, with many off-target and unintended on-target changes occurring alongside the intended ones. These include off-target cuts by the gene-editing tool, as well as repair errors: small and large deletions, insertions, and rearrangements of the DNA, and even chromothripsis, a catastrophic shattering of a chromosome and its random rejoining.
The CRISPR/Cas gene-editing tools and derivatives thereof (e.g. base and prime editing) are artificial adaptations of a system in bacteria which these organisms use as a defence mechanism against infection from pathogens, especially viruses. Therefore the CRISPR/Cas enzymes used in gene editing are not the same as the natural versions. While scientists started working with the natural system, they heavily modified both the enzyme and its guide RNA to make them functional inside human, animal, and plant cells.
What Doudna’s team did
Doudna’s team used AI tools to try to design more efficient, non-natural “synbio” nucleases that would work better with the RNA guide, in terms of being more precisely targeted and producing fewer unintended changes. To design the nucleases, the team defined which parts of the Cas protein needed to remain fixed, according to evolutionary data from this family of nucleases, which showed highly conserved and thus functionally crucial regions that needed to remain intact. Then they asked the AI model to redesign the rest of the protein, generating new variants of nucleases.
The resulting synbio nuclease variants were tested in bacterial, plant, and human cells. The researchers reported that some of the synbio nucleases had editing ability equal to or greater than their “unimproved” (“wild type”) counterparts. The wild type Cas in this research was a naturally occurring enzyme taken from a bacterium. The synbio nuclease with the most editing activity shared only 77% sequence identity with the wild type, suggesting that the AI-assisted design had made a significant alteration.
These findings, published in the top-flight journal Science, are intended to demonstrate that CRISPR-related nucleases can be extensively redesigned while preserving function. The AI model platform attempts to restrict changes only on parts known to be vital for function (see above), leaving other regions free for extensive redesign.
The authors foresee that possible end uses for such AI-assisted nuclease design could include personalised medicine applications. But as they tested the synbio nucleases in plants, it is clear that improving the efficiency of gene editing in plants is an additional goal.
What’s novel about the Doudna team’s approach?
Many scientists now use AI to design and improve CRISPR tools. But in the “old” way of designing gene editing nucleases, scientists took an existing natural nuclease and tried changing a few parts to see if it improved. In contrast, Doudna’s team told the AI model what the final, perfect shape of the nuclease needed to look like. The AI then worked backwards to write a completely new recipe (amino acid sequence) to build that exact shape.
Hype
Predictably, given Doudna’s pioneering hero status in the world of gene editing, her team’s findings have been uncritically hyped in the media. The science and technology website BioDigital hailed the results as a “breakthrough” that “signifies a major advancement in protein design and generative biology, offering novel tools for genetic therapies and agricultural innovation”.
Fierce Biotech said, “Unlike other [gene editing] methods that repurpose enzymes found in nature, this new platform is capable of crafting proteins that are far afield from those forged by millennia of evolution.” The outlet quoted Doudna as saying, “We were able to develop non-natural nucleases that were active in human, plant and bacterial cells. This approach opens the door to the possibility of designing an enzyme on demand for a particular problem, whether it’s for treating a genetic disease or helping crop plants adapt to a changing climate.”
However, a closer look at the new research suggests that the significance of the Doudna team’s innovation may be exaggerated. To see why, it’s necessary to put aside the media reactions and dive into the detail of the published paper. What follows is technical, but it explains why the published paper does not fully justify the headlines it has attracted.
The shortlisted variants
Out of many thousands of synbio nuclease sequences generated by the AI model, Doudna’s team found just 466 that were biologically active enough in bacterial assays to target genetic sequences, though only around 8% were better than the “wild type”. Out of these, just nine of the “most diverse and active variants” (v1 to v9) were shortlisted for testing in human and plant gene editing assays.
The findings in human cells – and the problems with them
The core data pertaining to targeting gene editing in human cells is presented in Fig. 3 of the paper. This figure shows the efficiency of the editing of the nine new AI-designed variants in human cells (HEK293T). Initial experiments targeting a blue fluorescent protein (BFP) transgene in these cells showed that only the v1 and v5 nuclease variants were able to knockout this gene with an efficiency greater than the wild type (up to 50% compared with 28% for the wild type). The other seven variants were either no better or worse.
Next, the authors targeted four host cell genes with the same nine novel nuclease variants, again comparing them to the wild type. As in the case of targeting the BFP transgene, only the v1 and v5 novel nuclease variants gave gene editing efficiencies that were greater than that with the wild type Cas nuclease.
However, the data shown in Fig. 3 shows that the efficiency of targeting with the v1 and v5 Cas variants varies with the gene being targeted. In the case of v1, it was more efficient to a variable degree in three out of five targeted genes, whereas with v5, efficiency of targeting was greater than wild type in just two out of five of the targeted genes.
Furthermore, according to our calculations based on Fig. 3, the degree of increased efficiency over wild type varied from 1.39- to 3.8-fold for v1 and 1.79 to 3.1-fold for v5. Statistical analysis showing significant increased editing over wild type is only described in the main text for the targeting of the EMX1 gene, which was 3.8- and 3.1-fold greater with v1 and v5 respectively compared to the wild type nuclease.
Since no data are provided, are we to conclude that for the other three genes targeted, there was no statistically significant difference between the action of the v1/v5 variants and the wild type nuclease? Notably, the statistical analysis presented in Fig. 3C compares the gene editing efficiency of the v1-v9 AI model variants and wild type nucleases to cells treated with a Cas nuclease-guide RNA combination that does not target any of the test genes. Therefore the crucial statistical analysis, which compares the efficiency of the v1-v9 variants with the wild type nuclease for all four genes targeted, is missing.
Based on this information, it appears that it is not possible to predict how efficient any AI-model Cas variant will be in targeting a given gene.
The findings in plant cells
In Arabidopsis protoplast plant cells, only one of the synbio nuclease variants (v1) was reported to perform better than the wild type. The other eight performed considerably worse than the wild type, an important fact that was relegated to the online Supplementary Materials accompanying the main paper.
Nevertheless, even the star performer, v1, showed only questionable improvement over the wild type. A single gene (AtPDS3) was targeted at four different locations with the v1 variant and wild type nucleases in combination with four different guide RNAs. In the main text, the authors state that consistent with the human cell results, v1 outperformed the wild type at nearly all tested targets (Fig. S13 and Table S1). However, on scrutiny of the data in Fig. S13, it is evident that there is a large overlap in the standard deviation (experimental variation between separate experiments) at three of the target sites in the AtPDS3 gene, suggesting no difference in efficiency of gene editing between v1 and wild type nucleases. At the fourth target site there was a clear decrease in efficiency of 53% with the v1 variant.
Furthermore, the data presented in Table S1 is also revealing. First, the baseline efficiency of editing in the Arabidopsis protoplast cells by the wild type nuclease is very low at all four targeted sites of AtPDS3, ranging from 0.31% to 1.4%. Second, the fold changes in editing efficiencies between v1 and wild type presented in the table are unimpressive, with editing at just one target site possibly being greater with v1 compared to the wild type, with a marginal increase from an average of 1.4% to 2.08%. On this basis, the statement in the main text that “v1 outperformed [wild type] at nearly all tested targets” is not supported by the data presented.
In addition, the v1 variant had undesirable off-target activity at a comparable level to the wild type (see main text).
Taken together, these findings question the value of this approach in a plant context.
Success? Not so fast
Some might see the generation of one synbio nuclease that marginally outperforms the wild type as a success, on the grounds that it could avoid gene editing scientists having to go down a thousand dead ends in the lab before they come up with one nuclease that efficiently “edits” genomes. However, there are major caveats.
1. The AI model is not a whole biological system, so what the computer thinks will work well doesn’t always pan out in practice. This may explain why so many of the Doudna team’s synbio nuclease designs showed promise according to the computer predictions but very few worked well in the actual targeted cells.
One gene editing scientist we contacted for their reaction responded: “You can’t predict the activity of the whole by manipulating a part, since the holistic functioning of any system, even at a complex molecular level such as proteins, is always greater than the sum of its parts. I am therefore not surprised by the high failure rate of the AI model designed Cas nucleases.”
They also asked, “How is the AI model of Doudna and colleagues superior to the directed iterative (repeating) and high-throughput evolutionary selection process in living organisms? In this process, you trial an initial large pool of randomly generated Cas nuclease variants in the type of cells (plant, human, or bacterial) you are targeting for editing. Having identified a few candidates of Cas nuclease variants with greater efficiency, these are further mutagenised to generate variants with even greater efficiency.
“This process is repeated (hence ‘iterative’) until you arrive at the desired efficiency end point. You can build a rapid readout (e.g. fluorescence) so it’s easy to see which cells have been edited successfully and with greater efficiency than your wild type control.
“This is a proven approach that is fast and easy, and which earned its inventors the Nobel Prize in 2018. And you have the benefit of knowing from the start that the winning candidates work successfully in the whole biological organism you intend to target for gene editing.
“Is it surprising that most AI model-designed Cas nuclease variants, selected for greater efficiency of targeting in the simple genomic environment of bacteria, which is uncluttered with structural proteins, turned out to be ineffective when targeting within the complex chromatin environment of mammalian and plant cells?”
Extending this point, the Doudna team’s AI model has not been shown to work in developing an actual gene-edited product that functions as desired.
2. AI is “trained” by humans, and what comes out of it is governed by what is put in. Since there is a lot we still don’t know about the genome and how gene editing works within this context, the products of an AI system will necessarily be limited by our ignorance.
3. Nuclease targeting is supposed to be improved by the Doudna team’s AI-assisted model, but this does not avoid the genetic errors caused by the rest of the gene editing process. However precisely the initial CRISPR is designed and targeted, all events after the CRISPR has cut the DNA occur independently of the CRISPR and are due to the cell's innate DNA repair machinery. So however much the gene editing scientist tweaks the CRISPR to make it more specific, they will be faced with the same spectrum of downstream problems, such as unintended small and large deletions, insertions, and rearrangements of the DNA. In addition to unwanted changes caused by the activity of the gene editing tool, gene editing of plants involves processes that cause additional unintended mutations, such as tissue culture and the GM transformation procedure. Standard screening methods often miss these structural changes. Doudna’s team did not succeed in reducing off-target effects, though they did succeed in improving editing efficiency in human cells.
4. Major divergence from natural constraints means major risks. In the view of the authors of the paper, divergence from natural gene-editing nucleases is a bonus, as it frees them up from natural limitations – “extending their functions beyond natural sequence constraints through generative biology”. But this will also increase risks beyond the natural constraints and even beyond the constraints of gene editing done with “unimproved” nucleases. An AI method designing proteins with a high number of changes increases the likelihood of unexpected outcomes.
Improving on natural evolution?
In her book A Crack in Creation, Doudna described natural evolution as a “deaf, dumb, and blind system”. She argued that CRISPR technology allows humans to replace unguided natural selection with intentional, conscious, and human-directed genetic design.
But is natural selection an unguided process? Research findings in bacteria and plants suggest otherwise, with the finding that the genetic variation arising from rounds of natural reproduction is not random, but a directed evolutionary adaptation response.
Now let’s apply this knowledge to the Doudna team’s new research. Their published paper, the Innovative Genomics Institute’s press release, and the media coverage all suggest that AI-assisted gene editing improves on the natural genetic sequences produced by evolution. But the results presented in their paper are patchy at best and it’s not yet clear how much of an improvement in gene editing efficiency their new AI model will offer.
We must be alert to the possibility that using AI systems to design CRISPR/Cas editing tools is just the latest example of the “precise until it isn't” narrative. Back in the 1990s we were told that the older-style genetic engineering, used for crops like GM soy and maize, was more precise than conventional breeding. Over the years, a massive body of evidence emerged showing it wasn’t precise at all. Then, when gene editing became the latest thing, the story became “older style genetic engineering wasn’t precise, but gene editing is”. But with gene editing too, there is already much evidenceto the contrary.
Only time will tell if the Doudna team’s AI model will solve some of the problems of the gene editing venture or whether it will turn out to be yet another over-hyped development.
The new paper: Skopintsev P et al (2026). Structure and evolution-guided design of minimal RNA-guided nucleases. Science 393(6808), 16 Jul. https://www.science.org/doi/abs/10.1126/science.aed6123?af=R
*GMWatch gratefully acknowledges technical assistance and input for this article from several scientists, who prefer to remain anonymous for the usual reasons.
