Journal · Systems
CRISPR and Cas9: editing behaviour you did not write
The invention worth stealing from CRISPR is not the cutting. It is the split between the address and the worker. The first approved CRISPR medicine shows what that split buys you, a precise edit, and what it does not buy you, a predictable result.
We are an ERP practice reading public regulatory documents, not clinicians. Nothing in this post is medical advice. It discusses an approved therapy for a disease people live with, and anyone making a decision about treatment should be talking to a haematologist rather than reading an essay about software architecture.
CRISPR is usually introduced as a pair of molecular scissors. That is the least interesting thing about it, and it is not the part that transfers to any other kind of work.
The invention was a separation. In a CRISPR editing system the thing that decides where and the thing that does the work are two different components, and only one of them changes between jobs. Cas9 is the worker, a protein that cuts double stranded DNA, and it has no opinion about location. The address arrives separately, carried by a short guide RNA whose sequence matches the target site. Swap the guide and the same unmodified protein acts somewhere else entirely, within a limit worth keeping in mind: Cas9 only cuts beside a short sequence it recognises, so the separation is real but not total. The 2020 Nobel Prize in Chemistry went to Emmanuelle Charpentier and Jennifer Doudna, in the committee's words, "for the development of a method for genome editing". A method, not a molecule.
You can see how literal that separation is by reading a drug label. The first CRISPR based therapy approved in the United States is Casgevy, cleared by the FDA on 8 December 2023. The agency's own review document describes the product as the patient's own blood stem cells
genome edited ex vivo using clustered regularly interspaced short palindromic repeats-associated protein 9 (CRISPR/Cas9) and the SPY101 single guide RNA to disrupt the GATA1 transcription factor binding domain of the B-cell lymphoma/leukemia 11A (BCL11A) gene erythroid enhancer.
SPY101 is the address. It is named on the regulatory paperwork as a component of the medicine, sitting beside the protein that acts on it. The worker is generic and the aim is a listed ingredient. That is the whole idea, shipped.
Why you edit the instruction and never the product
A protein's behaviour is not written down anywhere inside the protein. It is a consequence. A gene specifies a sequence of amino acids, that chain folds into a shape, and the shape presents a small number of positions that do nearly all of the useful work: a pocket that admits one molecule and refuses another, a surface that docks onto a partner, a site that changes the whole shape when something binds there. Behaviour is what that geometry does in a crowd.
Which leaves two ways to change what a protein does. You can design a different protein, which means choosing a sequence that folds the way you intended, and the field has spent decades establishing how hard that is. Or you can change the instruction and let the cell build the altered version using machinery it already has. The second is incomparably cheaper, and gene editing is simply that: an edit to the instruction, with the existing factory left in place.
It is worth noticing how the field handled the part it could not predict. The 2018 Nobel Prize in Chemistry was divided, one half to Frances Arnold "for the directed evolution of enzymes". Directed evolution is, in effect, an admission: rather than reason out which changes will produce the behaviour you want, you make many variants, measure them, keep the ones that work and repeat. The prize went to a method for getting useful proteins without being able to predict them.
The approved therapy does not repair the broken gene
Sickle cell disease has about the simplest cause a genetic disease can have. The FDA's review states it plainly: the disease is caused by "a point mutation substituting valine for glutamic acid in the sixth codon of the beta-globin gene", which produces the sickling form of haemoglobin. One gene, one codon, one substituted amino acid, although the trial population also included people carrying a second, different mutation on the other copy. The obvious plan is to go and fix that codon.
Casgevy does not do that. It never touches the faulty gene. The cut is made in a regulatory region of an entirely different gene, BCL11A, whose job in developing red cells is to switch off fetal haemoglobin, the oxygen carrier a person makes before birth and stops making in infancy. Disrupt that switch and fetal haemoglobin comes back. The FDA's description of the chain is mechanical: reduced BCL11A gives "an increase in gamma-globin expression and HbF protein production in erythroid cells", and in patients with severe disease that fetal haemoglobin "reduces intracellular HbS concentration, preventing the red blood cells from sickling and addressing the underlying cause of disease, thereby eliminating VOCs".
We have quoted that sentence to its full stop deliberately. An earlier draft of this post stopped it at "sickling", which was the convenient place to stop, because the rest of the sentence is the regulator using language very close to what a reader means by a cure. An adversarial review of this post caught it. The argument below survives the full quotation and is better for it, but you should know the cut was there and who found it.
Read that again as an engineering decision. The patient already knows how to make a working oxygen carrier. They stopped as a baby, on instruction. So the intervention does not repair the broken part, and does not install a new part. It deletes an instruction to stop, and an existing capability returns.
The best available edit was not at the broken place. It was at a control point that governed the behaviour. Finding that point was the hard part, and no amount of precision in the cutting would have found it.
What the measurement actually says
Here the house habit of reading the primary document earns its keep. The approval rests on a single arm trial with a defined primary outcome, called VF12: the absence of protocol-defined severe vaso-occlusive crises for at least twelve consecutive months inside a twenty four month follow up. Of the patients treated, the FDA's review reports 44 treated, 31 with enough follow up to be evaluable, and 29 of those 31, or 93.5%, meeting the endpoint, with a one sided 98% confidence interval running from 77.9% to 100%.
The denominator has a detail in it. One further patient died at month 8.9, of respiratory failure that the investigator attributed to COVID-19 pneumonia and busulfan lung injury, busulfan being part of the preparation rather than the edited cells. Under the trial's own statistical plan a death related to the whole treatment regimen before the endpoint was reached would have counted as a non-response. The FDA agreed to the sponsor's request to exclude that patient from the efficacy analysis, and gave its reason: the difficulty of determining, among multiple comorbidities, how much the busulfan had contributed to the death. The 93.5% is correct, and it is 29 of 31, and that is how it was built.
Now the word everyone reaches for. Big Think published a video with Doudna whose page summary describes the therapy as "functionally curative" and notes the roughly two million dollar price. The video has no transcript on the page, so we cannot tell you whether that is her phrase or the editor's, and we are not going to attribute it to her. What we can do is count. In the FDA's 6,889 word Summary Basis for Regulatory Action, the word "cure" appears exactly once, and it is not about this medicine; it refers to donor stem cell transplantation being a possible cure for the minority of patients who have a matched donor.
An earlier draft of this post leaned on that number. It should not have, and we
are leaving the correction visible rather than quietly deleting the paragraph. A
Summary Basis for Regulatory Action is a review memorandum organised around
endpoints, benefit and risk. It is not a document in which the agency would call
any product a cure, so the absence is a feature of the genre and discovers nothing
about this therapy. The word count also used a flag that hides a second mention:
one line earlier the same document calls existing drugs "noncurative", which
grep -w cannot match inside. And roughly 540 of those 6,889 words are
trade-secret redaction markers.
pdftotext casgevy_label.pdf casgevy_label.txt grep -o -i -w 'cure\|cured\|curative' casgevy_label.txt | sort | uniq -c wc -w casgevy_label.txt
So here is the honest version, which is harder and more interesting than the word count. The same regulator, in the same document, wrote both things: a mechanistic claim that the therapy addresses the underlying cause of the disease, and a measured endpoint of twelve crisis-free months in 29 of 31 evaluable patients. The mechanism is the reasoning. The endpoint is the evidence. They are not the same kind of statement, and when a result travels outward it is the mechanism that survives the retelling and the denominator that gets lost. That is the gap worth caring about, and it is not a gap between the regulator and the press. It runs through the middle of the regulator's own document.
Three things the precision does not buy
🔴 It does not buy you a local consequence. The FDA names the main danger without hedging: "The major risk of treatment with CASGEVY is the potential for off-target, unintended genome editing by CRISPR/Cas9." The approval came with postmarketing requirements to study exactly that, and long term safety "including risk of malignancy". A precisely aimed edit can still land somewhere nobody enumerated, and at approval that possibility could not be ruled out, so it was converted into an obligation to keep watching. Be careful how you read that: the confirmatory testing the agency did run found no off-target editing, and the postmarketing requirement exists because only a limited number of people were sequenced, not because anything was observed. The risk is unquantified rather than demonstrated, and those are different things.
🔴 It does not buy you a cheap installation. The edit is small. Getting it into a person is not. The patient's own stem cells are collected, edited outside the body, and returned only after full myeloablative conditioning, which means the existing bone marrow is destroyed first. Serious adverse reactions were seen in 45% of the sickle cell patients. Vertex set a US list price for Casgevy, the wholesale acquisition cost rather than what any particular patient or insurer pays, of 2.2 million dollars; bluebird bio set 3.1 million for Lyfgenia, the other therapy approved the same day, which uses a viral vector rather than CRISPR. The elegance of the edit says nothing about the cost of installing it.
🔴 It does not buy you a predictable behaviour. Precision in the edit and predictability of the outcome are different properties, and only the first one improved. That is why directed evolution won a Nobel Prize, and why the approved therapy required a trial with a measured endpoint rather than an argument from mechanism.
The same shape, in software you did not write
Everything above has a counterpart in the least glamorous work there is: changing the behaviour of a business system somebody else built.
Open source ERP platforms are built on exactly the CRISPR separation. In
Frappe, which is what ERPNext runs on, you do not modify the application to make
it behave differently. You register an address and a handler: a document type and
an event, such as "before a Sales Invoice is submitted, run this". Frappe's
documentation calls the mechanism doc_events, declared in your own
app's hooks.py, which means the platform's code is never touched.
The framework's own dispatcher is the unchanged worker. Your hook is the guide. The
upgrade next quarter replaces the application and leaves your edit standing, for
as long as the document type and the fields it reads are still there, because
your edit was never inside the application. That caveat is not decoration. A hook
survives an upgrade only while the names and the event contract it depends on
survive, and some jobs cannot be done in a hook at all.
The alternative is to fork: take the source, change it, and run your own version. That is designing a new protein. It works once, and then every upstream improvement becomes a merge you pay for, forever.
| In the therapy | In software you did not write |
|---|---|
| A named guide RNA, SPY101, listed as an ingredient | A hook: one document type plus one event name |
| Cas9, one worker, unchanged between jobs | The framework's own dispatcher, unchanged |
| The instruction in the cell, edited once | Your customisation layer, which upgrades survive |
| The protein, built by the patient's own cells | The behaviour your staff actually see on screen |
| Designing a new protein from first principles | Forking the application and owning it forever |
| Off-target editing, named by the regulator as the major risk | The report that breaks three modules away |
| Myeloablative conditioning, required before the infusion | The migration, the downtime window and the retraining |
| A 12 month measured endpoint, with a stated denominator | Real transactions run through a copy before go-live |
The three limits transfer as cleanly as the architecture does, and they are the ones we watch for.
The consequence is not local. A validation hook added to one document type to enforce one rule is precise in exactly the way a guide RNA is precise, and just as capable of surprising you. The invented but entirely typical case: a firm adds a rule that a sales order cannot be submitted without a customer purchase order number. Correct, cheap, three lines. Six weeks later the monthly revenue figure is wrong, because a recurring subscription job had been submitting orders with no purchase order number for two years, and now fails silently in a background queue. Nobody enumerated that caller. That is off-target editing with a different vocabulary.
The installation is the project. The edit is an afternoon. The migration of historical records, the week where both systems run, the people who have to be shown the new screen, and the month where the old report and the new report disagree: that is the cost, and it is the part that gets left out of the estimate.
The fix is often not at the fault. This is the one worth sitting with, because it is the one that saves real money. When a business says its margins are wrong, the margin report is often not where the edit belongs. Frequently something upstream is switched off: nobody records scrap, or landed freight never reaches the item cost, or two part numbers mean the same thing. Editing the report makes the number change. Editing the switch makes the number true. Casgevy is the best illustration we have come across of the difference, and it was worth a Nobel Prize and twenty years of work to find the switch.
What this does not establish
An analogy is not evidence. Nothing in the FDA's review of a medicine tells you anything about your software, and we are not claiming it does. The narrow claim is that the same architecture appears in both, that the architecture brings three specific limits with it, and that naming them in the domain where they were studied carefully makes them easier to see in the domain where they are usually discovered by accident.
This is not medical advice and we are not qualified to give any. We are an ERP practice reading a public regulatory document. Nothing here is a recommendation about any treatment. Anyone living with sickle cell disease, or deciding about a therapy, should be talking to a haematologist rather than reading an essay about software architecture.
We did not review the trial data. Every clinical figure above is read out of the FDA's own Summary Basis for Regulatory Action, which is linked below, and we have quoted it rather than paraphrasing where the wording carries weight. We did not watch the Doudna video; the phrase "functionally curative" is from the summary copy on the page, which has no transcript. The ERP example is invented, with round numbers, and is drawn from the general shape of the problem rather than from any client of ours. No client's figures appear in this post.
If this is your problem
Most requests for a customisation are really a request to change a behaviour, and the useful first question is never "where do we put the code". It is "which control point governs this, and how will we know it worked". Those are the two questions this post is about, and they are worth answering before anyone writes code.
That work has a name and a price on our services page: discovery and scope, which starts at $19,700 and is credited in full against the implementation that follows. It is paid work, not a free assessment, and it exists because scoping a system blind is how projects go wrong. If that is the shape of what you are facing, tell us what the system is doing that it should not.
Questions people ask
What is CRISPR, in plain terms?
CRISPR is a way of making a targeted change to DNA. It uses two parts that stay separate: a short guide molecule whose sequence matches the place you want to change, and a protein, commonly Cas9, that cuts DNA wherever the guide takes it. Because the targeting is carried by the guide rather than built into the protein, the same protein can be aimed at a different site by supplying a different guide. Emmanuelle Charpentier and Jennifer Doudna shared the 2020 Nobel Prize in Chemistry for developing the method.
Does the approved CRISPR therapy cure sickle cell disease?
The FDA did not approve it on a claim of cure. The approval rests on a measured outcome: the absence of severe vaso-occlusive crises for at least twelve consecutive months, met by 29 of 31 evaluable patients out of 44 treated. In the FDA's Summary Basis for Regulatory Action the word cure appears once, about donor stem cell transplantation rather than about this product. Researchers and the press often describe the result as a functional cure. We are not qualified to judge whether that description is right, and this post does not try to. What we can say is what was measured, which is the twelve month endpoint above. Note also that the same FDA document states the therapy addresses the underlying cause of the disease. This is not medical advice.
Which gene does Casgevy edit?
Not the faulty one. Sickle cell disease is caused by a point mutation in the beta-globin gene, and Casgevy leaves it alone. The edit disrupts a transcription factor binding site in the erythroid enhancer of a different gene, BCL11A, which normally switches off fetal haemoglobin after infancy. With that switch disrupted, fetal haemoglobin returns. What was measured in the trial was freedom from severe pain crises for at least twelve consecutive months in 29 of 31 evaluable patients, not the elimination of sickling as such. This is not medical advice.
Does CRISPR edit proteins?
No. CRISPR systems act on nucleic acid, not on a folded protein. You change the instruction and the cell builds the altered product with the machinery it already has. A protein's behaviour changes as a consequence, which is why gene editing is a practical lever on protein behaviour even though it never touches a protein. Changing a protein directly is a different discipline, and its best known method, directed evolution, works by making many variants and measuring them rather than by predicting which one will work.
What does gene editing have to do with business software?
They share an architecture. A hook in a system like ERPNext is an address plus an action, registered outside the application, which is the same split that makes CRISPR programmable. The useful part is that the same three limits follow in both: the consequence of a precise edit is not necessarily local, the cost of installing the edit dwarfs the cost of making it, and precision in the edit does not give you predictability in the behaviour.
Should we customise our ERP or change how we work?
Ask which control point governs the behaviour you want, then ask what the edit there would cost to install and how you would measure whether it worked. Sometimes the control point is in the software and a hook is the cheapest honest answer. Often it is upstream of the software, in what gets recorded and by whom, and no amount of configuration will fix it. The mistake is editing the report that shows the problem rather than the thing that causes it.
Sources
- U.S. Food and Drug Administration, Center for Biologics Evaluation and Research. Summary Basis for Regulatory Action, CASGEVY (exagamglogene autotemcel), BLA STN 125787, 8 December 2023. Every clinical figure and quotation in this post is from this document, read on 2 November 2026. https://www.fda.gov/media/175179/download
- U.S. Food and Drug Administration. FDA Approves First Gene Therapies to Treat Patients with Sickle Cell Disease. News release, 8 December 2023. https://www.fda.gov/news-events/press-announcements/fda-approves-first-gene-therapies-treat-patients-sickle-cell-disease
- U.S. Food and Drug Administration. CASGEVY product page, approval letters and package inserts. Read 2 November 2026. https://www.fda.gov/vaccines-blood-biologics/casgevy
- The Nobel Prize in Chemistry 2020, awarded jointly to Emmanuelle Charpentier and Jennifer A. Doudna for the development of a method for genome editing. NobelPrize.org, read 2 November 2026. https://www.nobelprize.org/prizes/chemistry/2020/summary/
- The Nobel Prize in Chemistry 2018, one half awarded to Frances H. Arnold for the directed evolution of enzymes. NobelPrize.org, read 2 November 2026. https://www.nobelprize.org/prizes/chemistry/2018/summary/
- Big Think, The Well. CRISPR is capable of more than we initially expected, a video with Jennifer Doudna. The page carries no transcript; the phrase quoted in this post is from its own summary copy, read 2 November 2026. https://bigthink.com/the-well/crispr-is-capable-of-more-than-we-initially-expected/
- Reuters, via Yahoo Finance. US FDA approves Vertex/CRISPR gene therapy for an inherited blood disorder, 8 December 2023, for the 2.2 million and 3.1 million US list prices. https://ca.news.yahoo.com/us-fda-approves-vertex-crispr-201916253.html
- Frappe Framework documentation. Hooks: registering handlers against document events without modifying the application. https://docs.frappe.io/framework/user/en/python-api/hooks