25.4 — What Is Happening Right Now

In December 2023 an American regulator approved a treatment for sickle cell disease that works by cutting a patient's own DNA. Cells are taken out of the person's bone marrow, an editing tool snips a specific spot in the genome to switch a fetal haemoglobin gene back on, and the corrected cells are put back. The disease described in Chapter 2.8 — one wrong letter, inherited, lifelong — became, for those patients, something that could be edited.

That is the frontier this chapter maps. Some of it is genuinely delivering, some of it is early, and some of it is being sold far ahead of the evidence. Knowing which is which is a large part of being useful to a client, because your company will be asked to build systems for all three.

Cell and gene therapy

These are treatments where the product is either living cells or a piece of genetic material intended to change how the patient's own cells behave. There are three main shapes and they are frequently confused.

Gene therapy delivered into the body. A working copy of a gene is packaged inside a harmless virus shell, usually an adeno-associated virus, or AAV, and infused. The virus enters cells and delivers the gene, which then makes the missing protein. This is how treatments for some inherited retinal blindness and for spinal muscular atrophy work. One dose, potentially permanent, and priced accordingly — several of these are among the most expensive medicines ever sold.

Cell therapy made from the patient. The best-known is CAR-T, short for chimeric antigen receptor T-cell therapy. The patient's own T cells (Chapter 13.2) are collected, genetically modified in a laboratory so that they carry a receptor recognising a marker on their cancer cells, grown in number, and infused back. In some blood cancers this produces lasting remissions in patients who had run out of options.

Gene editing. Rather than adding a gene, the tool changes the patient's existing sequence. The mechanism is CRISPR, explained in Chapter 2.9. The sickle cell therapy above is the first approved example.

Now the part that matters for your work, because it is unlike anything else in this industry.

When the therapy is made from one patient's own cells, the batch size is one person. The chain runs: collect cells at a hospital, ship them frozen to a manufacturing site, modify and grow them over one to several weeks, test them, ship them back, and infuse them into the same person — who may be seriously ill and waiting. Get one identifier wrong anywhere in that chain and you infuse one patient's engineered cells into another, which is potentially fatal.

So these products need what the industry calls chain of identity and chain of custody: an unbroken, auditable link between the patient, every container, every process step and the final infusion, with temperature history attached. This is a software problem, it is a highly regulated software problem, and it is one of the most active areas of new system build in the whole industry. Chapter 25.12 covers the traceability rules these systems must satisfy.

mRNA and the idea of a platform

In 2020, vaccines using messenger RNA went from a sequence to authorised use in about eleven months. Understanding why that was possible tells you where a lot of investment is now going.

The mechanism is in Chapter 2.3: mRNA is the working copy of a gene that your ribosomes read to build a protein. An mRNA vaccine hands your cells the instructions for one viral protein, your cells make it, and your immune system learns to recognise it.

The obstacle for two decades was that injected RNA triggered a violent inflammatory response. Katalin Karikó and Drew Weissman showed in 2005 that swapping in a chemically modified building block avoided that reaction — work that won the 2023 Nobel Prize in Physiology or Medicine. The second obstacle was delivery, solved by wrapping the RNA in a lipid nanoparticle, a fat bubble that carries it into cells.

The strategic point is the word platform. In a conventional vaccine, changing the target means changing the whole manufacturing process. With mRNA, the chemistry of manufacture is largely the same whatever protein you encode: you change the sequence, not the factory. That is why the same technology is being pushed towards influenza, other infections, and personalised cancer vaccines where the encoded proteins are taken from a specific patient's own tumour mutations.

For a services company the interesting consequence is that a platform turns a manufacturing problem into an information problem. The sequence, its design history, the analysis that chose it, and the link between a patient's tumour data and their individual product all have to be handled under GxP rules. Chapter 25.21 is where the rules for that kind of computer system live.

Where the money actually went: obesity and metabolic drugs

The largest commercial story of recent years is not a frontier technology at all. It is a class of injectable drugs — the GLP-1 receptor agonists described in Chapter 22.8 — that imitate a gut hormone which signals fullness and improves insulin release.

They were developed for type 2 diabetes and turned out to produce weight loss on a scale nothing else had matched, which converted a diabetes drug class into a treatment for obesity, a condition affecting a large fraction of the population. Deloitte's 2024 analysis of large pharmaceutical companies attributed a substantial part of the improvement in industry returns to this class.

Two lessons come out of this that apply well beyond obesity drugs. The first is that in this industry the biggest wins often come from finding a second use for something you already have, which is why "lifecycle management" in Chapter 25.24 is a serious discipline rather than paperwork. The second is that demand at that scale breaks supply chains — these products need injection devices, sterile filling capacity and cold storage, and the shortages were manufacturing shortages. Capacity planning software is not a glamorous project, and it has been worth a great deal of money.

Artificial intelligence in drug discovery, honestly

This is the area where a services engineer is most likely to be asked "can we do that?" so it deserves a careful, non-promotional account.

What has genuinely changed is protein structure prediction. Determining the three-dimensional shape of a protein used to take a laboratory months or years. AlphaFold, from DeepMind, predicts structures from sequence with accuracy good enough to be useful, and the 2024 release, AlphaFold 3, extended prediction to complexes — a protein together with a small molecule bound to it, or with DNA or RNA. Since a drug works by fitting into a pocket on a protein (Chapter 22.1), a good predicted structure changes where a chemist starts. In early 2026 the model was being integrated into large pharmaceutical partnerships, and Isomorphic Labs released a combined design engine, IsoDDE, in February 2026, aiming for its first human trial of an AI-designed candidate by the end of that year.

What has partly changed is molecule generation and target selection. Generative models propose new chemical structures predicted to bind a target and to have workable drug-like properties. The furthest-advanced public example is rentosertib, from Insilico Medicine, where the target was identified using their AI system and the molecule was designed generatively. Its Phase IIa results in idiopathic pulmonary fibrosis were published in Nature Medicine and showed an improvement in lung function at twelve weeks in the higher-dose arm, and a Phase III trial was started, enrolling 320 patients across 47 centres in China.

What has not changed is the part that kills programmes. The 6.7 percent success rate from Chapter 25.1 is dominated by Phase II failures — drugs that hit their target exactly as designed and still do not help patients, because the target was the wrong idea about the disease. No current model reliably predicts that. A model can propose a molecule that binds; only a trial can tell you the disease cares.

So the honest summary to give a client is this: AI has compressed the discovery end, from years to months in places, and has not yet moved the clinical end, which is where most of the time and nearly all of the cost sits. Saying exactly that, rather than repeating a vendor's claim, is the difference between being trusted and being managed.

And the nearer-term AI work in this industry is much less glamorous and much easier to sell: reading adverse-event narratives to help code them (Chapter 25.22), finding patients who match a trial's eligibility criteria (Chapter 25.9), drafting sections of regulatory documents for human review (Chapter 25.24), reading scans as a decision aid (Chapter 25.16), and coding a clinical note for billing (Chapter 25.28). Every one of those sits inside a regulated process, which means the interesting engineering problem is not the model — it is proving to an auditor that the model's output was controlled, reviewed and traceable.

Precision medicine, and what it really means

The phrase gets used loosely. In practice it means one thing: treatment chosen using a measurement of the individual patient's biology rather than the average patient's.

Cancer is where it is furthest along. A tumour is sequenced, the specific mutations driving it are identified, and a drug is chosen for those mutations rather than for the organ the tumour started in. Some approvals are now written for a molecular feature across any tumour type, which was a genuine change in how a label can be defined.

It works through the companion diagnostic idea from Chapter 25.3: the drug and the test that selects patients for it travel together. The consequence for data systems is significant — genomic results have to reach the prescribing decision, be stored, and be interpretable years later when the interpretation of a variant may have changed.

Pharmacogenomics is the everyday version of the same idea. People carry different versions of the liver enzymes that break drugs down (Chapter 22.2), so the same dose produces very different blood levels in different patients. Testing before prescribing is already standard for a handful of drugs and slowly spreading.

Trial design and evidence are changing too

Two shifts are worth knowing because they generate a lot of system work.

Decentralised and hybrid trials. Parts of a trial that used to require a hospital visit — consent, symptom diaries, some monitoring — now happen at home through applications, connected devices and video visits. This lowers the burden on patients and widens who can take part, and it creates a hard regulatory question: how do you prove the person who entered that reading was the enrolled patient? Chapters 25.9 and 25.20 cover the controls this demands.

Real-world evidence. Regulators increasingly accept evidence from routine care — insurance claims, hospital records, registries — alongside trials, particularly for safety questions and for rare diseases where a large randomised trial is impossible. That turns the health data of Chapters 25.31 to 25.33 into regulatory evidence, which raises its quality requirements enormously.

How to judge the next thing you are shown

A simple set of questions separates a real advance from a pitch, and they work for any technology in this chapter.

What clinical evidence exists, and in how many patients? A mechanism, an animal study, or a computational result is not evidence of benefit.

Has it changed the failure rate, or only the speed of the early steps? Faster discovery of candidates that still fail in Phase II changes little.

Who pays, and does the payment system have a route for it? A one-time therapy costing a large sum is a problem for insurers built around monthly costs, and Chapter 25.27 explains why that alone can stall a genuinely effective product.

And can it be manufactured reliably? The penicillin story in Chapter 25.2 is the permanent reminder that the biology being right is not the same as patients getting it.

Next: Chapter 25.5, the beginning of the pipeline — how anyone works out the root cause of a disease in the first place, which is the step everything in this Part depends on.