22.15 — How a Drug Is Discovered, Tested and Approved

Around 90 percent of drugs that enter human trials never reach a pharmacy.

Most of the failures happen after years of work and hundreds of millions of dollars — and the commonest reason is not that the drug was unsafe. It is that it did not work well enough, despite looking convincing in cells, in animals and in early human studies.

Which is the deepest lesson of the whole process: biology is complicated enough that the only way to find out whether something works in people is to try it in people, carefully, in a way that can prove you wrong.

Finding a candidate

Target identification — deciding what molecule to aim at.

This usually comes from understanding a disease mechanism. Genetics has become the most productive source, because a gene variant that protects people from a disease points directly at a target worth blocking.

PCSK9 is the model case. People with a rare loss-of-function variant had very low cholesterol and no apparent harm. That observation led directly to PCSK9 inhibitors (Chapter 22.7). The human experiment had already been run by nature; the drug just copied it.

And drug targets supported by human genetic evidence are roughly twice as likely to succeed in trials, which is why genetics now shapes so much of early discovery.

Finding a molecule that hits the target:

High-throughput screening — testing enormous chemical libraries automatically for anything that binds.

Rational design — building a molecule to fit a known protein structure, which structural biology and now protein structure prediction have made far more feasible.

Natural products — plants, fungi, soil bacteria and marine organisms. Penicillin from a mould, statins from a fungus, metformin from a plant, aspirin from willow bark, paclitaxel from yew tree bark.

Biologics — antibodies and engineered proteins, designed rather than screened.

Repurposing — finding a new use for an existing drug. Cheaper, faster, and it has produced some of the biggest recent wins. Sildenafil was developed for angina and turned out to do something else. Thalidomide, after its catastrophe, became an effective treatment for myeloma. Dexamethasone became a COVID treatment (Chapter 22.12).

Then lead optimisation — chemists modify the molecule hundreds of times to improve potency, selectivity, absorption, half-life and safety. This alone can take years.

Preclinical testing

Laboratory and animal studies, establishing:

Whether it does what it should in living tissue.

How the body handles it — absorption, distribution, metabolism, excretion (Chapter 22.2).

Toxicology — in at least two species, including effects on organs, on reproduction, and on DNA.

And this stage exists for a reason with a name attached. Thalidomide, marketed in the late 1950s for morning sickness, caused severe limb malformations in an estimated 10,000 babies (Chapter 4.6). It had not been adequately tested for effects on the developing fetus.

The modern regulatory system — mandatory preclinical safety testing, proof of efficacy before approval, formal pregnancy categories — is largely the response to that disaster.

Preclinical results are also where most optimism comes from and where most of it should be discounted. A drug that cures cancer in mice is a very common event; a drug that cures cancer in people is not.

The human trials

Phase 1 — is it safe, and what does the body do to it

20 to 100 people, usually healthy volunteers. Except in cancer and some other serious diseases, where it would be unethical to give a toxic drug to a healthy person, so patients are enrolled instead.

Starting at a very low dose and escalating slowly, watching for toxicity and measuring blood levels.

Roughly 6 to 12 months.

And the starting dose is chosen conservatively for a reason. In 2006, a first-in-human trial of an antibody called TGN1412 gave six volunteers a dose based on animal studies. All six suffered catastrophic multi-organ failure within hours. The dose was a small fraction of what animals had tolerated, and the human immune system responded completely differently. Trial design rules were rewritten afterwards, including dosing volunteers one at a time rather than together.

Phase 2 — does it work, and at what dose

100 to 500 patients with the condition.

Testing several doses to find the one that sits in the therapeutic window (Chapter 22.3), and looking for a signal of efficacy.

1 to 3 years.

This is where most drugs die. The molecule works exactly as designed, and it turns out that the target was not driving the disease as much as the theory said.

Phase 3 — does it work better than what we already have

1,000 to 5,000 or more patients, across many centres and often many countries.

Randomised, controlled and, wherever possible, double-blind.

And each of those words is doing specific work:

Randomised — who gets which treatment is decided by chance. Which means the two groups are balanced on everything, including the things nobody thought to measure. This is the single feature that separates a trial from an observation.

Controlled — compared against placebo, or against the best existing treatment. Because people improve on their own, and because placebo effects are real (Chapter 22.1).

Double-blind — neither patient nor assessor knows who got what. Because expectation changes both how people report symptoms and how clinicians assess them, in the same direction.

3 to 5 years.

The outcome measured matters enormously.

A hard endpoint is death, a heart attack, a stroke, a fracture. A surrogate endpoint is a number that stands in for one — cholesterol level, blood glucose, tumour shrinkage on a scan.

Surrogates are faster and cheaper, and they can mislead badly.

The clearest example is the CAST trial. Drugs that suppressed abnormal heartbeats after a heart attack were widely used, because suppressing those beats was assumed to prevent sudden death. The trial found the drugs increased mortality. The surrogate improved and the patients died more often.

Which is why regulators increasingly require hard endpoints, and why an impressive change in a number is not the same as a benefit.

Phase 4 — what happens in the real world

After approval, and it is not optional decoration.

Phase 3 trials enrol a few thousand carefully selected patients. A rare serious side effect occurring in 1 in 10,000 people will not appear. Nor will effects in the people trials tend to exclude — the very old, the pregnant, those on multiple other drugs.

So post-marketing surveillance collects reports of suspected reactions from doctors, pharmacists and patients directly.

And it works. Rofecoxib was withdrawn after post-marketing data showed increased cardiovascular events (Chapter 22.5). Rosiglitazone was restricted for the same reason. Several drugs have been withdrawn over rhythm risks found only after wide use.

Reporting a suspected reaction is something patients can do directly in most countries, and the systems depend on it.

Approval

A regulatory agency reviews everything — the FDA in the US, the EMA in Europe, the MHRA in the UK, and their equivalents elsewhere.

They assess whether benefit outweighs risk for a defined use, and they approve the specific indication, dose and population studied.

And that last point explains off-label prescribing. A drug approved for adults may have no paediatric approval simply because no trial was done in children — not because it is unsafe there. Off-label use is legal and common and appropriate in many situations, and it means the evidence base is thinner than for the licensed use.

Accelerated pathways exist for serious conditions with no alternatives, allowing approval on surrogate endpoints with confirmatory trials required afterwards.

And the honest problem with them: those confirmatory trials are sometimes delayed for years, and occasionally show no benefit, by which time the drug has been in wide use. Faster access and stronger evidence pull against each other, and there is no arrangement that gets both.

The money

Estimates of the cost of bringing one drug to market range from around $1 billion to over $2.5 billion, and the range is wide because the calculation includes the cost of all the failures and assumptions about the cost of capital.

10 to 15 years from discovery to approval.

Patents last 20 years from filing, which usually happens early in development — so the effective period of market exclusivity after approval is often only 8 to 12 years.

Then generics enter and prices fall by 80 to 90 percent.

Generic drugs contain the same active ingredient and must demonstrate bioequivalence — that they deliver the same amount of drug into the blood at the same rate. They are not inferior; they are the same molecule without the development cost attached.

And biosimilars are the equivalent for biological drugs, which cannot be copied exactly because they are large proteins made in living cells, so they must be shown to be highly similar with no clinically meaningful differences.

The problems worth knowing about

Publication bias. Trials with positive results are more likely to be published, which skews the apparent evidence for a drug.

The response has been mandatory trial registration before the trial starts, so an unpublished trial can be identified as missing. This has helped and has not solved it.

Industry funding. Industry-sponsored trials report favourable results more often than independently funded ones on the same drugs. The mechanisms are subtle rather than fraudulent — choice of comparator, choice of dose, choice of outcome, and choice of what to publish.

Outcome switching — changing which outcome is reported after seeing the data. Registration exists partly to prevent it.

Trial populations that do not match real patients. Older people, pregnant women, children and people with several conditions at once are systematically under-represented, which means the evidence is weakest exactly where prescribing is most complex.

And the neglected disease problem. Diseases affecting mainly poor populations attract little development, because the market cannot repay it. Product development partnerships and public funding exist to address this, and they cover a fraction of the need.

How to read a claim about a drug

And this is the part worth carrying into ordinary life, because these claims arrive constantly.

Ask for the absolute numbers, not the relative ones. "Reduces risk by 50 percent" means something entirely different if the risk went from 2 percent to 1 percent than if it went from 40 percent to 20 percent. Both are a 50 percent reduction.

Ask what it was compared against. Better than placebo is a much weaker claim than better than the current standard treatment.

Ask what was measured. A surrogate endpoint is a promise, not a result.

Ask how long and how many. A six-week trial in 40 people tells you very little about a drug you would take for twenty years.

And ask the number needed to treat — how many people have to take the drug for one to benefit. It converts a percentage into something you can picture, and it is the single most honest way to express a treatment effect.

What the next Part covers

Part 23 is the emergency chapters — what to do in the minutes that decide an outcome, before any medicine in this Part becomes relevant. It is the most important Part of this volume.