How a Drug Is Made, Tested, and Approved

TL;DR. Roughly 10 to 15 years and, by industry estimates, over a billion dollars to bring one drug to market, with around 90 percent of candidates that reach human trials failing. A tablet is typically 5 to 50 percent active drug and the rest is engineering: fillers, binders, disintegrants, lubricants, and coatings that control where and how fast it dissolves. Generics are the same molecule at the same dose and must prove bioequivalence, and they cost a small fraction of the original. The system has real failures, and the response to those failures is most of why modern approval is so slow.

1. From molecule to medicine

Discovery (3 to 6 years). Identify a target, usually a protein implicated in a disease. Screen enormous compound libraries, often hundreds of thousands of molecules, using automated assays, or design candidates computationally from the target's structure. Machine learning has become genuinely useful here in the last few years, particularly for protein structure prediction. A hit becomes a lead, which is then optimised over hundreds of synthesised variants to improve potency, selectivity, solubility, and metabolic stability.

Preclinical (1 to 2 years). Cell studies, then animal studies for pharmacology, toxicology, carcinogenicity, and reproductive effects, in at least two species. Determines a safe starting dose for humans.

Clinical trials (6 to 8 years).

Only about 1 in 10 drugs entering Phase I reaches approval. The failures are roughly half efficacy (it did not work), a third safety, and the rest commercial.

Phase IV matters more than its position suggests. A Phase III trial of 3,000 people cannot detect a side effect occurring once in 10,000. Rofecoxib (Vioxx), withdrawn in 2004 after its cardiovascular risk became apparent in wider use, is the standard example.

2. Why trial design is the whole ballgame

Randomisation allocates patients to treatment or control by chance, so the groups differ only in what they receive. This is the single most important idea in clinical evidence: it removes confounding, including confounders nobody thought of.

Blinding. Single-blind means the patient does not know; double-blind means the investigators do not either. Without it, expectations shape both reporting and assessment.

Control group. Placebo where no effective treatment exists; active comparator where one does. A drug shown to beat placebo has not been shown to beat what people already take, which is a recurring gap.

Endpoints.

• Hard endpoints: death, heart attack, stroke, fracture, hospitalisation. What actually matters.
• Surrogate endpoints: blood pressure, HbA1c, cholesterol, tumour shrinkage. Faster and cheaper, and sometimes misleading. Drugs have lowered a surrogate and increased deaths: the CAST trial found antiarrhythmics that successfully suppressed ectopic beats after heart attack increased mortality, and several diabetes drugs have improved HbA1c without improving outcomes.

Relative versus absolute. "Reduces risk by 50 percent" means nothing without the baseline.

A drug cutting risk from 2 in 1,000 to 1 in 1,000 is a 50 percent relative reduction and a 0.1 percentage point absolute reduction, with a number needed to treat of 1,000: a thousand people take it for one to benefit. Press releases quote the first number and the second is the one that should decide anything.

Number needed to treat (NNT) and number needed to harm (NNH) are the most useful numbers a patient can ask for, and they are rarely offered.

3. Manufacturing: what is actually in a tablet

The active pharmaceutical ingredient (API) is often 5 to 50 percent of the tablet by weight, and sometimes far less. A 100 mg tablet of a drug active at 1 mg is 99 percent something else.

API synthesis is multi-step organic chemistry, increasingly concentrated in India and China, which is a real supply chain vulnerability that COVID-19 exposed. Impurity limits are extremely tight, and this is where several major recalls originated: nitrosamine contamination (NDMA) was found in valsartan, ranitidine, and metformin from 2018 onward, caused by specific synthesis routes and solvent recovery practices, and led to worldwide withdrawals.

The excipients, which is the part nobody reads:

Two excipients cause real problems for real people. Lactose is the commonest filler, and while the quantity is small (usually well under the threshold that troubles most lactose non-persistent people), it matters for the highly sensitive. Gelatin capsules are animal-derived, which matters for vegetarians, vegans, and several religious traditions; HPMC capsules are the vegetarian alternative and are usually stated.

Coatings do more than look nice:

• Film coating: taste, appearance, ease of swallowing.
• Enteric coating: dissolves only above pH 5.5, so the tablet passes the stomach intact. Used to protect the drug from acid (some antibiotics), to protect the stomach from the drug (enteric aspirin), or to release in the intestine (peppermint oil for IBS). This is why enteric tablets must not be crushed or chewed.
• Modified release (MR, SR, XL, XR, LA, CR, Retard): the drug is embedded in a matrix or reservoir that releases over 12 to 24 hours. Crushing one delivers the entire day's dose at once, which has killed people with opioids and calcium channel blockers.

The single most important formulation rule: if a tablet has letters after its name (MR, SR, XL, XR, CR, LA) or is described as enteric-coated or gastro-resistant, do not crush, split, or chew it unless the leaflet or a pharmacist says you may. If someone cannot swallow tablets, ask the pharmacist for a liquid or a different formulation.

Quality control covers content uniformity, dissolution rate, hardness, friability, stability under heat and humidity, and impurity profile, all under Good Manufacturing Practice inspection.

4. Generics and biosimilars

A generic contains the same active ingredient, at the same dose, in the same form, for the same indication. It may differ in excipients, colour, shape, and name.

Bioequivalence is what it must prove: in a crossover study, usually in healthy volunteers, the 90 percent confidence interval for the ratio of peak concentration and total exposure against the original must fall within 80 to 125 percent. That range sounds wide and is standard internationally, and in practice most generics land within a few percent.

Generics cost 80 to 90 percent less and are equivalent for the overwhelming majority of drugs. Two caveats:

• Narrow therapeutic index drugs (levothyroxine, ciclosporin, some antiepileptics, lithium, warfarin) are where the 80 to 125 percent band can matter clinically, and prescribing guidance in several countries advises staying on the same brand for antiepileptics in particular.
• Excipient differences occasionally matter for allergy or intolerance.

Biologics are different in kind: large, complex proteins made in living cells (antibodies, insulins, vaccines). They cannot be copied exactly, so copies are biosimilars, which must demonstrate no clinically meaningful difference through analytical, pharmacokinetic, and often clinical comparison. They cost less than the originator and considerably more than a small-molecule generic.

Patents typically run 20 years from filing, of which perhaps 8 to 12 remain after approval, with extensions available. Companies argue this window funds the failures; critics point to evergreening (patenting minor modifications, new salts, or new formulations to extend exclusivity) and to the fact that a substantial share of basic research is publicly funded. Both observations are accurate.

5. Regulation, and why it looks the way it does

The main agencies: FDA (US), EMA (EU), MHRA (UK), PMDA (Japan), plus national bodies. The WHO prequalification programme matters enormously for global supply. Standards are broadly harmonised through ICH guidelines, and decisions still diverge.

The system was built out of disasters:

• 1937, elixir sulfanilamide: a US manufacturer dissolved a sulfa drug in diethylene glycol (antifreeze) with no safety testing. Over 100 people died, most of them children. This produced the 1938 US Food, Drug and Cosmetic Act requiring safety testing.
• 1957 to 1962, thalidomide: marketed for morning sickness, it caused severe limb malformations in around 10,000 babies worldwide. The US largely escaped because FDA reviewer Frances Kelsey refused approval over inadequate data. The result was the 1962 Kefauver-Harris Amendment requiring proof of efficacy, not just safety, and rigorous reproductive toxicity testing.
• 1976 onward, DES: a synthetic oestrogen given in pregnancy caused rare vaginal cancers in daughters decades later, establishing that harms can appear a generation downstream.
• 2004, rofecoxib (Vioxx): withdrawn over cardiovascular risk, strengthening post-marketing surveillance requirements.
• Ongoing, the opioid crisis: aggressive marketing of oxycodone on the basis of weak evidence about addiction risk contributed to hundreds of thousands of deaths in North America, and produced multi-billion dollar settlements (Chapter 68).

The structural tension is real and unresolved: approve too slowly and people die waiting; approve too fast and people are harmed by drugs that should not have been licensed. Accelerated and conditional approval pathways exist for serious diseases and are regularly criticised in both directions. The aducanumab approval for Alzheimer's disease in 2021, over the objection of the FDA's own advisory committee, is the most-discussed recent case.

Publication bias is the field's persistent problem: trials with positive results are more likely to be published, which inflates apparent effectiveness. Trial pre-registration (mandatory since 2005 for publication in major journals) and results-reporting requirements were introduced to address this, and compliance remains imperfect.

6. How to read a drug claim

A short checklist that applies to a press release, a news article, or a leaflet:

1. Randomised and blinded, or observational?
2. Compared with what? Placebo, or the current standard?
3. Hard endpoint or surrogate?
4. Absolute numbers, or only relative? Ask for the NNT.
5. How many people, for how long? Rare and late harms need size and time.
6. Who was in the trial? If it enrolled 60-year-old men and you are an 80-year-old woman with kidney disease, the result may not transfer.
7. Who funded it, and was it pre-registered?

7. The bottom line

• Ten to fifteen years, and around 90 percent of drugs entering human trials never reach market. Phase II is where most die.
• Randomisation and blinding are what make a trial worth reading. Surrogate endpoints have repeatedly misled, and relative risk reductions without absolute numbers tell you almost nothing.
• A tablet is mostly not the drug. Excipients control size, disintegration, stability, and where it dissolves.
• Never crush or chew a modified-release or enteric-coated tablet. It converts a day's dose into a single hit.
• Generics are the same molecule, must prove bioequivalence within 80 to 125 percent, and cost a fraction. The exceptions worth care are narrow-therapeutic-index drugs.
• The regulatory system was built out of specific disasters, and its slowness is the deliberate legacy of thalidomide.

Sources and notes

Development timelines, attrition rates by phase, and cost estimates follow DiMasi et al., Journal of Health Economics, 2016, and the subsequent critiques of that methodology by Prasad and Mailankody. Trial design principles follow the ICH E9 statistical guidance and standard epidemiology texts. The CAST antiarrhythmic mortality result is Echt et al., New England Journal of Medicine, 1991. Excipient functions follow pharmaceutical formulation texts. Nitrosamine contamination of valsartan, ranitidine, and metformin follows EMA and FDA recall documentation from 2018 onward. Bioequivalence criteria follow FDA and EMA guidance. Biosimilar requirements follow EMA guidelines. Regulatory history follows the 1938 Food, Drug and Cosmetic Act, the 1962 Kefauver-Harris Amendment, and the thalidomide record documented by the FDA and by Stephens and Brynner, Dark Remedy. The aducanumab approval controversy follows the FDA advisory committee record and the resulting resignations. Publication bias and trial registration follow the ICMJE requirements and the AllTrials campaign documentation.

Open questions. Drug development cost estimates vary by an order of magnitude depending on methodology and who funded the analysis, and none of the figures should be treated as settled. Whether accelerated approval pathways have net benefit is genuinely contested.

👉 Next: how to read a medicine label.