A Drug That Works in Cells Isn’t a Cure: How to Read a Biotech Breakthrough

Bleach kills cancer cells in a dish. So does heat, so does the wrong concentration of salt, and so do thousands of experimental molecules that will never become medicines. That is not a joke at science’s expense; it is the single most useful thing to understand about biotech news. Killing a diseased cell on a plastic plate is the easy part. Killing it inside a living person, without harming the healthy cells around it, at a dose the body can tolerate, for long enough to matter, is the hard part—and it is where most promising ideas quietly stop.

So when a headline says a compound “destroys tumours” or “reverses” a disease, the first question is not “Is this real?” It is “Real where?” In cells? In mice? In a dozen patients? In three thousand? Each answer can be legitimate science. Only the last kind comes close to meaning a treatment you might one day be offered.

The ladder every medicine has to climb

Drug development is best pictured as a ladder, and each rung answers a different question. The US Food and Drug Administration (FDA) lays out the sequence: discovery, preclinical research, clinical research in people, regulatory review, and safety monitoring after approval.

• In vitro (cells or tissue in a lab). Latin for “in glass.” These experiments ask whether a molecule does anything at all to a biological target. They are fast and cheap, and they strip away most of the complexity of a real body.
• In vivo (animals). Latin for “in the living.” Animal studies ask whether the drug reaches the right tissue, which doses are toxic, and whether the effect survives contact with a whole organism. FDA notes that preclinical studies are generally small but must still provide detailed information on dosing and toxicity.
• Phase 1. Typically 20 to 100 healthy volunteers or people with the condition, focused on safety and dosage. FDA estimates that roughly 70% of drugs move on.
• Phase 2. Up to several hundred people with the condition, looking at efficacy and side effects. About a third move on.
• Phase 3. Roughly 300 to 3,000 people, testing whether the benefit is real and watching for adverse reactions at scale. FDA puts the share that moves forward at about 25–30%.
• Approval and Phase 4. If regulators judge that benefits outweigh risks for a specific use, the drug is approved—and monitoring continues in many more people once it is on the market.

Notice what the ladder implies: “works in cells” sits on the bottom rung. That is where discoveries begin, not where they end. Our interactive Drug Development Journey lets you walk each step yourself.

Plain English: Preclinical means any research done before a drug is tested in people—in cells, tissues or animals. Clinical means research in human volunteers. Approved means a regulator has judged that, for a specific use, the benefits outweigh the risks.

Why so much that works in a dish fails in people

The drop-off is steep. A 2022 review in Acta Pharmaceutica Sinica B by Duxin Sun and colleagues estimated that about 90% of drug candidates that enter clinical trials fail. The authors grouped the failures into four broad causes, which are worth remembering because each one maps onto something early studies cannot see:

1. It doesn’t work well enough in people (lack of clinical efficacy)—roughly 40–50% of failures.
2. It causes unmanageable harm (toxicity)—about 30%.
3. It behaves badly as a drug—poor absorption, breakdown or distribution in the body—around 10–15%.
4. Business and strategy—lack of commercial need or poor planning—around 10%.

Every one of those failure modes can hide inside a cell-culture experiment. A dish has no liver to break the drug down, no kidneys to clear it, no immune system to react to it and no blood–brain barrier to keep it out. Cells grown in labs for years can drift far from the tumours they were originally taken from. Mice help, but they are not small humans: their metabolism, immune systems and lifespans differ, and many animal “models” of disease are engineered approximations rather than the human condition itself.

Dose is another quiet trap. A molecule may kill cancer cells at a concentration that would be impossible—or dangerous—to reach in human blood. When a press release leaves out the concentration, it is leaving out one of the most important numbers. The newer computational tools described in our piece on AI drug discovery aim to predict some of these problems earlier, but their candidates still have to climb the same ladder.

A cure is not a molecule. It is a molecule, at a dose, in a body, over time, compared with something else.

The replication problem

There is a second, less comfortable reason early findings can mislead: some do not hold up even in the lab. In 2012, C. Glenn Begley and Lee M. Ellis reported in Nature that when scientists at the biotech company Amgen tried to confirm the findings of 53 “landmark” preclinical cancer papers, the findings were confirmed in only six.

A more systematic effort followed. The Reproducibility Project: Cancer Biology, published in eLife in 2021, set out to repeat 193 experiments from high-profile cancer papers published between 2010 and 2012. It completed only 50 of them, in part because the original papers often lacked the detail needed to repeat the work. Among completed replications, the median effect size was 85% smaller than in the original experiments.

None of this means preclinical science is broken or that researchers are careless. Small experiments, pressure to publish and the natural tendency for striking results to attract attention all push early effect sizes upward. The practical lesson for readers is simple: one exciting result is a hypothesis with good lighting. Independent replication is what turns it into knowledge.

What was actually measured?

Even when a study is done in humans, check what it counted. FDA distinguishes a clinical outcome—whether patients feel better, function better or live longer—from a biomarker, a measurable signal in the body such as blood pressure or blood sugar. When a biomarker stands in for the outcome that actually matters, it is called a surrogate endpoint.

Surrogates are genuinely useful: they let trials finish sooner, and some, such as lowering blood pressure to reduce stroke risk, are well validated. But FDA itself cautions that improvement on a surrogate can sometimes mislead about a treatment’s overall balance of benefit and risk. A tumour that shrinks on a scan is encouraging; it is not the same thing as a patient living longer or living better. And because cancer is a group of many different diseases, a result in one tumour type rarely transfers automatically to another.

Our explainer Inside a Clinical Trial goes deeper on what each phase can and cannot prove.

A seven-question field test for any breakthrough headline

1. Where was it tested? Cells, animals or people? If people, how many?
2. Was there a comparison group? Without a control group—ideally randomised—improvements may reflect chance, the placebo effect or the natural ups and downs of a disease.
3. What was measured? A lab marker, a scan, or how patients actually fared?
4. How big was the effect, and at what dose? “Statistically significant” means a result is unlikely to be due to chance alone. It does not mean the effect is large or important.
5. What were the harms? A story that mentions only benefits is telling half of it.
6. Has anyone else reproduced it? Independent confirmation is worth more than a bigger headline.
7. Who is saying it? A peer-reviewed paper, a conference abstract, a company press release and a social media post are different kinds of evidence. Company statements can be accurate, but they are written by an interested party.

We turned these questions into a longer guide: read The Breakthrough Reality Check before you share your next medical headline.

Why early results still deserve attention

Scepticism is not cynicism. Every approved medicine began as a result in cells or animals, and many of today’s most important therapies spent years on the lower rungs before they reached patients. Early studies are how science decides which ideas deserve the expense, time and risk of human trials. Without them, there would be nothing to test.

The healthiest way to read a breakthrough, then, is to locate it on the ladder. Say “promising in mice” rather than “cure.” Say “early safety data in a small group” rather than “it works.” Those phrases are less thrilling, but they are more honest—and they leave room for the moment, sometimes years later, when a result has earned the bigger word.