Preclinical results do not reproduce

Each batch of cells used in an experiment would carry a small digital record showing when it was authenticated, tested for contamination, and how many times it had been grown, attached to the published result.

Every failed cancer drug, experiment and trial gets recorded in one open ledger, so nobody repeats a failure that has already cost years and millions.

Failed laboratory experiments are rarely published, so other teams repeat them. A searchable place to deposit them would save years of duplicated work.

Companies and public funders would jointly pay for standardised experiments that confirm or refute new cancer targets, sharing all results openly, so nobody wastes years on a target that does not hold up.

Thousands of cancer biomarkers are published; almost none reach patients because nobody validates them fairly. Create a public service that tests any candidate blind against stored samples.

Most lab experiments that fail are never written up, so other labs repeat them. A simple, structured registry with a citable record for each failed experiment would stop the waste.

Cancer AI models are usually tested on data from the same hospital they were built on. A registry of independent test datasets, and a rule that every model reports performance on at least one, would show which models really work.

When you look up a paper, you should immediately see whether anyone has tried to repeat it and whether they succeeded.

Getting a cell line, mouse model or antibody from another lab can take six months of paperwork. Funders would require a standard agreement that goes through automatically unless someone objects within thirty days.

Secure online workrooms where approved researchers can analyse cancer records without downloading them, with the data already cleaned and organised for cancer questions.

A reported prevalence of 0 to 84 per cent for the same late effect is a measurement failure. Core outcome sets are cheap, need no new biology, and would unlock the studies the field has already paid for.

When a study is retracted, everything built on it should get a warning. Today, retracted cancer papers keep being cited and used for years.

Many cancer lab results cannot be reproduced, and trials built on them fail. Fund an independent institute that re-runs important experiments before anyone spends millions on humans.

Most cancer proteins have never been tested to see whether a small molecule can attach to them at all. A public map of what is chemically reachable would tell the field where to aim.

Map every state a cancer cell can be in, and how drugs and the surrounding tissue move it between states, into an open computational model anyone can query and improve.

Most mouse studies of cancer drugs do not randomise animals or blind the people measuring tumours, which inflates results. Checking and publishing which institutions do it properly would change behaviour.

When a study is retracted or corrected, every guideline and software tool that relied on it would be alerted automatically, so wrong evidence stops influencing care.

Results can change when a supplier changes a batch of serum, antibody or growth factor. Recording which batch was used in each experiment, in a shared ledger, would let these effects be spotted.

Pay a fixed reward to any lab that pre-registers and carefully repeats a heavily cited preclinical cancer finding, whatever the outcome, with a bonus for the first documented non-replication that passes methodological review. Today that work is unpaid and unpublished.

Scientists are promoted for first-time discoveries and journal prestige, not for replicating others' work or sharing data. A structured section in tenure and promotion dossiers for replications conducted, data and code shared, and registered reports, weighted explicitly in decisions, would change what scientists spend their time on.

Published figures are cropped and processed. Requiring the original, uncropped image files to be deposited lets anyone check that the figure shows what it claims.

Journals and funders already ask trialists to share patient-level data; almost nobody checks. Make it a checked condition with real consequences.

Make the first project of every cancer biology doctoral student a funded, pre-registered attempt to repeat a published finding chosen from a curated list of translationally relevant results, with the outcome published in a replication registry. Students learn power analysis, blinding and reporting, and the field gets thousands of replications a year.

Drug sensitivity results for the same cell line and drug differ substantially between large screens. Every published cancer drug screen should include a defined panel of reference compounds with published expected activity ranges per reference cell line, reported in a standard format, so results from different labs can be calibrated against each other.

Almost no research money goes to checking whether published cancer findings hold up: one replication project could complete only 23 of 50 planned experiments. Requiring 3 to 5% of every funder's research budget to go to independent replication, published whatever the result, would build the missing feedback loop.

Lab-grown mini-tumours are already being sold to guide treatment, but the tests are not validated like other medical tests. They should be.

Before a new cancer drug from a university is given to people, a separate laboratory should have repeated the main experiment showing it works.

Papers say 'data available on request' or link to files that no longer exist. Journals should verify data access at publication and periodically afterwards, and mark papers whose data have disappeared.

Before testing a drug in people, someone should systematically gather all the animal and laboratory evidence, including the studies that failed. Almost no cancer trial does this.

A drug that works in one laboratory's mice often fails elsewhere. Running the confirmatory animal study as a randomised, blinded, multi-laboratory trial, as stroke and amyotrophic lateral sclerosis research has done, would catch this before a human trial; oncology has no such standing infrastructure.

Instead of one lab's mouse study deciding whether a drug goes to patients, several labs run the same protocol independently, like a multi-centre clinical trial for mice.

A large fraction of commercial research antibodies fail when tested against cells engineered to lack their target, so they do not bind what the label says. Journals and funders should require knockout-validated antibodies for the claims a paper rests on, and fund public validation of the most-used cancer targets.

Every trial writes its protocol and analysis plan from scratch. A shared library of well-written templates and ready-to-run analysis code would let teams start from the best version rather than a blank page.

Clinical trials get expert statistical review; the laboratory studies that justify them usually do not. Paying statisticians to review these papers before they influence a trial would catch errors early.

Clinical trials must be registered before they start so that failures cannot be hidden. Animal studies used to justify human trials should follow the same rule.

Drug trials must be registered before they start so results cannot be hidden or reshaped. Studies that claim a biomarker predicts outcome should be registered too.

Just as clinical trials must be registered before they start, studies using hospital data should be registered too, so the failed or unwelcome ones cannot quietly disappear.

Underpowered mouse experiments give exaggerated positive results and uninformative negatives, and papers often show one 'representative' result out of several attempts. Funders and journals should require a pre-specified power calculation, the number of independent repeats performed, and reporting of every repeat rather than the best one.

Universities and cancer centres would change how they promote scientists, giving credit for finishing trials, sharing data, replicating others' work and publishing failures, not just for papers in famous journals.

A troubling share of published cancer experiments use cell lines that are contaminated or mislabelled. Requiring a simple identity check before publication would stop this.

A large share of cancer research has been done on cells that were mislabelled or contaminated. A cheap DNA fingerprint test can prove identity; journals and funders should require it.

There is no agreed ruler for measuring how strong a CAR-T product is. Shared reference materials would let hospitals, companies and regulators compare products fairly.

If public or charity money paid to build a cancer AI model, the model itself (not just a paper about it) must be released so others can test, improve and use it.

Funders should randomly select a small fraction of the papers they paid for and check the raw data, analysis and records, with public results. The possibility of an audit changes behaviour.

Journals agree to publish a study based on the quality of the question and plan, before anyone knows the answer. That removes the pressure to make results look positive.

Software can already spot impossible statistics, mismatched p-values and duplicated images in a paper. Journals should run these checks on every submission, as spell-check runs on every document.

Scientists hide results for fear of being beaten to publication. If journals and funders guaranteed that a preprinted finding cannot be scooped, and encouraged rival groups to publish side by side, sharing would become safe.

No one keeps score of which laboratory models actually predicted what happened in patients. A public scoreboard would show which models to trust.

Robotic labs guided by AI that design experiments on tumour models, run them, read the results and design the next ones, around the clock, with every result published openly.

Before spending millions to turn a lab finding into a drug, spend a little to have an independent lab check it is real. Funders would reserve a small slice of money for exactly this.

If every lab had access to the same set of well-characterised tumour models, results could be compared directly instead of each lab using its own private models.

Electronic notebooks record when each experiment was done and what the raw result was. Submitting them with the paper would show whether the analysis was planned or fitted after the fact.

For the biggest claims, journals would require that a second, independent laboratory repeated the central experiment before the paper is accepted.