AI in Drug Discovery: A Shortlist Is Not a Drug

Three names recur whenever the press writes about artificial intelligence and the diseases medicine cannot yet cure: a new antibiotic, a protein that misfolds in Parkinson's disease, an old transplant drug that kept its discoverer alive. A recent BBC Future feature gathers all three and lets the scientists behind them speak. It is good, fair journalism. It is also, if you read it for what each story actually shows rather than what it promises, three accounts of the same narrow thing — and the size of that thing is the appraisal.

Start with the obvious point that the format tends to blur: this is a feature, not a paper, and the work it describes sits almost entirely at the preclinical and early-translational tier. Promising candidates, working platforms, an n of one who treated himself — not completed trials with clinical endpoints. None of that is a criticism of the reporting. It is simply the difference between the headline and the evidence, and a careful reader keeps both in frame at once.

What AI is doing, and where it stops

In every one of the three stories, the machine does its work at the same place: the beginning. James Collins's group at the Massachusetts Institute of Technology has used machine learning to screen more than 45 million chemical structures for activity against resistant bacteria — the kind of search, as Collins describes it, that now takes days or hours rather than years. At the University of Cambridge, Michele Vendruscolo's group screens vast molecular spaces for compounds that might stabilise proteins before they misfold into the clumps associated with neurodegeneration. And Every Cure, the non-profit founded by physician David Fajgenbaum, runs approved drugs against catalogued diseases to surface repurposing matches no one has tested. Search, propose, rank: that is the front of the pipeline, and AI is genuinely fast there.

The back of the pipeline is untouched. Preclinical safety and toxicology, Phase I, II and III trials, regulatory review — these take years and cost hundreds of millions, and they cost the same whether a candidate came from a model or from a graduate student with a hunch. AI shortens the time to a promising molecule. It does not shorten the time from a promising molecule to a licensed medicine, because that time is bought in trials, not in compute.

The three stories, by what they show

The antibiotics thread is the most concrete and the easiest to overstate. Resistant infections already kill more than a million people a year, projected to rise several-fold by mid-century, while only a dozen genuinely new antibiotics reached approval in the years 2017 to 2022. Against that backdrop, surfacing fresh candidates from a 45-million-structure library is a real contribution. But a candidate that acts against resistant bacteria in a screen is a shortlist entering the laboratory, not a drug leaving it.

Parkinson's is the hardest case, and the feature is honest about why. The condition was first described in 1817; more than two centuries on, no licensed treatment slows its course, because the underlying biology is not settled. Vendruscolo's framing is that stabilising the relevant proteins before they misfold would amount to preventing the disease rather than curing it — which is exactly the right ambition, and exactly why it is early-stage. You cannot reliably drug a target you do not yet fully understand.

Repurposing is where AI's speed can most plausibly reach a patient soonest, and the origin story explains why. Fajgenbaum, diagnosed with the rare Castleman disease around 2010, studied his own condition and identified an already-approved drug, sirolimus, that has held him in remission for more than a decade. Every Cure now matches thousands of approved drugs against thousands of diseases. Because those drugs are already licensed and their safety profiles known, the slow back end of the pipeline is, for once, partly pre-paid — which is precisely what makes this the most realistic near-term route.

There is a quieter caution running underneath all three. A model that predicts which compound binds a target, or which existing drug suits a new disease, produces hypotheses — ranked guesses, not findings. Most will not survive the wet-lab experiment or the trial that tests them, and that attrition is not a defect; it is what early discovery has always looked like. The danger lives entirely in the retelling: a shortlist reported as a breakthrough, a preclinical candidate reported as a cure.

Read the numbers as claims, and why it matters

The specific figures are worth a second glance, because across the feature and its many republications they drift — the size of the screened library, the resistance death toll and its 2050 projection, the count of recently approved antibiotics, the exact scale of the repurposing effort all appear in more than one form. The direction of every figure is sound and none of this undermines the reporting. But a clinician reading for decisions, not for reassurance, should treat the individual numbers as claims to confirm against the primary sources rather than as fixed facts.

For European health systems the sober reading is also the useful one. AI's most credible near-term contribution to therapeutics is not a wave of cures but a faster, cheaper way to generate well-grounded candidates and repurposing hypotheses — most valuable where the back end is already shorter, as with approved drugs of known safety. The bottleneck was never the idea. It is the evidence: trials still have to run, regulators still have to be satisfied, and the discipline that turns a plausible molecule into a safe medicine is the part no model has learned to skip. The honest promise is acceleration at the start of the road, not a shortcut to its end.

Source: "Can AI cure the incurable?", BBC Future, March 2026. This is a journalistic feature surveying primary research, not a study; the underlying work it describes sits largely at the preclinical and early-translational stage, and several of its headline figures vary between the original and its republications and should be checked against the primary sources before being relied upon.