Skip to main content

Unfortunately we don't fully support your browser. If you have the option to, please upgrade to a newer version or use Mozilla Firefox, Microsoft Edge, Google Chrome, or Safari 14 or newer. If you are unable to, and need support, please send us your feedback.

Elsevier
Publish with us
Connect

The undruggable isn’t just a chemistry problem

November 17, 2026

By Steve Korver

Scientists-discussing-cropped-2140x1070

Targets once written off are being hit, but usually only after a burst of insight and years of groundwork. What if the bottleneck was never the chemistry?

One unifying feature of those scientists who are making the undruggable druggable – be it the once-ungraspable KRAS, the shapeless MYC, or the broken p53 that needs to be rebuilt rather than blocked – is a willingness to call out what’s wrong. As a pioneering KRAS researcher, Dr. Kevan Shokat, saidopens in new tab/window: “Many of the assumptions we’ve made, about proteins we’ve studied a lot in cancer, are wrong, and these assumptions are hindering our ability to make discoveries.”

At the ASCO plenary in Chicago in May 2026opens in new tab/window, an oncology crowd gave a standing ovation. The once-daily pill daraxonrasib had roughly doubled median survivalopens in new tab/window in previously treated metastatic pancreatic cancer: 13.2 months versus 6.7 for chemotherapyopens in new tab/window. In a disease where progress is measured in weeks, that is a step change. Sotorasib had cracked one KRAS variant back in 2021, but not the ones driving most pancreatic tumours. Daraxonrasib hits them all – including the mutation found in some 90% of pancreatic cancers, long written off as undruggable.

sherry winter

Sherry Winter, Senior Director of Biomedical & Regulatory at Elsevier

It was the most elegant of hacks. Daraxonrasib can’t grip KRAS by itself. It first binds a chaperone protein, cyclophilin A, and the two together form a new surface that exists nowhere in nature – one shaped to clamp KRAS in its active state. Neither half does anything alone.

And additional breakthroughs involving p53opens in new tab/window (mutated in roughly half of all cancers) and MYCopens in new tab/window (dysregulated in around 70% of cancers) are making the “undruggable list” seem quaintly dated. Scientists are abandoning the lock-and-key model – the assumption that a target must offer a stable pocket to hold a drug – in favor of shims, glues, decoys, and other cellular workarounds.

More please.

Undruggable is a fancy word for “gave up”

George Georghiouopens in new tab/window, former Senior Knowledge Strategy Manager at Elsevier, who is now at Syngenta, knows someone who was at that conference, a colleague from his PhD years with a long family history of pancreatic cancer. It is the only time George has known him to admit he cried.

But the received wisdom about why KRAS took so long, George says, is wrong. “KRAS was never really an undruggable target. There were other challenges to overcome as well. Part of the difficulty was getting a drug into the solid tumour. It’s a nightmare: everything compacts very tightly, the environment becomes incredibly acidic, and delivering drugs to where the core cells are actually replicating and spreading is like, ‘Oh, it went in there; oh, sorry, it disintegrated.’”

“Undruggable” simply means a field has concluded that all methods have been exhausted. “More importantly, the funding bodies said, ‘I’m not going to sponsor that; it doesn’t work.’ And the structure reinforces this. Most innovation is no longer done inside pharma companies. It’s done in universities and startups, and big pharma buys.”

Hints, not predictions

Could today’s information and AI technologies work to flag these insights more quickly? George spent years at Novartis building a dashboard that tracked a decade of PubMed literature looking for emerging signals. Point it at KRAS research in 2012 and would it have seen what was coming? “Really unlikely, because that depended on mass volume, the same terms mentioned multiple times next to each other. Some interesting trends would have come out of it, but my gut feeling is I’d probably have missed it.” Sherry Winteropens in new tab/window, Senior Director of Biomedical & Regulatory at Elsevier, says the same from the biology-mapping side. In the case of KRAS, the original breakthrough hinged on a cryptic binding pocket, and a tool like EmBiologyopens in new tab/window tracks biological connections, not protein structures.

That raises a more useful question: not whether the tools can predict, but what they can’t see. Sherry’s job is to help determine whether a target is worth developing a drug against – and the answer is rarely in the papers everyone has already read. “You look for things reported a few years ago, buried in the results section, that give you a hint to follow up on. It might be in new data, but it’s more likely in the published literature. It’s about finding what’s innovative and unique there.”

Technology will get better at spotting those patterns, she adds, “but it still comes down to having someone who has the imagination to take it forward – along with the money. And in the end, you will need to validate it. Nothing’s going to change that.”

Targeting the silos that matter

For years the complaint has been that biology and chemistry sat in separate systems – the same target named one way in a chemistry database and another way in a biology one. No questions could dance between the two. That is improving, but it hasn’t got as far as it sounds. “Chemists are chemists and non-chemists are non-chemists,” says Sherry. “You have to figure out what language and content to give a person so it fits their needs.”

For George, the more damaging silo isn’t the wall between departments. “Where you run into data silos is by never doing those updates or never sharing that information. By not staying up to date, you’ve stifled your own innovation.” An aligned map that nobody refreshes just becomes a slower silo. “It needs to be treated as its own product. It needs a lifecycle.”

More failure, please

The oldest problem in R&D is the evidence that never gets published. Every one of these breakthroughs is based on decades of failure, and almost none of it is on the record. “No one publishes ‘this experiment failed,’” says Sherry. “No one gets grants because their experiment failed. And pharma companies are not going to say anything about stuff that failed, because that would give others a competitive advantage.”

George felt it at the bench. “During my PhD I’d probably have murdered somebody for a list of things that had been tried. The only way you find out what doesn’t work is literally over beers at a conference, because you’re all drunk enough to admit your failures,” says George. “Sometimes the investigation into why something doesn’t work is more interesting than the thing you were trying to prove.”

The gap matters most for systems that act on their own, as with agentic AI, because a system that only sees what worked will confidently repeat what didn’t. “But if you combine the failures with the literature, suddenly you’ve got some stochasticity you can filter and tweak, and you go, hang on, nobody’s actually tried that yet,” George says.

The field has stopped waiting for failures to be published and is now recovering them elsewhere. Sherry runs through the routes: adverse event reporting, drugs pulled from the market, and ClinicalTrials.gov “because you can see, oh, this trial started but it never finished, I wonder why,” says Sherry. “But I don’t think we’re there yet.”

Private data is the other source: tools that let customers run a shared ontology across their own records, or fold proprietary reactions – including poor-yield routes – into retrosynthesis planning.

Sherry’s related worry is the volume of poor positive data. “The real information you need is buried under stuff you don’t need, and you get a bias that maybe is sending people in the wrong direction.”

AI may speed the cycle. It can’t supply the curiosity.

Meanwhile AI-designed drugs are not exactly flooding the market. Partly that’s arithmetic: the earliest candidates are only a few years into a decade-long pipeline. But George expects the machine-human split to hold. “With science, the majority of the real meat and findings is never in the text. It’s in the pictures and figures. And the figure legends lie to you too.” Systems trained on text work from the part of the paper that summarizes the result, not the part that contains it.

Plus, sources disagree constantly. One database says gene A causes Parkinson’s; another says gene B. “It’s not for an AI to decide; it’s for the human to decide, based on the evidence.” Current models are built the other way. “They never want to tell you, ‘I don’t know.’”

“It’s a tool,” says Sherry. “AI is only as smart as the content it’s ingested and the tasks it’s been trained to perform.” Her advice is to use it while staying alert: “You’ve got to be willing to pivot and follow where it leads. That doesn’t mean following it into nonsense.”

The next breakthrough

Ask both where they want the next breakthrough to come from, and they pin their hopes on dementia. The bottom line is certainly thereopens in new tab/window: US dementia care is projected to cost $409 billion in 2026 and nearly $1 trillion by 2050, with almost 200 trials running on more than 150 novel drugsopens in new tab/window. But money was never the bottleneck. Nobody agrees on what causes Alzheimer’s, and at conferences, George notes, people still get up and yell at each other in the Q&A.

The contests that do get settled are settled by somebody standing where nobody else was standing – and many never get settled at all.

George watched a version of it two desks away during his PhD, while working on drugs that all aimed at the same well-behaved pocket. His neighbour was collaborating with D. E. Shaw Researchopens in new tab/window – David Shaw, a “Bruce Wayne equivalent” – to build supercomputers that simulate how proteins move. Shaw’s people were computer scientists, not pharmacologists, and they asked what the field had stopped asking: why not find somewhere else on the protein to stick something, and see what happens?

Simulations said yes, and the lab confirmed it. “I discovered this interesting little network within the protein, a little disruption point, that opens up new possibilities, but nobody’s developed it any further. It’s probably still sitting in the publications. And that happens a lot. These cool little side experiments don’t get picked up until somebody a lot more clever than the rest of us looks at it and goes, ‘Hang on, that has legs.’”

The tools have never been in better shape: systems connecting, failures being dug back out, archives retagged and resurfaced. But none of this manufactures the outsider who looks at a written-off target, asks why everyone is standing in the same place, and has the patience to find out.

***

Hear more from Sherry Winter by watching the recent Elsevier webinar “From Molecules to Markets: Leveraging Embase AI in Life Science Research” hereopens in new tab/window.

Related Articles

The microbiome: how the bubble burst and what came back

Beyond the miracle cure: scaling rare disease research into a system

When every drug becomes a possibility

Every Cure: turning old drugs into new hope

Contributor

SK

Steve Korver