2026

AI Can Predict a Protein's Shape Down to the Atom. It Still Can't Tell You Which Patient Will Respond to Your Drug.

That gap sums up where AI in drug discovery actually stands today, and it's the opposite of what most headlines suggest.

Most AI effort in this space is still concentrated at the easiest, earliest stage: finding and optimizing candidate molecules. That's understandable, this is where the data is cleanest and most abundant, exactly the conditions machine learning thrives on. But it's also, in a sense, looking for the keys under the streetlight, not where they were actually dropped.

The keys are dropped later. At phase II, the first real test of whether a drug works in an actual patient population rather than a small, uniform group of healthy volunteers. That's where trials shift from measuring safety to measuring efficacy across people with wildly different genetics, disease severities, and everything else that makes a real patient different from a clean dataset. It's the single most expensive point of failure in the entire drug development process, more costly than any earlier or later stage.

And the fix isn't a better model. It's better matching. In one dataset of roughly 21,000 drug candidates, trials that used biomarkers to select the right patients cost, on average, half as much per successful launch as trials that didn't. Nothing about the drug changed. Who got enrolled did.

That's the real gap in this market. Molecule discovery is crowded and well-funded. The layer that actually decides whether a trial succeeds, correct patient selection and a trial design built around the drug's actual mechanism, is comparatively empty. It's a less glamorous problem than "AI discovers a new drug," but it's the one with a proven, measurable link to whether that drug ever reaches patients at all.

Reference: Bender, A., Cortés-Ciriano, I. et al. “Artificial intelligence in drug discovery — what it is, where we stand and the path forward.” Nature Reviews Drug Discovery (2026).

2026

A Phase 3 Melanoma Result Just Showed What Personalized Cancer Vaccines Can Do

Moderna and Merck's Phase 3 trial of intismeran autogene (an individualized neoantigen therapy, mRNA-4157/V940) combined with Keytruda in resected high-risk melanoma met its primary endpoint of recurrence-free survival, along with the key secondary endpoint of distant metastasis-free survival, at interim analysis.

What makes this notable isn't just the result, it's the approach. Instead of a one-size-fits-all treatment, this therapy is manufactured individually for each patient, based on the specific mutations found in their own tumor. That's a fundamentally different model of drug development than the traditional single-molecule-for-everyone approach clinical trials are usually built around.

The market reaction was immediate and outsized: Moderna's share price and market capitalization moved sharply higher on the news, and Merck also gained. That kind of reaction reflects how much weight the industry is putting on personalized, biomarker-driven approaches as the next real lever for improving trial success rates, not just in melanoma, but as a template for other solid tumors.

One caveat worth keeping in view: this is interim Phase 3 data, not yet peer-reviewed. Confirmatory follow-up and full publication will matter for how durable this signal turns out to be.

2026

A New AI Model Can Show Exactly Which Breast Cancer Patients Chemotherapy Will Actually Help

When someone is diagnosed with the most common form of breast cancer, doctors often order a genomic test called Oncotype DX. It looks at 21 genes in the tumour and produces a single score, built to answer one question: how likely is this cancer to come back? Over time, doctors started using that same score to answer a different question: will chemotherapy actually help this patient? But those are two different things. A cancer can be high-risk without being one chemotherapy will help. Mixing the two up means some patients endure chemotherapy that was never going to work, while others miss out on a treatment that would have.

A team of researchers at Ataraxis AI built a model that tells these apart. It's called CTX. It uses a routine tissue sample already taken at diagnosis, plus basic clinical details already on file, and predicts how much a specific patient's five-year recurrence risk would drop if chemotherapy were added. No new biopsy. No extra genomic panel.

The strongest part is how it was tested: against two real clinical trials it had never seen. In TAILORx (NCT00310180), where doctors had used Oncotype DX to decide who got chemotherapy, CTX found an 11% subgroup who genuinely benefited, a real, measurable five-year survival gain that Oncotype DX had missed. In UNIRAD (NCT01805271), CTX identified patients unlikely to benefit from chemotherapy, and for that group, switching to a different drug, everolimus, significantly improved five-year outcomes, something Oncotype DX simply cannot do, since it only ever gives one risk number. Modelled across the full dataset, the same outcomes could be reached while treating roughly a quarter as many patients with chemotherapy. In plain terms, about three out of four women currently getting chemotherapy for this cancer might not need it.

It didn't stop at breast cancer either. Applied unchanged to sixteen other cancer types, it still worked for cancers treated with similar drugs, and failed on cancers treated with unrelated ones, evidence that it learned something real about treatment response, not just how aggressive a cancer looks.

This is what personalised medicine is supposed to mean: not custom-built drugs for everyone, but knowing, patient by patient, whether the treatments we already have are worth giving.

Reference: Biswas, D. et al. “Causal multi-modal AI for personalised chemosensitivity prediction.” Ataraxis AI (2026). Full article: lnkd.in/e4wfqt39. Trials: TAILORx (NCT00310180), UNIRAD (NCT01805271). #PrecisionMedicine #AIinHealthcare #BreastCancer #ClinicalTrials #DrugDiscovery