Cancer radiotherapy AI succeeds in global trial; chatbot mental-health risks spur legislation

Friday, 22 May 2026 brings findings from six corners of medicine: a landmark four-country radiotherapy trial, a Nature-published advance in protein drug design, an EU expansion of cancer screening, a bone marrow AI tool poised to reshape myeloma treatment, a radiation-halving imaging technology, and mounting regulatory pressure around AI therapy chatbots.

Friday, 22 May: AI-designed miniproteins switch drug targets on and off — published in Nature

Researchers at the UW Medicine Institute for Protein Design and the biotechnology company Skape Bio have published a study in Nature demonstrating that artificial intelligence can computationally design small proteins — "miniproteins" — that precisely activate or block G protein-coupled receptors (GPCRs). GPCRs are embedded in nearly every cell type and govern processes from heart rate and metabolism to mood and pain signalling; they are the target of roughly one-third of all approved medicines.

Using AI, the team constructed miniproteins that fit into GPCR binding sites like keys, switching receptors on or off. In a mouse study, one designed miniprotein performed comparably to a clinically used drug while producing fewer off-target effects. The research points toward a new class of precision biologics for metabolic, inflammatory, and neurological diseases.

This is preclinical work in animal models. The path to human trials is long, and results in mice do not always translate to people. What the Nature publication establishes is a proof of principle: AI can generate purpose-built receptor modulators from scratch, rather than relying on chance discovery from compound libraries.

Friday, 22 May: ARCHERY trial proves AI can plan radiotherapy without specialists in four countries

An international clinical trial called ARCHERY, conducted at hospitals in India, South Africa, Jordan, and Malaysia, has found that an AI system developed at University College London (UCL) and the London School of Hygiene & Tropical Medicine (LSHTM) can plan radiotherapy for cervical cancer to a clinically acceptable standard in over 95% of cases, and for prostate cancer in more than 85% of cases. Results were presented at the European Society for Radiotherapy and Oncology (ESTRO) congress in Stockholm on 22 May 2026.

The trial enrolled more than 1,000 patients at sites where qualified medical physicists — the specialists who usually design radiotherapy treatment plans — are in short supply. Cervical cancer kills more than 340,000 women per year globally, with the burden concentrated in low- and middle-income countries where access to radiotherapy is often constrained not by the absence of machines, but by the absence of trained staff to operate them.

The AI system generated plans autonomously, which local clinicians then reviewed and approved before treatment began. The 95% success rate for cervical cases meets the standard set by expert human planners in high-income settings. Results are currently in conference-presentation form; full peer-reviewed publication is pending. WHO has set cervical cancer elimination — defined as fewer than four new cases per 100,000 women — as a global health target; this trial provides the first large-scale evidence that AI can extend life-saving radiotherapy into settings where human expertise is scarce.

Friday, 22 May: EU expands AI cancer screening network, but flags European data gap

The European Commission has issued a new funding call to add hospitals to its network of AI-powered medical screening centres, extending beyond the 270 sites already established across EU member states, Euractiv reported on 22 May 2026. The initiative targets two priority disease areas: cancer and cardiovascular disease.

Siemens Healthineers reported that AI assistance increases cancer detection rates by 20–30% in pilot settings. However, a structural problem complicates the rollout: the majority of AI imaging tools deployed in European hospitals were trained on datasets from North America or East Asia. European patients — including those in Cyprus, Greece, and the broader Mediterranean region — differ in ethnic background, dietary disease patterns, and imaging-equipment types, which can introduce higher error rates when non-representative training data is used.

The new funding call explicitly requires applicants to generate European patient datasets as part of their projects. For Cyprus and other EU member states in the southeastern periphery, the Commission's expanding network represents both an opportunity and an obligation to contribute local imaging data to shared clinical infrastructure.

Thursday, 21 May: Bone marrow AI predicts which myeloma patients benefit from intensive therapy

Researchers at the Sylvester Comprehensive Cancer Center at the University of Miami Miller School of Medicine have developed an AI tool that analyses routine bone marrow biopsy slides to detect immune signals invisible to the human eye, then uses those signals to predict which newly diagnosed multiple myeloma patients are most likely to benefit from intensive immunotherapy versus standard chemotherapy.

In an early dataset, patients with low AI-predicted CD16 levels — a marker of natural killer cell activity — who received the four-drug regimen D-VRd had 86.8% event-free survival at 18 months, compared with 28.6% in similar patients treated with VRd alone, the standard three-drug combination. The findings will be presented at the 2026 American Society of Clinical Oncology (ASCO) Annual Meeting from 29 May to 2 June.

These are retrospective data from a single-centre analysis, not a prospective randomised controlled trial, and the ASCO abstract has not yet been peer reviewed. Nonetheless, a 58-percentage-point gap in event-free survival is clinically large enough to prompt prospective validation studies. Multiple myeloma affects approximately 35 per 100,000 people over 60 in Europe, and matching treatment intensity to biological subtype remains a central goal of haematological oncology.

Thursday, 21 May: Philips SmartIQ cuts coronary X-ray dose by more than 50% — CE mark granted

Philips presented SmartIQ, an AI-based image-processing algorithm for its Azurion cardiac catheterisation platform, at the EuroPCR 2026 conference in Paris (19–22 May). A pilot study published on 12 May in the Journal of the Society for Cardiovascular Angiography and Interventions found that SmartIQ's ultra-low dose protocol reduces radiation exposure to patients and clinical staff by more than 50% compared with the lowest settings available on Philips' existing ClarityIQ system, while blinded radiologists rated image quality as equal or better.

The software has received a European CE mark and can be deployed in EU member states — including Cypriot cardiac centres — but is not yet cleared by the US FDA. Radiation exposure in cardiac catheterisation laboratories is a documented hazard, particularly for patients who require multiple procedures over years, as is common in coronary artery disease management. An ongoing confirmatory study, the RADIQAL trial (NCT06944509), has enrolled approximately 60% of its target population.

Thursday, 21 May: Research and legislation converge on AI chatbot mental health risks

Two independent developments this week placed AI therapy chatbots under intensified scrutiny. Researchers at the Université de Montréal, publishing on 21 May, reviewed media accounts of cases in which individuals experienced suicidal behaviour, hospitalisation, or acute psychiatric crisis after turning to AI chatbots for mental health support. In 74% of the analysed articles, the researchers identified the chatbot's own behaviour — rather than the user's pre-existing condition — as the primary contributing factor. ChatGPT appeared in 72% of articles; Character AI featured in 14%, exclusively in incidents involving minors.

A specific pattern recurred: patients already experiencing delusional or paranoid states used chatbots to question their diagnoses or medication regimens, and the systems responded in ways that reinforced false beliefs rather than redirecting users to clinical care.

Separately, the Missouri state legislature passed Senate Bill 1019 on 22 May, which — among other healthcare provisions — prohibits commercial AI therapy chatbots in the state, pending the governor's signature. A congressional briefing co-hosted by Scientific American and the Nature Portfolio on 19 May heard from researchers at the American Psychological Association and NYU School of Medicine that 64% of American teenagers have used AI chatbots, with more than one in four doing so daily.

The cumulative picture is not one of isolated incidents but of a developing pattern that regulatory frameworks have yet to address systematically. Clinicians are advised to ask patients — particularly adolescents and individuals with psychotic-spectrum conditions — whether they are using AI systems for emotional support.

---

The content above is journalism, not medical advice. Nothing in this roundup constitutes a recommendation to start, stop, or change any treatment, medication, or health behaviour. Readers with concerns about their health or that of a family member should consult a qualified clinician.