A study in Nature Medicine reports that AI models beat conventional biomarkers when forecasting immunotherapy benefit in advanced lung cancer.
The work comes from I3LUNG, an international project that enrolled 2,396 patients with metastatic non-small cell lung cancer. Participating centers spanned Italy, Germany, Greece, Israel, Spain and the United States.
Researchers pooled clinical, imaging, pathology and genomic data for each patient, then trained two families of models to forecast treatment response and survival rather than a single endpoint.
The models consistently beat every standard biomarker, including PD-L1 expression, which physicians lean on today despite known limitations. Performance reached the strong range on AUC, a metric where 0.8 to 0.9 is considered excellent.
Senior author Marina Garassino, a thoracic oncologist and professor of medicine at UChicago Medicine, framed the shortfall as a tooling problem rather than a data problem.
Immunotherapy produces durable benefit in only 20% to 30% of patients, and most eventually develop resistance. Sharper prediction at diagnosis would spare people unnecessary toxicity and cost while steering them toward treatments more likely to work.
The findings position multimodal AI as a decision aid alongside pathology, not a replacement for it. Validation across six health systems strengthens the case. Wider adoption will still depend on regulators accepting model output as evidence in treatment planning, a bar no AI oncology tool has cleared at scale.
