AI identifies markers to avoid unnecessary chemotherapy

Researchers from RCSI University of Medicine and Health Sciences and University College Dublin (UCD) published a study in Nature Communications (DOI 10.1038/s41467-026-73432-2) on June 23, 2026 showing that AI-based analysis of the tumour microenvironment can extract immune markers that better stratify risk for patients with early-stage ER+HER2- breast cancer, a subtype that accounts for around 70% of all breast cancer diagnoses annually, per RCSI. Patients in this group are currently assessed with a genomic risk score, but most receive an intermediate result, so chemotherapy is often prescribed as a precaution. The researchers analysed tissue from the Irish arm of the randomized TAILORx trial, which compared hormone-blocking therapy alone against hormone-blocking combined with chemotherapy in patients with intermediate risk scores. They found that a high density of CD8+ cytotoxic T-cells in the stromal tissue surrounding a tumour predicted poorer outcomes when patients were treated with chemotherapy, indicating those patients are less likely to benefit from it. Research lead Professor Darran O'Connor said: "For patients with an intermediate genomic risk, the decision around chemotherapy is often difficult and uncertainty frequently leads to treatment that may not have been necessary, impacting on quality of life... because this approach works from tissue samples processed as standard, it has the potential to improve both the precision and the equity of treatment." First author Dr Zak Kinsella said the T-cell density measure "proved to be a remarkably strong predictor of treatment response."

The team applied digital pathology and an adapted open-source AI tool to standard histology slides to quantify immune-cell density in the stromal regions adjacent to tumours, rather than requiring a bespoke assay. RCSI and UCD have jointly filed a patent for the technology and are now seeking to commercialize it to support translation into clinical practice, per RCSI.

Clinical AI projects built on digitized histology and cell-density metrics have moved over the past several years from retrospective signal-finding toward hypothesis-driven validation in trial-derived cohorts, and this study follows that pattern by using archived randomized-trial tissue rather than a purely retrospective dataset. That distinction matters: coupling an AI-derived spatial immune metric with outcomes from a randomized trial (TAILORx) is a stronger form of evidence than a survival correlation drawn from an uncontrolled cohort, and could eventually complement existing genomic assays like the 21-gene test used in TAILORx itself.

Two features of this approach lower the bar for eventual clinical use: it works from formalin-fixed tissue already processed as part of standard pathology workflows, and it reuses an adapted open-source AI pipeline rather than a proprietary black box. Both authors are explicit that this is not yet a deployable test. Senior author Professor William Gallagher said: "Before this approach can be implemented in clinical practice, further validation in larger studies will be required."

• •Prospective validation in independent cohorts that replicate stromal CD8+ density as a predictor of chemotherapy benefit.
• •Reproducibility across labs, including staining variability and concordance for the adapted open-source AI pipeline.
• •Progress on the RCSI/UCD patent and commercialization effort, and whether pathology workflows can integrate the tool with pathologist oversight.

The study's strongest evidentiary feature is also its main current limitation: it draws on real randomized-trial outcomes, but from a single national trial arm, so the cohort size is modest and the authors themselves call for larger-scale, multicentre validation before this could inform actual chemotherapy decisions. Readers should treat this as a promising predictive-biomarker candidate, not a test that is ready to change treatment today.