AI Reads Standard Pathology Slides to Predict Lung Cancer Mutations

AI extracts genomic and immune signatures from routine tissue slides

Artificial intelligence models can identify specific genomic mutations and predict immunotherapy responses in non-small cell lung cancer by analyzing standard pathology slides, according to research from medical AI company Lunit scheduled for presentation at the 2026 World Conference on Lung Cancer in Seoul.

The studies demonstrate that computational pathology applied to hematoxylin and eosin (H&E) stained slides—the most common tissue preparation method in pathology labs—can reveal tumor microenvironment characteristics that correlate with specific genetic alterations and treatment outcomes. The approach could enable labs to extract molecular insights from existing slide archives without additional testing.

Why it matters

Genomic testing for lung cancer biomarkers typically requires separate molecular assays that add cost and turnaround time. If AI models can reliably predict mutation status and treatment response from slides already produced during standard diagnosis, pathology labs could triage cases more efficiently, prioritize molecular testing for ambiguous cases, and potentially identify trial-eligible patients faster. The research suggests computational pathology may bridge the gap between morphology and molecular characterization.

EGFR mutation subtypes show distinct immune profiles

In the first study, Lunit's SCOPE IO software analyzed 494 whole-slide images from patients with EGFR-mutant NSCLC. The AI identified distinct tumor microenvironment patterns across mutation subtypes. Tumors with exon 19 deletions showed significantly lower tumor-infiltrating lymphocyte density, while L858R mutations correlated with enriched lymphocyte and macrophage infiltration. Exon 20 insertion tumors displayed higher endothelial cell density.

The findings suggest that clinical outcome differences among EGFR subtypes may reflect not only the mutations' effects on kinase activity but also their influence on the surrounding immune landscape.

Spatial analysis predicts immunotherapy response

A second study, conducted with Paola Nistico of the Regina Elena National Cancer Institute in Rome, examined tissue from 32 NSCLC patients who received neoadjuvant chemo-immunotherapy. Researchers combined AI image analysis with spatial transcriptomics and spatial proteomics.

Patients who achieved pathological complete response showed highly inflamed tumors with organized immune structures, including prominent tertiary lymphoid structures. Non-responders exhibited immune exclusion and stroma dominated by activated fibroblasts. The spatial patterns identified by AI correlated with treatment efficacy.

TP53 mutation prediction from H&E alone

The third study validated an AI model that predicts TP53 mutation status directly from H&E slides in lung adenocarcinoma. In an independent cohort of 462 cases, the algorithm achieved an area under the curve of 0.759, with 82% sensitivity and 63% specificity.

Spatial analysis revealed that TP53-mutant tumors more frequently displayed an immune-inflamed phenotype, while wild-type tumors showed immune exclusion, higher endothelial cell density, and elevated stromal fibroblast counts. The researchers suggested the model could serve as a screening tool to enrich clinical trial cohorts targeting specific TP53 variants.

"By broadening the range of insights that can be derived from routinely available pathology images, we aim to advance AI-powered biomarker discovery and patient stratification," said Brandon Suh, CEO of Lunit, in a company release.

The research was first reported by CLP Magazine and is scheduled for presentation at the World Conference on Lung Cancer.