A Comprehensive AI-Enabled Pre-Screening Panel for Non-Small Cell Lung Carcinoma (NSCLC) Using Computational Pathology Applied to H&E Whole Slide Images (WSI)
ESMO 2026
C. Bifulco, G. Oakley, J. Bernhard, D. Miller, J. Berndt, S. Senthilnathan, S. Whitman, J. Gordon, B.K.B. Arens - Chow, B. Withey, S. Moulis, N. Lioutas, D. Lee, S. Hariharan, C. Mcwilliams, S. Willingham, R. Yousfi, B. Piening, A. Cummings, D. Carbone
Background
Timely identification of actionable biomarkers in NSCLC patients improves outcomes; however, guideline-recommended NGS-based testing and IHC testing rates remain low. Rapid pre-screening of routine hematoxylin and eosin (H&E) WSI at the time of diagnosis could facilitate follow-through testing and equitable access to timely treatment.
Methods
We developed a machine learning (ML) powered pre-screening panel to simultaneously predict 12 clinically relevant biomarkers using H&E WSI from advanced NSCLC. The underlying ML models were trained using multi-modal DNA sequencing (MSK-IMPACT or Tempus xT) and IHC staining (Dako 22C3, Dako 28-8, Ventana SP263) and the corresponding H&E WSIs. We assessed a regionally diverse multi-center, multi-test cohort of 60,459 cases across 56,969 patients with DNA sequencing available and a subset of 41,680 cases with IHC results also available. DNA sequencing was positive if a case contained a known driver oncogenic mutation for a target gene. We split the data into training (70%), tuning (15%), and holdout (15%) datasets. We selected a prediction threshold to target 90% sensitivity for each biomarker based on the tuning dataset predictions. The selected tuning dataset threshold applies to holdout dataset predictions to compute the reported metrics.
Results
In the holdout set, the NSCLC pre-screening panel demonstrates strong performance for predicting all target biomarkers (see Table). Sensitivity greater than 80% was obtained for all markers, with sensitivity greater than 90% for some rare events such as ALK fusion. These results in the holdout set demonstrate validation of the assay for pre-screening use in NSCLC.
| Biomarker Name | ROC AUC (%) | Sensitivity (%) | Specificity (%) | Prevalence of Alteration in Holdout Dataset (%) | Total Sample Count |
|---|---|---|---|---|---|
| ALK fusion | 90.06 | 92.49 | 63.27 | 2.24 | 9495 |
| BRAF | 84.76 | 89.92 | 56.64 | 1.40 | 9214 |
| EGFR | 88.37 | 89.38 | 69.40 | 10.32 | 9482 |
| ERBB2 (HER2) | 78.50 | 87.06 | 50.99 | 0.90 | 9488 |
| ERBB2 (HER2) AMP | 82.24 | 90.32 | 54.64 | 0.89 | 6980 |
| FGFR1 AMP | 90.33 | 86.01 | 78.26 | 1.64 | 8704 |
| KRAS | 83.43 | 89.63 | 57.33 | 26.52 | 9494 |
| MET | 84.89 | 92.16 | 48.01 | 1.63 | 9415 |
| NTRK1 AMP | 87.85 | 92.27 | 64.4 | 2.53 | 7165 |
| RET fusion | 81.49 | 80.77 | 64.54 | 0.82 | 9495 |
| ROS1 fusion | 85.68 | 85.51 | 66.42 | 0.73 | 9495 |
| PD-L1 >=1% TPS | 78.76 | 89.08 | 41.66 | 64.28 | 5627 |
Conclusions
This ML-based pre-screening panel offers rapid identification of therapeutically relevant biomarkers, and may reduce burdens of time, tissue and cost to improve definitive NSCLC screening rates and outcomes.
Related publications