EGFR mutations are a major lung cancer subtype. Mutations in the epidermal growth factor receptor (EGFR) gene drive 40-60% of lung cancer cases in Southeast Asian populations and 10-20% in Caucasians. Targeted drugs called EGFR tyrosine kinase inhibitors (TKIs) - such as gefitinib, erlotinib, and osimertinib - are the first-line standard of care for these patients and typically work very well initially.
All patients eventually develop resistance. Despite their early effectiveness, nearly every EGFR-mutant lung cancer patient eventually develops acquired resistance to TKI therapy. At that point, the standard fallback is platinum-based chemotherapy, which offers limited benefits. There is an urgent need for better treatment strategies after TKI failure.
Why immune checkpoint inhibitors usually fail in EGFR-mutant lung cancer. Drugs that unleash the immune system (immune checkpoint inhibitors, or ICIs) have transformed treatment for many lung cancer patients. However, EGFR-mutant tumors are notoriously immune-cold - they have few immune cells infiltrating the tumor and are largely unresponsive to immunotherapy. This cold immune environment makes ICIs largely ineffective in this population.
The key question this study addressed. Does the immune landscape of EGFR-mutant tumors change after TKI resistance develops? If TKI treatment causes the tumor microenvironment (TME) to warm up immunologically, some patients might then respond to ICI-based therapies. Identifying which patients undergo this transformation could help select the right people for immunotherapy after TKI failure.
Reading pathology slides with deep learning. The study used Lunit SCOPE IO, an AI system trained on data from over 26 tumor types. When applied to standard hematoxylin and eosin (H&E) stained tissue slides, the system automatically identifies and spatially maps tumor cells, tumor-infiltrating lymphocytes (TILs - immune cells attacking the cancer), fibroblasts (structural cells that often suppress immunity), endothelial cells (blood vessel lining cells), and tertiary lymphoid structures (organized immune cell clusters associated with better prognosis).
Dividing each slide into spatial tiles. Each whole-slide image was divided into 0.5 mm x 0.5 mm tiles. For every tile, the AI measured TIL density within the cancer area and within the cancer stroma (surrounding connective tissue). Based on these density thresholds, each tile was classified as inflamed (high intratumoral TILs), immune-excluded (TILs present in the stroma but excluded from the tumor), or immune-desert (low TILs everywhere). The slide-level immune phenotype was then determined by which pattern predominated across all tiles.
Validation against gold-standard molecular data. To confirm that the AI was correctly identifying cell types, the researchers compared its predictions against spatial transcriptomics data - a technology that measures gene expression at specific locations within tissue. AI-predicted endothelial cells strongly expressed known EC marker genes (VWF and PECAM1), TILs expressed CD3E (T-cell marker), and fibroblasts expressed ACTA2 and COL5A2. This molecular validation confirmed that the AI's visual cell-type assignments were biologically accurate.
143 patients across two cohorts. The study analyzed 143 NSCLC patients with acquired TKI resistance: 89 received ICI monotherapy in a real-world clinical cohort, and 54 participated in the ATTLAS phase III randomized trial comparing the combination regimen of atezolizumab plus bevacizumab, paclitaxel, and carboplatin (ABCP) against standard platinum-doublet chemotherapy. Critically, 89 patients also had paired tissue samples collected both before and after TKI treatment, enabling direct comparison of how the immune landscape changed.
TKI treatment reduces immune cell infiltration. Comparing paired pre-TKI and post-TKI tissue samples revealed that tumor-infiltrating lymphocytes (TILs) within the cancer area significantly decreased after TKI treatment (p = 0.045). This means that as tumors develop resistance to targeted therapy, they become even more immunologically cold - less infiltrated by immune cells that could potentially fight the cancer.
Blood vessel cells increase with TKI resistance. At the same time, endothelial cells (the cells lining blood vessels) within the cancer area significantly increased after TKI resistance (p = 0.006). This finding suggests that TKI resistance may drive increased angiogenesis - growth of new blood vessels - within the tumor. This was confirmed by a significant correlation between AI-measured endothelial cell density and gene expression signatures of angiogenesis.
Fibroblasts and lymphoid structures were unchanged. In contrast, cancer-associated fibroblasts and tertiary lymphoid structures did not change significantly between pre-TKI and post-TKI samples. This selective change in TILs and endothelial cells suggests specific biological mechanisms driving immune escape rather than a global reorganization of the tumor microenvironment.
PD-L1 expression increased after TKI treatment. The proportion of patients with high PD-L1 expression (tumor proportion score of 50% or more) doubled after TKI treatment, rising from 21.9% to 43.8%. Paradoxically, this increase in a standard immunotherapy biomarker did not reliably predict response to ICI therapy in this cohort - suggesting that in EGFR-mutant tumors, PD-L1 alone is an insufficient guide for treatment decisions.
Not all EGFR mutations respond the same way. Two of the most common EGFR mutations are exon 19 deletion (19del) and L858R point mutation. A third important feature is whether patients develop the T790M resistance mutation after first-line TKI therapy. The study found that these molecular subtypes drive strikingly different patterns of immune landscape change after TKI resistance.
19del and T790M mutations drive immune exclusion. Patients with exon 19 deletion mutations showed significant reductions in TILs within the cancer area (p = 0.045) after TKI resistance. Similarly, patients who acquired T790M resistance mutations had significantly lower TILs after TKI treatment (p = 0.033). These patients appear to transition toward a more immune-desert state after TKI failure, making immunotherapy even less likely to work.
L858R mutations show a different pattern - preserved TILs and more blood vessels. Patients with L858R mutations showed a different response: their TIL levels were largely maintained (p = 0.625, not significant), while endothelial cell density within the cancer area significantly increased (p = 0.009). This pattern - preserved immune presence alongside increased vascularity - may explain why L858R patients showed more favorable outcomes in the ATTLAS trial when treated with ABCP, a regimen that includes bevacizumab (an anti-angiogenesis drug) alongside immunotherapy and chemotherapy.
Post-TKI TME profiles differed by mutation subtype. In head-to-head comparisons of post-TKI samples, L858R tumors had significantly higher endothelial cells and TILs in the cancer area than 19del tumors (both p less than 0.025). T790M-positive tumors had significantly lower TILs than T790M-negative tumors (p less than 0.001). These differences suggest that sequencing a patient's EGFR mutation subtype should inform subsequent treatment decisions after TKI failure.
Inflamed immune phenotype predicts better outcomes with ICI monotherapy. Among the 89 patients who received ICI monotherapy after TKI resistance, those classified as having an inflamed immune phenotype (high intratumoral TILs) had an overall response rate of 38.5%, compared to only 9.9% for patients with non-inflamed phenotypes (p = 0.006). Their progression-free survival was also significantly longer (4.8 vs 1.8 months, hazard ratio 0.48, p = 0.019).
Higher TILs in the cancer area predict response. When TIL density in the cancer area was analyzed continuously rather than categorically, higher TIL levels were associated with a 41.7% overall response rate versus 9.7% for lower TIL levels (p = 0.003), and better progression-free survival (4.9 vs 1.8 months, hazard ratio 0.41, p = 0.006). This relationship between intratumoral immune cell density and response to immunotherapy is biologically intuitive: more immune cells already present in the tumor means a more active immune response ready to be amplified.
An unexpected finding: endothelial cells also predict ICI response. Higher endothelial cell levels in the cancer area were also associated with better ICI response - an overall response rate of 19.3% versus 3.7% (p less than 0.01), and significantly longer progression-free survival (hazard ratio 0.44, p less than 0.001). This counter-intuitive finding may reflect that increased vascularity facilitates immune cell trafficking into the tumor in some patients.
High fibroblasts predict worse outcomes. Patients with higher fibroblast levels in both the cancer area and cancer stroma had significantly shorter progression-free survival on ICI therapy (hazard ratio 1.68, p = 0.026). Cancer-associated fibroblasts are known to create physical and chemical barriers that exclude immune cells from tumors, which aligns with this negative prognostic signal.
The ATTLAS trial provided a prospective validation opportunity. The ATTLAS phase III trial compared ABCP (atezolizumab plus bevacizumab, paclitaxel, carboplatin - a regimen combining immunotherapy, anti-angiogenesis, and chemotherapy) against standard platinum chemotherapy in patients with EGFR-TKI resistant lung cancer. The researchers analyzed post-TKI tissue slides from 54 ATTLAS trial participants to see whether AI-derived TME features predicted which patients benefited from ABCP versus chemotherapy alone.
TILs predicted benefit from ABCP combination therapy. In the ATTLAS cohort, patients with higher TIL levels in the cancer area showed significantly better progression-free survival with ABCP than with chemotherapy alone (hazard ratio 0.42, p = 0.027). This means that patients with more immune-infiltrated post-TKI tumors - a rarer and better group - were the ones most likely to benefit from adding immunotherapy and bevacizumab to chemotherapy.
Endothelial cells showed marginal predictive value. Higher endothelial cell levels also showed a trend toward better outcomes with ABCP versus chemotherapy (hazard ratio 0.29, p = 0.067 - just above the conventional significance threshold). This marginal association may reflect the fact that bevacizumab, an anti-VEGF drug targeting angiogenesis, would be particularly effective in tumors with high vascular activity.
These findings support routine post-TKI biopsy. Since the immune phenotype of a tumor can change substantially after TKI resistance develops, using pre-TKI tissue samples to make treatment decisions after resistance would be misleading. The study makes a strong case for obtaining new biopsy samples at the time of TKI failure to accurately profile the current TME status and guide subsequent treatment choices.
PD-L1 is the standard immunotherapy biomarker - but it fails here. PD-L1 expression, measured as the tumor proportion score (TPS), is the main FDA-approved biomarker for selecting patients for immune checkpoint inhibitor therapy. In this study, PD-L1 TPS increased significantly after TKI treatment - yet it showed no significant predictive value for ICI response in the EGFR-mutant post-TKI setting.
Why PD-L1 can be misleading in EGFR-mutant tumors. EGFR-mutant tumors can upregulate PD-L1 through compensatory signaling pathways after TKI treatment (such as through PD-L1 induction by transforming growth factor-beta or sustained EGFR signaling) without this reflecting genuine immune activation. Additionally, increased regulatory T-cell activity and CD73 expression can maintain an immunosuppressive environment even when PD-L1 appears high, limiting actual immune activation despite the apparent positive biomarker signal.
Spatial distribution captures what single-marker tests miss. The spatial analysis approach used in this study - measuring where immune cells, fibroblasts, and blood vessel cells are located within the tumor versus the stroma - provides information about the functional immune state that a simple PD-L1 staining score cannot. A tumor can have high PD-L1 but still be immune-excluded (TILs trapped in the stroma, unable to reach tumor cells) or functionally suppressed by fibroblasts and regulatory cells.
AI-derived spatial biomarkers offer complementary and superior information. The fact that AI-measured TIL levels and endothelial cell density predicted ICI outcomes while PD-L1 did not demonstrates that these spatial features capture biologically distinct and clinically relevant information. Incorporating spatial AI analysis alongside or instead of PD-L1 testing could substantially improve patient selection for immunotherapy in the post-TKI resistance setting.
AI pathology analysis could change how post-TKI patients are selected for treatment. Currently, treatment selection after EGFR-TKI resistance relies heavily on identifying the resistance mechanism (such as T790M) and then choosing the next targeted drug or chemotherapy. The findings here suggest that routine AI-powered spatial analysis of post-TKI biopsy slides could guide additional decisions about whether to add immunotherapy or anti-angiogenesis drugs to the next treatment line.
Practical advantage: H&E slides are already routinely collected. A key strength of the AI-powered approach is that it works on standard H&E-stained slides - the most common type of pathology preparation already performed at every cancer center worldwide. No additional expensive molecular testing is required. The AI analysis simply extracts more information from slides that would otherwise show only basic tumor morphology to the pathologist.
L858R patients may benefit most from ABCP-type regimens. The mutation-specific TME patterns suggest a practical clinical algorithm: patients with L858R mutations, who showed preserved TILs and increased vascularity after TKI resistance, may be particularly suitable for combination regimens like ABCP that simultaneously target the immune system and blood vessel growth. Patients with 19del or T790M mutations, who trend toward immune-desert tumors, may require different strategies to generate an immune response first.
The technology is validated but needs scaling. The Lunit SCOPE IO system used in this study is already commercially deployed and used at over 200 institutions. The results show that it can provide clinically meaningful predictions beyond its originally validated uses, but prospective multicenter trials are needed to formally establish this approach as a standard-of-care companion test.
Small sample size limits definitive conclusions. With 143 total patients (89 for ICI monotherapy and 54 in the ATTLAS cohort), this study was exploratory in nature. The cutoffs used to classify patients as high or low TIL or EC were derived from the same cohort and need independent prospective validation. The findings are promising but should not change clinical practice until confirmed in larger, multicenter studies.
Heterogeneity in biopsy timing and site introduces variability. Post-TKI biopsies were taken at different time points after resistance, from different tumor sites, and from different tissue types (primary tumor vs. metastasis). Each of these variables can affect the measured TME composition. While this reflects real-world clinical practice and actually strengthens the generalizability argument, it also introduces noise into the analysis.
H&E images cannot fully characterize immune function. Standard stained slides can identify which cells are present and where, but cannot assess functional immune states like T-cell exhaustion - a critical factor in whether immune cells can actually kill tumor cells even when they are present. Future AI models incorporating spatial transcriptomics or multiplex protein imaging could provide richer functional characterization of the immune landscape.
Spatial omics integration is the next frontier. The study validated AI predictions against spatial transcriptomics, demonstrating strong concordance. Combining AI-based spatial histology analysis with spatial gene expression or spatial proteomics in larger cohorts would enable development of more comprehensive and biologically grounded predictive models - potentially capturing functional immune states, treatment-specific vulnerabilities, and patient-specific molecular heterogeneity within a single analytical framework.