The first cohort was a public dataset downloaded from SYNAPSE (https://www.synapse.org/Synapse:syn26722053) and recently published (19) comprising 182 patients. The inclusion criteria for this cohort were: patients with stage IV NSCLC who initiated treatment with anti-PD-(L)1 blockade therapy between 2014 and 2019 at the study institution, and who had a baseline CT scan, baseline PD-L1 IHC assessment and next-generation sequencing by MSK IMPACT. The second cohort comprised 108 NSCLC tumor biopsies collected between 2015 and 2024 in the Department of Pathology of the Georges François Leclerc Cancer Center in Dijon, France. The inclusion criteria for this cohort were: patients with stage IV NSCLC who initiated treatment with anti-PD-(L)1 blockade therapy between January 2017 and December 2023 at the study institution.

Only patients from whom informed consent was obtained were included in this retrospective study. The present study was approved by the CNIL (French national commission for data privacy) and the Georges François Leclerc Cancer Center (Dijon, France) local ethics committee, and was performed in accordance with the Helsinki Declaration and European legislation. This study falls within the scope of the biobanking authorization registered under the registration number AC-2014-2260.

MSK Cohort: IHC was performed on 4-μm FFPE tumor tissue sections using a standard PD-L1 antibody (E1L3N; dilution 1:100, Cell Signaling Technologies) validated in the clinical laboratory at the study institution. Staining was performed using an automated immunostaining platform (Bond III, Leica) using heat-based antigen retrieval employing a high pH buffer (epitope retrieval solution-2, Leica) for 30 min. A polymeric secondary kit (Refine, Leica) was used for detection of the primary antibody.

CGFL Cohort: PD-L1 protein expression in tumor cells was assessed using immunohistochemistry with a ready-to-use PDL1 commercial kit with QR1 or 22C3 antibodies. Tonsil tissue served as positive control tissue.

MSK Cohort: PD-L1 IHC-stained diagnostic slides were digitally scanned at a minimum of ×20 magnification using an Aperio Leica Biosystems GT450 v.1.0.0.

CGFL Cohort: PD-L1 IHC-stained diagnostic slides were digitalized with an Evident VS200 (Evident) at 20× magnification to generate a whole slide imaging (WSI) file in vsi format.

For all tumor slides, tumor area zones were manually selected, then these areas were separated into 100µm square tiles. Colors were normalized using MACENKO algorithm (20) and processed using UNI deep learning model (21) to extract high dimensional feature vectors (Figure 1A).

To ensure consistency across datasets, we performed feature-wise normalization using the MSK cohort as a reference. For each of the 1,024 features, we calculated its mean and standard deviation across all MSK patients. These feature-specific statistics were then used to normalize the data in the CGFL cohort: for each patient and each feature, the corresponding MSK cohort mean was subtracted and the result divided by the corresponding MSK cohort standard deviation. This procedure ensures that each feature is scaled relative to its distribution in the MSK cohort.

Quantitative variables are described as median and Interquartile Range (IQR), and qualitative variables as number and percentage. Patient characteristics were compared by cohort (whole cohort, MSK and CGFL) using the Chi-2 or Fisher’s exact test for qualitative variables, and the Wilcoxon rank sum test for continuous variables, as appropriate.

Survival analysis was performed using the survival R library. The prognostic value of the different variables was tested using univariate or multivariate Cox models for PFS when conditions of the model validity were applicable. Proportional hazards assumptions were tested based on Schoenfeld residuals. When the proportionality assumption was not verified, we fitted an extended Cox model, with time dependent coefficients for relevant variables; the time varying coefficient was described with a parametric time function. Survival probabilities were estimated using the Kaplan–Meier method and survival curves were compared using the log-rank test when appropriate. When the proportional hazards assumption was not checked, the estimated restricted mean survival time (RMST) for DFS at 24 months was assessed to compare groups of interest (SurvRM2 R library (22)). P-values less than 0.05 were considered statistically significant.

Statistical analyses were performed using the R software (http://www.R-project.org/) and graphs were drawn using GraphPad Prism version 9.0.2.

We used a public data set from patients treated for NSCLC at Memorial Sloan Kettering (MSK) Cancer Center and who received PD-(L)1-blockade-based therapy. These patients were treated between 2014 and 2019 (cohort characteristics are shown in Table 1). The second data set is constituted of patients treated in France for NSCLC at Center Georges Francois Leclerc between 2015 and 2024 with PD-(L)1-blockade-based therapy or chemoimmunotherapy; this cohort was used as a validation cohort. In the total population, there were more male than female patients. Most patients were smokers or former smokers, and the main histological type was adenocarcinoma. When pooling both cohorts, in first line, 212 patients were treated with anti PD-1-blockade-based therapy and 76 with chemoimmunotherapy. Immunotherapy was used in first line for 146 (66%) patients. PD-L1 TPS status is 0% for 66 patients, between 1 and 49% for 74 patients and greater than 50% for 150 patients.

Comparison of the clinical variable between the two cohorts showed differences for all available characteristics, except for histological type and PD-L1 TPS status with a cutoff at 50%, thus demonstrating the substantial heterogeneity between the two data sets.

10 patients from the MSK cohort were then isolated to train the model. We selected the five patients with the longest Progression-Free Survival (PFS) who did not progress, and the five patients with the shortest PFS who progressed. This corresponds to 361 tiles associated with response and 350 tiles associated with absence of response. Using Principal Components Analysis (PCA) followed by Hierarchical Clustering, tiles were separated into 3 clusters. Cluster 1 was constituted of responders only, cluster 2 was a mixture of responders and non-responders and cluster 3 was enriched in non-responders (Figures 1A, B).

To illustrate which histological patterns distinguish DL^High from DL^Low groups, Figure 2A provides representative tiles of each cluster. Morphologically, Cluster 1 matched tiles with low-cohesive epithelial cells that displayed a negative or an extremely weak stain for PD-L1. Cluster 2 matched tiles that mixed tumor epithelial cells with or without adjacent connective tissue. In this cluster PD-L1 staining was either low or quite strong, localized on tumor cells (TC) or immune cells (IC). Finally, cluster 3 was mainly represented by tiles displaying epithelial tumor cells with strong PD-L1 staining.

These observations were concordant with quantitative evaluation of PD-L1 through staining (Figure 2B).

For the remainder of the patients, we projected each new patient’s tile onto the training PCA space (Figures 1C, D). We looked at which centroid this tile was closest to, and assigned it the label of the corresponding cluster. We then counted the total number of tiles assigned to each of the three clusters and, by majority voting, assigned the patient to the cluster with the most tiles. The same process was then applied in the remaining 172 patients from MSK cohort and on the validation set from 108 patients from CGFL (Supplementary Figure S1).

Clusters 1 and 2 exhibited similar PFS rates (results not shown) and were thus grouped together: in the so-called DL^High group; cluster 3 constituted the DL^Low group. In the training set, 67 patients were attributed to the DL^High group and 115 patients to DL^Low. When looking at response rates, there were 2 Complete Responses (CR), 14 Partial Responses (PR) in the DL^High group and 3 CR and 18 PR in the DL^Low group (Chi-2 test p-value=0.01). When using PFS as an endpoint, patients in the DL^High group had better PFS than patients classified as DL^Low (HR = 0.63 [0.44, 0.89; p=0.01) with a median PFS of 5.7 vs 2.5 months for training cohort. Overall survival was not available for this cohort(Figures 3A, B).

When applying the DL model in the validation cohort, 58 patients were attributed to the DL^High group and 50 patients to DL^Low. When looking at response rates, there were 11 CR and 25 PR in the DL^High group, and 6 CR and 16 PR in the DL^Low group (Chi-2 test p-value =0.22). When using PFS as an endpoint, patients classified as DL^High had better PFS than patients classified as DL^Low (HR = 0.59 [0.36, 0.96]; p=0.03) with median PFS of 15.2 vs 6.2 months for the validation cohort. When looking at Overall Survival (OS), patients classified as DL^High did not have significantly better OS than patients classified as DL^Low (RMST: DL^Low 14.55[12.03;17.07] vs DL^High 16.98 [14.51;19.45]; p = 0.17) with median OS of 37.7 vs 15.2 months (Figures 3C-E).

To complete the analysis, all patients were grouped together and divided according to their treatment. The DL model successfully identified significant subgroups with distinct survival, offering a more refined stratification for patients treated with immunotherapy alone (Figure 3F). For patients treated with chemoimmunotherapy, the DL model did not distinguish patients’ outcome (Figure 3G).

We examined the association between DL model groups and PD-L1 TPS score. In each cohort and in the pooled cohort, PD-L1 TPS score was significantly higher in patients in the DL^High group (Figures 4A–C). However, there was not complete agreement between the two scoring systems: PD-L1 TPS score 0% was detected in the DL^High group, while PD-L1 TPS score >50% were also detected in the DL^Low group.

Moreover, when stratifying patients into high (≥50%) and low (<50%) PD-L1 TPS score groups, the DL model successfully identified significant subgroups with distinct survival, offering a more refined stratification for patients with high PD-L1 TPS score (Figure 4D). In the low (<50%) PD-L1 TPS score group, the DL model did not significantly distinguish patients’ outcome (results not shown).

Clinical variables associated with PFS were selected based on univariate Cox models, and a multivariate clinical model was then estimated based on variables with p-values<0.1 (Figure 5A; Table 2). Because PD-L1 TPS score is correlated with the DL model, this variable was excluded from the multivariate model. WHO performance status, smoking status, treatment information and line of therapy were retained in the model. Variables selected in the clinical model and the DL group variable were combined in a unique multivariate survival model, named the “combined model”. A combined score was then estimated using the linear predictor of the combined model. Using the median as a cut-off, patients with a low score had better PFS than those with a high score (HR = 0.50 [0.33; 0.75]; p<0.001, Figure 5B). Similar observations were made in the validation cohort, using a threshold adapted to the cohort (HR = 0.54 [0.31; 0.91]; p=0.02 (Figure 5C). In the pooled cohort, AUCs of the DL model, clinical and combined model were respectively 0.36, 0.66 and 0.71. The likelihood-ratio test showed that our DL score significantly added prognostic value to the clinical model (p=0.03 when comparing clinical and combined model).