WHAT IS ALREADY KNOWN ON THIS TOPICWHAT THIS STUDY ADDS

We applied ENLIGHT-DP, a novel transcriptomic-based digital pathology platform to scanned high-resolution H&E slides from pretreatment tumor-tissue samples from 50 patients with newly diagnosed lung adenocarcinoma (LUAD). In this proof-of-concept study, ENLIGHT-DP significantly predicts response and progression-free survival to first-line ICI and platinum-based chemotherapy combinations and outperforms both programmed death ligand 1 and tumor mutational burden.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

ENLIGHT-DP offers an accurate transcriptomic-based biomarker for first-line treatment of LUAD with ICI and chemotherapy relying solely on easily accessible H&E scanned slides. Further studies assessing the predictive power of ENLIGHT-DP on different tumor types, as well as prospective bigger NSCLC cohorts are warranted.

Introduction

Non-small cell lung cancer (NSCLC) is the leading cause of cancer-specific mortality worldwide. In the USA alone, approximately 1 25,000 patients are estimated to have died of lung cancer in 2023.1 Immune checkpoint inhibitors (ICI), mainly have dramatically improved the clinical outcomes of patients with NSCLC and have become the standard of care in first-line treatment of metastatic NSCLC who do not harbor actionable driver mutations. Tumor-tissue immunohistochemistry of programmed death ligand-1 (PD-L1) expression predicts response to ICI and directs the clinical application of ICI as monotherapy or in combination with platinum-based chemotherapy.2 3 Nevertheless, even after stratification with PD-L1, the clinical benefit of ICI treatment in NSCLC remains widely varied. In cases of PD-L1 positive tumors the objective response rate (ORR) ranges from 42% to 65% and the median overall survival (OS) ranges from 16 to 28 months. In cases of PD-L1 negative tumors, treatment with ICI remains significantly superior to chemotherapy alone with ORR of 30–35% and median OS of 15–18 months.4–7 Application of other potential predictive markers for ICI treatment in NSCLC, such as tumor mutational burden (TMB), have demonstrated promising results in one prospective trial but, conflicting data in several large retrospective cohorts.7–11 The presence of immune modulating or driver oncogenes, such as STK11, KEAP-1 or KRAS has also revealed equivocal predictive value, and are mostly relevant for specific subgroups (e.g., STK11 mutations are prevalent in 10% of NSCLC cases).12–15 Finally, up to 13% of patients with NSCLC treated with ICI will develop grade 3 or higher immune-related adverse events (irAE), which even if managed properly, warrant cessation of ICI treatment.16 Hence, better and more accurate predictive markers for ICI treatment in NSCLC are greatly needed.

ENLIGHT, developed by Pangea Biomed, is a pan-cancer transcriptome-based computational platform that predicts individual response to ICI and targeted therapies. ENLIGHT is based on the analysis of functional interactions between genes across the human genome, derived using big data from in vitro, in vivo and clinical sources. ENLIGHT uses these interactions to form drug-specific networks of genes (in this case a 10-gene network), whose activation levels, commonly measured through messenger RNA (mRNA) microarrays or next-generation sequencing (NGS), are predictive of response for each drug. In a previous retrospective study, ENLIGHT was able to accurately predict response to specific treatments in 21 different clinical trial databases, achieving an overall OR of 2.59 for response among 697 patients, and 2.39 for ICI.17 Deep-PT is a novel tool which employs deep learning of high-resolution scanned histopathology slides, stained with H&E, to infer full mRNA expression. In a recent study, we showed how combining Deep-PT with ENLIGHT (in a combined pipeline we termed ENLIGHT-DP), allows prediction of response to several targeted therapies and ICI with high accuracy directly from histopathology slides.18

In this study, we present a blinded retrospective analysis demonstrating and validating the predictive value of ENLIGHT-DP in a real-world clinical setting of ICI in first-line treatment of metastatic lung adenocarcinoma (LUAD) and compare it to PD-L1 expression and TMB.

Materials and methodsStudy design

This is a retrospective real-world, single-center observational pilot study, conducted at Hadassah Hebrew University Medical Center (HMC) in collaboration with Pangea Biomed. We first determined categorical thresholds for ENLIGHT-DP classification using previously collected data. We then established a cohort of patients with metastatic NSCLC, scanned high-resolution H&E slides retrieved from tumor-tissue samples from initial diagnostic biopsies and retrospectively applied the ENLIGHT-DP pipeline to generate an individual ENLIGHT-DP prediction score to programmed death 1 (PD-1) inhibitors for each sample in a blinded manner. We then unblinded the clinical outcomes, calculated the predictive power of ENLIGHT-DP and compared it to that of PD-L1 and TMB.

Participants

We retrospectively evaluated our oncology and pathology databases for all patients diagnosed with NSCLC treated at HMC, who underwent comprehensive NGS (Foundation-one Oncomine or Gaurdant). In this initial study, we focused on the more common adenocarcinoma subtype, for which DeepPT showed good performance in predicting mRNA expressions.18 We included only patients who were treated for at least 2 months with first-line ICI (with or without chemotherapy) for newly diagnosed metastatic or locally advanced disease, not amenable for definitive therapy (eg, surgery or chemoradiation), had sufficient data for follow-up for at least 6 months for surviving patients (or earlier in case of death) and had available archived H&E histopathology slides from pretreatment biopsies (primary tumor or metastasis). We excluded cases harboring actionable driver mutations who received targeted therapies for first-line metastatic disease: classical EGFR—exon 19del or exon 21 L858R; rearrangements in ALK, ROS1 or RET; BRAF V600E; MET exon 14 skipping.

Data collection and outcome measures

For each patient, we collected clinical and pathological data from electronic medical records including age, smoking status, performance status, comorbidities, metastatic sites at diagnosis, PD-L1 immunohistochemistry (IHC) expression (assessed by Dako’s pembrolizumab immunohistochemistry test for PD-L1, clone 22C3, using automated Ventana stainer) divided into three ranges: <1%, 1–49% and >50%, somatic pathogenic or likely pathogenic mutations from NGS and TMB, defined as number of mutations per megabase and divided into three ranges: low (<5), intermediate (5–10), or high (>10) mut/megabase. We collected major clinical outcome measures including ORR (by RESICT version .1.1), ICI progression-free survival (PFS), defined as the interval between the start of treatment and progression of disease or death and OS, defined as the interval between the date of start of treatment and death due to any cause. Surviving patients were censored at the date of the last follow-up. Finally, H&E-stained tumor tissue slides, retrieved from initial pretreatment diagnostic biopsies, were converted to high-resolution whole slide images by a Leica Biosystems Aperio AT2 brightfield slide scanner (40× magnification; 0.25 µM/pixel resolution) for ENLIGHT-DP analysis. We note that in a previous study (DeepPT), we demonstrated the predictivity and generalizability of ENLIGHT-DP on histopathological slides originating from different scanners and protocols, thus other scanners and protocols other than the one used here would also allow for the analysis of ENLIGHT-DP.

ENLIGHT-DP Matching Score

The ENLIGHT-DP pipeline is comprised of three steps for generating an individualized ENLIGHT-DP Matching Score (EMS) for response to ICI from a histopathological slide: (1) the indication-specific model of DeepPT is used to predict whole-transcriptome profiles from the H&E slide (in this case NSCLC); (2) individual gene-activation states (over, under or normally expressed) are inferred per gene for each sample based on the gene-wise percentiles of predicted expressions, relative to the respective gene-wise distribution calculated over the entire population (in this case, overall 50 patients), as described in previous study17; (3) the inferred gene-activation states, together with the previously derived ENLIGHT biomarker for response to ICI (consisted of the genes CXCL16, IL15RA, CD27, TNFRSF13C, TNFRSF13B, ICAM4, CD8A, LTBR, CD4, IFITM2), are used to calculate an individual EMS.

Determining EMS response categories

The EMS takes values between 0 and 1. Since a classifier of response has a greater clinical utility than a continuous score, we analyzed the performance of the EMS both as a continuous score and as a response classifier. To enable comparison to two standard biomarkers, namely PD-L1 expression and TMB that are usually divided into high, intermediate, and low categories, we also divide the EMS into three categories, by defining two thresholds using an independent tuning cohort consisting of 161 samples from five data sets of patients treated with PD-1 inhibitors, for which ENLIGHT was previously demonstrated to be predictive17 (see online supplemental table 1). For the upper threshold, we used the same criterion used in our previously published work17 and selected the threshold that maximized PPV (positive predictive value—the fraction of true responders out of those predicted as responders) while achieving at least 40% sensitivity (the fraction of actual responders who were classified as such) in the tuning cohort. This criteria resulted in an upper threshold of 0.6 which coincided with an inflection point of the PPV (see figure 1A). The EMS of 57 of the 161 patients (35.4%) in the tuning cohort passed this upper threshold. Then, since unlike TMB and PD-L1, there is no natural way to define a lower threshold on the EMS, we set a lower threshold that creates balanced groups, here 0.3, which is the 33 percentile of the EMS.

Figure 1

Predictive Value of ENLIGHT-DP Matching Score (EMS). (A) Isotonic regression curve depicting the sensitivity (X axis) and positive predictive value (PPV, (Y axis) of the EMS in an independent cohort of 161 patients treated with PD-1 inhibitors. The left and right asterisks denote the upper and lower EMS thresholds, respectively. (B) ROC AUC (Y axis) comparing the performance of the EMS, PD-L1 expression and TMB. (C) PPV (Y axis) and sensitivity (X axis) for binary classification of responders/non-responders for the three markers: EMS, PD-L1 and TMB. For each marker, two categorizations are examined: (1) denoting only the high category as predicted responders, and (2) denoting both the intermediate and high categories as predicted responders. Dashed horizontal line denotes the baseline response rate of the entire cohort. (D) ROC AUC (Y axis) comparing the performance of the EMS and three transcriptomic signatures (PD-1/PD-L1 mRNA expression, IFN-γ signature and combined biomarker). **p<0.05. AUC, area under the curve; IFN, interferon; IHC, IHC, immunohistochemistry; mRNA, messenger RNA; PPV, positive predictive value; PD-1, programmed death -1; PD-L1, programmed death ligand-1; ROC, receiver operating characteristic curve; TMB, tumor mutational burden.

Comparison to other biomarkers

We compared the predictive value of the EMS to PD-L1 and TMB, as well as to other transcriptome-based biomarkers: (1) the mean mRNA expression of PD-1 and PD-L1 (PD-1/PD-L1 mRNA expression); (2) a transcriptomic biomarker score based on an interferon (IFN)-γ signature19; and (3) a combination of EMS, PD-1/PD-L1 mRNA expression and the IFN-γ signature, which we term the combined transcriptome biomarker, which was previously demonstrated to be favorable for predicting response to ICI.17 18 For all three scores, we used mRNA expressions as predicted by DeepPT.

Statistical analysis

Descriptive analyses were carried out using median and 95% CI, range or quartiles for quantitative variables and percentages for qualitative variables. To measure and compare the continuous predictive power of the different biomarkers the area under the receiver operating characteristic curves (ROC AUC) was used. The p values for ROC AUC>0.5 were calculated using a one-sided permutation test. To measure and compare the categorical predictive power of the biomarkers the following measures were used: (1) ORR were compared between the different biomarkers, demographic or molecular categories using Fisher’s exact test. (2) Survival curves were analyzed using the Kaplan-Meier method and compared between biomarker category groups using the log-rank test. HRs were estimated using Cox proportional hazard regression.

ResultsPatient characteristics

Between January 2020 and December 2023, we evaluated 544 patients with newly diagnosed NSCLC using NGS from tumor-tissue (468) or liquid-based (82) biopsies. Of the 468 patients evaluated using tumor-tissue samples, 252 had adenocarcinoma histology, of which 156 presented with metastatic or locally advanced disease not amenable to definitive therapy. In total we identified 50 patients, with newly diagnosed metastatic LUAD, who were treated with first-line ICI (with or without chemotherapy), had available H&E histopathology slides from pretreatment biopsies and had sufficient outcome and follow-up data. These patients were included in the final analysis.

The median age at diagnosis was 68 years old (range 40–86), 32 patients (64%) were men and 42 (84%) had a history of smoking. Results from NGS revealed pathogenic or likely pathogenic mutations in 46 cases (92%), the most prevalent were TP53 (28 cases, 56%), KRAS (15 cases, 30%), STK11 (6 cases, 12%) and BRCA 1/2 (6 cases, 12%). Further patient characteristics including TMB and PD-L1 expression are detailed in table 1.

Table 1

Clinical characteristics of 50 patients with metastatic non-small cell lung cancer, adenocarcinoma subtype, treated with first-line immune checkpoint inhibitors (ICI)

Treatment and patient outcomes

A total of 46 patients (92%) were treated with combination ICI and platinum-based chemotherapy, of them 28 patients (56%) were treated with a PD-1 inhibitor and 18 patients (36%) were treated with a PD-1 inhibitor and a cytotoxic T-lymphocyte associated protien 4 (CTLA-4) inhibitor. The remaining 4 patients (8%) all of which had PD-L1>50%, were treated with PD-1 inhibitor monotherapy. At the time of data cut-off (May 2024) and a median follow-up of 17.5 months (range 6–52, interquartile 9–24), 29 patients (58%) remained alive. Of them, 13 patients (26%) were continuing to receive first-line ICI treatment, 12 patients (24%) were receiving subsequent treatment and 4 patients (8%) were under observation and best supportive care. The median PFS for first-line ICI was 8 months (95% CI: 7 to 11 months). ORR for first-line ICI (defined as partial or complete response) was 68% (34 patients), of whom 9 patients (18%) achieved complete response. A total of 5 patients (10%) suffered from high-grade irAE which led to discontinuation of treatment. Of them, 2 patients (4%) achieved complete response, discontinued any oncological treatment, and remained without evidence of disease for at least 2 years.

EMS predictive value and comparison to PD-L1 and TMB

The EMS is predictive of response to ICI with an ROC AUC of 0.69 (p=0.01), and outperforms both TMB as a continuous predictor, and PD-L1 expression as measured by IHC with ROC AUC of 0.52 and 0.46, respectively. Using the high cut-off of EMS 0.6, in patients expected to respond to ICI, the EMS achieves 100% PPV and 44% sensitivity. In comparison, predicting response according to PD-L1>50% achieves 65% PPV and 38% sensitivity, thus exhibiting no predictive power since the PPV is lower than the baseline response rate. Predicting response based on high TMB (>10mut/mega-base) has 82% PPV and 26% sensitivity, showing lower PPV and lower sensitivity than the EMS. Overall, the EMS achieves stronger patient stratification than both PD-L1 and TMB. Figure 1B,C depicts the ROC AUC, PPV and sensitivity of the three biomarkers when using thresholds for classification. Table 2 summarizes the ORR of patients in the different response categories defined by the three biomarkers.

Table 2

Objective response rates (ORR) among patients in different ENLIGHT-DP Matching Score (EMS), PD-L1 expression and TMB and categories (low, intermediate, or high). The overall response rate in the cohort is 68%. %lift high versus rest: the difference between the ORR in the high subgroup versus the ORR of the rest of the cohort, divided by the latter

Reassuringly, when focusing on the more homogeneous cohort of 46 patients who received combination therapy, the performance of ENLIGHT-DP remains the same (AUC of 0.68 compared with 0.69 on the entire cohort and PPV and sensitivity for the “high EMS” group of 100% and 42% compared to 100% and 44% on the entire cohort). EMS is the only marker significantly correlated with first-line ICI PFS (HR: 0.45, 95% CI: 0.2 to 0.99 p=0.048) and borderline significant with OS (HR: 0.31, 95% CI: 0.09 to 1.08 p=0.066). For all three markers, the relationship between the marker and PFS or OS is not monotonic: in most cases, the intermediate category has the worst outcomes. Figure 2 compares the PFS and OS according to EMS, PD-L1 expression and TMB, each divided into high, intermediate and low categories.

Figure 2

First-line progression-free survival (PFS) and overall survival (OS) according to ENLIGHT-DP Matching Score (EMS), PD-L1 expression and TMB— Kaplan-Meier curves for PFS (top) and OS (bottom) when stratifying the cohort to low, medium and high-biomarker, based on PD-L1 (left), TMB (middle) and EMS (right). P values and HR were compared between the high group and the rest of the cohort (intermediate and low). PD-L1, programmed death ligand-1; TMB, tumor mutational burden.

EMS predictive value in PD-L1 and TMB outlier groups

To further analyze the clinical utility of ENLIGHT-DP, we evaluated the EMS predictivity in PD-L1 and TMB outlier groups. We observed that the EMS was particularly predictive at stratifying patients with PD-L1<1% (n=18) or TMB-low (n=15) with an ROC AUC of 0.88 and 0.80, respectively (p value<0.05), exceeding that observed in the entire cohort. In the PD-L1>50% (n=20) and TMB-high (n=11) outlier groups, EMS showed a meaningful yet non-significant trend for prediction with ROC 0.64 and 0.69, respectively. Nevertheless, the four patients in this cohort who had PD-L1>50% but suffered rapid disease progression within 3–4 months on first-line ICI had an EMS below 0.5 (below the upper predefined threshold), illustrating the possible application of ENLIGHT-DP for this outlier group. Figure 3 depicts the ROC AUC of EMS in different PD-L1 and TMB outlier groups.

Figure 3

ENLIGHT-DP matching Score (EMS) according to other biomarkers outlier groups ROC AUC of the EMS (Y axis) according to PD-L1 expression and TMB outlier groups. “All” denotes the entire cohort. **p value<0.05. AUC, area under the curve; PD-L1, programmed death ligand-1; ROC, receiver operating characteristic curve; TMB, tumor mutational burden.

Of note, in further univariate analysis (online supplemental table 2), we did not find any demographic factors (ie, age, smoking status), nor the presence of immune modulating or driver oncogenes, such as STK11, KEAP1 or KRAS, that are associated with response to ICI.

Comparison of EMS predictive value to transcriptome-based biomarkers

Since the EMS is a transcriptomic-based biomarker, we further compare it to other transcriptome-based biomarkers, namely the expression of the ICI target (the mean of PD-1 and PD-L1 expression), the IFN-γ signature and a combination of these (see Methods). We found that the EMS was also superior to these three transcriptomic signatures. Figure 1D depicts the ROC AUC of the four transcriptome-based biomarkers.

Discussion

We present the significant predictive value of ENLIGHT-DP for real-world clinical outcomes in first-line treatment of metastatic LUAD with ICI and platinum-based chemotherapy, relying on existing high-resolution H&E slides and surpassing that of commonly used PD-L1 and TMB. ENLIGHT-DP has especially high predictive value in the high binary cut-off groups compared with PD-L1 and TMB, and was the only biomarker significantly correlated with PFS. The fact that ENLIGHT-DP does not only accurately predict response to ICI treatment but relies solely on readily available H&E histopathology slides is making this tool highly applicable for clinical use. EMS categorical cut-offs were predetermined according to analysis of previous data sets and only then applied to this validation cohort. Future studies with larger sample sizes are needed to establish the EMS classification thresholds suggested here.

Importantly, the application of ENLIGHT-DP was able to better stratify response to ICI in the PD-L1 outlier subgroups, which represent a major clinical challenge regarding the decision of whether to combine a PD-1 inhibitor with a CTLA-4 inhibitor or chemotherapy in the first-line treatment of PD-L1<1% or PD-L1>50% or cases, accordingly. One typical example is patients with PD-L1>50% with a high burden of, in which the addition of chemotherapy is considered to achieve a prompt response. In this cohort, one patient in this cohort with PD-L1>50% and high-burden symptomatic disease suffered from rapid disease progression when treated with a PD-1 inhibitor as monotherapy but achieved a durable response when platinum-based chemotherapy was added. A review of this patient’s EMS revealed a score of 0.45 (below the upper threshold), suggesting that the application of ENLIGHT-DP could have better-directed addition of chemotherapy to first-line treatment.

The clinical challenge of first-line ICI combinations in NSCLC is even more emphasized when considering the high per cent of irAE in NSCLC, as many irAE occur before clinical benefit from ICI can be determined, rendering ICI harmful for non-responders.16 20 Of note, one patient in this cohort with PD-L1<1% was treated with first-line combination of PD-1 and CTLA-4 inhibitors with platinum-based chemotherapy and after 2 months experienced severe irAE pneumonitis leading to hospitalization. She was later treated with platinum-based chemotherapy without ICI with a durable complete response lasting over 30 months. While it is difficult to determine whether the durable complete response was the result of chemotherapy alone or an enduring response to chemotherapy and ICI combination, a review of this patient’s EMS revealed 0.3, placing her in the low category and suggesting that the application of ENLIGHT-DP could have prevented unwarranted irAE. Conversely, the two patients who suffered severe irAE, discontinued any oncological treatment but remained without evidence of disease for at least 2 years had an EMS of 0.625 and 0.7, placing them in the high category, indicating the potential use of ENLIGHT-DP in complex clinical scenarios. While the data on PD-L1 and TMB outlier groups in this cohort was highly suggestive, the small sample size renders it to be preliminary at this stage and larger cohorts are needed to validate the utility of ENLIGHT-DP in these clinically challenging subgroups.

ENLIGHT-DP was able to significantly predict PFS for this cohort but only showed a trend in predicting OS. This is probably due to a small sample size, high variability of OS (range 3–52 months), and the effect of subsequent and relatively effective chemotherapies and targeted treatments for patients with actionable driver mutations (ie, ERBB2—four patients, KRAS G12C—three patients, EGFR exon 20ins—two patients) which are not integrated into the ENLIGHT-DP prediction. Additional, larger sample sizes are needed to examine the true benefit of ENLIGHT-DP in terms of OS predictivity, specifically in cohorts where PD-L1 and TMB point to unfavorable responses.

Response to ICI in NSCLC has been shown to be affected by an array of factors including neoantigen presence and recognition, tumor microenvironment and immune cell infiltration, exhaustion and functionality. It is therefore not surprizing that biomarkers that reflect one-dimensional aspects of immune response such as TMB and PD-L1 do not accurately stratify response to ICI leading to the development of more advanced biomarkers.21 Previous retrospective analyses on large cohorts have already revealed the significant predictive value of transcriptomic-based biomarkers in response to ICI treatment in NSCLC. However, most of these studies focused on gene expression subsets encompassing one or two factors of immune response namely, IFN-γ, tumor-infiltrating lymphocytes and DNA damage repair.19 21–25 Ranganath et al found a 27-gene tumor microenvironment expression assay to be significantly associated with ORR and PFS in 67 NSCLC cases receiving ICI; however, this result was not repeated in a separate cohort.26 Immune Response Score, developed by Tomlins et al which combines TMB and expression of four target genes, including PD-1 and PD-L1 was recently shown to retrospectively predict PFS and OS in a large pan-cancer cohort including 110 patients with NSCLC.27 28 Mautofi et al found ribosomal RNA genes to be associated with resistance to ICI in NSCLC when applying molecular-compartment spatially informed high-throughput transcriptomics.29 More recently, Ravi et al conducted a genome-wide analysis of 393 cases of NSCLC, 81% of which were treated with ICI monotherapy. Expression of immunoproteasome, proteasome and IFN-γ associated genes was found to be significantly correlated with response to ICI with immunoproteasome genes as key predictors of response.30 While this novel transcriptomic-based approach is promising, it is still unavailable nor validated for common clinical use. Compared with these markers, the main advantage of ENLIGHT-DP is that it requires no dedicated information other than easily accessible H&E slides, making it a practical tool for clinical use and further prospective research.

The most important limitation of this study is its small sample size which resulted from the exclusion of patients who did not have available archived H&E slides, died before completing 2 months of ICI treatment or were treated with definitive therapies (ie, chemoradiation or surgery), as we could not retrospectively assess radiologic tumor response to ICI. This entailed a relatively small sample size especially for some subgroup analyses. Indeed, while results for the entire cohort were significant, the performance of the intermediate category based on each of the three biomarkers had untypical worst outcomes. In addition, a comparison to landmark studies incorporating ICI treatment in first-line metastatic LUAD, revealed that ORR was relatively high for this cohort at 68%, compared with 47.6% in KEYNOTE-189 and 38% in CheckMate-9LA, while OS and PFS remained relatively comparable.5 7 While ENLIGHT-DP obtained significant results for predicting response and PFS, additional studies with larger sample sizes are warranted to validate its predictive value and demonstrate its clinical relevance.

Another major limitation of this study is that this cohort included a majority of patients (46 of 50, 92%) who received ICI in combination with platinum-based chemotherapy, hence we cannot determine for certainty whether the response predicted by ENLIGHT-DP was to ICI, chemotherapy or the combination. However, the following should be considered: (1) it was shown before that ENLIGHT’s predictivity for targeted therapies does not, seem to be affected by the addition of chemotherapy17 (2) the calculation of ENLIGHT-DP prediction to PD-1 inhibition is based on gene-expression network analysis derived to predict response specifically to PD-1 inhibitors; and (3) the thresholds for this cohort were set based on the tuning cohort in which 161 patients received PD-1 inhibitors. Further analyses on additional cohorts should be done on patients with NSCLC who were treated with ICI monotherapy to validate ENLIGHT-DP’s predictive value of ICI and determine whether chemotherapy should be added to ICI regimens for patients with PD-L1>50%.

Conclusion

In this study, we provide evidence and proof-of-concept of the predictive value of a transcriptome-based approach for response to ICI with platinum-based chemotherapy in first-line treatment of metastatic LUAD. Application of ENLIGHT-DP not only accurately predicts response and PFS to first-line treatment, outperforming commonly used predictive markers, but also allows for a practical and accessible approach relying solely on scanned H&E slides. We believe this work to be of high importance as the first real-world application and validation of this novel biomarker in the rapidly evolving fields of digital pathology and transcriptome-based biomarkers. Further studies assessing the predictive power of ENLIGHT-DP on different tumor types, other lung histologies and treatments including ICI monotherapy as well as a bigger prospective NSCLC cohort are warranted to validate the predictive value of ENLIGHT-DP.

Data availability statement

Data are available upon reasonable request. Data generated in this study are available upon request from the corresponding author.

Ethics statementsPatient consent for publication

Not applicable.

Ethics approval

The study was approved by the HMC Institutional Review Board, reference number: HMO-0398-23. Patient written informed consent was waived as this was a retrospective analysis based on clinical data taken from medical files and already exciting histopathology slides.