Predictive modeling of immunotherapy efficacy in driver gene-negative non-small cell lung cancer.

Introduction
In many nations and areas, lung cancer has the greatest incidence and mortality rates among any malignant tumor. About 80–85% of lung cancer cases are non-small cell lung cancer (NSCLC), and individuals with driver gene-negative NSCLC respond poorly to conventional treatment modalities like chemotherapy. In immunotherapy, immune checkpoint inhibitors (ICIs), which are represented by programmed cell death protein 1 (PD-1) and programmed death-ligand 1 (PD-L1), have been evolving significantly. The quick development of immunotherapy has led to significant advancements in the treatment of lung cancer. Immunotherapy plus chemotherapy has clear advantages over chemotherapy alone and is now a suggested course of treatment for patients with locally or driver gene-negative advanced NSCLC (1). Meanwhile, an increasing amount of evidence shows that immunotherapy and chemotherapy combined also increase early-stage patients’ chances of survival (2).
Unscreened NSCLC patients receiving immunotherapy still have a poor total response, nevertheless, which makes treatment prediction crucial. Although PD-L1 expression has been demonstrated to be a useful biomarker for evaluating the effectiveness of immunotherapy, anti-PD-1/PD-L1 therapeutic response rates in PD-L1-positive NSCLC patients are still just 15.6–48%, while immunotherapy can also be beneficial for some PD-L1-negative individuals (3-5). As PD-L1 testing is still not widely used in many primary medical organizations, it has proven challenging to forecast the effectiveness of ICIs using a single biological indicator in the context of precision tumor therapy (3,4).
There is currently plenty of research on the prediction of NSCLC immunotherapy success, and the primary indices are computed tomography (CT) imaging histology (6,7), tumor metabolic indexes of positron emission tomography (PET)-CT (8-10), tumor markers (11), inflammatory indexes (12) and tumor mutation burden (TMB) (13). It is now challenging to generalize to the clinic because of the wide variations in picture histology characteristics across various research institutes caused by the various parameters of CT machines and the non-uniformity of image segmentation standards. PET-CT (8) is a crucial test for assessing tumor spread and staging since it is characterized by repeatable monitoring and strong consistency of tumor metabolic markers. However, because of its expensive cost, many patients choose not to have the test. On the contrary, tumor markers, inflammatory markers, and TMB are readily available, highly prevalent, and have all been demonstrated to be associated with the results of immunotherapy in patients with lung cancer. Nevertheless, the current research also predicts immunotherapy based on a single or single type of indication, which still has significant limits. Consequently, more accurate prediction findings may be obtained from a study of multidimensional and multi-indicator combined prediction of the effectiveness of immunotherapy for NSCLC. In populations with weak immune responses, further immunotherapy and chemotherapy combinations with other forms of treatment (such as radiation) might increase the benefit of survival.
In this study, we gathered relevant markers that are easier to get during the diagnosis and treatment process in order to build a prediction model of immunotherapy achievement for NSCLC patients with negative driver genes. Meanwhile, more scientific direction for precision therapy of NSCLC may be provided by model-based screening of the population with inadequate immune response and evaluation of the therapeutic benefit of early imposition of combined therapies. We present this article in accordance with the TRIPOD reporting checklist (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-923/rc).

Methods

Predictive modeling of immunotherapy efficacy in driver gene-negative NSCLC
Retrospective gathering of case data using an electronic medical record system from patients with driver gene-negative NSCLC treated at the First Affiliated Hospital of Army Medical University between January 1, 2017 and December 31, 2024. The following were the requirements for inclusion: (I) a person who is at least 18 years old; (II) NSCLC with a negative driver gene that has been pathologically verified; (III) all first-line therapy, including or excluding combination chemotherapy, with ICIs for at least four cycles; (IV) Eastern Cooperative Oncology Group (ECOG) score of 0-1. The following were the exclusion criteria: (I) contemporaneous primary malignant tumors in other locations; (II) concurrently undergoing additional therapies (anti-angiogenic therapy, radiation therapy, etc.) while undergoing first-line treatment; (III) serious infections; autoimmune diseases; significant physical or mental illnesses; or any condition interfering with treatment protocols or impacting short-term life prognosis; (IV) insufficient clinical information (over 20% absent); (V) individuals had surgical indications but no surgical intervention. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of the First Affiliated Hospital of Army Medical University (No. KY2025114). Medical records obtained in previous clinical treatment were collected and analyzed retrospectively, therefore, individual consent for this study was waived. Classification criteria and inclusion variables (all baseline data collected prior to treatment): (I) gender; (II) age; (III) smoking index; (IV) kind of pathology (non-squamous or squamous); (V) stage of tumor (stage I, II, III, IV); (VI) if it was a recurrence following surgery; (VII) chemotherapy protocol (no combination of chemotherapy, single-agent chemotherapy, or dual-agent chemotherapy); (VIII) TMB; (IX) Ki-67; (X) carcinoembryonic antigen (CEA); (XI) neutrophil-to-lymphocyte ratio (NLR, the ratio of neutrophil count to lymphocyte count in peripheral blood. It is used to assess systemic inflammatory, infectious, or immune status). The Response Evaluation Criteria in Solid Tumors (RECIST) 1.1 criteria defined durable clinical benefit (DCB) as partial-remission (PR) or stable disease (SD) for more than 6 months. In this study, no durable benefit (NDB) was utilized as an endpoint.

Study on the efficacy of first-line combination regimen therapy in patients with poor immune response
Gathering of patients with driver gene-negative NSCLC who visited the First Affiliated Hospital of Army Medical University between January 1, 2017, and December 31, 2024, had their immunotherapy efficacy predicted using the best model that was constructed. The study population was the group that was screened for NDB. The following were the criteria for inclusion: (I) a person who is at least 18 years old; (II) NSCLC with a negative driver gene that has been pathologically verified; (III) receiving ICIs as a first-line treatment for at least four cycles, either alone or in combination with chemotherapy; (IV) ECOG score of 0–1. The following were the exclusion criteria: (I) concurrent primary malignant tumors in other locations; (II) a confluence of severe illnesses that impact the course of treatment; (III) a missing value over 20%.

Statistical analysis
Data analysis was conducted using SPSS 26.0 and R 4.2.2. Missing data was filled up via multiple interpolation, and all patients were divided into training and validation sets at random in a 7:3 ratio. Univariate plus multivariate logistic regression methods were used to screen the training set for variables that would eventually be included in the model, and P<0.05 was considered statistically significant. A total of 10 machine learning algorithms were used to construct and validate the model by incorporating the screened final variables with the ending of whether or not durable benefit is lasting, and 5-fold cross-validation and hyperparameter tuning were used to optimize the accuracy rate. Plotting receiver operating characteristic (ROC) curve and determining their area under the curve (AUC), precision-recall (PR) curve and determining their AUC, calibration curves for measuring model accuracy, and decision curve analysis (DCA) for determining the model’s level of therapeutic benefit. The best model was chosen following an extensive evaluation and deployed as an interactive web calculator via Shinyapps.io.
Patients treated with combination radiation therapy were assigned as the combination therapy group based on their treatment of immunotherapy (with or without chemotherapy) in the first line, whereas patients not receiving any other treatment regimens were designated as the control group. Both groups’ proportion of patients who experienced NDB (disease advancement within 6 months) was determined, and the difference was compared using the Chi-squared test (or Fisher’s exact test). The Kaplan-Meier method constructed subgroup the progression-free survival (PFS) curves, with the log-rank test evaluating differences, with P<0.05 deemed statistically significant.

Results

Predictive modeling of immunotherapy efficacy in driver gene-negative NSCLC
A total of 161 patients were included in the study based on the inclusion and exclusion criteria. Figure 1 displays the case screening flow chart. Patients were 60.7±9 years old on average, with 143 (88.8%) males and 18 (11.2%) females. Of them, 46 (28.6%) were NDB. Following a random division of the aforementioned cases into a training set (n=113) and a validation set (n=48) in a 7:3 ratio, the baseline data differences between the two groups were found to be not statistically significant (Table 1), and there was substantial data uniformity between the training and validation sets. All variables in the training set were subjected to univariate logistic regression analysis; those that were significant (P<0.05) were then included in the multivariate logistic regression. The final results indicated that tumor stage {P=0.002, odds ratio (OR) =4.294 [95% confidence interval (CI): 1.693–10.890]}, Ki-67 [P=0.007, OR =1.028 (95% CI: 1.008–1.049)], and TMB [P=0.032, OR =0.913 (95% CI: 0.840–0.992)] were independent influences on the outcome (NDB) (Table 2).
While the average ROC-AUC and PR-AUC of the 5-fold cross-validation are compared with the ROC-AUC and PR-AUC of the training and validation sets of all models in Table 3 and Figures 2,3, it is evident that the light gradient boosting machine (LightGBM) model ranks in the top three for each of the evaluation dimensions listed above. In meanwhile, the model’s calibration curves for both the training and validation sets demonstrated high accuracy (Figure 4). Additionally, the DCA curves for both sets made it clear that the LightGBM model’s net clinical benefit for prediction was the highest of all models (Figure 5). In conclusion, LightGBM was ultimately determined to be the best model in this investigation.
The LightGBM-based prediction model was further generalized by creating an online calculator (https://lightgbmpredict.shinyapps.io/LGBM/). By visiting the website and entering the values of the relevant variables in the option box, it is possible to predict the possibility that driver-negative NSCLC patients will experience no lasting benefit following immunotherapy. For example, in order enter tumor stage = stage IV, TMB =8, and Ki-67 =70%, and then click “Calculate”, the result will be high risk (probability 54.80%), meaning that patients with the aforementioned conditions are more likely to experience early advancement (NDB) following immunotherapy (Figure 6).

Study on the efficacy of first-line combination regimen therapy in patients with poor immune response
Baseline characteristics were collected prior to treatment initiation. Using the web-based predictive calculator, prognostic probabilities were computed for each case. Patients with >50% probability of poor immune response were classified as high-risk. Ultimately, 52 patients meeting this risk threshold were identified as potentially deriving limited durable benefit from therapy. All of these patients received immunotherapy and chemotherapy as part of their first-line treatment; 27 of these patients also received radiation therapy (combination therapy group) and 25 did not receive any additional treatments (control group). There was no statistical difference in the distribution of baseline information between the two groups. In the combination therapy group, there are 21 (77.7%) males and 6 (22.3%) females, and 8 (29.6%) patients advanced within 6 months of treatment (NDB), while in the control group, there are 23 (92.0%) males and 2 (8.0%) females, 16 (64.0%) patients had NDBs after treatment. The Chi-squared test results indicated (Figure 7A) that the combination treatment group had a significantly lower incidence of NDB than the control group (P=0.03). According to the survival analysis (Figure 7B), the two groups’ median PFSs were 9.6 (95% CI: 7.882–11.318) and 5.6 (95% CI: 4.988–6.212) months, respectively, and the combination treatment group had a better PFS than the control group (log-rank test P=0.04).

Discussion
In clinical research, predicting the effectiveness of immunotherapy in driver gene-negative NSCLC has always been an important subject. Several predictive markers have been identified in recent years, such as imaging histological features (6,7), tumor metabolic indexes (8-10), tumor markers (11), inflammation indexes (12), TMB (13), and circulating tumor DNA (ctDNA) (14). All of the aforementioned research, however, was done from the standpoint of a single indicator or type of indicator, which had limited predictive value. As of yet, no prediction model based on the combination of several indicators has been created. We included easily accessible metrics that are commonly tested in clinical practice in this study. After variable screening, we ultimately identified three metrics—tumor stage, TMB, and Ki-67—as independent factors influencing outcome. Using a machine learning approach, we built a prediction model for NDB (advancement within 6 months) of immunotherapy in driver gene-negative NSCLC patients. LightGBM was ultimately determined to be the best model following a number of screenings and comparisons. It had high discriminability in both the training and validation sets, with ROC-AUC values of 0.858 (95% CI: 0.783–0.933) and 0.852 (95% CI: 0.727–0.977), respectively. Additionally, calibration and DCA curves demonstrated the model’s high accuracy and clinical application value. Notably, the PR curve—a curve made up of Recall and Precision with an AUC between 0 and 1—is also used in this study as an evaluation device. More attention should be paid to the PR curve’s performance based on the traditional ROC curve because it focuses more on the ability to identify the minority class (i.e., positive examples), which is more applicable in the case of unbalanced data. The results show that the PR-AUC of LightGBM is also higher in all models, with the training and validation sets being 0.722 (95% CI 0.554-0.871) and 0.723 (95% CI 0.458-0.927), respectively.
Simultaneously, we developed a web-based online calculator to better support clinical decision-making. Clinical professionals can access the corresponding prediction conclusions by obtaining the information of the aforementioned indicators and entering it into the webpage. Following cohort studies have shown that the model works well for identifying individuals with poor immune responses, for whom early radiation therapy plus first-line immunotherapy plus chemotherapy can effectively prevent the progression of the cancer.
One of the key elements influencing tumor patients’ chances of survival is staging (15), which is also a crucial component of lung cancer diagnosis, treatment, and research. Based on three factors—tumor size, lymph nodes, and distant metastasis—TNM staging is currently a widely recognized tumor staging approach that offers a crucial foundation for therapeutic treatment guidance.
TMB is the density of non-synonymous mutations in a tumor cell’s coding region, expressed in mutations per million bases and calculated as the number of non-consensual mutation sites in the coding region divided by the total length of the coding region (16). Snyder’s research in 2014 discovered the first association between the quantity of tumor mutations and immunotherapy response (17). Later, in 2015, a study published in Science (18) established the link between TMB and immunotherapy effectiveness in NSCLC, revealing that patients with TMB above the median had longer PFS. Immune checkpoint molecules have the ability to suppress cytotoxic T cell function, which prevents them from attaching to immature antigens on the tumor surface and producing an immune response against the tumor (19). By binding to immune checkpoint molecules, PD-1, PD-L1, and cytotoxic T lymphocyte antigen 4 (CTLA-4) can assist relax the inhibition of cytotoxic T cells and restore anti-tumor activity. Tumor neoantigens are proteins or products of protein degradation that are presented to the surface of tumor cells by the major histocompatibility complex. These proteins are produced by tumor-specific somatic mutations. TMB can be utilized as a biomarker for the effectiveness of immunotherapy since a high TMB indicates that a tumor has more neoplastic antigens (20).
The MK167 gene encodes for the protein Ki-67, which is produced during the G1, S, G2, and M phases of the cell cycle but not during the G0 phase (stationary phase), and is directly linked to cellular value-added activity (21). Ki-67 is tightly associated with tumor value-added, metastasis, and prognosis. Numerous studies have shown the association between Ki-67 expression and the prognosis of patients with NSCLC. According to two meta-analyses, patients with higher Ki-67 had significantly shorter disease-free survival (DFS) and overall survival (OS), including 1,931 patients from 15 studies and 14,732 patients from 108 studies, respectively (22,23).
Numerous fundamental studies have confirmed the synergistic effect of radiotherapy and immunotherapy. The mechanism of this effect consists of four main components (24): radiotherapy’s enhancement of tumor cell antigen presentation; radiotherapy’s modulating effect on the tumor immune microenvironment (TIME); immunotherapy’s sensitizing effect on radiation therapy; and radiotherapy’s distant effects. Simultaneously, a number of extensive clinical trials have been carried out to assess the effectiveness of immunotherapy and radiotherapy in advanced and locally advanced NSCLC. These studies offer significant evidence in favor of the combined regimen. 709 patients with stage III NSCLC were included in the PACIFIC study (25), 473 of these patients received simultaneous radiation therapy in addition to dulvarizumab, and 236 received simultaneous radiotherapy in addition to a placebo. According to the findings, the dulvarizumab group’s median PFS was 16.8 (95% CI: 13.0–18.1) months, while the placebo group’s was 5.6 (95% CI: 4.6–7.8) months, the hazard ratio (HR) was 0.52 (95% CI: 0.42–0.65; P<0.001). As the first randomized controlled trial (RCT) evaluating combined radiation therapy and immunotherapy in advanced NSCLC, PEMBRO-RT demonstrated higher objective response rate (ORR), median PFS, and median OS in the combination therapy group vs. controls—though these differences did not reach statistical significance (26).
Immunotherapy and chemotherapy are sometimes hard to maintain in the early stages for patients with a poor immune response, which leads to tumor growth. Therefore, early radiotherapy can both control the growth of the tumor through local radiation for these individuals and promote the release of antigens from the tumor cells, improve antigen presentation, and up-regulate the expression of PD-L1. This may result in the transformation of “cold tumors” that primarily responded poorly to immunotherapy into “hot tumors” and increase the effectiveness of immunotherapy (27).
This study has the following limitations: (I) there is currently no external validation data for the model, therefore its extrapolation cannot be verified. In the follow-up study, data from other centers will be used as external validation sets. Their corresponding model evaluation metrics (ROC-AUC, PR-AUC, calibration curves, etc.) will be calculated, and the pertinent conclusions will be further adjusted after being compared with the outcomes of the current models. (II) There are no standard criteria for assessing the effectiveness of immunotherapy, and the studies that are currently available use the following outcome indicators: whether progression occurs at 4 cycles of treatment, whether progression occurs at 6 months of treatment, PFS time, and OS time. Immunotherapy is often administered to NSCLC patients every 21–28 days for 4–6 cycles, with an average length of roughly 4.4 months. Our study identified the outcome indicator as whether or not progression occurs within 6 months after treatment (NDB), taking into account the possibility of delayed response and pseudo-progression of immunotherapy, etc. However, the same data may yield different results with different outcome indicators, more large samples are needed in the future, and studies are required to determine the most scientific outcome indicators. (III) The predictive model’s variables were not yet complete; many indicators [such as PD-L1 expression, maximum standardized uptake value (SUVmax), lymphocyte subpopulations, C-reactive protein (CRP), lactate dehydrogenase (LDH), etc.] were eliminated because of missing values that exceeded 20%, and indicators like ctDNA have been left out because the test is not yet generally accessible. Future prospective studies will be required to include more thorough indicators in future analyses.

Conclusions
In this study, we constructed a prediction model for the effectiveness of immunotherapy in NSCLC. We found that the combination of three indicators—tumor stage, TMB and Ki-67—can more accurately predict the population with NDB. For the patients mentioned above, we suggest additional radiotherapy in the first-line treatment stage to control tumor progression and improve the state of immune suppression to obtain better therapeutic effects.

Supplementary
The article’s supplementary files as