The training cohort (n = 125) were retrospectively reviewed and collected from the stage-III NSCLC patients with plain CT scans (Jul 2012 to May 2016), including 22 CR patients and 103 non-CR patients (PR, n = 62, SD, n = 22; PD, n = 19). Follow-up information was collected from the hospital’s information system.

Following the same criteria for the training cohort, an independent testing cohort (R422) cohort including 104 stage III NSCLC patients receiving CCR was downloaded from the Cancer Imaging Archive (TCIA, https://www.cancerimagingarchive.net/, 2020). NSCLC-Radiogenomic (RG211) cohort, including 113 patients regardless of treatment strategy, had pre-treatment CT scans and RNA sequencing profiles (generated by an Illumina HiSeq 2500 system) before treatment, was used for biological analyses. The baseline characteristics of patients used in this study are listed in Table 1.

The workflow of this study is shown in Fig. 2. Firstly, in the training cohort, 371 candidate response-related features were identified as their feature values were significantly different between CR and non-CR patients (Wilcoxon rank sum test, P < 0.05). Then, we constructed FP matrix, and identified 139 response-related FPs, whose rank pattern (eg. Ra > Rb) were potentially significantly associated with response status (Fisher exact test, P < 0.05) and OS of patients receiving CCR (univariate Cox model, P < 0.05). Finally, the GA method outperformed the other three methods (FA, LASSO and BA), with the highest average C-index value of 0.74 in the inter-validation datasets after 10 rounds of sampling (Fig. 3A). Therefore, GA method was selected to develop a radiomic signature based on all the 125 samples in the training cohort. The predictive signature for CCR benefit included 15 FPs (15-RiFPS, Table 2), whose cut-off was set as 7.5 for classification; that is a sample was predicted as a responder or non-responder if more than 7.5 FPs in the signature voted for CCR responder or non-responder.

The flowchart of this study

Development of the predictive qualitative radiomic signature in the training cohort. A The average C-index of four methods in training cohort derived. B The Kaplan–Meier curves of OS for patients received CCR (n = 125). C Calibration curves of the qualitative radiomic signature (15-RiFPS) for 1-year OS, 2-year OS and 3-year OS. D The time-dependent receiver operating characteristic (ROC) curve of 15-RiFPS. E Multivariate COX analysis of 15-RiFPS, T stage, N stage, overall stage, smoking status, degree of differentiation and lymph vascular invasion. F The Kaplan–Meier curves of OS for SCC patients received CCR (n = 19). G The Kaplan–Meier curves of OS for ADC patients received CCR (n = 106)

In the training cohort, according to the cut-off, 60 LA-NSCLC patients were stratified into a CCR-response group, and 65 patients were stratified into a CCR-non-response group with significantly different OS when they received CCR treatment (log-rank P = 0.0009, HR = 13.79, 95% CIs: 1.83–104.10, C-index = 0.74, Fig. 3B). The 3-year survival rate for patients predicted to have a CCR-response was 88.89%, which was obviously higher than the rate (68.36%) for patients predicted to have a CCR-non-response. Calibration curves of 15-RiFPS for 3-year survival, were displayed in Fig. 3C. Figure 3D shown the time-dependent ROC curve of 15-RiFPS under 3 years (AUC = 0.65). We performed multivariate COX analysis for 15-RiFPS and clinical prognostic factors (P < 0.05, Univariate Cox analysis) together and found that 15-RiFPS was an independent factor after adjusting T stage, N stage, overall stage, Lymph vascular invasion, smoking status and degree of differentiation (Table 3, Fig. 3E). Additionally, 15-RiFPS could distinguish responders and non-responders with significantly different OS in squamous cell carcinoma (log-rank P = 0.0072, Fig. 3F) and adenocarcinoma patients receiving CCR (log-rank P = 0.0598, Fig. 3G), respectively.

One public independent cohort (R422) with 104 stage III patients receiving CCR was used for validation. In R422 cohort, 77 stage III patients who were classified into a response group by 15-RiFPS had significantly longer OS than the 27 patients classified into the non-response group (log-rank P = 0.004, HR = 2.40, 95% CIs: 1.30–4.40, Fig. 4A). Calibration curves of 15-RiFPS for 3-year survival, in R422 cohort, were displayed in Fig. 4B. The AUC values of 15-RiFPS at 3-year were 0.67 in R422 cohort (Fig. 4C). The risk scores of patients and their OS in R422 cohort are displayed in Fig. 4D. 15-RiFPS also exhibited favorable performance in predicting OS of ADC (log-rank P = 0.014, HR = 4.55, 95% CIs: 1.22–16.93, Fig. 4E) and SCC patients receiving CCR in R422 cohort (log-rank P = 0.0600, HR = 1.98, 95% CIs: 0.96–4.10, Fig. 4F), respectively. Univariate COX analysis showed that the clinical factors recorded in this cohort was not significantly associated with patients’ OS (all P > 0.05, Table 4).

Validation of 15-RiFPS in the testing cohort. A The Kaplan–Meier curves of OS for patients with plain CT scans (n = 104). B Calibration curves of the 15-RiFPS for 1-year OS, 2-year OS and 3-year OS. C The time-dependent receiver operating characteristic (ROC) curve of 15-RiFPS in the validation cohort. D Risk score distribution and survival status of patients. E The Kaplan–Meier curves of OS for ADC patients with plain CT scans (n = 23). F The Kaplan–Meier curves of OS for SCC patients with plain CT scans (n = 81)

First, we applied 15-RiFPS to the RG211 cohort with paired radiomic and transcriptomic data to estimate the risk scores of patients. Using Spearman rank correlation analysis, we identified 747 genes significantly correlated with risk scores and found that they were enriched in six GO terms (Hypergeometric distribution model, FDR < 0.05, Fig. 5A), including “glutathione metabolic process”, “cellular oxidant detoxification”, “cellular detoxification”, “cellular response to toxic substance”, “xenobiotic metabolic process” and “detoxification”. Details of functional enrichment analysis of genes related with 15-RiFPS are presented in Supplementary Table 1. Next, we constructed a protein–protein interaction (PPI) network of the signature-related genes in the enriched terms, and demonstrated that MGST1, GPX4, GSTA1 which have been reported to be related with CCR (Zhang et al. 2023; Gisbergen et al. 2016), were significantly associated with the scores of our signature, supporting our signature reflecting the intra-tumoral molecular characteristics.

Functional enrichment analysis of genes related with 15-RiFPS. A Biological analyses for 15-RiFPS based on GO pathway database. The transcriptomic analyses provided biological pathways enriched in six GO terms underlying the 15-RiFPS. B A protein–protein interaction netwok of the signature-related genes in the enriched terms