Each patient provided detailed informed consent to the investigator. PD-L1 expression and EGFR/ALK mutation status were not necessary conditions for enrollment. The key eligibility criteria were as follows: (i) patients were at least 18 years old and willing to provide routine peripheral blood (2 mL) for immune cell analysis; (ii) Eastern Cooperative Oncology Group performance status was 0–1; (iii) a tissue biopsy of the tumor was confirmed to be lung adenocarcinoma (LUAD) or lung squamous cell carcinoma (LUSC) before any treatment; and (iv) patients with resectable stage Ib-IIIa NSCLC were examined using whole-body computed tomography (CT) or positron emission tomography-CT. The exclusion criteria were: (i) patients who could not tolerate treatment, such as those allergic to albumin-bound paclitaxel; (ii) patients undergoing CAPD but whose NSCLC was progressing rapidly or showed organ metastasis and were not suitable for resection treatment. Between September 2019 and June 2021, 211 patients who were receiving neoadjuvant treatment before surgery and met the criteria were recruited at the Cancer Hospital of the University of Chinese Academy of Sciences (CHUCAS) ( Supplementary Figure 1 ). The patients were randomly divided in a 3:2 ratio into a discovery set (n = 127) and validation set (n = 84). The institutional review board of the Second Affiliated Hospital of Guizhou Medical University and CHUCAS approved our clinical research design. We have been conducted in accordance with the World Medical Association’s Declaration of Helsinki.

A flowchart of the study design is shown in Figure 1 . Before treatment, whole blood samples were collected from patients and rapidly tested by multicolor flow cytometry before initial treatment. The detailed results of peripheral immune cells and clinical characteristics were recorded. Two hundred and eleven patients with NSCLC received 2–4 cycles of CAPD. According to the MPR status, the doctors then performed radical surgery of lung cancer. The pathological response of tumor tissues was estimated by a senior pathologist. We then analyzed the association between clinical factors and MPR using univariate analysis of both cohorts (discovery and validation sets). The immune cells in peripheral blood were chosen by recursive feature elimination (RFE) and least absolute shrinkage and selection operator (LASSO) methods. After integrating these two selection methods, the final immune cell subtypes were confirmed. An SVM model based on liquid immune profiling (LIP-SVM) was developed and validated in the two cohorts. Multivariate analysis for MPR was performed for patients with NSCLC using logistic regression. The integration of LIP-SVM, clinical factors, or radiological response was evaluated in terms of predictive accuracy for MPR. Contrast-enhanced CT images were examined to clearly identify the primary tumor. The radiological response, including complete response (CR), partial response (PR), stable disease (SD), and progressive disease (PD), was estimated before radical surgery by a senior thoracic radiologist using Response Evaluation Criteria in Solid Tumors (RECIST) version 1.1.

MPR is defined as the reduction of viable tumors to clinically defined significant margins, depending on the particular histological type of lung cancer and the specific treatment type. All histological types of lung cancer had an MPR with a histological definition of less than or equal to 10% of the viable tumor. The MPR was calculated by dividing the size of the viable tumor by the size of the tumor bed. Here, this was used to establish the threshold of the number of clinical trials. The pathology report recorded the total number of masses in the tumor bed, including some uninvolved lungs, even if these masses were not entirely composed of the tumor bed. The MPR can also be classified as a primary pulmonary tumor, where little or no viable metastatic carcinomas were found in the lymph nodes (ypT0, N1, 2, 3).

Patients with LUAD preoperatively received a folic acid, vitamin B12, and glucocorticoid pretreatment that is prescribed according to the local guidelines for pemetrexed. All patients received intravenous injections of cisplatin (75 mg/m², d1) or carboplatin (under the concentration-time curve, 5 mg/mL/min, d1) and pemetrexed (500 mg/m², d1) for 2–4 cycles. Patients with LUSC preoperatively received intravenous injections of cisplatin (75 mg/m², d1) or carboplatin (under the concentration-time curve, 5 mg/mL/min, d1) and nab-paclitaxel (135 mg/m², d1, d8) for 2–4 cycles. The anti-PD-1 regimen included camrelizumab (3 mg/kg, Q2W), pemetrexed (500 mg/m², Q3W), nivolumab (3 mg/kg, Q2W), toripalimab (3 mg/kg, Q2W), tislelizumab (200 mg/m², Q3W), or sintilimab (200 mg, Q3W), whereas the anti-PD-L1 regimen included durvalumab (10 mg/kg, Q2W); the patients received both injections following each chemotherapy cycle. Chemotherapy and anti-PD-1/PD-L1 regimens were administered every 3 weeks to patients with LUAD and LUSC.

We evaluated four types of circulating immune cells: B, T, natural killer (NK), and natural killer T (NKT) cells. B and T cells were defined by CD19 expression (CD19⁺ B cells) and CD3 expression (CD3⁺ T cells), respectively. The presence of CD8 and CD4 was used to identify T-lymphocyte subsets (CD3⁺CD8⁺ T cells and CD3⁺CD4⁺ T cells). Memory (CD4⁺CD45RO⁺) T cells and CD4⁺ naïve (CD4⁺CD45RA⁺) T cells were identified by CD45RA and CD45RO expression. A combination of CD56 and CD3 was used to identify NKT (CD3⁺CD56⁺) and NK (CD3^-CD56⁺) lymphocyte subsets. Activated CD8⁺ T cells (CD8⁺CD38⁺ T cells) were recognized by CD38 expression. BD Biosciences (San Jose, CA, USA) provided the following antibodies in Supplementary Table 1 : CD8-FITC (#555366), CD4-FITC (#550628), CD3-FITC (#555332), CD56-FITC (#55664), CD19-FITC (#555412), CD45RO-APC (#559865), CD38-PE (#555460), CD45RA-PE (#555489), and FITC/APC/PE controls (#55749; #555748; #5555776). For lymphocyte staining, 4 mL of peripheral blood was collected into a blood tube with ethylenediaminetetraacetic acid and an anticoagulant. Next, 20 μL of CD3-FITC/CD56-PE, CD19-FITC, CD4-FITC/CD45RO-APC/CD45RA-PE, CD8-FITC/CD38-PE, and FITC/PE/APC isotype controls was separately added to five flow cytometry tubes and 100 μL of every blood sample was added to every test tube. The tubes were sufficiently mixed with corresponding antibodies in the dark and incubated for 30 min at room temperature (20°C). A hemolytic agent (up to 2 mL; #70-LSB3; BD Biosciences) was then added to every tube. The supernatant was then removed by centrifugation (6 min), washed twice with phosphate-buffered saline (#SH300256; Hyclone, Logan, UT, USA), and resuspended in paraformaldehyde. Flow cytometry (FACSVia; BD Biosciences) was used to examine the cells. More than 2,000 cells were detected at the lymphocyte gate in every sample. CellQuest Pro software (BD Biosciences) was used to analyze the percentages of positively labeled lymphocytes. The staining procedure was completed and analyzed within 24 h after blood collection.

Two feature selection methods (RFE and LASSO) were used in this study. RFE recursively reduces the size of the examined feature set to select features. The prediction model based on the original features is trained and a weight is assigned to each feature. Features with a minimum absolute weight are recursively removed until the desired number is reached. Random forest function and 5-fold cross-validation (CV) sampling were used for RFE. In addition, we used LASSO to select the most important immune cells from the discovery set. A log partial likelihood subject based on the LASSO method is minimized to add the absolute values of the parameters. Here, the standardized constraint parameter was set to -1.434. RFE and LASSO were performed using the “caret” package in R version 3.5.1.

The SVM is a classical model of machine learning with important value in tumor classification, prognosis, and treatment response predictions (23). Radial basis function, the most popular kernel function of SVM for nonlinear classification, can significantly improve the classification ability of the SVM by mapping the original input space to the feature space. The original nonlinear input space is transformed into the linear separability space and classified linearly in the feature space. The equation we used was as follows:

where γ is greater than 0 and the parameters need to be adjusted. The tuning parameters were set as sigma = 0.035, C = 100, and cross = 10. The “kernlab” library was implemented using R software.

R software and GraphPad Prism were used to perform statistical analysis. A correlation heatmap was generated using the “pheatmap” package to depict the relationships between immune cells in the discovery and validation sets. Correlation scatter plots were used to indicate associations between immune cells. The scatter dot and box plots were used to represent the median and 95% confidence intervals (CI). The differences between groups were analyzed using the Mann–Whitney U test. The receiver operating characteristic curves (ROC) were plotted using the “pROC” package. The frequencies of two groups were compared using the chi-square test. Univariate and multivariate logistic analyses for MPR were performed in the two sets. Statistical significance was defined as P-value < 0.05.

The baseline characteristics of 211 patients from two independent cohorts are presented in Table 1 . Most patients (92.42%) were male, old (> 60 years; 60.67%), smokers (78.67%), and with squamous cancer (82.46%). Stage IIIa was assigned to 68.72% of patients, 89.57% received anti-PD-1 immunotherapy, 70.61% underwent two treatment cycles, and 45.50% of cases experienced radiological PR ( Table 1 ). The pathological response was MPR or no-MPR in the discovery (46.46% or 36.90%) and validation (53.54% or 63.10%) sets, respectively. No variables significantly differed between the two cohorts (P > 0.05).

By analyzing the association between clinical data and MPR in the discovery and validation sets, we found that old age (> 60 y) was significantly correlated with MPR (P = 0.023) ( Supplementary Figure 2A ), patients with squamous cancer presented with a higher MPR than patients with adenocarcinoma (P = 0.025) ( Supplementary Figure 2B ), and patients who underwent an anti-PD-L1 regimen showed a higher MPR than those who underwent an anti-PD-1 regimen (P = 0.023) ( Supplementary Figure 2C ). Moreover, patients with radiological CR and PR showed a higher MPR than patients with radiological SD or PD (both P < 0.001) ( Supplementary Figure 2D ). The detailed results of the two cohorts are presented in Supplementary Table 2 .

The results of the discovery set were similar with those of the validation set ( Figures 2A, B ). The relative abundance of CD4⁺CD45RA^- and CD3⁺CD4⁺ T cells showed a positive correlation (r = 0.621, P < 0.001; r = 0.534, P < 0.001), whereas that of CD3^-CD19⁺ B cells and CD3^-CD56⁺ NK cells showed a negative correlation (r = -0.373, P < 0.001; r = -0.390, P < 0.001) ( Figures 2C, D ).

To find suitable predictors for MPR, the immune cells before neoadjuvant therapy were identified, and then RFE and LASSO were used to perform feature selection. Based on 5-fold CV analysis, root mean square error (RMSE) showed that a combination of six variables had the smallest errors, and thus these immune cells were selected ( Figure 3A ). Then, LASSO coefficient analysis of 12 features of immune cells was performed, after which eight coefficients were chosen based on a 5-fold CV analysis of minimum criteria ( Figure 3B ). To find robust features, we selected five immune cell types with overlapping features for the SVM model ( Figure 3C ). CD3⁺CD56⁺ NKT cells were found significantly associated with MPR, and can thus be a predictor for MPR (P < 0.001, Figure 3D ). In the discovery set, the percentage of CD3^-CD19⁺ B cells was higher in the no-MPR group (P = 0.011, Figure 3D ), whereas that of CD3^-CD56⁺ NK cells was higher in the MPR group (P = 0.032). Moreover, patients in the MPR group had a higher percentage of CD4⁺CD45RA^- T cells than those in the no-MPR group (P = 0.017). The percentage of CD4⁺CD45RA⁺ T cells was higher in the no-MPR group than in the MPR group.

To develop the SVM model, fine-tuning was performed during the training process. LIP-SVM was calculated as the output score of machine learning in the discovery and validation sets. We found that patients in the MPR group showed significantly higher LIP-SVM scores than patients in the no-MPR group in both cohorts (both P < 0.001; Figures 4A, B ). To compare the accuracy of different models or predictors in our study, we first used three meaningful clinical features, including age (≤ 60 vs. > 60 years), histology (squamous vs. adenocarcinoma), and immunotherapy (anti-PD-1 vs. anti-PD-L1), to build a clinical model using logistic regression [area under the curve (AUC) = 0.663, 95% CI: 0.568–0.756, P = 0.001; AUC = 0.680, 95% CI: 0.565–0.796, P = 0.005, respectively; Figures 4C, D ]. Moreover, the radiological response (CR, PR, SD, and PD) was also a strong indicator for predicting MPR of the two sets (AUC = 0.840, 95% CI: 0.772–0.908, P < 0.001; AUC = 0.820, 95% CI: 0.727–0.913, P < 0.001, respectively; Figures 4C, D ). Comparing the above two predictors, the LIP-SVM signature showed the higher accuracy in the discovery and validation sets (AUC = 0.886, 95% CI: 0.823–0.949, P < 0.001; AUC = 0.874, 95% CI: 0.791–0.958, P < 0.001, respectively; Figures 4C, D ). To better understand the utility of LIP-SVM, pathological images from four patients in validation cohorts are presented. After 2–4 cycles of neoadjuvant CAPD before resection of lung cancer, patients 1 and 2 with MPR had lower LIP-SVM scores than patients 3 and 4 with no-MPR (0.379 and -0.049 vs. 0.936 and 1.049; Figure 4E ). The model is now available for free online testing (https://pengjie.shinyapps.io/lipsvm/).

Based on the multivariate logistic regression analysis in the discovery set, age, radiological response, and LIP-SVM signature were independent risk factors for MPR [P < 0.016, P < 0.001, and P < 0.001; OR = 0.169 (0.034–0.649), 7.318 (3.147–20.588), and 29.788 (1.572–56.574), respectively; Table 2 ]. Age, radiological response, and LIP-SVM signature were also independent risk factors for MPR in the validation set [P = 0.046, P = 0.002, and P < 0.001; OR = 0.162 (0.019–0.990), 15.352 (3.487–27.805), and 28.186 (1.096–58.460), respectively]. To further improve the predictive accuracy, we integrated a combination of clinical factors or radiological response for multivariate analysis; we found that the LIP-SVMRC (LIP-SVM signature + radiological response + clinical model) model had the highest accuracy in the discovery and validation sets (AUC = 0.951, 95% CI: 0.916–0.986, P < 0.001; AUC = 0.943, 95% CI: 0.912–0.993, P < 0.001, respectively; Figures 4F, G ). Two models (LIP-SVM signature + radiological response and LIP-SVM signature + clinical model) also exhibited high predictive accuracy for MPR in the discovery (AUC = 0.935, 95% CI: 0.895–0.975, P < 0.001; AUC = 0.907, 95% CI: 0.855–0.959, P < 0.001, respectively) and validation (AUC = 0.931, 95% CI: 0.899–0.994, P < 0.001; AUC = 0.905, 95% CI: 0.837–0.934, P < 0.001, respectively; Figures 4F, G ) sets.