2.1 Ethics approval and consent to participate

This study was approved by the Ethics Committee of Jiangsu Cancer Hospital (No. 2020129), Xuzhou Central Hospital (XYFM2020002), and Taixing People's Hospital (KD2020069). All participants provided written informed consent before phlebotomy and surgery. The study was carried out in accordance with the Declaration of Helsinki.

2.2 Study design and participants

The flowchart of the study is shown in Figure 1. A multicenter, retrospective cohort was enrolled using plasma samples collected from patients with malignant and benign pulmonary nodules from three thoracic surgery departments in China, between May 2020 to August 2021. Samples from the discovery cohort (n = 100) and internal validation cohort (n = 83) were collected from The Affiliated Cancer Hospital of Nanjing Medical University, while specimens collected between March 2021 and January 2022 from Xuzhou Central Hospital and Taixing People's Hospital served as the external validation cohort (n = 66). It should be emphasized that our research did not interfere with the treatment of patients. The characteristics of the 249 participants are shown in Table S1.

Inclusion criteria included 18–85 years old, single pulmonary nodule, detected by LDCT with maximum diameter ≤2 cm, presented with solid, GGNs, pure GGNs, receiving surgery, and postoperative pathological diagnosis. Exclusion criteria included pregnant or lactating women, other cancer history, multiple pulmonary nodules, metastasis nodules, and receiving neo-adjuvant chemotherapy or radiotherapy, or receipt of transfusion within 30 days prior to enrollment. In addition, clinical information including LDCT imaging reports and pathology reports were collected. Blood samples were centrifuged within 1 h of collection for 10 min at 3500 rpm at 4°C to harvest plasma samples before storing at −80°C for analysis.

2.3 High-throughput sequencing

According to inclusion criteria, five temporarily untreated LUAD patients were selected and their tumor tissues and adjacent tissues were collected for tRF and tiRNA sequencing. An Agilent Bioanalyzer 2100 (Agilent Technologies) was used to quantify the sequencing libraries. The sequencing analysis was carried out using the Illumina NextSeq 500 platform according to the manufacturer's protocol.

2.4 RNA extraction

Total RNAs of blood samples and tissues were extracted using TRIzol LS Reagent (Invitrogen) according to the manufacturer's protocol, and 25 fmol of Caenorhabditis elegans cel-miR-39-3p RNA was added to each sample as a spike-in control.^{14, 15} Total RNAs were extracted from 200 μL of each plasma sample and eluted with 10 μL RNase-free water. RNA concentration and quality were measured on an Agilent Bioanalyzer 2100. All total RNAs were stored at −80°C within 3 days before analysis.

2.5 Reverse transcription-qPCR

Total RNA (100 ng) was used for reverse transcription using Bulge-Loop MicroRNA qRT-PCR Primer Sets (one RT primer and a pair of qPCR primers for each set), specially designed for tsRNAs (RiboBio). According to the protocol, the reverse transcription reaction assays were carried out at 42°C for 60 min followed by 70°C for 10 min. Then quantitative real-time PCR was carried out to analyze tsRNA expression on a Q6 RT-qPCR machine. In brief, the RT-qPCR assay was run in 96-well and 384-well plates at 95°C for 1 min, followed by 40 cycles at 95°C for 10 s, 60°C for 30 s, and finally 95°C for 15 s, 60°C for 60s, 95°C for 15 s. The expression of tsRNA was normalized to cel-miR-39-3p. Relative expression levels were analyzed by the −ΔCt method.

2.6 Intraoperative FS diagnosis

Two pathologists with different levels of clinical experience (senior pathologist has worked as a cancer subspecialty pathologist for over 20 years, junior pathologist has worked as a cancer subspecialty pathologist for 2 years) were instructed to independently diagnose IA by intraoperative FS diagnosis during operations. Neither pathologist was involved in previous evaluation of patients.

2.7 Comparison with predictive models

The Mayo Clinic model for malignancy in pulmonary nodules calculated the malignancy probability as follows¹⁸: Probability of malignancy = e^x/(1 + e^x), where x = −6.8272 + (0.0391 × age) + (0.7917 × smoking) + (1.3388 × cancer) + (0.1274 × nodule diameter) + (1.0407 × speculation) + (0.7838 × upper lobe).

The specific Brock University cancer prediction equation is as follows¹⁹: Cancer probability = 100 × (e^(Log_odds) / (1 + e^(Log_odds))), where Log_odds = (0.0287 × (age − 62)) + sex (female = 0.6011, male = 0) + family_history_lung_Ca (0.2961) + emphysema (0.2953) − (5.3854 × ((nodule_size/10) − 0.5–1.58113883)) + nodule_type + nodule_upper_lung (0.6581) − (0.0824 × (nodule_count − 4)) + speculation (0.7729) − 6.7892.

Our previous study constructed the CT-based Lung-DL model²⁰ based on CT information to predict MIA and IA. To further evaluate the diagnostic performance of our nomograms, the Lung-DL model was also used for comparison.

2.8 Statistical analysis and in silico bioinformatics analysis

Statistical analysis was performed using R (version 4.1.2) and GraphPad Prism 8 statistical software. Most graphs contain graphs for each data point and show the mean ± SD. To test the significance, a t-test was carried out, and the p value was denoted by an asterisk. Point-biserial correlation is used in order to reflect the correlation between continuous variables and binary categorical variables. Analysis of differential expression of tsRNA between five tumor tissues and five adjacent tissues was undertaken with the DESeq2 R package. The circulating tsRNA score was constructed according to the randomForest R package.²¹ The ROC analyses were carried out using the pROC R package. Construction of the malignant nodules probability nomogram and invasive adenocarcinoma probability nomogram, as well as calibration curves for the above, was undertaken using the rms R package.²²

3.1 Development and validation of circulating tsRNA score for malignant nodule diagnosis

The tRF and tiRNA sequencing was carried out in five tumor samples and five adjacent tissues from five temporarily untreated early-stage LUAD patients, who were strictly selected according to inclusion criteria, to explore tsRNAs specifically expressed in early-stage LUAD patients. Through differential expression analysis of the tsRNA expression in tumor tissues and adjacent tissues, a total of nine dysregulated tsRNAs were selected based on fold changes and p values (p value < 0.001 and fold-change > 10; Figure 2A). To exam the diagnostic performance of malignant nodules, the expression levels of nine dysregulated tsRNAs were validated using RT-qPCR in the discovery cohort, including 86 cases of malignant nodules and 14 cases of benign nodules confirmed in the subsequent diagnosis, assessed by two senior pathologists independently. Each RT-qPCR product of the five tsRNAs was confirmed by cloning sequences that contained the full-length sequence (Figure 2B). Among the nine tsRNA candidates, tRF-Ser-TGA-003, tRF-Val-CAC-005, tRF-Ala-AGC-060, tRF-Val-CAC-024, and tiRNA-Gln-TTG-001 were found significantly upregulated (p < 0.05, Student's t-test) in the plasma of the malignant nodules group (Figure 2C).

In order to accelerate the capability of circulating tsRNAs to detect malignant pulmonary nodules, we focused on the five upregulated tsRNAs in malignant pulmonary nodules and established a five-tsRNA diagnostic model for detection of malignant nodules. Combining the five upregulated tsRNAs, the circulating tsRNA score was calculated using the following formula: Circulating tsRNA score = 10.631148 × relative expression of Ser-TGA-003 + 2.310571 × relative expression of Val-CAC-005 + 6.588001 × relative expression of Ala-AGC-060 + 6.588001 × relative expression of Val-CAC-024 + 2.242256 × relative expression of Gln-TTG-001 using the randomForest R package (Figure 2D). The results indicated that circulating tsRNA scores were significantly different between patients with benign and malignant pulmonary nodules (Figure 3A) and was associated with malignant nodules (Tables 1 and 2; OR 17.00; 95% CI, 7.348–36.300). To further evaluate the accuracy of detection of the circulating tsRNA score, ROC curves were generated and AUCs were calculated. The AUC for the circulating tsRNA score was 0.890, with a sensitivity of 79.1% and a specificity of 92.9%, which was better than any identified tsRNA. The AUCs were 0.873, 0.767, 0.887, 0.682, and 0.686 for tRF-Ser-TGA-003, tRF-Val-CAC-005, tRF-Ala-AGC-060, tRF-Val-CAC-024, and tiRNA-Gln-TTG-001, respectively (Figure 3B).

To validate the value of the circulating tsRNA score for malignant lung nodule detection, we prospectively collected clinical data and peripheral blood samples from 83 patients with pulmonary nodules in The Affiliated Cancer Hospital of Nanjing Medical University as an internal validation cohort. The enrollment criteria are the same as the discovery cohort. The specific expression level of the five tsRNAs are shown in Figure S1. Consistently, the circulating tsRNA score achieved a robust performance in the internal validation cohort (Figure 3C,D, Tables 3 and 4; AUC = 0.801; 95% CI, 1.588–15.177; sensitivity 69.2%, specificity 83.3%). Furthermore, an external validation cohort was collected from Xuzhou Central Hospital and Taixing People's Hospital, containing clinical data and peripheral blood samples of 66 patients with pulmonary nodules. The enrollment criteria were the same as the discovery cohort. As expected, the external validation cohort also revealed robust detection ability (Figure 3E,F, Tables 5 and 6; AUC = 0.765; 95% CI, 2.416–39.576; sensitivity 83.0%, specificity 76.9%).

Additionally, the AUCs for the circulating tsRNA score were both superior to any single tsRNA in the internal validation and external validation cohorts (Figure 3D,F). These results indicated that the circulating tsRNA score had potential translational value in distinguishing malignant nodules from benign nodules by liquid biopsy.

Chest CT is the most commonly used noninvasive method to determine the type of pulmonary nodules in preoperative examination of pulmonary nodules with a maximum diameter <2 cm.^{23, 24} In the assessment of risk factors for malignant pulmonary nodules in the discovery cohort, internal validation cohort, and external validation cohort, we found that GGO components in pulmonary nodules, especially mixtures (with both solid and GGO components), were associated with the risk of malignant pulmonary nodules. We constructed a Malignant Nodules Probability nomogram by integrating tsRNA and CT feature vectors to improve the accuracy of malignant lung nodule detection (Figure 4A). The calibration curve showed that malignant nodules were predicted with good accuracy in the three cohorts (Figure 4B). The results of the ROC analyses (Figure 4C) indicated that our nomogram could accurately predict malignant nodules in the discovery cohort (AUC = 0.953, sensitivity 88.4%, specificity 100.0%), internal cohort (AUC = 0.930, sensitivity 100.0%, specificity 73.8%), and external cohort (AUC = 0.943, sensitivity 100.0%, specificity 86.8%).

When compared with the Mayo Clinic Model and Brock University cancer prediction equation, our Malignant Nodules Probability nomogram outperformed both clinical models in the three cohorts. The AUCs of the Mayo Clinic model were 0.637, 0.634, and 0.813 and the AUCs of the Brock University cancer prediction equation were 0.660, 0.673, and 0.745, respectively (Figure 4C). The Malignant Nodules Probability nomogram enhances the detection accuracy of the circulating tsRNA score for malignant lung nodules by integrating CT feature vectors.

3.2 Prediction of invasive adenocarcinoma

According to the International Association for the Study of Lung Cancer/American Thoracic Society/European Respiratory Society IASLC/ATS/ERS classification, LUAD was classified as AIS, MIA, or IA.⁵ Sublobar resection is currently considered to be sufficient for AIS and MIA, while lobectomy is the preferred therapy for IA. However, surgeons can only determine the subsequent operation method and the scope of resection according to the solid component in CT images and intraoperative FS diagnosis, which is widely used to distinguish MIA from IA during surgery and considered as the gold standard in clinical practice. A previous study showed that the diagnostic accuracy of the intraoperative FS for tumors ≤1 cm in diameter was 79.6%,⁸ while this accuracy rate was even lower for junior pathologists or medical centers with less experience. Thus, it is necessary to develop a new, noninvasive method that provides a reference for the invasiveness degree before surgery, so as to reduce the occurrence of inappropriate surgical plan choices and optimize the distribution of medical resources. Next, we evaluated the ability of our circulating tsRNA score to identify tumor categories in order to help the diagnosis of NSCLC. The specific expression of each tsRNA is shown in Figure 5A. Interestingly, most selected tsRNAs were highly expressed in IA patients' plasma. We then tested whether the panel was able to identify patients with IA. As expected, the circulating tsRNA score was significantly higher in the IA group compared to the MIA group in the discovery cohort (Figure 5B) and was associated with IA (Table 7; OR 4.253; 95% CI, 1.502–17.864). In addition, point-biserial correlation was carried out and indicated that high circulating tsRNA score was highly correlated with IA (p = 0.001542; Figure 5C). To further confirm the performance of the circulating tsRNA score in IA prediction, the internal validation cohort and external validation cohort were also involved. Consistently, the circulating tsRNA score show good potential in the prediction of IA (Figure 5E,F,H,I, Tables 8 and 9). The AUC of preoperative circulating tsRNA score for detecting invasiveness of NSCLC were 0.6816 (discovery cohort), 0.6638 (internal validation cohort), and 0.7186 (external validation cohort), better than the performance of any individual tsRNA (Figure 5D,G,J).

However, the circulating tsRNA score is insufficient to detect the invasiveness of NSCLC. Compared with the postoperative pathological examination of the majority opinion of pathologists, the AUCs of preoperative circulating tsRNA scores for detecting invasiveness of NSCLC were 0.682 (discovery cohort), 0.664 (internal validation cohort), and 0.719 (external validation cohort). Additionally, the AUCs of the diagnostic accuracy for pathologists were 0.860, 0.862, and 0.773 in the senior pathologist and 0.534, 0.569, and 0.621 in the junior pathologist, respectively, using intraoperative FS diagnosis (Figure 6A–C). These results showed that the diagnostic accuracy of the circulating tsRNA score was slightly higher than that of junior pathologists but significantly lower than that of senior pathologists.

To improve the accuracy of IA prediction, CT image data of the participants was involved. As maximum diameter and GGO composition were correlated with NSCLC invasiveness (Tables 7–9), a nomogram was established to explore more precise prediction for invasive adenocarcinoma probability controlling for the GGO component and maximum diameter (Figure 6D). The calibration curve showed that the nomogram achieved a good performance in predicting invasive adenocarcinoma (Figure 6E). Once again, ROC analysis was carried out. The five-tsRNA signature robustly distinguished IA patients from MIA patients in the discovery cohort (Figure 6F; AUC = 0.850, sensitivity 86.0%, specificity 81.4%), internal validation cohort (Figure 6G; AUC = 0.784, sensitivity 78.8%, specificity 78.1%), and external validation cohort (Figure 6H; AUC = 0.837, sensitivity 85.7%, specificity 84.0%). To further evaluate the diagnostic performance of the Invasive Adenocarcinoma Nodules Probability nomogram, the CT-based Lung-DL model, which was constructed based on CT information in our previous study, was used for comparison. The AUCs of diagnostic accuracy for the Lung-DL model were 0.840 (sensitivity 87.8%, specificity 77.1%, discovery cohort), 0.786 (sensitivity 48.9%, specificity 96.9%, internal validation cohort), and 0.759 (sensitivity 50.0%, specificity 96.0%, external validation cohort). Through comparison, the diagnostic accuracy of the Malignant Nodules Probability nomogram could achieve a similar level to the Lung-DL model. The specificity of the Malignant Nodules Probability nomogram was slightly lower than the Lung-DL model, but achieved better sensitivity.

In general, our five-tsRNA diagnostic model showed prospective capacity in predicting invasive components in malignant lesions, which could aid surgical decision-making and provide hints in FS diagnosis.