With the intent of building a miRNA-based biomarker classifier for SCLC diagnosis ( Figure 1 , see method section), we performed RNA-seq on a discovery cohort ( Supplementary Table 1 ) of 38 RNA plasma samples. We compared control (CTR), Adenocarcinoma (ADENO), Squamous Cell Carcinoma (SCC), and SCLC groups and profiled the expression of circulating cell-free miRNAs (constituted by both cell-free and EVs included miRNAs). The sequencing was focused on small RNA profiling and revealed deregulation of several miRNAs in SCLC patients compared to the other histological groups. We specifically found that plasma miR-375-3p was significantly upregulated in the SCLC group compared to the other groups ( Supplementary Table 2 and Supplementary File 1 ).

To improve the distinction between SCLC and the other histological types, a second analysis was conducted on the RNA-seq data to identify additional deregulated miRNAs. For this purpose, we chose those miRNAs with low P-value (<0.05) in at least one comparison (SCLC vs CTR, SCLC vs ADENO, SCLC vs SCC, SCLC vs NSCLC, SCLC vs OTHER) and/or with low correlation with miR-375-3p. Among these, we elected to test miR-122-5p, miR-144-3p, miR-145-5p, miR-200a-3p, miR-200b-3p, miR-205-5p, and miR-320b ( Table 1 and Supplementary File 2 ).

In parallel, we profiled EV-included miRNAs by NanoString in a second discovery cohort of 24 patients ( Supplementary Table 1 ). The bioinformatic analysis identified few miRNAs upregulated in SCLC. Among these, we selected only miR-1285-5p, which was present in all the comparisons and showed a fold change higher than 1.5 in at least one comparison ( Supplementary File 3 ). Moreover, considering the relevant upregulation of miR-375-3p in plasma, which contains also EVs, we decided to investigate its expression specifically in EVs.

We evaluated the expression of the selected miRNAs by qRT-PCR in a training cohort of 111 samples ( Table 2 ), from which we extracted both circulating cell-free miRNAs in whole plasma and EV-included miRNAs.

Due to the high variability existing between human samples, we used three different endogenous miRNAs to normalize the expression of the circulating cell-free miRNAs, resulting in three features in the selection pool for each of these miRNAs. An exogenous spike-in was used as normalizer for EV-included miRNAs. Circulating cell-free miR-200b-3p was removed from the analysis because of its low expression (CT values >35 in more than 10% of samples).

Based on qRT-PCR data, we identified 7 circulating cell-free miRNAs significantly upregulated in the SCLC group in at least one comparison ( Figure 2A ). miR-375-3p, was significantly upregulated in SCLC group compared with the other groups (SCLC vs CTR: p<0.0001; SCLC vs SCC and SCLC vs ADENO: p<0.001); miR-122-5p was significantly upregulated in SCLC vs all the other groups (p<0.05); miR-144-3p was upregulated compared to CTR (p<0.01) and ADENO (p<0.001); miR-145-5p was upregulated in SCLC vs ADENO group (p<0.05); miR-200a-3p was upregulated in SCLC vs all the other groups (SCLC vs CTR and SCLC vs SCC: p<0.01; SCLC vs ADENO: p<0.001), miR-205-5p was upregulated in SCLC vs CTR (p<0.05) and SCLC vs ADENO (p<0.001); miR-320b was upregulated in SCLC vs all the other groups (SCLC vs CTR: p<0.001; SCLC vs SCC: p<0.05; SCLC vs ADENO: p<0.01).

By contrast, none of the tested EV-miRNAs showed significant deregulation in SCLC group ( Figure 2B ).

With the intent of building a classifier capable of distinguishing SCLC from the other histological subtypes (SCC, ADENO and CTR), we employed LASSO regression for selecting the variables (miRNAs expression and clinical data). Acknowledging the high expression of circulating cell-free miR-375-3p, and its significant upregulation in SCLC, we forced its expression into the model. We also forced the batch effect in the classifier, to consider the variability between the sample processing placed in the two different medical centers from which we obtained the samples. We first tested the association between clinical variables (age, sex, race, smoking history, histology, stage) and miR-375-3p expression to ensure that the miRNA expression was independent from the clinical data. We found that the only association of miR-375-3p expression was related to histology and stage.

We estimated the model on the complete training dataset (see methods), and then we evaluated its diagnostic performance on the qRT-PCR results for both the circulating cell-free and EV-included miRNA of the validation data set ( Supplementary Figure 1 ). For each comparison, we estimated the probability of being SCLC by using area under the receiver operating characteristics (ROC) curve (AUC), setting a specific cutoff, calculating specificity, sensitivity, positive predictive value (PPV), and negative predictive value (NPV) at the optimal cutpoint derived from the training data ( Table 3 ).

We also evaluated the predicted probability of SCLC association with the group in the validation dataset. Our data show that the variables selected comprise circulating cell-free miRNAs, whereas none of the EV-included miRNAs were retained in the model. As shown in Figure 3 , the model built on the training cohort provided a significant classification between SCLC patients and NSCLC (SCC + ADENO) and between SCLC and SCC patients in the validation dataset, highlighting its diagnostic relevance. In the SCLC vs SCC classification, miR-375 turned out to be an effective predictor, with AUC= 0.833 and with sensitivity and specificity 0.375 and 1, respectively. The applied cutoff was 0.5494 which discriminated, in the validation dataset, 6 out of 16 test-positive SCLC patients and 0 out of 6 test-positive SCC patients ( Supplementary Table 3 ), ruling out any SCC from having a positive test. As well, in SCLC vs NSCLC classification, miR-375-3p was effective predictor of SCLC, with AUC= 0.793 with sensitivity and specificity 0.375 and 1, respectively. The set cutoff was 0.4012 which correctly discriminated 6 out of 16 test-positive SCLC patients, and 0 out of 13 test-positive NSCLC patients. The results of the remaining comparisons show a significant classification in the training dataset, however, the performance of the classifier in validation data was not significant with a low AUC ( Supplementary Figure 2 , Table 3 ). Taken together, these results demonstrate that levels of circulating cell-free miR-375-3p can provide a significant classification between SCLC and NSCLC patients.

To evaluate the usefulness of our classifier in informing stage, we compared early stage (stages I+II), non-metastatic (stages I+II+III), and metastatic (stage IV) SCLC with the CTR ( Supplementary Figure 3 ). We found that miR-375-3p, miR-320b, and miR-144-3p combined with race and age could discriminate between stage IV SCLC patients and the CTR group (AUC=0.882, sensitivity=0.2857, and specificity =0.9412).

A total of 166 patients were selected for this study. The patients were recruited from two medical centers (Virginia Commonwealth University, Richmond, VA, USA, and Vanderbilt University, Nashville, TN, USA) in accordance with the institutional review board (IRB) for equivalent guidelines for each institution. Informed consent was obtained from all subjects and/or their legal guardian(s).

Based on previous approaches (56, 57), we conducted a three-phase study ( Figure 1 ): Discovery (phase 1), Training (phase 2) and Validation (phase 3). The Discovery phase aimed to identify cell-free and EV-included miRNAs present in circulation that were deregulated in the SCLC group compared to the other histological types: control (CTR) which includes normal and granuloma histology, squamous cell carcinoma (SCC), and adenocarcinoma (ADENO). We used plasma samples from a first discovery cohort collected at Virginia Commonwealth University (VCU) (n=38, Supplementary Table 1 ) to extract circulating cell-free miRNAs that were then analyzed by Illumina Next Generation Sequencing (NGS). These samples include both circulating free miRNA as well as EV-included miRNA. In parallel, we isolated EVs from a second discovery cohort (n=24, Supplementary Table 1 ) and extracted EV-included miRNAs, which were then analyzed through NanoString.

Accurate isolation of EVs was confirmed by Nanoparticle Tracking Analysis (NTA) and cryo-electron microscopy (cryo-EM), as suggested by the International Society of Extracellular Vesicles (58) ( Supplementary Figure 4 ).

In the Training phase, we extracted both circulating cell-free and EV-included miRNAs from a larger cohort (training cohort, Table 2 , left), which included part of the samples from the discovery cohort. From these samples, we tested, by qRT-PCR, candidate miRNAs in both plasma and EVs.

A biomarker classifier for detecting SCLC was established by using the expression of circulating cell-free miRNAs and EV-contained miRNAs from paired samples, as well as six clinical variables (histology, stage, sex, race, smoking history, and age), in the selection pool for the LASSO model in the training cohort.

In the Validation phase, we first evaluated the expression of circulating cell-free and EV-included miRNAs by qRT-PCR on an independent validation cohort ( Table 2 , right), which included part of the samples from the discovery cohort. Finally, we applied of the model generated on the training data set to the validation data set.

The plasma samples were divided into two clinically homogeneous groups, identified as training (n=111) and validation (n=55) (59, 60). To ensure comparable training and validation sets, splits of 111/55 were performed randomly 10,000 times, and group (training/validation) was tested for association with each of six clinical variables (histology, stage, sex, race, smoking history, and age) at each split. The split with the highest minimum p-value (p>0.9) across the six variables was retained for the training and validation analyses.

Plasma samples were spun at 2,000 X g for 20 minutes and subsequently at 10,000 X g for 20 minutes to remove cellular debris. 200 µl of supernatant were used for whole plasma miRNAs extraction and 500 µl were used for EVs isolation.

EVs were isolated from plasma using the Total Exosome Isolation Kit (INVITROGEN #4484450) following the manufacturer’s protocol. EVs were thoroughly resuspended in 200 µl of PBS with 1/10 reserved for EVs characterization.

Nanoparticle tracking analysis was performed using the NanoSight NS300 system (Malvern, Great Malvern, UK) as previously described (61).

EV characterization by cryogenic transmission electron microscopy (cryo-EM) was performed at the Advanced Materials and Liquid Crystal Institute (AMLCI), Kent State University, OH, USA. A FEI Vitrobot (Mark IV) plunge freezer was used to prepare vitrified cryo-TEM specimens from the solution sample (62). Cryo-TEM observation was performed on a FEI Tecnai F20 transmission electron microscope. The basic experimental setup and procedure can be found in (63).

Circulating cell-free RNA was extracted by using the miRNeasy micro kit (QIAGEN #217084) according to the manufacturer’s protocol. RNA from EVs was extracted using TRIzol reagent (Invitrogen #15596018) and supplied with 35 picograms of non-human miRNA spike-ins (ath-miR-159a; cel-miR-248; osa-miR-414), for normalization purposes. RNA purification was performed with RNA Clean-Up and Concentration Kit (NORGEN #43200). 3 µl of RNA from plasma and EV samples were retrotranscribed using the TaqMan® Advanced miRNA cDNA Synthesis Kit (#A28007) and qRT-PCR was performed using TaqMan® reagents (TaqMan™ Advanced miRNA assay #4444964; TaqMan™ Fast Advanced Master Mix # 4444557, Thermo Fisher). All the probes used were obtained from TaqMan® (#A25576). MiRNAs 24-3p, 93-5p and 126-3p were used as endogenous controls for normalizing whole plasma PCR data, as suggested from the manufacturer; ath-miRNA-159a was used as a unique normalizer for EV PCR data. All the assays were carried out in three technical replicates. Samples and technical replicates in which the Ct value was >35 were excluded from the statistical analysis. The 2^-dct values were used as relative expression of the biostatistical analysis.

NGS was performed at the Institute for Systems Biology, Seattle, Washington as previously reported (64). One of the control samples was considered an outlier and removed from the NGS analysis. Raw fastq files were initially pre-processed by trimming Illumina adapter sequences via Cutadapt tool (65). and then qualitatively filtered by Condetri tool (66) (parameters: -pb=fq -lq=20 -hq=30 -minlen=15 -sc=33). For each sample, all post-processed reads were mapped into the human genome (GRCh37) by using miARma-Seq tool (67), which includes Bowtie tool (v1) (68) for aligning reads and featureCounts tool (69) to annotate them into microRNA space [miRBase v20 (70)]. A minimum expression filtering is applied prior to normalization, retaining all those miRNA molecules with at least 10 reads in 50% of samples. Filtered miRNA were normalized by applying the TMM method (trimmed mean of M-values). Differentially expression analysis was carried out by using EdgeR package (71). For each comparison, we took into consideration all those miRNAs with a P-value<.05 and |Linear Fold-Change|>1.5 ( Supplementary File 1 ). For a second more restricted analysis in the discovery data, features were filtered by percent present >90%, and at least 15 read counts in a least 50% of the samples in at least one group. Count data were analyzed across groups with the omnibus Kruskal-Wallis test, followed by pairwise by group analysis with the Wilcoxon rank sums tests. Spearman correlation was used to determine miRNA/miRNA associations. Additional miRNAs were selected for inclusion in the classification stage by association with group and low correlation with miR-375 ( Supplementary File 2 ).

Total exosomal RNA from a cohort of 24 samples (8 CTR, 4 SCC, 4 ADENO, 8 SCLC) were profiled through NanoString nCounter Human v3 miRNA Expression Assay as previously described (29).

Raw data (.rcc files) produced via NanoString nCounter Human v3 miRNA Expression Assay were analyzed with nSolver™ (provided by NanoString Technologies). Negatives controls were considered to perform background noise subtraction, while positive controls were considered to perform technical normalization, adjusting lane-by-lane variability due to differences in hybridization, purification, or binding. Finally, data were biologically normalized by calculating the geometric mean of the top 100 miRNAs in all samples, as recommended by NanoString company. Differential expression analysis was carried out by using Limma R package (72) from the Bioconductor R project. For each comparison, we took into consideration all those miRNAs with a P-value<.05, |Linear Fold-Change|>1.25 and an average expression of >30 count in at least one condition ( Supplementary File 3 ).

Training and validation analyses were performed in R (73). For qRT-PCR data, hierarchical clustering using the Ward method (74) on correlations was performed using differentially expressed miRNAs. For the training data, LASSO (glmnet package) was used to select variables and produce the final model. Data were produced in two batches, and therefore a batch variable was forced into the LASSO models. Additionally, the initial strongest association in the discovery data was miR-375-3p, and, therefore, this feature was also forced into the LASSO models. The variables considered were the log10(2^-dct) of EV- included miR-1285-5p, miR-375-3p, and circulating cell-free miR-122-5p, miR-144-3p, miR-145-5p, miR-200a-3p, miR-205-5p, miR-320b, miR-375-3p vs the normalizers miR-24-3p, miR-93-5p and miR-126-3p, as well as clinical data (race, sex, age, and smoking history).

The samples with missing features were not considered for the predictive modeling and the analysis was performed on a total of 94 (Training cohort) and 46 (Validation cohort) paired EVs and PLASMA samples.

The predicted probability of being SCLC from the model was used as a score for each sample. The optimal cutpoint by minimum p-value (OptimalCutpoints package) (75), was determined for each comparison to classify samples. The training set derived model was then applied to the validation set. The probability of SCLC was tested (Wilcoxon rank sums) for association with group (control vs. SCLC, for example), AUC calculated, and using the cutpoint determined from the training data, sensitivity, specificity, PPV, and NPV calculated to evaluate the utility of the model for classification in the validation set.

For the real-time PCR data, normal distribution was first assessed and non-parametric Mann Whitney test was performed using GraphPad Prism version 9.5.1 for Windows, GraphPad Software, San Diego, California USA, www.graphpad.com.