Landscape of extracellular small RNA and identification of biomarkers in multiple human cancers

Summary

Extracellular RNAs (exRNAs) in biofluids, sourced from diverse tissues, exhibit various biological functions and diagnostic potential. Small non-coding RNAs, such as rsRNAs and tsRNAs, are abundant in tissues and likely secreted into biofluids, contributing to exRNA profiles. To comprehensively evaluate exRNAs, we employed traditional and enzymatic treatment RNA sequencing to systematically profile exRNAs across six human biofluids including serum, ascites, urine, milk, seminal plasma and saliva as well as sera from mice, rats, rabbits, and bovines. rsRNAs were identified as the most abundant exRNA species in human biofluids, with rsRNAs and tsRNAs showing high expression and species-specific profiles across animals. In serum samples from 51 healthy individuals and 69 cancer patients, exRNA-based machine learning model achieved 94.1% sensitivity and 100% specificity in cancer detection and accurately classified tumor origin. Altogether, this study reveals distinct rsRNA-abundant exRNA landscape across multiple biofluids and support the potential of exRNA signatures in pan-cancer diagnostics.

Graphical abstract

Highlights

• •Enzymatic-treatment RNA-seq broadens exRNA landscape beyond traditional RNA-seq
• •rsRNAs and tsRNAs are abundantly detected across diverse biofluids
• •exRNA expression profiles display distinct biofluid-specific and species-specific signatures
• •exRNA signatures from 120 human serum samples achieve high-accuracy cancer detection

Oncology; Diagnostics; Transcriptomics

Introduction

Once thought to exist only within cells, RNA is now known to be exported from cells and to play vital roles in cell-to-cell communication as extracellular RNA (exRNA).^1,2 Recent studies conducted by our team and others have shed light on the remarkable stability and abundance of extracellular small RNAs, predominantly microRNAs (miRNAs), in a spectrum of bodily fluids, such as serum, plasma,^3,4 saliva,^5,6,7 urine,^8,9 and milk.^10,11 Extracellular miRNAs have thus emerged as noteworthy candidates for noninvasive, cost-effective diagnostics, prognostics, and theranostics across a spectrum of diseases.^{12,13,14,15,16} Although several commercial kits are available for clinical disease diagnosis through extracellular miRNA detection, the full potential spectrum of exRNA biomarkers is not fully utilized. This limitation primarily results from the focused nature of existing identification methods, which predominantly target approximately 1000 species of extracellular miRNAs.^17,18

Small RNA sequencing (sRNA-seq) remains a high-throughput platform for both the quantification and “de novo” discovery of exRNAs.¹⁹ Nonetheless, recent studies have revealed the occurrence of sequencing bias, which arises from two primary challenges encountered during cDNA library preparation. First, terminal modifications in small noncoding RNAs (sncRNAs) hinder the successful ligation of sequencing adaptors to these molecules. Second, internal RNA modifications in small RNAs pose obstacles during the reverse transcription (RT) process, which is responsible for converting RNA into cDNA.^20,21 In recent years, novel methods employing certain enzymatic treatments have been used to efficiently resolve RNA termini and eliminate RT-blocking RNA modifications, leading to the successful detection of previously elusive sRNAs and unveiling a more comprehensive landscape.^22,23 One noteworthy example is PANDORA-seq, which has demonstrated the abundance of tRNA-derived small RNAs (tsRNAs) and rRNA-derived small RNAs (rsRNAs) in various tissues and cells.^18,24 Emerging from diverse cell origins, exRNAs present in biofluids have been hypothesized to possess additional modifications or modifications that may hinder their detection. Thus, despite considerable research efforts dedicated to exploring circulating exRNAs, including investigations into their biomarker potential and biological functions, comprehensive profiling of exRNA characteristics across various biofluids has not been performed.

To comprehensively explore the exRNA landscape, we applied both traditional RNA-seq and enzymatic treatment RNA-seq techniques to a range of biofluids, including human serum, ascites, urine, milk, seminal plasma, and saliva, as well as serum from mice, rats, rabbits, and bovines. Traditional RNA-seq analysis highlighted miRNAs or tsRNAs as predominant exRNAs in all six human biofluids. In contrast, enzymatic treatment RNA-seq revealed a broader diversity of RNA species in exRNAs across various bodily fluids, with rsRNAs exhibiting the highest abundance. By investigating the serum exRNA profiles of cancer patients and comparing them to those of healthy controls (HCs), we identified cancer-specific exRNA signatures. Notably, serum exRNAs exhibited the ability to differentiate among cancer types in patients. These findings will enhance our understanding of unexplored exRNA species and drive further progress in the research of exRNA biomarkers and biological function.

Results

Landscape of rsRNAs enrichment in exRNAs across diverse human biofluids

To determine the exRNA landscape in human biofluids, we employed both traditional RNA-seq and enzymatic treatment RNA-seq (with reference to PANDORA-seq¹⁸). The overall study design is shown in Figure S1. Enzymatic treatment RNA-seq utilizes a dual enzymatic treatment strategy designed to address specific RNA modifications that might impede reverse transcription (via AlkB treatment²²) or adapter ligation (via T4PNK treatment²⁵).Commercially available T4PNK was used according to established protocols, while the AlkB enzyme was purified after expression from a codon-optimized plasmid, as previously reported (Figure S2A). The enzymatic efficacy of AlkB in removing RNA methylation was evaluated using dot blot and liquid chromatography-tandem mass spectrometry (LC‒MS/MS) analyses. The treatment efficiently removed m1A and m3C, as evidenced by the results (Figures S2B and S2C). ExRNAs were isolated from six human biofluids using TRIzol LS reagent, including serum, urine, ascites, seminal plasma, and milk. The integrity and quantity of each RNA sample with or without the two-step enzymatic treatment were assessed through agarose gel electrophoresis, NanoDrop spectrophotometry, and the Agilent 2100 platform (Figure S3A). No significant shift in fragment peaks were observed under enzyme treatment. Additionally, similar sequence mapping rate and read summaries (except for urine and saliva) of exRNAs between the two protocols in the six human biofluids are presented in Figure S3B. It is worth mentioning that human saliva contains abundant oral microbiomes which contributed approximately 34% of reads. The distribution of exRNA sequences, as exemplified in human serum, indicated that miRNAs are the predominant exRNAs detected using traditional RNA-seq. However, the application of the AlkB and T4PNK treatments significantly increased the read counts of the rsRNAs, revealing rsRNAs as the primary exRNAs (Figure 1A). Similarly, in ascites, milk, urine and saliva, traditional RNA-seq identified tsRNAs as the primary exRNAs, whereas enzymatic treatment substantially enhanced the RPM of the rsRNAs (Figures 1B–1F and S4). In seminal plasma, both traditional RNA-seq and enzymatic treatment RNA-seq revealed rsRNAs as the primary exRNAs, with enzymatic treatment RNA-seq partially increasing the RPM of rsRNAs (Figures 1D and S4). We further separately analyzed the relative rsRNA/miRNA ratios in six human biofluids under both two RNA-seq protocols. Notably, milk exhibited the highest concentration of rsRNAs, correlating with the highest rsRNA/miRNA ratio, while ascites and milk demonstrated the highest concentrations of tsRNAs, associated with the highest tsRNA/miRNA ratio across all samples analyzed with enzymatic treatment RNA-seq (Figure 1G). In addition, in all six human biofluids, the rsRNA/miRNA ratio was significantly greater than the tsRNA/miRNA ratio. The unexpectedly abundant expression of rsRNAs revealed by enzymatic treatment RNA-seq in human biofluids underscores the critical importance of further investigating the physiological and clinical implications of these findings.

Figure 1.

Validation and characterization of enzymatic treatment RNA-seq and traditional RNA-seq-derived sncRNA profiles in six different human biofluids

(A–F) Characterization of sncRNA categories and length distributions from two sequencing methods in six different biofluids: serum (A), ascites (B), milk (C), seminal plasma (D), urine (E) and saliva (F). The length distribution data are shown as the mean ± SEM (n = 2 biologically independent samples).

(G) Relative ratios of tsRNAs/miRNAs and rsRNAs/miRNAs determined via two sequencing methods across six human biofluids. The data are shown as the mean ± SEM (n = 2 biologically independent samples).

Characterization of rsRNAs and tsRNAs in human biofluids under enzymatic treatment RNA-seq

Since the enzymatic treatment RNA-seq had identified abundant novel rsRNAs and tsRNAs in the biofluids, these two types of exRNAs were further characterized. Comprehensive comparative expression profiles of rsRNAs across six biofluids are depicted based on their origin from individual ribosomal RNAs (rRNAs), including 5S, 5.8S, 18S, 28S, 45S, and mitochondria-encoded 12S and 16S rRNAs (Figures 2A and S5A). The consistency of the abundance of rsRNAs originating from the same rRNA source was evident across the six biofluids (Figure 2A). Notably, rsRNAs derived from 28S to 18S rRNA demonstrated the highest abundance in exRNA across the six biofluids, whereas rsRNAs derived from 12S to 16S rRNA exhibited the lowest abundance (Figure S5A). These findings suggested that the observed rsRNA levels in biofluids may closely correlate with the expression levels of rRNA within cells. We utilized an UpSet plot to visualize the coexpression patterns of the rsRNAs across the six biofluids. A total of 15,033, 91,287, 94,846, 133,665, 338,664, and 116,293 rsRNAs were identified as being specifically expressed in serum, urine, ascites, seminal plasma, milk and saliva, respectively, highlighting the biofluid specificity of the rsRNA distribution. Furthermore, 17,077 rsRNAs were identified as shared among all biofluids, indicating that only a small proportion of rsRNAs have a consistent abundance pattern across these diverse biofluids (Figure 2C). The comprehensive mapping of all rsRNAs on their respective precursor rRNA length scales (28S, 5S, and 18S) revealed distinctive patterns within the six biofluids (Figures S5B–S5D).

Figure 2.

Characterization of tsRNAs and rsRNAs in six human biofluids via enzymatic treatment RNA-seq

(A) Heatmap showing the relative expression levels of the six biofluids stratified by the origin of the rsRNAs (genomic) (n = 2 biologically independent samples). Heatmap was under column normalization using Z Score and clustered using centroid distance.

(B) Radar plots illustrating the genomic tRNA origins of tsRNAs derived from enzymatic treatment RNA-seq across six different biofluids. Log values are shown for the mean RPM.

(C and D) UpSet plot showing the unique and shared rsRNA (C) or tsRNA (D) types based on their sequences. The vertical bars in the plot show the type and number of rs/tsRNAs. Overlapping types are illustrated by connected black circles below the histogram. Red bars and circles represent rs/tsRNA types that overlap among the six biofluids.

(E and F) Expression levels of the top 10 rsRNAs (E) and tsRNAs (F) determined via enzymatic treatment RNA-seq. The data are shown as the mean ± SEM (n = 2 biologically independent samples).

Radar plot depicting the expression patterns of the tsRNAs across the six biofluids according to their origin from genomic tRNA (Figure 2B) and mitochondrial tRNA (Figure S6F) was generated. Overall, the amount of tsRNAs originated from genomic tRNA were much higher than tsRNAs from mitochondrial tRNA. Heatmaps of the tsRNAs further delineated their respective origins (Figure S6B). The consistency of the abundance of tsRNAs originating from the same tRNA source was high across the six biofluids and rare tsRNA was derived from tRNA-Sup and tRNA-Sec across all human biofluids, which suggested that tRNA carrying different amino acids had distinct tendency differences in generating tsRNA (Figures 2B and S6B). UpSet plot analysis revealed 1139, 1458, 2064, 2463, 3429, and 319 tsRNAs specifically expressed in serum, urine, ascites, seminal plasma, milk and saliva, respectively, highlighting the biofluid specificity of the tsRNA distribution. Moreover, 1648 tsRNAs were detected as shared among all six human biofluids, suggesting that a considerable proportion of tsRNAs share a common expression pattern across these diverse biofluids (Figure 2D). For further characterization of serum tsRNAs, MINTmap was used to classify them based on their distinct origins. Specifically, the tsRNAs were categorized into six types: 5′-half, 3′-half, 5′-tRF, 3′-tRF, i-tRF, and tRF-1 (Figure S6A). Among these, i-tRF consistently exhibited the highest expression abundance across all six biofluids, potentially linked to its expression levels in cells or tissues and the stability of various tsRNA species (Figure S6B). Intriguingly, large amount of tsRNA from all tested body fluids tended to be derived from either side of anticodon loop of their parental tRNA, implying that cleavage at this position of parental tRNA by nucleases such as ANG is one of the crucial mechanisms of tsRNA generation²⁶ (Figures S6D and S6E). We presented the top 10 expressed rsRNAs and tsRNAs in serum, urine, ascites, seminal plasma, milk and saliva samples, emphasizing the significant differences in RNA content among these biofluids (Figures 2E and 2F). These findings suggest considerable similarity in the sources of rsRNAs and tsRNAs across diverse bodily fluids in the human body; however, within each fluid, a distinct expression profile is evident.

Cross-species analysis of serum-derived exRNAs

We next employed enzymatic RNA-seq to elucidate the dynamics of serum exRNAs across different species, including mice, rats, rabbits, and bovines. The levels of miRNAs, tsRNAs, and rsRNAs exhibited dynamic changes across these species (Figure 3A). Comparative analysis with human serum indicated a significantly greater abundance of rsRNAs and tsRNAs than miRNAs in mouse, rat, rabbit, and bovine serum. The consistency of the rsRNA abundances originating from the same rRNA source showed significant inconsistencies across these species (Figures 3B and S7), suggesting a species-specific distribution of rsRNAs in serum.

Figure 3.

Characterization of serum RNAs from multiple species via enzymatic treatment RNA-seq

(A) Characterization of sncRNA categories and length distributions in enzymatic treatment RNA-seq data from mouse, rat, rabbit and bovine serum.

(B) Radar plots showing the origin of the genomic tsRNAs identified via enzymatic treatment RNA-seq across serum samples from multiple species. The log RPM is plotted. Heatmap was under row normalization using Z Score and clustered using centroid distance.

(C) Heatmap showing the tRNA origins of serum tsRNA in multiple species. Heatmap was under row normalization using Z Score and clustered using centroid distance.

(D) Heatmap showing the types of serum rsRNAs across species classified by their precursor tRNA origin detected by enzymatic treatment RNA-seq. Heatmap was under row normalization using Z Score and clustered using centroid distance.

(E and F) Expression levels of the top 10 rsRNAs (E) and tsRNAs (F) according to enzymatic treatment RNA-seq. The data are shown as the RPM.

Radar plots illustrating the expression patterns of the tsRNAs across the four species according to their origin from genomic tRNA was generated (Figure 3B). The consistency of tsRNA abundance originating from the same tRNA source was prominent across serum samples from mice, rats, rabbits, and bovines (Figure 3C). Heatmaps of the rsRNAs were generated to further detail their respective origins (Figure 3D). Lack of tsRNA-Sup and tsRNA-Sec was also observed in multiple-species serum with which was consistent in human biofluids (Figure 2B). Overall, mapping of tsRNAs on a rsRNA/tsRNA length scale revealed the loci from which the rsRNAs/tsRNAs were preferentially derived across multiple species (Figure S8B). The top 10 expressed rsRNAs and tsRNAs in the serum of mice, rats, rabbits, and bovines were detected, highlighting notable differences observed across these species (Figures 3E and 3F). These findings showed enrichment of both rsRNAs and tsRNAs in the serum across different species, suggesting conservation of this phenomenon. However, the observed variations in the expression abundance of rsRNAs and tsRNAs between different species also imply that the expression profile of exRNAs in serum may possess species-specific characteristics.

Analysis of serum-derived exRNAs across multiple cancers to identify tumor-associated exRNAs signatures

Serum/plasma is one of the most accessible media for liquid biopsy samples, and its use provides insights into tumor-associated exRNAs. To elucidate the characteristics and composition of these exRNAs, we focused on identifying small RNAs present in the serum of cancer patients. The sex and age of the participants were balanced among the enrolled patients and were consistent with those of the HCs (Figures S9A and S9B). The exRNA patterns of 51 healthy individuals and 69 cancer patients, including colorectal cancer, thyroid cancer, gastric cancer, prostate cancer, and hepatocellular carcinoma, were analyzed (Figures S9C and S9D). By applying the random forest classifier, which is known for its robustness, we identified a distinct set of exRNAs capable of accurately discriminating between HCs and tumor patients. The top 20 exRNAs for distinguishing between cancer patients and HCs were identified based on the weights assigned by the classifier (Figure 4A). Subsequently, we optimized the selection procedure to identify the best combination of six exRNAs, chosen to achieve the minimal number of markers for optimal discriminatory efficacy. The weighted expression levels for these six small RNAs were derived and effectively distinguished between samples from cancer patients and HCs (Figure 4B). For model training and testing, the samples were evenly divided into HC and cancer groups, with 75% of the samples comprising the training set and the remaining 25% serving as an independent test set. Analysis of 120 serum-derived exRNAs from 51 cancer patients with five different cancer types and 69 HCs yielded 100% sensitivity and 100% specificity in 10-fold cross-validation of the training set (Figure 4C). When the six small RNAs detected in the serum exRNAs were used to generate the model, 94.1% sensitivity and 100% specificity were achieved in the test set (Figure 4D). These results collectively suggest that serum-derived exRNAs could serve as valuable liquid biopsy tests for cancer detection.

Figure 4.

Identification of enzymatic treatment RNA-seq-derived serum exRNA signatures from patients with various cancers

(A) Heatmap illustrating the top 20 signature serum exRNAs with high predictive values revealed by the random forest algorithm for classifying tumor and nontumor serum samples. The red rectangles represent the top 6 serum exRNAs with the highest predictive values, used to construct the signature.

(B) Violin plot showing the distribution of PC1 (principal component 1) of weighted expression levels of the 6 serum sncRNAs.

(C and D) Classification error matrix for the 6 serum exRNA-based signature in the training set (75% of samples) and test set (25% of samples).

Tumor-derived exRNA profiles classify primary tumor of origin

We aimed to explore whether the exRNA signatures of patients could be linked to the specific cancer type. Balanced age and gender distributions were ensured. We analyzed exRNAs extracted from the serum of patients diagnosed with five distinct cancer types: colorectal, thyroid, gastric carcinoma, hepatocellular, and prostate cancer (Figure 5). Employing the least absolute shrinkage and selection operator (LASSO) classification to identify exRNA signatures distinguishing between the five tumor types, we discovered 11 exRNAs, including those associated with the uronic acid metabolic process, with the highest predictive value (AUC for ROC = 1.0) for discriminating hepatocellular cancer (Figure 5A). It has been indicated that a substantial proportion of tsRNAs and rsRNAs interacted with Argonaute (AGO) complex and may regulate gene expression in a manner similar to miRNA.^{27,28,29,30,31} Given this functional similarity and the lack of specialized prediction algorithms, we employed a widely used miRDB, miRNA target prediction tool to screen potential target genes of tsRNAs and rsRNAs. LASSO feature selection revealed 9 exRNAs related to the Wnt signaling pathway that had the highest predictive value (AUC for ROC was 1.0) for distinguishing thyroid cancer (Figure 5B). For prostate cancer, LASSO identified 7 exRNAs, including those related to the Wnt signaling pathway, with the highest predictive value (AUC for ROC was 1.0) (Figure 5C). Gastric cancer was distinguished by 12 exRNAs, including those associated with sodium ion transmembrane transport, with the highest predictive value (AUC for ROC was 1.0) (Figure 5D). Colorectal cancer was characterized by 6 exRNAs, some related to the Wnt signaling pathway, with the highest possible predictive value (AUC for ROC was 1.0) (Figure 5E). These findings serve as proof of principle that serum-derived exRNAs constitute genuine tumor-specific signatures capable of distinguishing cancer types.

Figure 5.

Characterization and selection of features in various cancers

(A–E) Tumor-derived exRNA profiles screened by LASSO are shown as a heatmap (left panel). The diagnostic value of tumor-derived exRNA profiles was analyzed by receiver operating characteristic (ROC) curve analysis (middle panel). GO analysis (right panel) of the exRNA combinations for hepatocellular carcinoma (A), thyroid cancer (B), prostate cancer (C), gastric cancer (D) and colorectal cancer (E).

Discussion

This study primarily explored the comparison between traditional RNA-seq and enzymatic treatment RNA-seq in identifying exRNAs in multiple biofluids and species. Utilizing enzymatic treatment RNA-seq, our analysis revealed a distinct exRNA landscape that mainly consisted of rsRNAs, in contrast to the traditional RNA-seq approach, which primarily yields miRNAs. Notably, we provide evidence of the outstanding performance of serum-derived exRNAs, including rsRNAs, tsRNAs, and miRNAs, in detecting cancer.

The investigation of various classes of sncRNAs, such as rsRNAs and tsRNAs, alongside the more well-established miRNAs, in cells and tissues is increasing in popularity.^32,33,34,35 However, intricate RNA modification patterns have posed challenges in high-throughput analyses of sncRNAs, hindering RNA-seq library preparation and the detection of modified tsRNAs and rsRNAs.^21,36 To address these issues, RNA-seq methods that incorporate enzymatic treatment, such as PANDORA-seq, have been developed to enhance adapter ligation and reverse transcription during library construction.¹⁸ In this study, we explored the exRNA profiles generated using both traditional RNA-seq and enzymatic RNA-seq across a range of biofluids. Rozowsky et al. and Murillo et al. also profiled the exRNA across diverse biofluids using traditional RNA-seq.^16,37 Their findings are comparable to our traditional RNA-seq results, particularly in the relative composition of exRNA species. Our comparative analysis revealed the following: (1) Serum: Both our traditional RNA-seq and exRNA-Atlas data confirmed that miRNAs dominate the serum exRNA profile, with a similar exRNA distribution observed in both datasets. (2) Seminal plasma: Our traditional RNA-seq identified more rsRNAs, while exRNA-Atlas data reported a predominance of tsRNAs. (3) Saliva: exRNA-Atlas data showed a rsRNA-dominated pattern, whereas our traditional RNA-seq indicated a more comparable expression level of tsRNAs and rsRNAs in saliva. Collectively, these comparisons demonstrate that our traditional RNA-seq results align well with previously published exRNA profiles across diverse human biofluids. However, our enzymatic treatment RNA-seq provides a unique perspective, revealing an rsRNA- and tsRNA-dominated exRNA landscape that was not fully captured by previous studies.

Interestingly, our analysis of serum samples from different species (mouse, rat, rabbit, and bovine) revealed that tsRNAs are the most abundant small RNA species across all four species, with rsRNAs closely related to each other, a pattern not mirrored in human serum. Additionally, while the sources of tsRNAs were largely consistent across the different species, the sources of rsRNAs showed notable variations, suggesting a potential species-specific distribution of rsRNAs in serum. This observation also implies that the rsRNAs present in human serum may exhibit unique biological functions distinct from those in other species.

Exploring biomarkers in extracellular body fluids offers a significant advantage for patients, given its accessibility, reduced invasiveness, and potential cost-effectiveness. Because current identification methods have a limited ability to discover and characterize extracellular small RNAs, mainly extracellular miRNAs (approximately 1000 species), only a small number of disease-specific small RNAs are available as potential clinical biomarkers. Thus, the lack of a comprehensive landscape of extracellular small RNAs presents a significant obstacle in this field. Our analysis identified the top 20 extracellular small RNAs via weighted random forest classification to distinguish between cancer patients and healthy individuals. Notably, 18 of the top 20 RNAs identified were rsRNAs, indicating the potential significance of serum rsRNAs as biomarkers for cancer. Additionally, an optimized set of six exRNAs (including five rsRNAs) was utilized to develop a classifier for discriminating between cancer patients and healthy individuals.

The determination of tumor types and their tissue origins is of significant importance for cancer treatment.^38,39,40 Our research revealed the ability to differentiate between various cancer types, such as colorectal cancer, thyroid cancer, gastric carcinoma, hepatocellular cancer, and prostate cancer, based on specific combinations of exRNAs, primarily rsRNAs. Such distinct exRNA signatures tailored to specific cancer types have the potential to serve as a liquid biopsy tool to assist in the diagnosis and treatment management of these patients. In this study, we used the miRDB to predict potential target genes of cancer-specific exRNAs and then examined their GO biological process enrichment using Metascape. The enrichment of target gene functions in the Wnt signaling pathway, which plays a key role in tumorigenesis regulation,^41,42 suggests the potential involvement of exRNAs in tumorigenesis via this pathway. It is worth noting that functioning in an miRNA-like manner is only one of the several mechanisms through which rs/tsRNAs exert their biological effects.^43,44 Recent studies provided evidence that tsRNA can act in an aptamer-like manner by directly binding to functional proteins and promoting protein oligomerization to promoting tumor metastasis.^45,46 Studies also revealed that rs/tsRNAs could interact with toll-like receptor 7/8 (TLR 7/8) and elicit TLR7/8 stimulation which immune and inflammatory responses in cancer.^47,48,49,50 In addition, although current study identified abundant exRNAs across multiple biofluids, the mechanism by which these exRNAs are able stably exist in an RNase-rich environment remains unclear. Evidence suggested that exRNAs could be protected from degradation by extracellular vesicles and circulating protein complexes, which may present distinct biological functions.^2,51,52 Thus, the exact form of existence in the biofluids of the specific exRNAs and corresponding functions need further investigation.

In summary, our findings unveil a unique rsRNA-abundant exRNA landscape using enzymatic treatment RNA-seq. This prevalence suggests the potential of these exRNAs as biomarkers for cancer and enhances prospects of exRNA-based clinical diagnostics.

Limitations of the study

Our study unveils a unique rsRNA-abundant exRNA landscape using enzymatic treatment RNA-seq. This prevalence suggests the potential of these exRNAs as biomarkers for cancer and enhances prospects of exRNA-based clinical diagnostics. However, the enzymatic treatment RNA-seq harbored repeated enzyme treatment and RNA extraction procedures which may make the experimental manipulation more complicated, thus impeding its further clinical application. Furthermore, due to the limitation amount of the substrates of AlkB and T4PNK employed in the sRNA treatment, potential sRNA modifications that interfere sRNA library construction cannot be completely removed by the current two enzymes. Also, the diagnostic capacity of these novel exRNAs revealed by enzymatic treatment RNA-seq will also needed to be further discovered and validated in large, independent cohorts.

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Yanbo Wang (ybwang@nju.edu.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• (1)Enzyme-treatment RNA sequencing data and traditional RNA sequencing data of human and other mammalian biofluids samples have been deposited on the China National Center for Bioinformation (CNCB) under the accession number CRA014466, HRA006520 and HRA010175 which can be accessed via https://ngdc.cncb.ac.cn/gsa/browse/CRA014466, https://ngdc.cncb.ac.cn/gsa-human/browse/HRA006520 and https://ngdc.cncb.ac.cn/gsa-human/browse/HRA010175 respectively.
• (2)Code for data analysis is available upon reasonable request to the corresponding author.
• (3)Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

We thank all the authors who contributed to this topic. pET28a plasmid was kind gifted by Dr. Xiaoyun Ji (NJU Advanced Institute of Life Sciences, Nanjing University).

This work was supported by National Natural Science Foundation of China (No. 32370629); National Natural Science Foundation of China (No. 32000549); National Natural Science Foundation of China (No. 82003024); Training Program of the Major Research Plan of the National Natural Science Foundation of China (No. 92049109); and Special Fund of Jiangsu Provincial Science and Technology Plan (Key Research and Development Program for Social Development, No. BE2023804).

Author contributions

Conceptualization, Y.B.W., Z.Z., and X.L.; methodology, S.B.Z., B.T.L., P.G., S.K.Z., and T.T.Y.; investigation, S.B.Z., Y.X., G.L.L., and X.B.K.; visualization: P.G., C.W.Z., and L.J.; supervision, Y.B.W., Z.Z., C.-Y.Z., and X.C.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
anti-m1A antibody	Abcam	208196
anti-rabbit IgG secondary antibody	Santa Cruz	sc-2357; RRID: AB_628497
Bacterial and virus strains
E. coli BL21 (DE3)	Vazyme	C504-02
Biological samples
Human serum	The Affiliated Drum Tower Hospital of Nanjing University Medical School	N/A
Human ascites	The Affiliated Drum Tower Hospital of Nanjing University Medical School	N/A
Human urine	The Affiliated Drum Tower Hospital of Nanjing University Medical School	N/A
Human milk	The Affiliated Drum Tower Hospital of Nanjing University Medical School	N/A
Human saliva	The Affiliated Drum Tower Hospital of Nanjing University Medical School	N/A
Human seminal plasma	The Affiliated Drum Tower Hospital of Nanjing University Medical School	N/A
Cancer patients’ serum	The Affiliated Drum Tower Hospital of Nanjing University Medical School	N/A
Chemicals, peptides, and recombinant proteins
Nuclease P1	New England Biolabs	M0660S
phosphodiesterase I (PDE1)	Millipore	9025-82-5
alkaline phosphatase (CIAP)	TAKARA	2250A
1 M HEPES	Gibco	15630106
Adenosine 5′-Triphosphate(ATP)	New England Biolabs	P0756S
RNase Inhibitor, Murine	New England Biolabs	M0314S
T4 Polynucleotide Kinase	New England Biolabs	M0201S
α-ketoglutaric acid	Sigma	K1128-25G
L-sodium ascorbate	Sangon Biotech	A500830-0100
ferrous ammonium sulfate	Sigma	09719-50G
Critical commercial assays
TRIzol LS	Thermol	10296028CN
TRIzol reagent	Thermol	15596018
Deposited data
sRNAseq of Cross-species serum-derived exRNAs	https://ngdc.cncb.ac.cn/gsa/browse/CRA014466	CRA014466
sRNAseq of human biofluids-derived exRNAs	https://ngdc.cncb.ac.cn/gsa-human/browse/HRA006520	HRA006520
sRNAseq of human saliva-derived exRNAs	https://ngdc.cncb.ac.cn/gsa-human/browse/HRA010175	HRA010175
Recombinant DNA
pET28a	Ji lab	N/A
Software and algorithms
SPORTS1.1	–	–
glmnet	–	–
SPSS 24.0	–	–
GraphPad Prism 8.0	–	–

Experimental model and study participant details

Clinical samples

Human biofluids, including serum, ascites, urine, milk, seminal plasma and saliva were obtained from healthy donors or patients. All human samples were acquired from consenting donors at The Affiliated Drum Tower Hospital of Nanjing University Medical School. All the experiments were performed in accordance with the Code of Ethics of the World Medical Association (Declaration of Helsinki) and the approved guidelines of Ethics Committee of Nanjing Drum Tower Hospital affiliated Nanjing University medical School (approval no. 2022-461-02). All participants were Asian. Refer to Table S1 for additional details on age, sex and clinical information.

Animal samples

Wild-type male C57Bl/6J mice (8 weeks of age) and male Sprague-Dawley rats (6 weeks of age) were purchased from Gempharmatech. All animal protocols followed the National Institutes of Health guidelines for the care were approved by the Institutional Animal Care and Use Committee of Nanjing University (IACUC-2104004). Rabbit serum was purchased from Nanjing SenBeiJia Biological Technology (SBJ-SE-RAB001) from 6-month-old New Zealand Rabbit. Bovine serum was purchased from Thermo Fisher Scientific (Fetal Bovine Serum, Premium Plus, A5669701).

Method details

RNA extraction from biofluids

The biofluid samples were pooled from 10 healthy donors, while the serum samples of pan-tumor patients and matched healthy donors were subjected to RNA extraction separately. For the serum or ascites samples, 1–2 mL of samples were centrifuged at 5000 rpm for 15 min at 4°C to remove potential cell fractions. For urine, samples were mixed from 10 healthy donors to a final volume of 25 mL and subsequently centrifuged at 2000×g for 20 min at 4°C to remove the cell component. For milk, 1 mL of each mixed sample was centrifuged at 1500×g for 20 min at 4°C, followed by another centrifugation at 3000×g for 20 min at 4°C, after which the supernatant was collected. For seminal plasma, 800 μL of mixed sample was centrifuged at 6000×g for 10 min at 4°C, after which the supernatant was collected. Saliva was collected followed the previous protocols.^7,53 Briefly, unstimulated samples were mixed from 6 healthy donors to a final volume of 20 mL followed by centrifugation at 2600×g for 20 min at 4°C. Supernatant was collected and RNase Inhibitor (New England Biolabs, USA) added. RNA of all biofluid samples were extracted with TRIzol LS reagent (Ambion, USA) following the manufacturer’s instructions. The integrity and quantity of each RNA sample were tested using agarose gel electrophoresis and a NanoDrop instrument.

Expression and purification of AlkB

AlkB purification and storage were performed according to previous methods.¹⁸ The AlkB gene with six histidines at the amino terminus was cloned and inserted into the pET28a plasmid, and its identity was confirmed by Sanger sequencing. The AlkB-pet28a plasmid was subsequently transformed into E. coli BL21 (DE3), after which the expressed protein was purified via Ni column affinity chromatography (GE Healthcare, USA) and Sephadex G-75 column purification (GE Healthcare, USA). The target peak fractions were pooled and concentrated to 250 mg/mL using 5-kDa filters (Millipore, USA). Purified proteins were stored in AlkB storage buffer (Tris-HCl (pH 8.0), 0.2 M NaCl, 2 mM DTT, and 50% glycerol). The purity of the AlkB was validated by SDS‒PAGE and Coomassie blue staining.

Quantification of modified nucleosides in RNA molecules by LC‒MS/MS

A total of 2 μg of RNA was incubated with 2 μL of Nuclease P1 (New England Biolabs, USA) or 5 μL of NH4OAc (pH 5.3) in a 50 μL reaction mixture at 50°C for 3 h, followed by phosphodiesterase I (PDE1) (Millipore, USA) digestion at 37°C for 2 h and 85°C for 10 min. After incubation with alkaline phosphatase (CIAP) (TAKARA, Japan) at 37°C for 2 h, the single ribonucleosides mixture was cleaned with Amicon Ultra 0.5 mL filters (Millipore, USA) and HyperSep Hypercarb SPE cartridges (Thermo Fisher, USA). The samples were subjected to LC-MS/MS on a 4600 UPLC/Triple TOF (AB, SCIEX) system. Ribonucleosides were quantified using mass transitions of 258.0 to 126.0 for 3-methylcytidine (m3C).

m1A dot blot

Purified RNA was denatured at 95°C for 3 min, followed by chilling on ice immediately. The RNA was spotted on nylon transfer membranes. After UV crosslinking, the membrane was washed with TBST at room temperature for 5 min and blocked with PBST supplemented with 5% BSA at room temperature for 1 h. The membrane was incubated with an anti-m1A antibody (1:500; 208196; Abcam, USA) overnight at 4°C, followed by incubation with an HRP-conjugated anti-rabbit IgG secondary antibody (1:1000; sc-2357; Santa Cruz, USA) and washing with TBST. The signal was detected with the ECL Western Blotting Detection System (Tanon, China).

Treatment of biofluid RNA with AlkB and T4PNK

The AlkB and T4PNK treatment procedure was as described previously.¹⁸ Briefly, total RNA extracted from biofluids was treated with an equal amount of AlkB enzyme, 2.5 μL of 1 M HEPES (pH 8.0) (Gibco, USA), 0.5 μL of 5 mg/mL BSA, 1 μL of 3.75 mM ferrous ammonium sulfate, 0.5 μL of 100 mM α-ketoglutaric acid (Sigma‒Aldrich, USA), 1 μL of 100 mM L-sodium ascorbate (Sangon Biotech, China), and 2.5 μL of RNase Inhibitor (New England Biolabs, USA), brought to a total volume of 50 μL with DEPC-treated water. After incubating at 37°C for 30 min, the mixture was purified with TRIzol reagent (Ambion, USA). Then, 5 μL of 10× PNK buffer (New England Biolabs, USA), 5 μL of 1 mM ATP (New England Biolabs, USA), and 1 μL of T4PNK (New England Biolabs, USA) were added, and the mixture was incubated at 37°C for 20 min. The mixture was subsequently purified with TRIzol reagent and then subjected to small RNA sequencing.

Small RNA sequencing and quality control

RNA samples were subjected to enzymatic treatment or traditional library construction, and deep sequencing was subsequently performed via BGI Genomics (Shenzhen, China). Sample quality was tested using an Agilent Bioanalyzer 2100 instrument. Standard 50-bp single-read sequencing on an Illumina NextSeq instrument was performed for small RNA sequencing. Library construction was performed according to the following procedure: 1) A 5′-adenylated and 3′-blocked single-stranded DNA adapter was ligated to the 3′ end of the RNA; 2) the UMI-tagged RT primer was hybridized with free or ligated 3’ adapter; 3) 5′ adapter ligation was performed; 4) reverse transcription and extension were conducted by using the UMI-tagged RT primer followed by cDNA amplification by utilizing both 3′ and 5′ adapters; 5) the library was size-selected, and circularization primers were used to form a circular single-stranded PCR product; and 6) the library was sequenced on the instrument. Clean reads were generated by SOAPnuke using the following parameters: “-n 0.01 -L 13 -q 0.1 -p 0.7 --minReadLen 18--maxReadLen 44 -- outQualSys 1 --ada_trim”. Subsequent data analyses were performed on the clean reads.

Small RNA annotation

The small noncoding RNA annotation pipeline was implemented with the SPORTS1.1 package. 18–45 bp sncRNAs were matched to the following independent noncoding RNA databases by SPORTS1.1 with a maximum of one base mismatch (SPORTS1.1 parameter setting-M 1): Small noncoding RNA reads were mapped to the following databases: the miRNA database miRbase 21: http://www.mirbase.org/index.shtml; the rRNA and YRNA databases from the National Center for Biotechnology Information: https://www.ncbi.nlm.nih.gov/nuccore; the tRNA databases GtRNAdb: http://gtrnadb.ucsc.edu/ and MINTmap: https://cm.jefferson.edu/mintmap/; the piRNA databases piRBase and piRNABank: http://www.regulatoryrna.org/database/piRNA/ and http://pirnabank.ibab.ac.in/index.shtml. Noncoding RNAs were annotated by Ensembl and Rfam 12.3: http://www.ensembl.org/index.html; http://rfam.xfam.org/. The unique and shared sncRNAs among specific biofluids were analyzed and visualized by the UpSetR package.⁵⁴ SPORTS1.1 package did not assign a unique name to the ts/rsRNA sequences which were only given information of their precursors such as 28S-rRNA. Due to the lack of acknowledged nomenclature system of tsRNA and rsRNA, we assigned a unique randomized number to distinguish different tsRNAs/rsRNAs from same precursor. Additional reads elimination procedure was performed on saliva samples using Human Oral Microbiome Database (HOMD) as a reference. The microbial genome sequences were downloaded in FASTA format and indexed using Bowtie2 (v2.3.4.1). The reads were aligned to the microbial reference genomes using the following Bowtie2 parameters: -N 0, -x, -U, -S, --un. The unmapped reads (those not aligning to the microbial genomes) were retained for further analysis as described above.

Small RNA length distribution analysis

The output results of SPORTS1.1 were calculated as counts per million (CPM) to standardize the sequencing depth. The ggplot package was subsequently applied, and bar-stacked charts and pie charts were generated to determine the origins of the tsRNAs and rsRNAs, respectively.

Characterization of the loci generating tsRNAs/rsRNAs

Overall, the preferential genomic loci of tsRNA or rsRNA precursors were identified using the original starting loci and sequence lengths generated from the SPORTS1.1 analysis procedure. The CPM values of the sequence fragments were calculated and aligned. In the visualization of the mapping plot, a solid line with a shaded band indicates the mean CPM and standard error of the mean (SEM).

High-performance classifier prediction via random forest analysis

A high-performance serum exRNA diagnostic model for multiple cancers was constructed from the enzymatic treatment RNA-seq derived sncRNA profiles (read counts >10) using the random forest algorithm. Grid search was used to further optimize the hyperparameters of the random forest model to obtain the best hyperparameter combinations. The random forest algorithm is a machine learning algorithm based on decision trees that predicts the value of a response variable. This method effectively reduces the risk of overfitting and enhances the robustness of the model to outliers and noise. Five different samplings were conducted to reduce the influence of the sampling process in the training set on the importance of features. Subsequently, 5-fold cross-validation was performed, and the top 20 most important features were identified. Furthermore, backward feature elimination was used to identify the optimal combination of biomarkers while minimizing the required number of features. Beginning with the initial set of top 20 features, the feature with the smallest average importance value was removed at each iteration. The Seurat package was used to standardize and reduce the dimensionality of the sncRNA profiles for subsequent visualization. We calculated the weighted expression levels of the six prioritized exRNAs based on their importance weights derived from the random forest classifier. The importance weights reflect the contribution of each exRNA to the classification model. The weighted expression levels were computed as follows: Weighted Expression = $\sum _{\mathrm{i}=1}^6($ wi × Expressioni). Where wi is the importance weight of the i-th exRNA, and Expressioni is the normalized expression level of the i-th exRNA. To visualize the distribution of tumor and nontumor samples. Then, the principal component analysis (PCA) on the weighted expression data of the six exRNAs and plotted the first principal component (PC1). Furthermore, we employed the confusion_matrix function in the sklearn.metrics library to generate confusion matrices for both the training and test sets. The ComplexHeatmap package was used to generate a heatmap based on the selected top 20 features.

Primary tumor signature selection by least absolute shrinkage and selection operator (LASSO) regression analysis

Due to the variation in the age and gender distributions of different cancer types, equal numbers of healthy controls were randomly enrolled, and it was ensured that the age and gender distributions were similar between the healthy controls and patients with each cancer type. The LASSO algorithm was used to screen and extract features from the enzymatic treatment RNA-seq derived sncRNA profiles of individual cancer types, resulting in potential primary tumor signatures. To visualize the LASSO screening process, the glmnet function of the glmnet package was used to construct a dynamic process diagram. Additionally, the cv.Glmnet function was used to plot the mean squared error (MSE) curve in relation to Log(λ) for the standardized sncRNA expression profiles. In the present study, the λ value that minimized the MSE was selected as the final λ of the LASSO algorithm. The features with nonzero coefficient values at this λ are considered potential tumor signatures. Subsequently, a heatmap was generated for each cancer type.

sncRNA target prediction and gene ontology enrichment

miRDB⁵⁵: https://mirdb.org/custom.html was used to predict potential target genes (Custom Prediction model) of the sncRNAs selected on the basis of a predictive target score greater than 70. For Gene Ontology (GO) enrichment, gene sets were collected and enriched for GO biological processes by Metascape⁵⁶: https://metascape.org/gp/index.html#/main/step1. The top 10 categories were selected and visualized as bubble plots via the ggplot R package.

Quantification and statistical analysis

All the statistical analyses were performed with SPSS 24.0 statistical software and GraphPad Prism 8.0 (GraphPad, San Diego, CA, USA). Student’s unpaired t-test for 2-group comparison was used. The data are presented as the means ± SEMs. A p-value <0.05 was considered significant.

Published: July 22, 2025