OM-sqPCR enables one-step multiplex quantification of circulating small-RNA subclasses for colorectal cancer classification

Summary

Accurate multiplex quantification of circulating small RNAs (sRNAs) enables comprehensive evaluation of their diagnostic potential. Here, we develop OM-sqPCR, a closed-tube, one-step multiplex RT-qPCR platform that simultaneously quantifies microRNAs, transfer RNA-derived sRNAs (tsRNAs), and ribosomal RNA-derived sRNAs (rsRNAs). By integrating optimized reaction chemistry for short RNA templates with computational primer-probe design, OM-sqPCR enables coordinated amplification of multiple sRNA subclasses in a single reaction. Applied to a multicenter cohort of 651 participants, OM-sqPCR identified a four-sRNA classifier that discriminated colorectal cancer from controls with an area under the ROC curve of 0.97, 85.9% sensitivity, and 93.1% specificity, with sustained performance in early-stage disease. The platform also supported measurement of previously reported sRNA panels for renal and thyroid cancer. These findings establish OM-sqPCR as a practical approach for multiplex circulating sRNA analysis and clinical biomarker evaluation.

Graphical abstract

Highlights

• •OM-sqPCR enables one-step multiplex quantification of diverse small-RNA subclasses
• •Small RNA-optimized chemistry enables one-step RT-qPCR detection of small RNAs
• •Algorithm-guided primer-probe design enables multiplex detection of small-RNA subclasses
• •Multicenter validation supports robust OM-sqPCR-based CRC classification

Health sciences; Medicine; Medical specialty; Internal medicine; Gastroenterology; Oncology

Introduction

Colorectal cancer (CRC) remains a major cause of cancer-related morbidity and mortality worldwide.¹ Early detection is critical for improving prognosis,^2,3,4 yet current screening methods, such as colonoscopy, are invasive, costly, and limited by patient compliance.⁵ Non-invasive alternatives like fecal immunochemical testing (FIT) and serum carcinoembryonic antigen (CEA) testing are more accessible but suffer from limited sensitivity, particularly in early-stage disease.^6,7 These challenges underscore the urgent need for accurate, scalable, and patient-friendly diagnostic tools that can facilitate early intervention.

Liquid biopsy has expanded interest in circulating small RNAs (sRNAs) as candidate cancer biomarkers.^8,9,10,11 Circulating microRNAs (miRNAs) have been studied most extensively because of their relative stability in biofluids and disease-associated expression patterns.^12,13 However, miRNAs represent only 1 component of a broader circulating sRNA repertoire.^14,15 Other subclasses, such as transfer RNA-derived small RNAs (tsRNAs)^16,17,18 and ribosomal RNA-derived small RNAs (rsRNAs),^19,20 are abundant in circulation and may provide complementary diagnostic information. Accordingly, combined measurement of multiple sRNA subclasses may improve assay performance relative to single-class approaches. A major obstacle, however, is that these analytes differ substantially in sequence, abundance, and structural features, complicating their simultaneous quantification in a single assay format.

Current detection platforms, including next-generation sequencing, microarrays, and reverse-transcription quantitative PCR (RT-qPCR), have limited capacity for multiplex sRNA detection. Sequencing and microarray methods offer high throughput but require complex workflows, specialized instrumentation, and substantial cost, limiting their routine clinical use. Conventional RT-qPCR remains the benchmark for sensitivity and specificity but requires separate reverse-transcription reactions for each sRNA, followed by individual qPCR assays, which increases sample consumption, reagent use, and hands-on time. These limitations have hindered the development of efficient, scalable, and clinically accessible sRNA-based diagnostic tests.

To address this analytical gap, we developed one-step multiplex small-RNA quantitative PCR (OM-sqPCR), a closed-tube, one-step multiplex probe-based RT-qPCR platform for unified measurement of circulating miRNAs, tsRNAs, and rsRNAs. The platform combines optimized enzyme-buffer chemistry for short RNA templates with a computational primer-probe design workflow based on thermodynamic and structural scoring. We applied OM-sqPCR to circulating biomarker discovery and validation in CRC and identified a 4-sRNA classifier with strong performance in blinded multicenter validation. We further evaluated the platform in proof-of-concept studies using previously reported circulating sRNA panels for renal cell carcinoma and thyroid carcinoma. These findings support OM-sqPCR as a practical multiplex assay format for circulating small-RNA analysis in translational and clinical laboratory settings.

Results

Identification of circulating small RNAs associated with colorectal cancer

To identify circulating small RNAs associated with CRC, we performed small-RNA sequencing (sRNA-seq) on serum samples from 10 patients with CRC and 10 healthy controls (HCs). After quality control filtering, 216,100 unique sRNA species were detected, including miRNAs (37.4%), tsRNAs (7.2%), and rsRNAs (17.4%). Differential abundance analysis showed broad alterations across these subclasses, with 210 sRNAs increased and 610 decreased in CRC relative to HCs (adjusted p < 0.05; |log₂ fold change| > 1) (Figure 1A).

Figure 1.

Differentially expressed small RNAs in patients with CRC and HCs

(A) Volcano plots showing the differential expression of miRNAs, tsRNAs, and rsRNAs between serum samples from patients with CRC and HCs. Red and blue dots indicate significantly upregulated and downregulated sRNAs, respectively, defined by a subclass-specific Benjamini-Hochberg-adjusted p value (FDR) < 0.05 and |log₂(fold change)| > 1, whereas gray dots indicate nonsignificant transcripts. The four selected candidates are highlighted.

(B) Screening identified 20 sRNAs that met the biomarker-selection criteria. A heatmap of these sRNAs in 10 CRC and 10 HC samples showed separation between the two groups. Expression values were Z score normalized after log10 transformation of transcripts per million values.

(C) Rank plots of the 200 most abundant sRNAs ordered by log2(fold change), with the 20 candidate sRNAs indicated.

(D) The 20 candidates were evaluated using conventional two-step RT-qPCR in an independent screening cohort comprising 30 patients with CRC and 30 HCs. Four sRNAs were significantly increased in CRC: hsa-miR-223-3p (p < 0.0001), tsRNA-Asp-GTC-1 (p < 0.0001), rsRNA-28S-12 (p < 0.0001), and hsa-let-7f-5p (p = 0.0012).

(E) ROC curve analysis of the four selected sRNAs for distinguishing patients with CRC from HCs. The AUCs were 0.8778 (95% CI, 0.7925–0.9631) for hsa-miR-223-3p, 0.8633 (95% CI, 0.7730–0.9536) for tsRNA-Asp-GTC-1, 0.9350 (95% CI, 0.8731–0.9969) for rsRNA-28S-12, and 0.7422 (95% CI, 0.6163–0.8682) for hsa-let-7f-5p. Comparisons in (D) are presented as the mean ± SEM, with each point representing one participant. Comparisons in (D) were performed using two-tailed unpaired student’s t tests. ROC AUCs are reported with their corresponding 95% confidence intervals. Asterisks denote the following significance thresholds: p ≥ 0.05, ns, ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001.

From the sequencing dataset, 20 candidate sRNAs (sequences are shown in Table S3) were selected for further evaluation based on abundance and differential expression, including 2 miRNAs, 2 tsRNAs, and 16 rsRNAs (Figure 1B). All 20 candidates were among the 200 most abundant circulating sRNAs detected in both CRC and HC samples (Figure 1C). These candidates were then examined by two-step RT-qPCR in an independent cohort comprising 30 patients with CRC and 30 HCs. Four sRNAs—hsa-miR-223-3p, tsRNA-Asp-GTC-1, rsRNA-28S-12, and hsa-let-7f-5p—showed significantly higher concentrations in CRC serum samples than in HC samples (Figure 1D). Receiver operating characteristic (ROC) analysis showed that each of these 4 sRNAs individually discriminated CRC from HCs, with areas under the curve of 0.878 for hsa-miR-223-3p, 0.935 for rsRNA-28S-12, 0.863 for tsRNA-Asp-GTC-1, and 0.742 for hsa-let-7f-5p (Figure 1E). These 4 markers were subsequently advanced for classifier construction.

Development of the OM-sqPCR platform

To address the incompatibility between the short length of circulating small RNAs (sRNAs; <50 nt) and the requirements of conventional RT-qPCR workflows, we developed a OM-sqPCR. The assay integrates reverse transcription and PCR amplification within a single closed-tube reaction, thereby reducing manual handling and potential between-tube variability while enabling simultaneous measurement of miRNAs, tsRNAs, rsRNAs, and other short RNA species (Figure 2).

Figure 2.

Overview of the OM-sqPCR platform

The OM-sqPCR platform is an integrated system for multiplex small RNA quantification, enabling simultaneous reverse transcription and real-time quantitative PCR of multiple small RNA markers within a single closed reaction tube. Algorithm-guided optimization of primer-probe combinations markedly improves quantification efficiency, assay reproducibility, and multiplex compatibility.

(A) The OM-sqPCR system employs optimized enzyme cocktails—including reverse transcriptase, hot-start DNA polymerase, and uracil-DNA glycosylase (UDG)—together with precisely tuned buffer compositions and temperature control tailored for short RNA templates.

(B) Algorithms calculate primer-probe parameters, including sequence length, GC content, secondary structure stability, dimerization potential, and annealing temperature, to ensure target-specific amplification while minimizing cross-interference among multiple sRNAs.

(C) The OM-sqPCR workflow comprises two coordinated steps: (i) algorithm-guided design and validation of primer-probe sets, (ii) one-tube integration of reverse transcription and qPCR reactions. The primer-probe sets include four sequences: ① stemloop primer; ② forward primer; ③ probe; ④ reverse primer. Sequence ① enables reverse transcription, while sequences ②–④ facilitate real-time fluorescence quantitative PCR. Compared with conventional two-step RT-qPCR, OM-sqPCR rapidly identifies optimal primer-probe configurations, eliminates iterative optimization, and performs all amplification steps in a single tube—thereby reducing assay time and cost (< US$1 per test) while streamlining workflows and improving data stability. This figure was created using BioRender under agreement no. WQ2A4QH3S7.

Development of OM-sqPCR involved 2 sequential engineering phases. In the first phase, we optimized enzyme-buffer chemistry to establish a one-step reaction compatible with structurally diverse sRNA templates. Candidate formulations incorporating different combinations of reverse transcriptase and hot-start DNA polymerase were evaluated under varying ionic conditions by using a reaction-screening workflow (Figure 2A).

In the second phase, we developed a computational workflow for multiplex primer-probe design. Because reliable multiplex detection requires coordination among primer-probe sets to minimize cross-hybridization and fluorescence-channel interference, the design framework incorporated sRNA sequence features, thermodynamic modeling, and fluorescence-channel assignment (Figure 2B). The resulting workflow was used to identify primer-probe combinations suitable for multiplex analysis and to support assay standardization across targets (Figure 2C). Detailed optimization procedures are described in the STAR Methods. For assay implementation, target sequences were entered into the design software, which generated ranked primer-probe combinations on the basis of the integrated scoring framework. This approach reduced empirical redesign during assay development and supported reproducible multiplex assay construction.

Optimization of one-step reaction chemistry for OM-sqPCR

To establish a one-step reaction system suitable for circulating small-RNA analysis, we optimized assay chemistry for sequential reverse transcription and real-time PCR within a single reaction mixture (Figure 3A). We first evaluated 4 commercially available one-step RT-qPCR kits (TakaRa, Vazyme, Fapon, and Novoprotein) by using 10-fold serial dilutions (0.1–10,000 fM) of synthetic hsa-miR-16-5p and hsa-let-7g-5p as representative circulating small-RNA targets.^21,22 Under manufacturer-recommended conditions, the TakaRa and Vazyme systems did not produce linear amplification across the tested range, whereas the Fapon and Novoprotein systems showed linear responses only at concentrations above 100 fM (Figure 3B). Because circulating small RNAs in serum and plasma are commonly present at low femtomolar concentrations,^23,24,25 these results indicated that existing commercial one-step systems had insufficient analytical sensitivity for the intended application.

Figure 3.

Screening and performance evaluation of one-step reaction chemistry in the OM-sqPCR detection system

(A) Workflow comparison between the one-step OM-sqPCR and conventional two-step RT-qPCR. OM-sqPCR integrates reverse transcription and real-time qPCR within a single closed tube, eliminating the need for intermediate handling and reducing total assay time.

(B) Four commercial RT-qPCR reagents were selected to detect 10-fold serially diluted standards of hsa-miR-16-5p and hsa-let-7g-5p (0.1–10,000 fM).

(C) Screening of optimized one-step RT-qPCR reaction mixtures. Four optimized one-step RT-qPCR reaction mixtures (mix 1–4) were evaluated using 10-fold serial dilutions (0.1–1,000 fM) of four representative circulating small RNA biomarkers—hsa-miR-223-3p, rsRNA-28S-12, tsRNA-Asp-GTC-1, and hsa-let-7f-5p. Amplification performance was assessed by ΔCt values across concentration gradients, where a ΔCt of approximately 3.3 per log₁₀ dilution represents theoretical amplification efficiency.

(D) Representative fluorescence amplification curves comparing one-step and two-step RT-qPCR systems.

(E) Standard curves generated for four sRNA targets using one-step single-channel RT-qPCR, showing strong linearity (R² > 0.99) across six orders of magnitude.

(F) Agarose gel electrophoresis (4%) of PCR products from the one-step single-channel assay. Lane M, 20 bp DNA ladder; Lane 1, hsa-miR-223-3p; Lane 2, rsRNA-28S-12; Lane 3, tsRNA-Asp-GTC-1; Lane 4, hsa-let-7f-5p.

(G) Correlation analysis between OM-sqPCR and traditional two-step RT-qPCR results across matched samples, showing high concordance (Pearson’s r = 0.95). Data are presented as the mean ± SEM in (B), (C), and (E).

We therefore formulated a series of optimized one-step RT-qPCR reaction mixtures based on the reverse transcriptases and hot-start DNA polymerases from the Fapon and Novoprotein systems. By varying reverse transcription primer concentrations, we generated 4 candidate enzyme-buffer formulations (mix 1–4) for further evaluation (Table S2). Synthetic hsa-miR-223-3p, tsRNA-Asp-GTC-1, rsRNA-28S-12, and hsa-let-7f-5p were then used as representative templates to compare amplification efficiency, linearity, and reproducibility across 10-fold serial dilutions ranging from 0.1 to 1,000 fM.

Among the tested formulations, mix 4 showed the best overall analytical performance. This formulation produced near-theoretical amplification behavior, with ΔCt values of approximately 3.3 per 10-fold dilution and linear quantification over the tested concentration range (R² > 0.99; Figures 3C–3E). Agarose gel electrophoresis confirmed the expected amplicon sizes (Figure 3F). Quantitative results obtained with the optimized one-step assay were also strongly correlated with those obtained by conventional two-step RT-qPCR (Pearson r = 0.95; Figure 3G). Relative to the two-step workflow, the optimized one-step assay reduced total assay time from approximately 4 h–2 h while maintaining comparable quantitative performance. These data established the reaction chemistry used for subsequent multiplex assay development and analytical validation.

Algorithm-guided primer-probe design and multiplex assay integration

To support multiplex amplification with minimal fluorescence-channel interference, we developed a computational primer-probe design workflow for OM-sqPCR. The framework incorporated small-RNA sequence characteristics, thermodynamic modeling, and fluorescence-channel assignment to facilitate coordinated assay design under multiplex conditions (Figure 4A).

Figure 4.

Analytical performance of the OM-sqPCR platform

(A) Development of the algorithm-guided assay-design framework. A parameter library for sRNA RT-qPCR was established, encompassing stem-loop motifs, primer-probe length ranges, melting-temperature windows, Gibbs free-energy (ΔG) thresholds, forward-primer 5′-extension lengths, and reverse-primer gap distances. Experimental feedback was used to refine the parameter network and identify high-performing primer–probe combinations. A scoring algorithm then ranked multiplex configurations by predicted amplification efficiency, specificity, and spectral compatibility.

(B) Real-time amplification curves for four representative sRNAs measured by OM-sqPCR across concentrations from 100 aM to 100 pM.

(C) Reproducibility analysis of serum samples with high, medium, and low RNA abundance.

(D) Sanger sequencing confirmed the identities of the four amplicons.

(E) Interoperator and interinstrument reproducibility of rsRNA-28S-12 measurements across 27 replicate measurements (nine measurements for each operator-instrument configuration).

(F) Comparison of OM-sqPCR and conventional two-step RT-qPCR measurements in the same 40 samples. No significant between-method differences were detected for hsa-miR-223-3p (p = 0.9992), tsRNA-Asp-GTC-1 (p > 0.9999), rsRNA-28S-12 (p = 0.9525), or hsa-let-7f-5p (p > 0.9999).

(G) OM-sqPCR measurements in patients with CRC (n = 15) and non-CRC controls (n = 25). Concentrations were higher in CRC for hsa-miR-223-3p (p < 0.0001), tsRNA-Asp-GTC-1 (p = 0.0002), rsRNA-28S-12 (p = 0.0007), and hsa-let-7f-5p (p = 0.0004). The coefficient of variation was calculated as CV (%) = 100 × SD/mean (E). Data in (F) and (G) are presented as the mean ± SEM. Comparisons used two-tailed paired student’s t tests in (F) and two-tailed unpaired student’s t tests in (G). Asterisks denote the following significance thresholds: p ≥ 0.05, ns, ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001.

During the initial design stage, we assembled a parameter set based on established small-RNA RT-qPCR design principles, including stem-loop sequence libraries,^26,27 primer and probe length ranges,²⁸ GC content, melting temperature ranges, predicted Gibbs free energy values for primer-template and primer-dimer interactions,²⁹ 5′ extension length of forward primers, and reverse-primer spacing. These parameters were used to generate candidate configurations for single-channel one-step assays. In the multiplex design stage, target sequences were further evaluated together with stem-loop structures and fluorophore/quencher combinations^30,31 to identify primer-probe sets with favorable predicted amplification performance, specificity, and spectral compatibility. Candidate designs were then ranked by the integrated scoring framework and selected for experimental evaluation (Figure 4A).

Using this workflow, we generated primer-probe sets for the four CRC-associated sRNAs: hsa-miR-223-3p, tsRNA-Asp-GTC-1, rsRNA-28S-12, and hsa-let-7f-5p. In single-channel testing, each assay generated detectable amplification at 100 aM and maintained linear quantification across a 10⁶-fold concentration range (Figure 4B). Assay performance remained stable across serum samples with varying RNA abundances (Figure 4C), and Sanger sequencing confirmed amplicon identity (Figure 4D). Interoperator and interinstrument imprecision was low, with coefficients of variation <10% for both log10-transformed concentrations and Ct values (Figures 4E; Tables S4 and S5). Across the four biomarkers, LOB values ranged from 0.15 to 2.85 fM and LOD values ranged from 0.29 to 4.37 fM (Table S6), confirming that the concentrations used in clinical validation were well within the validated analytical measuring range. Quantification results from OM-sqPCR were comparable to those obtained using conventional two-step RT-qPCR (Figure 4F), and both methods effectively distinguished patients with CRC from HCs (Figure 4G). Collectively, these results validate the utility of the design workflow for multiplex assay integration in the OM-sqPCR platform.

Construction of a four-sRNA diagnostic classifier

Following the establishment of a robust, multiplex-compatible detection platform, we applied OM-sqPCR to evaluate the clinical diagnostic potential of four circulating small RNAs—hsa-miR-223-3p, tsRNA-Asp-GTC-1, rsRNA-28S-12, and hsa-let-7f-5p—in serum samples from CRC patients. A training cohort of 250 participants from Nanjing Drum Tower Hospital was analyzed, including 101 CRC patients (18 stage I, 41 stage II, 41 stage III, and 1 stage IV), 49 colorectal polyp (CRP) patients, 41 individuals with other gastrointestinal (GI) disorders, and 59 HCs.

All four markers were significantly upregulated in CRC relative to both control groups (Figure 5A). Receiver-operating-characteristic (ROC) analyses demonstrated strong individual discriminative performance for distinguishing CRC from HCs, with AUCs of 0.91 (hsa-miR-223-3p), 0.91 (rsRNA-28S-12), 0.84 (tsRNA-Asp-GTC-1), and 0.88 (hsa-let-7f-5p) (Figure 5B). When differentiating CRC from CRP, the four markers showed similarly high accuracy (AUCs = 0.88–0.91; Figure 5C). These results indicate that each of the four sRNAs provides independent diagnostic value and collectively capture non-redundant molecular information from distinct sRNA subclasses.

Figure 5.

Diagnostic performance of the four-sRNA biomarker panel in the training cohort

(A) Serum concentrations of hsa-miR-223-3p, rsRNA-28S-12, tsRNA-Asp-GTC-1, and hsa-let-7f-5p in HC (n = 59), CRP (n = 49), and CRC (n = 101) groups. Tukey-adjusted p values for HC versus CRP, HC versus CRC, and CRP versus CRC, respectively, were 0.8399, < 0.0001, and < 0.0001 for hsa-miR-223-3p; 0.8653, < 0.0001, and < 0.0001 for rsRNA-28S-12; 0.9358, 0.0003, and 0.0002 for tsRNA-Asp-GTC-1; and > 0.9999, < 0.0001, and < 0.0001 for hsa-let-7f-5p.

(B) ROC analysis for distinguishing CRC from HC. AUCs were 0.9114 (95% CI, 0.8656–0.9572) for hsa-miR-223-3p, 0.9142 (95% CI, 0.8674–0.9611) for rsRNA-28S-12, 0.8397 (95% CI, 0.7761–0.9033) for tsRNA-Asp-GTC-1, 0.8856 (95% CI, 0.8334–0.9377) for hsa-let-7f-5p, and 0.9261 (95% CI, 0.8782–0.9739) for the combined panel.

(C) ROC analysis for distinguishing CRC from CRP. AUCs were 0.8917 (95% CI, 0.8417–0.9417), 0.8828 (95% CI, 0.8305–0.9351), 0.9147 (95% CI, 0.8690–0.9604), 0.8909 (95% CI, 0.8411–0.9406), and 0.9447 (95% CI, 0.9109–0.9786), respectively.

(D) ROC analysis of four machine-learning models in the training cohort. AUCs were 0.9659 (95% CI, 0.9432–0.9886) for SVM-Linear, 0.9630 (95% CI, 0.9383–0.9878) for LogReg-L2, 0.9518 (95% CI, 0.9238–0.9797) for LDA-lsqr, and 0.9506 (95% CI, 0.9229–0.9782) for random forest.

(E) SHAP-based ranking of the four sRNA features from the SVM-Linear model.

(F) Confusion matrix for the internal test set (non-CRC, n = 45; CRC, n = 30) at the prespecified classification threshold of 0.4603. The resulting sensitivity was 90.0% and specificity was 95.6%. Data in (A) are presented as the mean ± SEM. Group comparisons in (A) used one-way ANOVA followed by Tukey’s multiple-comparisons test. ROC AUCs are reported with 95% confidence intervals. Asterisks denote the following Tukey-adjusted significance thresholds: adjusted p ≥ 0.05, ns, ∗adjusted p < 0.05, ∗∗adjusted p < 0.01, ∗∗∗adjusted p < 0.001, ∗∗∗∗adjusted p < 0.0001.

To construct and evaluate diagnostic classifiers, a machine learning-assisted modeling framework was developed, incorporating four machine learning models: support vector machine with a linear kernel (SVM-Linear), random forest, linear discriminant analysis (LDA-lsqr), and logistic regression with L2 regularization (LogReg-L2). These models encompass linear and nonlinear learning approaches. The model training workflow was as follows: a stratified sampling approach was applied to a derivation cohort of 250 participants, which was split into a training set (n = 175) and an internal testing set (n = 75) at a 70:30 ratio. Model development was conducted exclusively on the training set using repeated nested cross-validation. Out-of-fold predictions from the outer loops were aggregated across repetitions to obtain unbiased performance estimates. The optimal hyperparameter configurations were selected based on maximizing the mean cross-validated AUC. The Youden index was applied to the aggregated training predictions to determine the optimal threshold, which was then fixed. The final models were retrained on the full training set with the selected hyperparameters and evaluated on the internal testing set and external validation set.

Preliminary pairwise comparisons using DeLong’s test for correlated ROC curves showed that SVM-Linear achieved a significantly higher AUC than LDA-lsqr (0.9659 versus 0.9518; p = 0.0403) and random forest (0.9659 versus 0.9506; p = 0.0179), whereas its performance was comparable to that of LogReg-L2 (0.9659 versus 0.9630; p = 0.2835) (Figure 5D). Across bootstrap resamples, SVM-Linear yielded a slightly lower Brier score (0.0728 versus 0.0748), a higher mean diagnostic odds ratio (223.3 versus 183.0), and a higher mean Youden index (0.8355 versus 0.8275) than LogReg-L2. Because the linear-kernel SVM used only four input features, its computational burden was minimal. Considering its overall discrimination, calibration, robustness, and computational efficiency, SVM-Linear was selected as the final classifier.

The final classification threshold of the SVM-Linear model was set at 0.4603, achieving a cross-validated sensitivity of 87.32% and specificity of 94.23% in the training cohort. Threshold stability was evaluated using 2,000 bootstrap resampling iterations based on training cohort predictions. Within the range of 0.4603 ± 0.1, sensitivity and specificity varied by ≤2%. Moreover, fold-specific thresholds independently derived from 50 outer-fold partitions showed low dispersion (quartile coefficient of dispersion: 0.1188–0.1338), with variations in diagnostic metrics generally below 0.05, supporting the robustness of the selected cutoff across resampling procedures (Figure S3). SHAP analysis identified tsRNA-Asp-GTC-1 and hsa-miR-223-3p as the major contributors to classifier predictions (Figure 5E). To determine whether the four biomarkers provided complementary diagnostic information, pairwise partial information decomposition analysis was performed in the training cohort (n = 175). Although all marker pairs exhibited significant redundant information (0.357–0.380 bits; permutation p = 0.000999), significant synergistic information was also observed (0.110–0.198 bits; permutation p ≤ 0.006993), indicating that the classifier integrates complementary signals from distinct small-RNA subclasses beyond shared information (Table S8).

In the internal test set (n = 75), SVM-Linear correctly classified 90.0% of CRC cases and 95.6% of non-CRC controls at the prespecified threshold of 0.4603 (Figure 5F). Its AUC exceeded those of the four individual sRNAs; DeLong’s test showed a significant difference relative to hsa-let-7f-5p (p < 0.05). Prespecified subgroup analyses evaluated discrimination of CRC from non-CRC according to control type (HC or CRP), sex, and age (< 45, 45–59, or ≥ 60 years). High specificity was retained across these subgroups.

Multicenter clinical validation of the four-sRNA classifier

To independently validate the diagnostic performance of the four-sRNA classifier, we conducted blinded, multicenter assessments across three external clinical cohorts: Nanjing Drum Tower Hospital (NJ), The First People’s Hospital of Lianyungang (LYG), and Zhejiang Cancer Hospital (ZJ). Upon unblinding, the NJ cohort included 154 participants (58 CRC, 44 CRP, 52 HCs); the LYG cohort comprised 136 participants (50 CRC, 24 CRP, 62 HC); and the ZJ cohort comprised 111 participants (48 CRC, 24 CRP, 39 HC).

Across all centers, CRC patients displayed significantly higher classifier-derived prediction probabilities than CRP and HC groups (Figure 6A). Notably, prediction probabilities remained elevated across all CRC stages (I–IV) relative to controls (adjusted p < 0.0001 for all comparisons). Among 78 early-stage cases (I/II), 66 were correctly identified (sensitivity = 84.6%), demonstrating the classifier’s ability to detect early CRC (Figure 6B). When benchmarked against the conventional serum biomarker CEA, the OM-sqPCR classifier consistently outperformed CEA across cohorts, achieving AUCs of 0.970 (NJ), 0.981 (LYG), and 0.943 (ZJ) (combined AUC = 0.968) compared with 0.587, 0.621, and 0.514 for CEA (combined AUC = 0.579) (Figure 6C). The classifier achieved pooled sensitivity of 85.9% (89.7%, 86.0%, 81.3% per site) and pooled specificity of 93.1% (94.8%, 94.2%, 88.9% per site) across centers (Figure 6D).

Figure 6.

Blinded multicenter validation of the SVM-Linear classifier at a prespecified probability threshold of 0.4603

(A) Predicted probabilities for HC, CRP, and CRC groups in three independent validation cohorts—Nanjing (NJ: HC, n = 52; CRP, n = 44; CRC, n = 58), Lianyungang (LYG: HC, n = 62; CRP, n = 24; CRC, n = 50), and Zhejiang (ZJ: HC, n = 39; CRP, n = 24; CRC, n = 48)—and in the combined multicenter cohort. Tukey-adjusted p values for HC versus CRP, HC versus CRC, and CRP versus CRC, respectively, were 0.8796, < 0.0001, and < 0.0001 in NJ; 0.9911, < 0.0001, and < 0.0001 in LYG; 0.9994, < 0.0001, and < 0.0001 in ZJ; and 0.9866, < 0.0001, and < 0.0001 in the combined cohort.

(B) Predicted probabilities in HC (n = 153), CRP (n = 92), and CRC groups stratified by pathological stage (stage I, n = 31; stage II, n = 47; stage III, n = 61; stage IV, n = 17). Each CRC stage differed from HC and CRP (all adjusted p < 0.0001), whereas HC and CRP did not differ (adjusted p > 0.9999).

(C) ROC analysis comparing the sRNA classifier with serum CEA for distinguishing CRC from non-CRC. AUCs for the classifier and CEA, respectively, were 0.9696 (95% CI, 0.9425–0.9968) and 0.5869 (95% CI, 0.4733–0.7006) in NJ; 0.9807 (95% CI, 0.9638–0.9976) and 0.6207 (95% CI, 0.5012–0.7401) in LYG; 0.9431 (95% CI, 0.9034–0.9829) and 0.5137 (95% CI, 0.3972–0.6301) in ZJ; and 0.9685 (95% CI, 0.9539–0.9830) and 0.5786 (95% CI, 0.5127–0.6444) in the combined cohort.

(D) Confusion matrices for each cohort and the combined analysis. Sensitivity and specificity were 89.7% and 94.8% in NJ, 86.0% and 94.2% in LYG, 81.3% and 88.9% in ZJ, and 85.9% and 93.1% in the combined cohort, respectively. Data in (A) and (B) are presented as the mean ± SEM. Comparisons in (A) and (B) used one-way ANOVA followed by Tukey’s multiple-comparisons test. ROC AUCs are reported with 95% confidence intervals. Asterisks denote the following Tukey-adjusted significance thresholds: adjusted p ≥ 0.05, ns; ∗adjusted p < 0.05, ∗∗adjusted p < 0.01, ∗∗∗adjusted p < 0.001, ∗∗∗∗adjusted p < 0.0001.

The SVM-Linear classifier was then evaluated at the fixed threshold of 0.4603 in three clinical centers. DeLong’s tests indicated that its AUC exceeded that of each individual sRNA. Subgroup analyses by sex, age, and control type are reported in Table S9. Performance estimates and their 95% confidence intervals were calculated using 2,000 stratified bootstrap resamples.

To formally assess inter-site heterogeneity, we performed Cochran’s Q test and calculated the I² statistic for AUC, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) across the three external validation cohorts. No statistically significant heterogeneity was detected for any performance metric (all Cochran’s Q test p > 0.05). Corresponding I² values were low (AUC: 0%; sensitivity: 0%; specificity: 8.4%; PPV: 0%; NPV: 28.5%), indicating that the observed differences among sites were compatible with expected sampling variability rather than systematic site-specific effects (Table S10).

To assess whether biomarker integration improved diagnostic performance beyond individual markers, we compared the four-sRNA classifier with each constituent biomarker across the training, internal testing, and multicenter external validation cohorts. The combined classifier consistently achieved higher overall diagnostic performance than individual biomarkers, based on AUC, sensitivity, specificity, and accuracy, demonstrating the added value of integrating multiple small-RNA signals for CRC classification (Table S11).

Technical applicability of OM-sqPCR across disease-associated small-RNA panels

To assess the generalizability of OM-sqPCR beyond CRC, we applied the platform to independent small cohorts of renal cell carcinoma (RCC)^32,33,34,35 and thyroid carcinoma (THCA)³⁶ using previously reported circulating small-RNA biomarker panels (RCC: hsa-miR-122-5p, hsa-miR-155-5p, hsa-miR-210-3p, hsa-miR-21-5p; THCA: hsa-miR-222-3p, hsa-miR-146a-5p, hsa-miR-451a). These experiments were designed to demonstrate the technical adaptability of the OM-sqPCR workflow to different disease-specific small-RNA panels, rather than to establish new diagnostic classifiers. Absolute quantification with standard curves was used for RCC detection, while relative quantification with U6 as internal reference was employed for THCA—the choice of quantification strategy was determined by the original published biomarker studies. OM-sqPCR successfully quantified all target molecules within a single-tube multiplex reaction without re-optimization of chemistry or thermal parameters. Analysis based on the current validation data revealed that all four markers were upregulated in RCC (hsa-miR-122-5p, hsa-miR-155-5p, hsa-miR-210-3p, hsa-miR-21-5p). In papillary THCA, hsa-miR-146a-5p was downregulated, while the remaining markers (hsa-miR-222-3p, hsa-miR-146a-5p, hsa-miR-451a) were consistently upregulated. Diagnostic classifiers developed from these markers exhibited significant differences between tumor and healthy groups (∗∗∗p < 0.001, Figures 7A and 7B) with combined diagnostic AUC values of 0.994 for RCC and 0.982 for THCA (Figures 7C and 7D), comparable to or superior to those achieved via conventional two-step RT-qPCR. These findings demonstrate that OM-sqPCR can accommodate diverse circulating small-RNA biomarker panels and support both absolute and relative quantification within a unified multiplex assay framework. However, given the limited sample sizes and exploratory nature of these analyses, the clinical utility of OM-sqPCR for other cancer types will require further validation in larger, disease-specific cohorts.

Figure 7.

Scalability validation of OM-sqPCR in RCC and THCA

(A) Absolute quantification of four sRNAs in serum from HCs; (n = 40) and patients with RCC (n = 40). All four markers were increased in RCC: hsa-miR-122-5p (p < 0.0001), hsa-miR-155-5p (p < 0.0001), hsa-miR-210-3p (p < 0.0001), and hsa-miR-21-5p (p < 0.0001).

(B) Relative quantification of three sRNAs in serum from HC (n = 40) and patients with THCA (n = 40), normalized to U6. hsa-miR-222-3p and hsa-miR-451a were increased in THCA (both p < 0.0001), whereas hsa-miR-146a-5p was decreased (p = 0.0002).

(C) ROC analysis for distinguishing RCC from HC. AUCs were 0.9938 (95% CI, 0.9830–1.0000) for hsa-miR-210-3p, 0.9594 (95% CI, 0.9171–1.0000) for hsa-miR-21-5p, 0.9550 (95% CI, 0.9135–0.9965) for hsa-miR-122-5p, 0.7978 (95% CI, 0.6922–0.9035) for hsa-miR-155-5p, and 0.9938 (95% CI, 0.9835–1.0000) for the combined panel.

(D) ROC analysis for distinguishing THCA from HC. AUCs were 0.9791 (95% CI, 0.9490–1.0000) for hsa-miR-222-3p, 0.9297 (95% CI, 0.8718–0.9875) for hsa-miR-451a, 0.7169 (95% CI, 0.6058–0.8280) for hsa-miR-146a-5p, and 0.9819 (95% CI, 0.9554–1.0000) for the combined panel. Data in (A) and (B) are presented as the mean ± SEM. Between-group comparisons used two-tailed unpaired student’s t tests. The combined RCC and THCA panels were fitted separately using L2-regularized binary logistic regression (scikit-learn LogisticRegression). Features were Z score standardized before fitting (C = 1.0; solver = “lbfgs”; maximum iterations = 1,000), and positive-class probabilities were used for ROC analysis. ROC AUCs are reported with 95% confidence intervals. Asterisks denote the following significance thresholds: p ≥ 0.05, ns, ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001.

Discussion

Multiplex measurement of circulating small RNAs remains a major analytical challenge in liquid-biopsy assay development. Although next-generation sequencing supports broad biomarker discovery and conventional two-step RT-qPCR provides sensitive quantification, neither approach is ideally suited to routine multiplex analysis across structurally distinct small-RNA subclasses in clinical laboratory settings.^14,37,38 In this study, we developed OM-sqPCR, a closed-tube, one-step multiplex RT-qPCR platform for unified measurement of circulating miRNAs, tsRNAs, and rsRNAs. By combining optimized reaction chemistry with a computational primer-probe design workflow, the platform was designed to address a central barrier in circulating small-RNA analysis: reliable co-quantification of targets with different sequence and structural properties.

From an assay-development perspective, OM-sqPCR provides several practical advantages. Integration of reverse transcription and PCR amplification into a single reaction reduces handling steps and may reduce variability associated with multistep workflows. In the present study, the assay showed attomolar analytical sensitivity, a broad dynamic range, and low interoperator and interinstrument imprecision. The computational design workflow further supported multiplex implementation by incorporating sequence features, thermodynamic constraints, and fluorescence-channel assignment during primer-probe selection. Together, these features enabled construction of multiplex assays for short RNA targets within a standardized one-step format.

Using this platform, we identified and validated a 4-sRNA classifier for CRC composed of hsa-miR-223-3p, tsRNA-Asp-GTC-1, rsRNA-28S-12, and hsa-let-7f-5p. In blinded multicenter validation, the classifier achieved a pooled area under the curve of 0.968, with 85.9% sensitivity and 93.1% specificity, and outperformed serum CEA. Sensitivity for stage I-II CRC was 84.6%, supporting the potential utility of this assay format for early disease detection. These findings indicate that OM-sqPCR can support both multiplex assay construction and clinically relevant classification of CRC in independent cohorts.

An additional advantage of OM-sqPCR is its technical adaptability. Once a small-RNA panel is defined, the platform can accommodate new disease-specific panels without re-optimization of reaction chemistry or thermal cycling; only the primer-probe sets require replacement via the algorithm-guided design pipeline. As proof-of-concept, we applied OM-sqPCR to previously reported panels for RCC and THCA and observed maintained classification performance. Although these small-sample validations demonstrate technical feasibility rather than full clinical utility, they confirm that OM-sqPCR is directly adaptable to other small-RNA panels, highlighting its broad applicability as a multiplex quantification platform.

Compared with several other emerging multiplex nucleic acid detection approaches, OM-sqPCR is intended to balance analytical performance with practical implementation. The assay was compatible with standard qPCR instrumentation, required a relatively small serum input volume, and reduced total workflow time compared with conventional two-step RT-qPCR. These characteristics may facilitate adoption in laboratories that already use real-time PCR platforms. In this respect, the potential value of OM-sqPCR lies not only in biomarker measurement performance, but also in its compatibility with existing laboratory workflows.

Beyond their diagnostic utility, circulating small RNAs have been implicated in disease-associated biological processes. Recent studies suggest that tsRNAs and rsRNAs may regulate innate immune signaling through structure-dependent mechanisms and Toll-like receptor pathways.^39,40,41 A study on inflammatory bowel disease identified coordinated alterations across multiple small RNA subtypes.⁴² These findings highlight the potential value of integrating multiple small-RNA subclasses to capture complementary disease-associated molecular signals.

In conclusion, OM-sqPCR provides a one-step multiplex RT-qPCR platform for unified measurement of multiple circulating small-RNA subclasses. In this study, the assay showed robust analytical performance and supported blinded multicenter classification of colorectal cancer. With further prospective and external validation, this approach may facilitate broader clinical laboratory implementation of circulating small-RNA biomarkers.

Limitations of the study

This study has several limitations. First, the small-RNA sequencing workflow used for biomarker discovery employed conventional library preparation without enzymatic pretreatment. Emerging approaches such as PANDORA-seq⁴³ and CPA-seq⁴⁴ may enable more comprehensive biomarker identification in future investigations. Second, the clinical cohorts were derived from Chinese centers, and broader validation in geographically and ethnically diverse populations will be needed to assess generalizability. Third, the present study used a cross-sectional design; prospective studies will be required to define performance in intended-use settings and to compare the assay directly with established noninvasive screening strategies. Fourth, although the computational design workflow supported assay development, its comparative performance against alternative primer-probe design strategies was not evaluated systematically in this study. Fifth, CEA was used as a reference biomarker in this study. However, CEA is primarily indicated for postoperative surveillance rather than CRC screening, whereas FIT represents a clinically relevant noninvasive screening comparator. FIT was not formally evaluated because of insufficient paired FIT data in the HC cohort (Table S13). Future prospective studies incorporating paired FIT and small-RNA classifier measurements in screening populations are warranted to enable direct head-to-head performance comparisons. As a minimally invasive blood-based assay, the small-RNA classifier may represent a potential complementary screening approach, particularly for individuals who are unwilling or unable to undergo stool-based testing; however, its clinical utility within routine screening workflows requires further prospective validation. Finally, the proof-of-concept analyses for renal cell carcinoma and thyroid cancer were limited by small sample sizes and the use of a shared healthy-control group. Although sex distributions were comparable, the ccRCC cohort was older than the HCs, and potential age-related confounding cannot be fully excluded. These findings should therefore be interpreted as preliminary technical validation of the broader applicability of OM-sqPCR rather than disease-specific clinical validation.

Resource availability

Lead contact

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

Materials availability

The study did not generate any unique reagents.

Data and code availability

• •All de-identified source data underlying the figures and tables are publicly available in Mendeley Data at https://doi.org/10.17632/t96tgd58pf.2. The raw sRNA-sequencing data generated in this study are deposited in the Gene Expression Omnibus under accession GSE342306.
• •Source code, computational workflows, and the trained classifier implementation are publicly available at https://github.com/RaceyYe/OM-sqPCR. Repository identifiers and accession information are also listed in the key resources table.
• •Any additional information required to reanalyze the data reported in this work is available from the lead contact upon reasonable request.

Acknowledgments

This work was supported by the National Natural Science Foundation of China (no. 32370629, 32000549, and 82003024), Training Program of the Major Research Plan of the National Natural Science Foundation of China (no.92049109), and Jiangsu Provincial Key R&D Program – Social Development Project (no. BE20230804).

Author contributions

Y.W. conceived and designed the research. D.G., X.L., X. Zhang, and R.Y. developed the OM-sqPCR framework. D.G. and X. Zhang performed experiments for OM-sqPCR model training, assisted by S. Zhou, G.L., X. Zhu; and S. Zhan, P.Y., X.W., L.Y., and X.B. collected human plasma samples and clinicopathological data. D.G., X. Zhang, and Q.Z. conducted detection assays on the training cohort. S.G., Q.Z., and Y.X. performed multicenter blind detection. S. Zhou and Q.Z. validated OM-sqPCR on other disease models. R.Y. and R.X. performed computational analyses. Y.W. drafted the manuscript with contributions from C.-Y.Z., L.L., D.G., and X.L., incorporating feedback from all authors.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Critical commercial assays
SuperM-MuLV Reverse Transcripyase	Fapon	MD028
HS Taq DNA Polymerase Ⅲ (UDG)	Novoprotein	E285-03
dN(U)TP	Novoprotein,	E110
RNase Inhibitor	Fapon	MD015
RNA Simple Total RNA Kit	Tiangen Biotech	DP419
VAHTS Small RNA Library Prep Kit	Vazyme	NR811
Deposited data
De-identified source data underlying figures and tables	Mendeley Data	https://doi.org/10.17632/t96tgd58pf.2
Raw small-RNA sequencing data	Gene Expression Omnibus	GSE342306
Software and algorithms
OM-sqPCR source code and trained classifier	GitHub	https://github.com/RaceyYe/OM-sqPCR
Computational workflows	GitHub	https://github.com/RaceyYe/OM-sqPCR
BioRender	BioRender	RRID: SCR_018361
GraphPad Prism 10.1.2	GraphPad Software	RRID: SCR_002798

Experimental model and study participant details

Human participants and clinical sample collection

In the colorectal cancer study, 731 participants aged 17–88 years were enrolled at three hospitals in China under a standardized recruitment and sampling protocol. Participants were allocated to four independent cohorts according to the study design: (i) a small-RNA sequencing cohort (n = 20), comprising 10 patients with colorectal cancer (CRC) and 10 healthy controls (HCs); (ii) a screening cohort (n = 60), comprising 30 patients with CRC and 30 HCs; (iii) a machine-learning cohort (n = 250), comprising 101 patients with CRC, 49 participants with non-advanced colorectal adenomatous polyps (CRP), 41 patients with other gastrointestinal diseases, and 59 HCs; and (iv) a blinded multicenter validation cohort (n = 401), recruited at Nanjing Drum Tower Hospital (n = 154), The First People’s Hospital of Lianyungang (n = 136), and Zhejiang Cancer Hospital (n = 111). All centers followed a uniform protocol for participant recruitment, serum collection, and sRNA processing to support intersite comparability. Demographic characteristics are provided in Table S1. Participants were 17–88 years old, with no significant differences in age or sex distribution among the relevant comparison groups in the CRC cohorts (all p > 0.05).

Eligibility criteria were predefined to ensure cohort homogeneity. CRC cases met the following inclusion criteria: (i) primary diagnosis of CRC confirmed by histopathological examination; (ii) absence of other colorectal lesions or extra-colorectal malignancies; and (iii) no prior treatment, including surgery, chemotherapy, or radiotherapy. HCs were recruited from the Physical Examination Center and were required to have (i) no abnormalities on colonoscopy; (ii) normal findings in routine blood, biochemical, and tumor-marker tests; and (iii) no evidence of intestinal or systemic disease based on imaging and histological examinations. Participants with non-advanced adenomatous polyps (CRP) were confirmed by colonoscopy and pathological examination, with concurrent intestinal disorders excluded. All CRC cases were staged according to the eighth edition of the American Joint Committee on Cancer (AJCC) classification. The 41-patient gastrointestinal disease interference group in the machine-learning cohort comprised liver hepatocellular carcinoma (n = 10), hepatitis B (n = 10), stomach adenocarcinoma (n = 10), and chronic atrophic gastritis (n = 11). All participants in the interference group had clinically confirmed diagnoses based on the corresponding standard diagnostic procedures and medical records at the recruiting institution.

For proof-of-concept cross-disease validation of the technical scalability of OM-sqPCR, an additional 120 participants were recruited at Jiangsu Cancer Hospital between 2024 and 2025, comprising 40 patients with clear cell renal cell carcinoma (ccRCC), 40 patients with thyroid carcinoma (THCA), and 40 HCs. The same HC group was used as the reference group for both disease cohorts because these experiments were designed to assess the technical adaptability of OM-sqPCR to previously reported disease-associated sRNA panels rather than to establish new disease-specific diagnostic classifiers. Demographic characteristics are provided in Table S12. Sex distributions were comparable between the disease and HC groups; however, patients with ccRCC were older than HCs. All patients with ccRCC or THCA were staged according to the eighth edition of the AJCC classification.

For serum preparation, 2 mL of peripheral blood was collected from each participant and allowed to stand at room temperature for 30–60 min, followed by centrifugation at 1,500 × g for 15 min. The supernatant was collected, aliquoted into 300-μL portions, and stored at −80°C without repeated freeze–thaw cycles. Serum samples were visually inspected, and only non-hemolyzed and non-lipemic specimens were included in subsequent experiments.

The study protocols were reviewed and approved by the institutional ethics committees of all participating hospitals before participant enrollment and sample collection: Nanjing Drum Tower Hospital (Approval No. 2022-461-02), The First People’s Hospital of Lianyungang (Approval No. 2020-178), Zhejiang Cancer Hospital (Approval No. IRB-2025-1263 [IIT]), and Jiangsu Cancer Hospital (Approval No. KY-2024-065). Written informed consent was obtained from all participants before sample collection. All procedures involving human participants were conducted in accordance with the Declaration of Helsinki and under the oversight of the respective institutional ethics committees.

Method details

Establishment of the OM-sqPCR detection platform

To enable scalable and clinically practical quantification of circulating small RNAs (sRNAs), we developed a one-step multiplex sRNA RT-qPCR (OM-sqPCR) assay that simultaneously detects miRNAs, tsRNAs, and rsRNAs in a single reaction tube. Development proceeded in two phases: (i) optimization of the one-step reaction chemistry and (ii) implementation of an algorithm-assisted framework for multiplex assay design.

In the first phase, the conventional two-step RT-qPCR workflow was replaced by a single-tube amplification system that performs reverse transcription and quantitative PCR within the same reaction mixture. This integrated design reduced reagent handling and instrument runs from two to one, eliminated cross-contamination risk, and shortened the total assay time from ∼4 h to 2 h. A standard one-step reaction contains two components: an enzyme–buffer mix (reverse transcriptase, hot-start DNA polymerase, reaction buffer, and dNTPs) and target-specific primers with fluorescent probes. To accommodate the unique structural and thermodynamic properties of short sRNAs (< 50 nt), a customized enzyme–buffer formulation was optimized for efficient reverse transcription and amplification of short RNA templates.

The second phase aimed to extend this single-tube strategy to multiplex detection across multiple fluorescent channels. Conventional small RNA assays rely on two-step RT-qPCR using stem-loop RT primers that convert ∼22nt sRNAs into cDNA molecules > 70 nt,^45,46 allowing primer and probe annealing during qPCR. Forward and reverse primers are designed with melting temperatures around 60 °C, and hydrolysis probes are labeled with 5′ fluorophores and 3′ quenchers for real-time signal acquisition.⁴⁷ Theoretically, multiplex detection is achievable by combining distinct RT primers, amplification primers, and probes labeled with spectrally separable fluorophores. However, developing a true one-tube, multi-channel, one-step RT-qPCR system poses substantial challenges: (i) stem-loop RT primers must retain high target specificity and folding stability; (ii) PCR primers must avoid cross-reactivity while maintaining matched GC content and Tm values for synchronous amplification; and (iii) fluorophore-quencher pairs must be spectrally compatible to minimize crosstalk and maximize signal-to-noise ratios.

To address these challenges, we established an algorithm-assisted, end-to-end optimization framework. For stem-loop RT-primer generation, sRNA sequence features, predicted secondary structures, and thermodynamic parameters were integrated into a machine-learning model trained on empirical efficiency and specificity data. Candidate primers were scored automatically for binding affinity, structural stability, and off-target potential, and primer sets with minimal mutual interference were selected for multiplex compatibility. Amplification primers were designed using a multi-objective genetic algorithm that simultaneously optimized GC content, melting temperature, primer length, and dimerization potential to generate primer combinations with synchronized amplification kinetics.

For probe design, an algorithm-guided fluorescence assignment module evaluated fluorophore–quencher pairs based on spectral overlap, signal-to-noise ratio, and inter-channel compatibility. A Bayesian scoring model trained on historical assay data was used to select fluorophore combinations that minimize spectral bleed-through while maximizing detection resolution. All design modules were implemented in a unified software interface that automates the complete workflow—from sRNA input sequence to optimized RT primers, amplification primers, and probe configurations.

Evaluation and optimization of one-step RT-PCR reagents

To identify a reaction chemistry compatible with short-RNA amplification, four commercially available one-step RT-PCR kits—originally designed for long RNA or DNA templates—were evaluated following manufacturer protocols:

• (i)One Step PrimeScript™ RT-PCR Kit (TaKaRa, Japan, Cat. no. RC801S),
• (ii)one-step RT-PCR Kit (Novoprotein, China, Cat. no. E295),
• (iii)HiScript II U+ One Step qRT-PCR Probe Kit (Vazyme, China, Cat. no. P612-01),
• (iv)HotStart HiTag & Super M-MuLV one-step qRT-PCR Mix (Fapon, China, Cat. no. MD027).

Primers and probes were synthesized by General Biol Co. (Chuzhou, China). Among the tested systems, those from Novoprotein and Fapon generated correct amplification curves for hsa-miR-16-5p and hsa-let-7g-5p standards at concentrations of 10² fM to 10⁴ fM, respectively. To further optimize a formulation suitable for small-RNA amplification, cross-combinations of reverse transcriptase and DNA polymerases from the two systems were tested, along with systematic variation of primer and probe concentrations. Four optimized reaction formulations (Mix 1 – 4) were established based on amplification efficiency and signal linearity. The compositions of Mix 1 – 4 are described in Table S2.

The reaction protocol consisted of incubation at 20°C for 5 min, 50°C for 15 min, and 95°C for 5 min, followed by 50 cycles of 94°C for 10 s and 55°C for 30 s. The amplification threshold was established within the exponential phase, and the baseline was automatically determined to yield the Ct values for each amplification curve.

Algorithm-guided primer and probe design software

A cross-platform software tool was developed to automate primer and probe design for the OM-sqPCR platform. The program, written in Python and released under the MIT license, is publicly available through the GitHub repository for academic and non-commercial use. The source code can be downloaded and executed locally, enabling full reproducibility of the algorithmic workflow.

The software supports automated design of primers and probes for up to four small-RNA sequences per batch, generating mutually compatible sets that can be co-amplified within a single qPCR reaction. For more than four targets, sequences can be processed in multiple batches. Cross-batch combinations are not recommended for simultaneous detection due to differences in reaction optimization.

Input sequences are uploaded in FASTA format, and optional custom stem-loop sequences can be specified. Parameter settings can be adjusted to balance sequence-specific customization with universal assay design rules. The pipeline integrates Primer3-py for secondary-structure prediction and applies the Premier Biosoft melting-temperature formula to evaluate thermodynamic compatibility. A scoring system derived from experimentally optimized parameters ranks candidate combinations by predicted amplification efficiency, structural stability, and spectral separation.

The program automatically outputs the top three primer–probe–vector sets for each small-RNA target based on composite performance scores. Under standard experimental conditions, the highest-ranked combination is recommended for primary use, providing an accurate and reproducible starting point for one-step multiplex assay development.

Quantification and statistical analysis

Study cohorts and statistical design

The screening cohort was designed to identify differentially expressed sRNA combinations from sequencing-verified profiles using conventional two-step RT-qPCR. This cohort comprised 60 samples (30 CRC and 30 HC samples). OM-sqPCR measurements from the training cohort were used to develop the machine-learning models. The minimum required training-cohort size was estimated as 146, and 250 participants were enrolled to improve statistical precision. The sample size for the multicenter validation cohort was calculated using the following formula:

$$N=\frac{{\left[Z_{1-\alpha /2}\right]}^2\ast P\left(1-P\right)}{{\Delta }^2}$$

The calculation assumed Δ = 0.05, Z1−α/2 = 1.96, and p = 0.90, corresponding to an expected sensitivity of 90% in the positive group and an expected specificity of 90% in the negative-control group.

This calculation yielded a target of 276 evaluable participants. Allowing for 10% attrition, the minimum recruitment target was 304 participants; the study exceeded this requirement by enrolling 401 participants.

Analytical methods for RNA sequencing

Reads that failed the quality-control criteria (quality score < 25 or length < 18 nucleotides) were excluded. Clean reads were mapped sequentially to the following noncoding-RNA databases: (i) miRBase v22 for miRNAs, (ii) GtRNAdb for human genome build hg19 tRNAs, (iii) NCBI rRNA databases, and (iv) piRBase v3.0 for piRNAs. sRNAs were annotated using these resources, and miRDeep2 was used to identify precursor–mature miRNA matches. tsRNA and rsRNA sequences were annotated using SPORTS1.1, an updated version of SPORTS1.0 that uses Bowtie. Candidate sRNAs were required to have a Benjamini–Hochberg-adjusted p value (FDR) < 0.05, a mean abundance > 1,000 TPM in CRC samples, and higher expression in CRC than in HC samples.

Comparative analysis between groups in PCR validation

To enhance sample quantification precision, absolute quantification was performed using calibration curves generated via 10-fold serial dilution. Log10-transformed concentrations and Ct values were plotted on the X-axis and Y-axis respectively to establish linear correlation. Calibration curve validity was determined by evaluating amplification efficiency and regression coefficients, with validity criteria set as amplification efficiency within 100% ± 10% and regression coefficient R² ≥ 0.99.⁴³ Sample Ct values were substituted into the standard curve to calculate biomarker concentrations, and amplification efficiency was derived using the relevant formula.

$$\text{Amplification}\text{Efficiency}={10}^{\wedge }\left(-1/\text{slope}\right)-1$$

ROC curves were generated to evaluate the diagnostic performance of the sRNAs.

Construction of machine learning models

The machine-learning workflow comprised model-development and testing phases. The derivation cohort was divided by stratified random sampling into a training set (70%) and an internal test set (30%). Model development, including hyperparameter optimization and model selection, was conducted exclusively within the training set.

To obtain less biased estimates of generalization performance, nested cross-validation was repeated 10 times, with a fivefold outer loop and a threefold inner loop. The outer loop was used to estimate model performance, whereas the inner loop was used for Bayesian hyperparameter optimization. Out-of-fold predictions from the outer folds were averaged across repetitions. Within each inner cross-validation loop, the hyperparameter configuration that maximized the AUC was selected. The optimal decision threshold was defined by the Youden index using the aggregated out-of-fold training predictions; this fixed threshold was then applied to the internal test set and external validation cohorts. Final models were refitted to the complete training set using the selected hyperparameters.

Given the limited sample size relative to the four features, candidate models were selected to reduce overfitting and included a support vector machine with a linear kernel (SVM-Linear), random forest, linear discriminant analysis with the least-squares solver (LDA-lsqr), and logistic regression with L2 regularization (LogReg-L2). These models represent linear, kernel-based, and nonlinear learning approaches. Model performance was compared using DeLong’s test for pairwise comparisons of correlated AUCs and McNemar’s test for paired differences in classification outcomes at the Youden-index thresholds.

After model selection, the final model was refitted to the entire training set using the selected hyperparameters. Performance was then evaluated in the held-out internal test set and three independent external validation cohorts. Discrimination was quantified by the area under the ROC curve (AUC); 95% confidence intervals were estimated using 2,000 stratified bootstrap resamples.

All preprocessing steps were embedded within scikit-learn pipelines to prevent information leakage across cross-validation folds. Feature attribution was performed using SHAP (v0.48.0) with LinearExplainer.

Models were implemented in scikit-learn v1.7.0 (Python v3.12.11), and hyperparameters were optimized using BayesSearchCV from scikit-optimize v0.10.2. The hyperparameters optimized for each model are listed in Table S7.

Statistical analysis

Statistical analyses of individual experiments were performed using GraphPad Prism v10.0 (GraphPad Software, USA) and Microsoft Excel (Microsoft Corporation, USA). Unless otherwise specified, data are presented as the mean ± SEM, and each point represents one biological replicate. Comparisons between two independent groups were performed using two-tailed unpaired Student’s t-tests. Matched measurements obtained using OM-sqPCR and conventional two-step RT-qPCR were compared using two-tailed paired Student’s t-tests. Comparisons among three or more groups were performed using one-way ANOVA followed by Tukey’s multiple-comparisons test, with multiplicity-adjusted p values reported. Age was compared using a two-tailed unpaired Student’s t-test for two groups and one-way ANOVA for three or more groups; sex distributions were compared using Pearson’s chi-square test. ROC performance is reported as the AUC with its 95% confidence interval. Statistical methods for machine-learning evaluation and model comparison are described above. Sample sizes (n) and biological-replicate definitions are provided in the corresponding figure legends. Exact p values are reported when available; values below the software reporting limit are reported as p < 0.0001. Statistical significance was defined as p < 0.05. In figures, p ≥ 0.05, ns; ∗p < 0.05; ∗∗p < 0.01; ∗∗∗p < 0.001; ∗∗∗∗p < 0.0001.