Single-Molecule Nanopore and Deep Learning–Based Platform for Probe-Free Multiplex RPA Detection

Introduction: : Recombinase polymerase amplification (RPA) is a rapid, sensitive, isothermal method for nucleic-acid amplification that has become popular in diagnostic applications. Compared with PCR and LAMP, RPA consumes less energy and offers greater freedom in primer design. However, extending RPA beyond single-target assays toward multiplex detection—crucial for differential diagnosis and incorporation of internal controls—remains technically challenging. Existing multiplex RPA schemes hinge on proprietary probes such as Exo, Fpg, or Nfo; these require intricate design, risk cross-reactivity, and normally rely on sophisticated optical instrumentation, limiting scalability. Probe-free solid-state nanopore (SSN) sensors present an attractive alternative by enabling single-molecule detection with digital counting and length discrimination, yet direct identification of short RPA amplicons is difficult. Recent advances in machine-learning-driven nanopore signal analysis mitigate this limitation. Here we introduce a deep-neural-network (DNN)-assisted platform that classifies multiplex RPA products through a probe-free, length-encoded electronic readout from an SSN sensor. Using Mpox as a model, we validated a duplex assay that simultaneously detects the Mpox target gene and the human RNase P gene as an internal control. Systematic tuning of nanopore diameter balanced high detection sensitivity with minimal clogging, and single-molecule translocation events were leveraged to train and iteratively refine the network. The resulting model achieved high accuracy at the individual-event level, enabling reliable population-scale identification. Although demonstrated for duplex detection, the architecture is inherently scalable, offering a general framework for multiplex RPA that eliminates complex probe chemistry and optics while advancing robust, portable molecular diagnostics for field-deployable, low-resource environments worldwide and clinics.

Materials and

Methods: : Multiplex assay. A 25 µL RPA reaction included 1.2 µL each of forward and reverse primers at ~500 nM final concentrations, 12.5 µL 2x-buffer, 2.5 µL 10x-E-mix, 1 µL dNTPs (400 µM), 1.25 µL core mix, 1 µL each of Mpox and RNase P targets, and 1.25 µL MgAcO (280 mM). The mix was incubated at 39 °C for 20 minutes and analyzed on 5% agarose gel-electrophoresis.
Nanopore setup. Quartz capillaries were Piranha-cleaned, laser-pulled, filled with KCl or LiCl/TE, and verified by I–V sweeps. Ionic currents were recorded with an Axopatch 200B at 100 kHz using a 6363 DAQ. Amplicons were diluted 1:20, and five voltage sweeps cleared clogs. Events with SNR > 3.5, dwell < 10 ms, and interarrival-time ≥200 ms spacing were extracted in Python.
DNN model. Nanopore data were collected for 2–5 min/sample using ~7 ± 1.6 nm pores, targeting 100–300 events. Events were extracted, labeled, and split into 2,450 events from 30 samples (60/20/20 %). A five-layer ReLU network (512→32) was trained using AdamW and a learning rate scheduler.
Event caller. The proportion P(T) was calculated by misclassifications, Perror and correct classification (1− Perror). Expected correct events were N1.(1− Perror) for T1 and N2.(1 −  Perror) for T2, while misclassifications were N1.Perror + N2.Perror. In pure samples, the dominant target nears one, the non-target zero. Corrected P(T) were computed as: P(T1)=N1.(1− Perror)/Nt and P(T2)=N2.(1− Perror)/Nt. A sample was classified as T1 if P(T1)>Perror and P(T2)< Perror, T2 if the reverse, and T1+T2 when both proportions exceeded Perror.

Results, Conclusions, and Discussions:: Test workflow: To enable rapid, probe-free discrimination of multiplex RPA amplicons, samples are amplified in a single-tube reaction, driven through sub-10 nm glass nanopores, and their I–t traces recorded. A pretrained DNN using five event-level features, IB, IRMS, ΔI, Δt, and inter-peak interval t, extracted from T1 and T2 samples (Figure 1a) processes each trace online. The Event-Proportion-Classification-framework then computes P(T) for every target and labels the sample as T1, T2, or multiplex T1 + T2.
Assay validation. To establish a reliable multiplex RPA assay for nanopore detection, it is essential to generate distinct amplicons for each target without cross-reactivity or background artifacts. Figure 1b shows the Mpox F3L gene (T1, 110 bp) and the human RNase P gene (T2, 309 bp) together with their primer-encoded length design, ensuring the products can be readily distinguished in a single-tube reaction. Figure 1c presents the validation gel: lanes 1–2 and 3–4 confirm clean singleplex amplification of T1 and T2, respectively; lanes 5–6 (T1 + T2) display both bands, and lanes 7–8 (NTC) remain blank, demonstrating successful multiplexing and assay readiness for nanopore application.
Nanopore sensitivity. Accurate detection requires pores suited to the full 80–500 bp size range. Representative translocations (Figure 2a) and ΔI/Δt distributions (Figures 2b–d) reveal partial signal overlap. Event-frequency mapping versus pore diameter and DNA length (Figure 2e) shows that ~7 nm pores simultaneously avoid clogging seen with < 5.2 nm apertures and the loss of short DNA above 8.6 nm, providing optimal resolution, high event throughput.
Blind Sample Classification Using DNN. The network (Figure 3a) contains five fully connected layers (512→32) with ReLU and SoftMax. Tested on 492 events (20% events) from 35 samples measured on 19 nanopores, it achieved 94.2 % accuracy with a 5.8 % Perror (Figures 3b–c). Using P(T) analysis (Figure 3d), it correctly called 11 T1, 9 T2, and 15 T1 + T2 samples, giving 100 % sample-level accuracy. ROC curves showed AUC = 1.00 for both targets (Figures 3e–f), and reproducibility across nanopores N1–N3 had %CV < 6 % (Figure 3g) consistently.

Acknowledgements and/or References (Optional):: This work was partially supported by the National Institutes of Health (R33AI147419) and the National Science Foundation (2045169, 2319913). Any opinions, findings, conclusions, or recommendations expressed in this work are those of the authors and do not necessarily reflect the views of the National Science Foundation and National Institutes of Health