Autonomous Molecular Sensing with a Chemically Stateful Solid-State Nanopore Authors: Makusu Tsutsui, Yuki Komoto, Kazumichi Yokota, Wei-Lun Hsu, Denis Garoli, Ali Douaki, Germán Lanzavecchia, Ryuichiro Abe, Hirofumi Daiguji, Tomoji Kawai Published: 2026 ABSTRACT Nanopores provide a direct electrical interface to biomolecules, yet most operate as passive conduits whose sensing state is imposed externally. Here we report an autonomous solid-state nanopore in which voltage-driven precipitation and dissolution repeatedly create and erase transient conductive pathways under constant bias. This reaction cycle generates stochastic ionic spikes and converts the pore into a chemically stateful nanofluidic element whose conductance reflects recent history. As nucleotides and amino acids traverse these transient openings, they reshape the evolving pore state and produce distinct spike amplitudes, durations, and firing statistics. Learning these state-dependent fingerprints enables discrimination of four nucleotides, quantitative readout of mixed populations, and extension to seven amino acids without active gating. Our results reposition the solid-state nanopore as a reaction-programmed molecular interface that integrates sensing, memory, and signal generation within chemically dynamic nanoscale environments for molecular analytics and iontronic information processing. Nanopores have emerged as one of the most powerful electrical interfaces to single molecules. − By monitoring ionic current through a nanometer-scale conduit, they provide real-time, label-free access to variety of biomolecules from DNA to proteins, while retaining an architecture compatible with miniaturization and integration. − Yet most nanopore platforms are still conceived as passive sensors, where the pore is treated as a fixed conduit, and performance depends on externally imposed control, such as motor enzymes, , feedback electronics, , or demanding fabrication of ultrasmall openings. , In living systems, by contrast, ion-conducting pores are not passive apertures but dynamic stateful elements. Biological ion channels couple transport, gating, and internal state, enabling excitability and spike generation. − This view has inspired the emerging field of nanofluidic memristors, in which confined ionic transport acquires memory through interfacial adsorption, ionic self-assembly, wetting transitions, or chemical transformations. − Such devices display memory effects to circuit-level ionic logic, suggesting that ionic conduits can process information as well as transmit it. − This conceptual shift opens a new direction for nanopore science. Once the pore is allowed to evolve chemically, − transmembrane voltage, pH, salt composition, and the arrival of biomolecules no longer simply perturb a sensing current, but instead write and erase the conductive state of the nanoconduit. In this regime, ionic current spikes are more than transient readout events, constituting state-dependent responses of a self-evolving nanofluidic element, in which sensing, actuation, and memory are physically colocalized. , For molecular inputs such as nucleotides and amino acids, whose charge, hydration, adsorption, and acid–base character can all perturb interfacial chemistry, such a pore is expected to function not only as a detector but as a chemically embodied spiking element with relevance to brain-inspired ionic information processing. Here, we report an autonomous solid-state nanopore membrane in which voltage-driven in-pore precipitation and dissolution repeatedly create and erase transient nanoconduits under constant bias. This self-healing reaction cycle generates stochastic ionic current spike trains while continuously renewing the sensing state of the pore. Because the spike waveform depends on both the in-pore reaction dynamics and the biomolecule translocation, single nucleotides and amino acids can be recognized without active gating or enzymatic stepping. This architecture positions the nanopore not only as a single-molecule sensor but also as a stateful nanofluidic element. Results/Discussion A Chemically Stateful Nanopore Generates Autonomous Ionic Spikes Our system operates through voltage-driven in-pore chemistry inside a 70 nm nanopore in a 50 nm-thick silicon nitride membrane (Figure ). When a positive transmembrane bias V b is applied across a membrane separating 2 M MnCl2 from phosphate-buffered saline (PBS), Mn2+ and phosphate species are electromigrated into the pore. Although phosphate speciation depends on local pH, the essential precipitation criterion is that the local ion-activity product for manganese phosphate exceeds its solubility product. Under this condition, manganese phosphate nucleates within the aperture and suppresses ionic conduction (Figure a,b). , Once the conductive pathway is sealed, ion flux collapses and precipitation is arrested. The acidic MnCl2 reservoir then shifts the local balance toward dissolution of the nanoprecipitate until a nanoscale conductive pathway re-emerges. The recovered current again delivers Mn2+ and phosphate species to the constriction, restoring supersaturation and driving precipitation that recloses the pathway. This transport-reaction feedback produces a self-limiting cycle rather than stable opening of the original 70 nm aperture. | Repetitive single-molecule detection by autonomous nanopore sensing. Schematic illustration of single-nucleotide identification from ionic current (I ion) traces. Under a constant transmembrane bias (V b), in-pore chemical reactions spontaneously drive the opening and closure of nanopores within a nanoprecipitate layer, generating periodic fluxes of ionic reactants. DNA nucleotides or amino acids introduced into the cis chamber are electrophoretically captured by the transiently formed pores, enabling repetitive single-molecule readout events. | In-pore chemistry-driven autonomous nanopore membrane. (a), Ionic current trace recorded in a MnCl2/PBS system at a transmembrane bias of 1.0 V. Stochastic conductive spikes in I ion indicate spontaneous formation and closure of nanopores mediated by in-pore chemical reactions. Inset shows a false-colored scanning electron micrograph of a 70 nm diameter nanopore in a 50 nm thick SiN x membrane. (b), Schematic of the autonomous nanopore formation mechanism. (I) The precipitated manganese phosphate layer gradually thins through dissolution in the acidic MnCl2 solution. (II, III) Once a pore opens, the applied voltage drives Mn2+ (purple) and PO4 3– (green) ions across the membrane, inducing local precipitation within the pore. (IV) The newly formed deposit recloses the pore. This cyclic sequence of dissolution and precipitation reactions gives rise to the transient ionic current spikes shown in the center. (c), Scatter plots of ionic spike height (I p) versus width (t d). (d), Log10 I p histogram. This recurring dissolution–precipitation cycle establishes the pore as a chemically stateful nanofluidic element (Supplementary Figure S1). Because reopening immediately restores the flux of both reactants, it also triggers reprecipitation; this negative feedback explains why the conductive states are short-lived rather than persistent. The spike amplitude thus reflects the transient conductance of the reopened small pore within the precipitated region, whereas the spike duration reports the lifetime of that conductive state before reclosure. We note that the low-current region between spikes (Supplementary Figure S2) should be interpreted as a sealed or near-sealed state of the precipitated nanopore, not as the open-pore baseline of the original 70 nm aperture. The effective pore size can be tuned coarsely through reactant concentration and pH, with lower concentrations and more acidic conditions favoring greater dissolution and hence larger transient openings. Under the 2 M MnCl2/PBS condition, the height I p of ionic spikes ranges over 10 nA at V b of 1 V (Figure c–d). This anticipates d pore of up to several nanometers within a framework of Maxwell-Hall analysis − giving the ionic resistance R pore = V b/I ion = ρ­(4L mem/πdpore 2 + 1/dpore) with the solution conductivity σ = 6.7 Ωm under an assumption of L mem ≪ d pore (Supplementary Figure S3). Each reopening event creates a brief high-field conduit through which an analyte can pass, after which the channel recloses and the sensing state resets spontaneously. The system thus combines signal generation, state reset, and analyte sensing within the same reaction-driven cycle. We note that the simultaneous formation of more than one transient conductive pathway within the 70 nm aperture cannot be excluded. However, such overlap is expected to be statistically rare because the spike widths are typically on the order of 0.1 ms, whereas the majority of interspike intervals Δt are much longer than 1 ms. Thus, a second opening would need to occur within the short lifetime of an already open pathway for multiple pores to coexist. Moreover, the conductance-derived effective diameters and the analyte-dependent shifts in spike amplitude and duration described below indicate that the dominant sensing states are molecular-scale constrictions. Therefore, although occasional multipathway events may occur, they are unlikely to dominate the observed signals or invalidate the molecular readout. Biomolecules Perturb the Evolving Pore State We examined whether molecules simply pass through these transient pores or instead reshape the evolution of the pore state itself. To test this, we investigated the four DNA nucleotides, adenosine monophosphate (AMP), thymidine monophosphate (TMP), guanosine monophosphate (GMP) and cytidine monophosphate (CMP), by adding them to PBS at concentration c while retaining MnCl2 on the trans side (Figure a). Under positive bias, the negatively charged phosphate group promotes nucleotide capture and translocation through the transient conductive state (Figure b and Supplementary Figures S4–S17). , In the absence of biomolecules, the spike-height distribution is broad, consistent with stochastic pore reopening over a range of effective diameters. The autonomous spike trains remained stable over 1 h, providing a reproducible baseline for molecular readout (Supplementary Figures S18–S22). At low nucleotide concentration (c = 10–5 M), this distribution is largely preserved because most transient pores reclose before capturing a molecule. With increasing concentration, however, a pronounced peak emerges in the I p distribution and becomes progressively sharper up to 0.05 M (Figure c and Supplementary Figures S13 and S14). This evolution indicates that captured nucleotides partially obstruct ionic transport during passage through the transiently open state. Single-nucleotide detection by autonomous nanopore sensing. (a), Scatter plots of spike height I p versus duration t d recorded at V b = 1.0 V in a MnCl2/PBS system, with 20 mM AMP (red) or TMP (dark yellow) added to the PBS. Inset, representative ionic current traces recorded without nucleotide (gray) and with AMP or TMP, showing nucleotide-dependent modulation of I ion spikes. (b), Schematic of the sensing cycle: dissolution of the manganese phosphate layer opens a transient pore. The applied bias drives Mn2+ and PO4 3– fluxes that trigger in-pore precipitation while electrophoretically capturing nucleotides. Subsequent precipitation reseals the pore, enabling repetitive single-nucleotide readout. (c), Histograms of I p for AMP, GMP, CMP and TMP in PBS at concentrations from 2 to 50 mM. (d), Gaussian fits to the 50 mM I p distributions in (c), highlighting distinct ionic blockade characteristics of the four nucleotides. The peak positions of the I p histograms differ systematically among the four nucleotides in the order CMP < TMP < GMP < AMP (Figure d), revealing analyte-specific modulation of ionic transport. These signals differ from conventional resistive-pulse readouts in stable biological or solid-state nanopores, where a fixed open-pore baseline is transiently reduced by analyte passage. − In the present system, the pore itself is created and erased during each event; the analyte therefore modulates not only the instantaneous blockade but also the lifetime and waveform of the conductive state. This distinction also explains why the Gaussian fits of the I p histograms are broader than blockade current distributions typically obtained from a fixed pore: the molecular signal is convolved with the stochastic distribution of transient pore states. In a purely steric picture, the Coulter principle predicts that the blockade current scales with d s 3 for a spherical object of diameter d s in an aperture. Because purines (AMP and GMP) are larger than pyrimidines (CMP and TMP), stronger current suppression would be expected for the former. Density functional theory calculations indeed estimate the molecular volume, V mol, of AMP and GMP to be approximately 7% greater than that of CMP and TMP (Figure a,b). Finite-element simulations based on coupled Poisson-Nernst–Planck–Stokes equations, , modeling the nucleotide as a rigid sphere of volume V mol in a 1.3 nm pore, predict current reductions of approximately 16% for purine-like and 15% for pyrimidine-like nanospheres relative to the open state (Figure c–e and Supplementary Figure 23), in broad agreement with steric exclusion. The modest difference between these values, however, indicates that steric effects alone cannot account for the experimentally observed ordering. Additional contributions likely arise from counterion transport, ,, which finite-element analysis suggests can alter the blockade current by approximately 12% over the relevant range of nucleotide surface charge density σn (Figure f), as well as from dehydration, ion pairing, concentration polarization, molecular conformation, and electroosmotic flow. − Consistent with this view, I p varies nonlinearly with V b, pointing to a strongly coupled transport regime in which steric, electrostatic, and interfacial effects mutually define ion flow through the nucleotide-in-pore system (Supplementary Figure 24; see also Supplementary Figures 25 and 26 for the spike widths and firing rates). Numerical simulations of nucleotide-induced ionic current blockades. (a), Model geometry used in finite-element simulations: a cylindrical nanopore (diameter 1.3 nm, length 1.3 nm) with a membrane surface charge density of −15 mCm–2. (b), Nucleotide structures obtained from density function theory calculations and equivalent-volume spheres used to approximate each nucleotide in the transport simulations. (c,d), Maps of electric potential and electroosmotic flow velocity with an AMP sphere positioned at the pore center. (e), Simulated ionic current as a function of the sphere axial position z (z = 0 at the pore center) for AMP spheres carrying different surface charge densities σ n (−0.10, −0.05, −0.02 and −0.01 Cm2–). Schematics indicate the corresponding sphere locations (I through V). (f), Ionic blockade current I p plotted against σn for AMP, GMP, CMP and TMP spheres. Nucleotide Identity is Encoded in State-Dependent Ionic Fingerprints The temporal features of the spikes further indicate that molecular identity is encoded not only in conductance amplitude but also in the lifetime of the conductive state. At concentrations below 0.01 M, the signal width t d varies little among nucleotides, consistent with infrequent capture before the pore reclosure (Supplementary Figures 27 and 28). At higher concentration, however, the durations separate clearly in the order GMP > TMP > CMP > AMP (Supplementary Figure 29). Because all four nucleotides contain a single phosphate group and thus experience comparable electrophoretic driving forces, this hierarchy is unlikely to originate from charge-driven drift alone. Instead, it indicates transient adhesion and interfacial interactions with the pore wall as key determinants of residence within the conductive channel. This interpretation is supported by diffusion-limited estimates. For nucleotide monomers, diffusion constants D are of the order of 10–10 m2 s–1, which gives a one-dimensional translocation time t 1D ≈ L mem 2/2D of only ∼5 ns for a pore depth L mem = 1.3 nm, far shorter than the measured dwell times. Surface interactions therefore dominate molecular residence. The four nucleotides also modulate the spike firing rate f cap in distinct ways, with f cap tending to increase with concentration in the order AMP > GMP > TMP ≈ CMP (Supplementary Figure S17). The autonomous nanopore membrane thus emerges as an adaptive ionic interface that spontaneously encodes small-molecule inputs, wherein the analogy to biological ion channels and stochastic ion-channel devices lies not in molecular structure, but in the colocalization of transport, gating, memory and spike generation within one nanoscale object. − , Learning the Spike History Enables Base Recognition and Counting Ionic spikes are demonstrated to constitute molecular fingerprints generated by the coupled dynamics of reaction-driven pore opening and analyte-dependent perturbation of ion transport under nanoscale confinement. To determine whether this state-dependent waveform space is sufficiently structured for sequencing-relevant molecular recognition, we extracted the ionic current time-series of each I ion spike and used these signals to train supervised learning models, , including XGBoost, Random Forest, and Logistic Regression, on data acquired in 50 mM solutions of the four nucleotides together with blank electrolyte controls. The trained models were then used to predict nucleotide identity within a 5-fold cross-validation framework. Classification performance was evaluated using the F-measure, F m = 2P pre P rec/(P pre + P rec), where precision, P pre = TP/(TP + FP), and recall, P rec = TP/(TP + FN), quantify assignment accuracy and detection completeness, respectively, assessed from the true positive (TP), false positive (FP), true negative (TN), and false negative (FN) outcomes. In XGBoost-based binary classification between blank electrolyte and each nucleotide solution, F m ranged from 0.93 to 0.98, confirming the pronounced difference between ionic spike waveforms recorded in the presence and absence of the biomolecules (Figure a–d). Using these trained classifiers, we counted the number of nucleotide-containing events, N x , at different concentrations c and estimated the translocation probability as p = N x /N total, where N total is the total number of detected spikes. The resulting positive correlation between p and c is consistent with electrophoretic capture of nucleotides in a diffusion limit (Figure e). , At concentrations above 0.02 M, p exceeds 0.5, indicating that most recorded spikes at 0.05 M arise from molecule-occupied conductive states. Although these measurements were performed at relatively high molecular concentrations, the required concentration could potentially be reduced by using multipore membranes, which would increase the overall event yield through parallelized molecular capture. Machine learning-assisted identification of four nucleotides. (a–d), Confusion matrices from waveform-based machine-learning classification of ionic current spikes recorded in blank electrolyte (0 mM) and 50 mM solutions of AMP (a), GMP (b), CMP (c) and TMP (d) in PBS at V b = 1.0 V. Matrices report event counts for blank versus nucleotide classes; the F-measure Fm exceeds 0.9 for each nucleotide-versus-blank task. (e), Translocation probability P versus nucleotide concentration c, defined as the fraction of events classified as nucleotide among all detected spikes; P approaches unity at high c. (f), Multiclass confusion matrix for the blank test and the four nucleotides at 50 mM each, yielding F m = 0.68. (g), Classification results for equimolar (50 mM each) nucleotide mixtures demonstrating quantitative counting: CMP + GMP binary mixture, CMP + GMP + TMP ternary mixture, and the four-nucleotide mixture, in which class fractions converge to approximately 0.25. We evaluated discrimination among the four nucleotides using translocation-mediated I ion signals acquired at 0.05 M. XGBoost yielded an F m of 0.68 for classification across the four nucleotides and the blank electrolyte condition, substantially above the chance-level value of 0.20 (Figure f). Comparable performance was obtained with Random Forest (F m = 0.65), whereas Logistic Regression gave lower accuracy (F m = 0.40) (Supplementary Figures 30 and 31). The autonomous nanopore also resolved mixtures containing multiple coexisting species (Figure g, see also Supplementary Figure 32). In a binary solution of GMP and TMP, the classifier identified events in the ratio GMP:TMP:CMP:AMP = 0.51:0.42:0.07:0, with the small CMP fraction likely reflecting misclassification between the two structurally similar pyrimidines. When CMP was added to form a ternary mixture, the fraction assigned to CMP increased accordingly, giving 0.38:0.17:0.38:0.07. Notably, AMP was rarely misclassified in either case. In a quaternary mixture containing all four nucleotides, in contrast, the system recovered GMP, TMP, CMP and AMP in the order of occurrence with quantitative ratios of 0.28:0.20:0.28:0.24. The threshold concentration under the present conditions can be estimated from the relationship between capture statistics and measurement time. In the binary classification between blank electrolyte and nucleotide-containing solutions, the machine-learning models distinguished translocation-associated spikes with an average accuracy of approximately 96% (Figure a–d), indicating that nucleotide-containing events can be separated reliably from intrinsic autonomous spikes. At 50 mM nucleotide concentration, most transient pore-opening events were assigned to molecular translocation, while the autonomous spiking rate was typically ∼20 s–1, corresponding to ∼18,000 detected spikes during a 15 min measurement with a single SiN x nanopore. Assuming, as a first-order approximation, that the molecular capture frequency scales linearly with nucleotide concentration, the concentration required to observe one translocation event within 15 min is estimated to be ∼50 mM/18,000, or approximately 3 μM. Meanwhile, we note that this value should be regarded as an order-of-magnitude estimate, because the practical detection limit will depend on background spike statistics, classification accuracy, molecular capture efficiency and measurement duration. State-dependent Readout Extends from Nucleotides to Amino Acids To examine the broader scope of this framework, we extended the measurements from nucleotides to seven representative amino acids: glycine, alanine, leucine, isoleucine, aspartic acid, glutamic acid and histidine. Each amino acid generated a reproducible yet distinct family of ionic spikes, with characteristic blockade amplitudes, dwell-time distributions and firing statistics (Figure a and Supplementary Figures S33–S38). For example, histidine produced a narrow and sharply peaked dwell-time distribution, whereas isoleucine gave smaller-amplitude blockade events. The amino acids also exhibited longer dwell times than nucleotides under similar conditions, consistent with their weaker electrophoretic driving force arising from lower net charge, while also pointing to distinct interfacial interactions within the evolving pore interior. The chemically active pore is therefore responsive to a wider set of molecular attributes, including size, charge, hydration and acid–base character, each of which can reshape the lifetime and conductance of the transient open state (Figure b). Identification of seven amino acids. (a), Typical I ion spike signals detected under 1.0 V in a MnCl2/PBS system with 0.02 M isoleucine (dark yellow), leucine (green), glycine (gray), histidine (skyblue), alanine (orange), glutamic acid (dark blue), and aspartic acid (pink). Signals found in a blank test are also shown (red). (b), I p versus t d scatter plots for isoleucine (dark yellow), histidine (skyblue), and aspartic acid (pink). (c), Confusion matrix obtained by machine learning classification of ionic signal waveforms within a framework of 5-fold cross-validation. Machine-learning analysis of the amino-acid waveforms yielded an F m of 0.37 for seven-class discrimination (Figure c). Although this performance remains below that obtained for nucleotides and does not yet capture the full chemical complexity required for proteomic readout, it nevertheless shows that the platform can transform chemically diverse molecular encounters into distinguishable state-dependent spike signatures. From the perspective of sequencing, this result highlights a key advantage of the autonomous pore: it repeatedly regenerates a molecularly informative readout state without active gating or inadvertent clogging. These results show that the autonomous pore operates not merely as a resetting detector, but as a reaction-programmed chemical spike generator whose waveform statistics encode both molecular identity and abundance, while the underlying nanofluidic dynamics furnish a computationally informative representation space for molecular recognition and counting. Conclusions The present work establishes the autonomous solid-state nanopore as a resetting molecular sensor. Under constant bias, precipitation and dissolution within the pore repeatedly create, erase, and recreate conductive nanoscopic pathways, such that ionic transport is governed by an evolving internal chemical state rather than a fixed aperture. Even in the absence of analytes, this intrinsic reaction cycle generates stochastic spike trains. Upon the introduction of biomolecules, the analytes not only traverse the channel but also perturb the evolving pore state itself, shifting spike-height distributions, altering dwell times and reshaping higher-order waveform features. The signal therefore emerges from coupled dynamics among confined ionic transport, in-pore chemistry, and molecular translocation. This behavior is particularly significant at the monomer level, where differences in molecular size alone are insufficient to explain the observed hierarchy of ionic blockade characteristics. Rather, the waveform reflects a richer interplay of steric, electrostatic, hydration, and surface-interaction effects operating within an adaptively reconfigured nanoscale conduit. − From the perspective of sequencing, the chemically evolving nature of the pore is an asset rather than a complication. Conventional nanopore readout often treats internal dynamics as noise to be minimized through precise fabrication, external feedback, or enzymatic stepping. − Here, by contrast, the transient pore state becomes an integral part of the sensing mechanism. Because each molecular event is encoded in a waveform shaped jointly by analyte identity and the contemporaneous pore state, the resulting signal carries a higher-dimensional fingerprint than a static blockade alone. This feature allowed the discrimination between blank electrolyte and nucleotide solutions, the above-chance multiclass identification of the four nucleotides, the quantitative counting of mixed populations, and the extension to amino-acid recognition. The present architecture thus suggests a route to molecular readout in which the in-pore chemistry continuously regenerates the sensing state, while learning extracts analyte identity from state-dependent ionic trajectories. The platform also points to a convergence of sensing, memory and computation within a single nanofluidic object. Ionic spikes already retain a short-lived history, as their amplitude and duration reflect the preceding chemical evolution of the pore. Molecular inputs, including pH-active species, nucleotides, and amino acids, can further bias this evolution by modulating local ion transport and interfacial interactions. This perspective positions the pore not only as a stochastic sensor, but also as a chemically programmable spiking element whose transition statistics can be tuned by voltage, electrolyte composition, pH, and biomolecular environment. In this view, analytes are not simply detected, but actively reshape the dynamical landscape of the pore itself. Such behavior is precisely what makes nanofluidic memristive systems attractive for neuromorphic and in-sensor computing, − ,− where information processing emerges from transport coupled to internal chemical dynamics rather than from a strict separation between sensing, memory, and computation. Several limitations should also be addressed before the approach can be translated into general molecular analytics. For example, the present single-pore experiments require millimolar nucleotide concentrations because the conductive state is present only transiently and the analyte-induced perturbation is small for monomers. Amino-acid classification remains moderate, indicating that additional chemical handles, longer residence times or improved waveform models will be required for proteomic readout. The reaction cycle also depends on a controlled balance among metal ions, phosphate, pH and evaporation; in complex biofluids, proteins, chelators, multivalent ions and particulate matter may irreversibly adsorb to the precipitate or shift the precipitation–dissolution equilibrium. These limitations are not fundamental to the concept, because the effective pore size and lifetime can be tuned by pH and reactant concentration, and the event yield can be increased by multipore arrays. , Future implementations should combine sealed or replenished reservoirs, surface passivation, sample pretreatment or buffer exchange, and alternative reaction chemistry chosen for compatibility with the target biofluid. In this context, biofluid operation represents a realistic future direction, particularly because endogenous inorganic phosphate in biological matrices may be leveraged to activate autonomous nanopore sensing without adding phosphate-containing buffer. Methods/Experimental Fabrication of Solid-State Nanopores A 4-in. Silicon wafer was coated on both sides with 20 nm-thick SiN x films by low-pressure chemical vapor deposition. The wafer was diced into 25 mm × 25 mm chips using a precision dicing saw (DISCO Corporation, Tokyo, Japan). On one side of each chip, photoresist AZ5206-A was spin-coated and prebaked on a hot plate. A microelectrode pattern was defined by photolithography, developed and metallized by radio frequency magnetron sputtering of 5 nm Ti (adhesion layer) followed by 20 nm Au. The patterned chip was immersed overnight in N,N-dimethylformamide (DMF) and subjected to ultrasonication for lift-off, then cleaned sequentially with ethanol and acetone. An electron-beam resist (ZEP 520A-7, Zeon Corporation, Japan) was spin-coated on the metal-patterned surface and baked at 180 °C. Using part of the metal pattern as an alignment marker, a circular feature was exposed at the chip center by electron-beam lithography (Elionix Inc., Tokyo, Tokyo, Japan). The exposed SiN x region was etched by reactive-ion etching (RIE) with CHF3 gas (Samco Inc., Kyoto, Japan), using the remaining resist as a mask to define the nanopore position. The remaining resist was dissolved in DMF and rinsed with ethanol and acetone. Subsequently, a 1 mm × 1 mm window was opened on the backside SiN x layer by LED photolithography followed by CHF3-based RIE (Samco Inc., Kyoto, Japan), aligned to the nanopore using the surface metal pattern as a reference. A 40 nm Cr protective layer was then deposited on the front, metal-patterned side. The underlying Si substrate was removed from the backside by anisotropic wet etching in KOH solution at 90 °C through the square SiN x window. The Cr layer acted as an etch stop and prevented nanopore enlargement due to overetching. Finally, the Cr coating was stripped in a Cr etchant, and the chip was rinsed sequentially with water, ethanol and acetone. Nanopore Sealing Polymeric blocks containing microfluidic channels were attached to both sides of the nanopore chip for single-molecule measurements. To fabricate the blocks, I-shaped channel molds were first patterned by photolithography on SU-8 photoresist spin-coated onto a 4-in. Si wafer. Sylgard 184 (Dow, Midland, MI, USA) was poured over the mold and cured at 90 °C for 6 h in an oven. The cured polydimethylsiloxane (PDMS) layer was cut into two blocks using a surgical knife, and three holes were punched in each to serve as solution inlets/outlets and electrode ports. The PDMS blocks were bonded to both sides of the nanopore chip by oxygen plasma activation using a Femto plasma cleaner (Femto Science, Gyeonggi-do, Korea), followed by immediate alignment and attachment. Ionic Current Measurements The 2 M MnCl2 and 5× PBS solutions were introduced into the cis and trans fluidic reservoirs of the PDMS measurement setup through the access holes (Supplementary Figure S39); the solutions were not injected directly into the nanopore itself. The reservoirs were typically filled with approximately 30 μL of solution on each side, sufficient to keep the Ag/AgCl electrodes immersed throughout the measurement and to avoid air bubbles at the membrane. Ag/AgCl electrodes were inserted into separate holes on each side of the chip to establish electrical contact. A transmembrane voltage was applied to one Ag/AgCl electrode using a battery-driven potentiostat, while the ionic current was measured through the opposing electrode without analog filtering. The current was amplified using a custom-built low-noise current amplifier and digitized at a sampling rate of 1 MHz with a high-speed digitizer (National Instruments, Austin, TX, USA). Ionic Current Spike Extraction The baseline of each ionic-current trace was corrected by subtracting a linearly fitted component within every 0.5 s segment of the data. This operation removed slow drift in the low-conductance state without altering the submillisecond to millisecond spike waveform. The phrase baseline therefore refers to the near-zero current level of the completely or nearly completely sealed nanopore, not to a stable open-pore current. The resulting traces were smoothed using a second-order Savitzky–Golay filter, which reduced the noise to below 300 pA while preserving the fine temporal features of the ionic current spikes. Spikes were identified by detecting local maxima exceeding 300 pA in the baseline-corrected traces, and for each detected event, 0.005 s of data before and after the spike apex were saved separately. All digital processing was carried out using custom programs written in Python and Visual Basic. Finite Element Analysis Numerical simulations of single nucleotide translocation through a SiN x nanopore were performed in a cylindrical coordinate system. The model comprised a cylindrical domain with a radius of 8 μm and a height of 16.05 μm, containing a SiN x membrane positioned at its center. The membrane was 50 nm thick overall, but thinned to 1.3 nm within a 70 nm-wide region around the center, where a nanopore of radius 0.65 nm was defined. The SiN x membrane was assigned an electrical conductivity of 1.0 × 10–12 S m–1 and a relative permittivity of εr = 9.7. The electric field and fluid velocity field were obtained by finite-element analysis through the coupled solution of the Poisson equation, the steady-state current continuity equation, the Nernst–Planck equation and the Stokes equation. All simulations were carried out using COMSOL Multiphysics v6.4. Machine Learning For each ionic pulse, 2000 data points preceding and following the peak apex were extracted. Dimensionality reduction was subsequently performed using Piecewise Aggregate Approximation (window size = 10), yielding a 400-dimensional representation of each signal. The resulting 400-dimensional vectors were used as input features for classification. To estimate the population fractions of individual molecular species in a mixture set, we employed a probability-based quantification approach using calibrated classification outputs. A classifier was trained on labeled reference data, where each class corresponds to a distinct molecular configuration, structural state, or nanoscale regime. For each measurement x, the classifier outputs a calibrated posterior probability vector p̂(y = i|x), representing the likelihood that the measurement originates from state i. Here, y denotes prediction labels. For the unlabeled data set, we computed the ensemble-averaged predicted probability q=1N∑n=1Np̂(y|xn) which represents the expected classifier response over the entire measurement population. Assuming that the class-conditional measurement characteristics are approximately preserved between the labeled reference data and the unlabeled data set, while the relative population fractions may differ, the vector q can be expressed as q≈Aπ Here, π denotes the unknown mixing ratio vector of the molecules in the mixture, and the matrix A represents the average classifier response conditioned on each true state. Each jth column of A was estimated from the labeled reference data as. A :,j = E[p(y|x)|y = j],using 5-fold cross-validated predictions to mitigate overfitting and optimiztic bias. Here, E denotes expected value. The ground-truth mixing ratio π were obtained by solving the constrained least-squares problem min⁡π∥Aπ−q∥22subjecttoπ≥0,∑iπi=1 The optimization was performed using projected gradient descent, where each update was projected onto the probability simplex to ensure physically meaningful population fractions. This formulation relies only on ensemble-averaged prediction statistics rather than instance-wise posterior corrections, making it robust to local measurement noise, imperfect probability calibration, and moderate deviations from ideal distributional assumptions. The estimated fractions therefore reflect the global composition of molecules present in the unlabeled measurement ensemble. Supplementary Material