Carbon-nanotube field-effect transistors for resolving single-molecule aptamer–ligand binding kinetics

Abstract

Small molecules such as neurotransmitters are critical for biochemical functions in living systems. While conventional ultraviolet–visible spectroscopy and mass spectrometry lack portability and are unsuitable for time-resolved measurements in situ, techniques such as amperometry and traditional field-effect detection require a large ensemble of molecules to reach detectable signal levels. Here we demonstrate the potential of carbon-nanotube-based single-molecule field-effect transistors (smFETs), which can detect the charge on a single molecule, as a new platform for recognizing and assaying small molecules. smFETs are formed by the covalent attachment of a probe molecule, in our case a DNA aptamer, to a carbon nanotube. Conformation changes on binding are manifest as discrete changes in the nanotube electrical conductance. By monitoring the kinetics of conformational changes in a binding aptamer, we show that smFETs can detect and quantify serotonin at the single-molecule level, providing unique insights into the dynamics of the aptamer–ligand system. In particular, we show the involvement of G-quadruplex formation and the disruption of the native hairpin structure in the conformational changes of the serotonin–aptamer complex. The smFET is a label-free approach to analysing molecular interactions at the single-molecule level with high temporal resolution, providing additional insights into complex biological processes.

Small molecules, such as neurotransmitters, are crucial in biological processes, including cellular signalling¹ and gut–brain communication^2–4. However, the real-time detection and analysis of neurotransmitter release remains challenging. Traditional detection methods such as ultraviolet–visible spectroscopy and mass spectrometry, which require sample collection and offline analysis, are not capable of real-time measurement^5,6. Fast-scan cyclic voltammetry, coupled with microdialysis, offers time-resolved measurement of neurotransmitter release⁷ with nanomolar sensitivity but suffers from sensitivity degradation over time due to electrode fouling^8,9. Although recent advancements in chemically modified electrodes^10–13 and engineered waveforms¹⁴ have improved selectivity, these approaches also still struggle with the oxidation of multiple species at overlapping potentials. Fibre-based optoelectronics exploiting genetically encoded fluorescent indicators, despite their high specificity and resolution^15,16, are hampered by photobleaching and sensitivity limitations^17–19.

Field-effect transistor (FET) biosensors that use aptamers have emerged as a promising solution for detecting small, negligibly charged or neutral molecules^20,21. Aptamers, DNA stem–loop structures that can undergo conformational changes upon binding with small molecules, have enhanced the sensitivity and selectivity of FET sensors^22–24. However, the readout in these devices requires an ensemble signal from densely packed probe molecules connected to the gate of the device^25–27. This signal lacks the kinetic information of the molecular interactions. Furthermore, it is easily corrupted by inherent transconductance drift in salt solution, making it difficult to measure temporal changes in analyte concentrations²⁸.

These challenges in field-effect detection can be circumvented by conducting measurements at the single-molecule level^26,27. Utilizing carbon nanotubes (CNTs) conjugated with a single DNA aptamer, single-molecule field-effect transistors (smFETs) offer label-free detection and prolonged measurement periods without device degradation²⁸. The temporal resolution in smFET is solely limited by the subsequent detection electronics, making it possible to observe fast kinetics. Compared to single-molecule fluorescence resonant energy transfer methods^29,30, this approach requires less complex instrumentation and avoids the limitations of photobleaching. Replicating sensors in smFET systems further lowers the detectable concentration range within the same measurement timeframe.

Another alternative for single-molecule electronic sensing is the exploration of peptides as a one-dimensional conducting channel³¹. Although peptides provide precise tethering of single-molecule targets and facilitate metal contact attachment, their conductivity is much lower than that of CNT FETs, affecting signal-to-noise performance. Additionally, the necessity of dielectrophoretically drawing down bridging molecules at measurement time reduces sensor yield.

In this study, we integrate a serotonin-specific aptamer with a CNT in an smFET system, enabling us to analyse neurotransmitter interaction kinetics at the single-molecule level. This system allows real-time observation of dynamic aptamer conformational changes through a random telegraph signal (RTS) time–current trajectory. Our analysis reveals a complex four-state aptamer–ligand system, differing from the traditional two-state models. We also demonstrate the modulation of these kinetics by varying target concentrations and applied gate voltage.

The smFET device

Our smFET device is a single-walled CNT interfaced with drain/source electrodes. Scanning electron micrographs confirm the presence of individual CNTs bridging the electrodes (Extended Data Fig. 1). To render the smFET sensitive to serotonin, we covalently functionalize the CNT with a specific DNA aptamer for serotonin (5-HT)²². Briefly, a single sp³ defect in the CNT is generated by a gate-voltage-driven diazonium reaction on individually contacted CNT devices (Fig. 1a and Supplementary Fig. 1)^32,33. This functionalization involves creating a single sp³ defect on the CNT through a controlled diazonium reaction (Extended Data Fig. 2), which, while slightly reducing device conductance, maintains the electrical properties essential for FET functionality. This site serves as a potential barrier within the CNT that dominates the device conductance and is sensitive to the local electrostatic environment³⁴. This point functionalization also provides an aldehyde end-group to which the amine group at the 3′ end of the aptamer can be linked by a covalent imine reaction.

Fig. 1 |. Device schematic and concentration-dependent current traces.

a, Schematics of the single aptamer immobilization onto the CNT and V_LG-controlled diazonium chemistry. A common V_LG is applied via a reference electrode in the buffer solution. A single functionalization site on the CNT is generated by sp³ addition controlled by V_LG-driven aryl radical generation from a diazonium salt (FBDP). The amine group of a functionalized DNA aptamer is covalently attached to the site by a Schiff base reaction. b, Representative baseline I_D–t trace of Device A after aptamer probe attachment in phosphate-buffered saline (pH 7.0). The V_LG was fixed at 200 mV, and a V_DS of 25 mV was applied. c–f, Representative I_D–t traces of Device A at different serotonin concentrations: 0.5 nM (c), 5 nM (d), 50 nM (e), 500 nM (f). The raw I_D–t traces (blue line) are overlaid with the idealized fit, revealing two conductance states (orange line). The histograms of I_D distributions are shown in the right panels. g, Concentration dependence for the fraction of time spent in the lower conductance state (P_low). The plots of P_low against serotonin concentrations are fitted to the Langmuir isotherm function. Data points are the mean probability of the low conductance state calculated from all dwell times by bootstrapping (N_boot = 2,000). Error bars represent the 90% confidence interval from the bootstrapped mean value of P_low.

The fabricated device operates in direct contact with an electrolyte within a microfluidic chamber (Supplementary Fig. 2), using a liquid gate potential (V_LG) to modulate electron and hole densities in the CNT’s semiconducting channel³⁵. The operational parameters of individual devices, including the optimal V_LG for maximum transconductance (g_m = dI_D/dV_LG) and signal-to-noise ratio, are carefully determined based on the drain–current transfer (I_D–V_LG) characteristics (Supplementary Table 1). Although oxygen and water molecules can adsorb on the CNT surface, the transfer characteristics of the CNT devices remain largely unchanged over long-term measurement times (Supplementary Fig. 3).

The ionic strength of the electrolyte solution determines the width of the double layer, as determined by the Debye length (λ_D), on the CNT surface^35–38. In phosphate-buffered saline (PBS, pH 7.0), the Debye length is estimated at 0.7 nm from the CNT surface (Supplementary Fig. 4). On this scale, the small size and ligand-induced conformational changes of the aptamer become detectable, leading to alterations in the electrical signal upon target molecule binding.

Target-specific conductance in aptamer–CNT FETs

The I_D time traces (I_D–t) directly show target-induced changes in the aptamer-conjugated CNT FET signals. We record baseline current signals in PBS before introducing serotonin (Fig. 1b). Upon serotonin introduction, we observe shifts in the current level and the emergence of discrete RTSs above the flicker noise floor (Fig. 1c–f). These signals, characterized by two distinct conductance states, are typically separated with an SNR of better than 4 (Extended Data Fig. 3). The fraction of the lower conductance state (P_low) exhibits a concentration-dependent increase (Fig. 1g and Extended Data Fig. 4), indicating that the low conductance state represents a serotonin-bound aptamer complex. These signals remain active in the presence of serotonin and are quenched when serotonin is removed from the solution (Extended Data Fig. 5 and Supplementary Fig. 5).

Control experiments further affirm the specificity of these signals. Introducing dopamine, a molecule structurally similar to serotonin, to serotonin-specific aptamer devices does not elicit the same two-state RTS activity (Extended Data Figs. 5 and 6). Unfunctionalized CNT devices also do not exhibit this activity upon serotonin introduction (Extended Data Fig. 7). However, both dopamine and serotonin cause a moderate baseline shift in device conductance. Dopamine and serotonin each have one positive charge at physiological pH and lead to a negative shift in the threshold voltage of the I_D–V_LG transfer curves due to non-specific adsorption on the CNT (Supplementary Fig. 6)³⁹. The degree of shift with neurotransmitter adsorption is relatively constant over the entire concentration range examined (100 pM to 1 μM).

Our system can detect serotonin in complex biological matrices, such as artificial cerebrospinal fluid (aCSF). We introduced serotonin (0–400 nM, final concentration) and a major serotonin metabolite, 5-hydroxyindoleacetic acid (100 μM), into aCSF (Supplementary Fig. 7). Despite the presence of other ions and metabolites, the system detects serotonin in the nanomolar range. The frequency of these signals increased proportionally with the concentration of serotonin, which aligns with our observations in the phosphate buffer solution.

The versatility of the smFET platform is further demonstrated by the detection of dopamine using a dopamine-specific aptamer (Supplementary Fig. 8). In the I_D–t traces, we observed RTSs representing two-conductance states at dopamine concentrations of 100 nM and 2,000 nM. These signals indicate dopamine-induced conformational changes of the aptamer near the CNT surface. Compared to the serotonin system, the high-conductance state becomes more prevalent as dopamine concentrations increase. This contrasting behaviour aligns with previous findings from ensemble FET experiments and can be attributed to different conformational changes occurring upon target binding²². These results highlight the sensitivity of our method to different aptamer–ligand interactions.

Kinetic analysis of aptamer conformation changes

The two-level signals originate from the conformational changes in the aptamer due to serotonin binding. The negatively charged aptamer backbone tends to conform to the CNT surface by π–π interactions. This aptamer orientation increases the device conductance by charge screening of the defect, locally enhancing the hole density in the CNT^39–41. In the presence of serotonin, the aptamer orientation shifts away from the CNT, reducing conductance^22,40. These kinetics are quantified by analysing the dwell times of the high (τ_high) and low (τ_low) conductance states (Fig. 2a). For both high (τ_high) and low (τ_low) conductance states, the dwell-time distributions were best represented by the sum of two exponential distributions (Fig. 2b and Supplementary Fig. 9), indicating multiple reaction pathways (Extended Data Fig. 8). We confirmed the statistical significance of the double-exponential fits over the single-exponential fits using an F-test, which was particularly pronounced when the serotonin concentration exceeded 1 nM (Supplementary Table 2).

Fig. 2 |. Single-molecule kinetic analysis.

a, Representative I_D–t trace at 50 nM serotonin concentration. The conformational state of the aptamer at the edge of the Debye layer (λ_D) on the CNT surface corresponds to the high-conductance (blue dashed line) and low-conductance states (orange dashed line). Two-state HMM analysis idealizes experimental I_D–t traces and defines the dwell times in each state (τ_high and τ_low). The rate constants of high-to-low transition (k_HL) and low-to-high transition (k_LH) are determined by 1/τ_high and 1/τ_low, respectively. b, Normalized dwell-time distributions for the low- and high-conductance states extracted from Device A. The distributions are normalized by 1 s and fitted with a double-exponential function (orange lines) and a single-exponential function (red lines). F-test confirmed that double-exponential functions provided a statistically better fit than the single-exponential functions. The most probable fast and slow dwell times are determined by double-exponential fit. c,d, Estimated rate constants: fast and slow rate constants of low-to-high state ($k_{\text{LH}}^{\text{fast}}$ and $k_{\text{LH}}^{\text{slow}}$) (c) and high-to-low state ($k_{\text{HL}}^{\text{fast}}$ and $k_{\text{HL}}^{\text{slow}}$) (d) are plotted as a function of serotonin concentration. Reported data points are mean values from the bootstrapped double-exponential fitting of the dwell-time distribution (N_boot = 2,000). The error bars indicate a 90% confidence interval from bootstrapping. e, The number of fast and slow high-to-low transitions in the recording time ($N_{\text{HL}}^{\text{slow}}$ and $N_{\text{HL}}^{\text{fast}}$) as a function of concentration. Reported data points are mean values from the bootstrapped double-exponential fitting of the dwell-time distribution (N_boot = 2,000) and error bars indicate the 90% confidence interval of the bootstrapping.

We determined the rate constants for the fast and slow transitions in each direction (low-to-high and high-to-low) at different serotonin concentrations. For low-to-high transitions, both the fast rate constant $\left(k_{\text{LH}}^{\text{fast}}\right)$ and slow rate constant $\left(k_{\text{LH}}^{\text{slow}}\right)$ were independent of serotonin concentration (19.5 ± 7.7 s⁻¹ and 1.9 ± 1.1 s⁻¹, respectively), as expected for aptamer–serotonin dissociation (Fig. 2c). In contrast, the fast rate constant for high-to-low transitions $\left(k_{\text{HL}}^{\text{fast}}\right)$ increased from 0.8 ± 0.2 s⁻¹ to 27.8 ± 5.5 s⁻¹ with serotonin concentration, while the slow rate constant $\left(k_{\text{HL}}^{\text{slow}}\right)$ remained unaffected (0.6 ± 0.3 s⁻¹) (Fig. 2d). Furthermore, the frequencies of both high-to-low transitions were similar across the range of serotonin concentrations, showing concentration-dependent increase (Fig. 2e). At low concentrations, the estimated rate constants display large standard deviations due to the smaller number of transition events available for dwell-time statistics²⁸.

The existence of two rate constants for each transition and the concentration-independent nature of the high-to-low transition suggest that the interaction is not a simple ligand-mediated folding mechanism. Rather, it suggests hidden conformational transitions within the observed kinetics⁴¹. These findings are consistent with the observation that the binding mechanism between serotonin and the aptamer is not a simple lock-and-key model^42,43. This complexity is further supported by the fractional occupancy of the low conductance state (P_low), which saturates below 50% even when the serotonin concentration is much higher than the reported dissociation constant (K_D) (30 nM)²² (Fig. 1g and Extended Data Fig. 4). This supports the notion of additional conformational states of the aptamer–serotonin complex on the CNT, which do not induce a change in smFET conductance.

Hidden Markov models (HMMs) are employed to decipher the reaction pathways. To determine the most probable kinetic representation, we calculated the Bayesian information criteria (BIC) score for various kinetic networks to fit the smFET data (Supplementary Fig. 10 and Supplementary Table 3)⁴². The analysis proposes a four-state model with two possible transitions from each state (Fig. 3a). Two states (State 0 and State 1) belong to the high-conductance state, while the other two (State 2 and State 3) correspond to the low-conductance state. Among the four states, State 0 represents the aptamer without bound serotonin, whereas the remaining three states involve serotonin binding. Transitions are allowed between State 0 and State 1, State 0 and State 3, State 1 and State 2, and State 2 and State 3.

Fig. 3 |. Serotonin binding and kinetic observables analysed by HMMs.

a, Representation of serotonin binding using four kinetic observables, consisting of two high-conductance states (States 0 and 1, shown as blue circles) and two low-conductance states (States 2 and 3, shown as orange circles). b, Plot illustrating the two rate constants (pink, on-rate; green, off-rate) obtained from the HMM analysis for each binding pathway as a function of serotonin concentration. The data from Device A’s smFET experiments are utilized. Data points are calculated from fitting dwell-time distributions for each state transition to a single-exponential distribution. The error bars indicate the 1σ confidence interval of the model fit.

All extracted dwell times for each pathway follow a single-exponential distribution, indicating that the four-state HMM analysis kinetically resolves experimental dwell-time distributions (Extended Data Fig. 9). The rate constants are determined by the model-assigned dwell-time distribution for each transition at each concentration (Fig. 3b and Extended Data Fig. 9). At low serotonin concentrations, the devices primarily reside in the high-conductance state, with State 0 predominantly occupied. As the serotonin concentration increases, the conversion rates from State 0 to 1 (0–1) and from State 0 to 3 (0–3) increase. The rates for other transitions remained constant regardless of serotonin concentration. Importantly, the four-state model not only explains the observed conductance transitions in terms of inter-state conversion rates but also accounts for the occurrence of both short and long dwell times in two-way transitions at similar frequencies.

By differentiating between State 0 (unbound) and State 1 (bound) within the high-conductance state, we calculated the dissociation constant (K_D) of the serotonin–aptamer complex. K_D is defined as the concentration of serotonin at which the aptamer spends an equal amount of time in the serotonin-unbound state (State 0, P₀) and the serotonin-bound states (States 1, 2 and 3, P_1,2,3 = 1 − P₀). Plotting 1 − P₀ as a function of concentration yields an adsorption isotherm (Fig. 4). Using a basic Langmuir model and assuming a single binding site, we estimated the K_D as 31.6 ± 7.4 nM at the ungated condition (V_LG = 0), where there are negligible electric field gradients around the CNT. The K_D is close to the value obtained from ensemble measurements (30 nM)²², confirming that our four-state model accurately captured the binding behaviour of serotonin to the aptamer under specific experimental conditions.

Fig. 4 |. Dose–response curves showing the bound state occupation estimated by HMM analysis.

Dose–response curves showing the bound state occupation (1 − P₀) estimated by HMM analysis for three independent devices: Device A (orange, V_LG = 200 mV), Device B (blue, V_LG = 100 mV) and Device F (green, V_LG = 0 mV). Inset: dissociation constants (K_D) obtained from Langmuir isotherm fitting, plotted as a function of V_LG. The measured K_D values are as follows: 31.5 ± 9.2 nM (V_LG = 0 mV, R² = 0.95, n = 6), 13.5 ± 7.4 nM (V_LG = 100 mV, R² = 0.87, n = 5) and 2.2 ± 1.3 nM (V_LG = 200 mV, R² = 0.96, n = 4). The error bars in the inset show the 1σ confidence interval of the Langmuir isotherm fitting.

Conformations of the serotonin–aptamer on the CNT

The different conformational states of the serotonin–aptamer complex associated with the four-state model involve energetic considerations and changes in the proximity to the CNT surface. The aptamer without bound serotonin (state 0) is predicted to maintain its native hairpin structure on the CNT surface, where positive ions screen the negatively charged CNT. A portion of the serotonin–aptamer complex moves away from the CNT surface in State 2 and State 3, while it remains close to the surface in State 1. The calculated rate constants for each pathway indicate that State 1 and State 2 are more energetically favourable, while State 3 is less favourable (Supplementary Fig. 11).

Single-molecule FRET experiments and thioflavin T (ThT) displacement assays are conducted to understand these conformational changes. The FRET results demonstrated transitions between two FRET states upon the introduction of serotonin (Supplementary Fig. 12). The increase in FRET efficiency with serotonin suggested that the distance between the donor (attached to a base of a loop) and acceptor (attached to the 3′ end) fluorophores became closer upon serotonin binding. These results cannot be explained by simple bending of the serotonin–aptamer, suggesting that the intramolecular arrangement of the aptamer is altered upon serotonin binding.

A previous study reported the formation of the G-quadruplex structure of the aptamer upon serotonin binding, as observed using circular dichroism spectroscopy and fluorescence assay^22,44. We performed ThT displacement assays to identify the bases involved in this binding. ThT binds to DNA and emits higher fluorescence compared with its unbound state⁴⁵. ThT is displaced from the serotonin–aptamer complex, resulting in a substantial decrease in fluorescence intensity. We employed various aptamers with modified base sequences in these assays. Truncated or bases-added aptamers indicated that the stem length does not strongly impact serotonin binding (Supplementary Figs. 13 and 14). Furthermore, we found that 14 out of the 19 guanines of the aptamer are involved with serotonin binding, consistent with the formation of G-quadruplex structure in the serotonin–aptamer complex (Supplementary Fig. 15)⁴⁴.

Our proposed model describes the involvement of G-quadruplex formation and the disruption of the native hairpin structure in the conformational changes of the serotonin–aptamer (Fig. 5). State 0 represents the serotonin-unbound aptamer, adopting its native structure in close proximity to the CNT surface. The serotonin–aptamer complex can undergo two different conformational changes to reach either State 1 or State 3. Although the transition rates are similar for both pathways, these two states are distinct in many aspects. State 1 is energetically more favourable, positioned near the CNT surface, and characterized by longer distance between 29T and 44T (3′ end). In State 1, a G-quadruplex stack is formed, involving eight guanines, while bases in a bulge loop remain within the Debye layer of the CNT. This conformation results in high conductance and low FRET efficiency. On the other hand, considering the kinetics of the transitions from State 1 to both State 0 and State 2 occurs slowly, while the transition from State 3 to both State 0 and State 2 is much faster, State 3 may act as an intermediate state in the context of ensemble measurements, not involving the formation of a G-quadruplex. Instead, the native hairpin structure of the aptamer is disrupted in State 3. State 2, which is another energetically favourable state, is characterized by the formation of a larger G-quadruplex stack with 12 guanines. In this state, a substantial portion of the aptamer extends out of the Debye layer, resulting in low conductance and a shorter inter-dye distance, leading to high FRET efficiency.

Fig. 5 |. Schematic conformations of the serotonin–aptamer complex on the CNT FET.

In the presence of serotonin in the solution, the CNT-conjugated aptamer undergoes dynamic transitions between different conformational states. The serotonin (depicted as a green circle) binding involves G-quadruplex formation (State 1 and State 2) and disruption of native structure (State 3), leading to their corresponding characteristics in terms of conductance and FRET efficiency. The transitions from state 0 to both State 1 and State 3 occur rapidly at high serotonin concentrations. In State 1, a G-quadruplex stack is formed, involving eight guanines (G2, G3, G4, G5, G7, G8, G10, G11) while State 2 exhibits a larger G-quadruplex stack with 12 guanines (G2, G3, G4, G5, G7, G8, G9, G11, G15, G16, G17, G18), resulting in altered conductance and enhanced FRET efficiency. The guanine bases used to form G-quadruplex stacks are numbered as described in Supplementary Fig. 14a. State 3 acts as a potential intermediate state in ensemble measurements.

Gate-voltage effect on the serotonin–aptamer binding

We explored the influence of gate voltage (V_LG) on the binding affinity and kinetic behaviours of serotonin–aptamer interactions using the smFET system. We observed a significant reduction in the estimated K_D value from 31.6 nM to 2.2 nM when V_LG was increased from 0 V to 0.2 V (Fig. 4). This increase in binding affinity at positive gate voltage can be attributed to a repulsive electrostatic force at the CNT–aptamer interface, which destabilizes the native secondary structure of the aptamer and promotes the binding-induced conformational change (Supplementary Fig. 16). The modulation of V_LG also affects the kinetic trends, including the k₀₁ pathway (Supplementary Fig. 17). In contrast, aptamer activity was silenced in the negatively gated device, causing the CNT surface to become positively biased (Supplementary Fig. 18). In this case, adsorption of negatively charged species, including the aptamer, to the CNT surface hinders serotonin–aptamer interactions^46,47. To prevent this, tests were conducted under ungated or positively gated conditions, reducing surface adsorption.

The kinetic behaviour of ligand–receptor interactions is primarily influenced by molecular structure⁴⁸. However, these findings demonstrate that aptamer–ligand interactions can be electrically modulated on the CNT FET, allowing real-time control over their kinetics and energetics²⁶, which is not possible with other single-molecule methods.

Conclusions

Compact nucleic acid receptors enable dynamic studies of single-molecule interactions within the Debye layer, illustrating the potential of smFET technology for detailed molecular analysis. For detection alone, reliance on G-quadruplex structures may make the kinetic pathways unnecessarily complex. Non-G-quadruplex sequences in the aptamer may be engineered in the future to produce more straightforward detection with enhanced signal transduction. Kinetics influenced by electrostatic forces in the smFET system provide an opportunity to tune the affinity for detection. Such tuning may be crucial for optimizing single-molecule detection across a large dynamic concentration range of various target molecules. High-density sensor arrays, utilizing integrated CMOS electronics⁴⁹, hold promise for both low-concentration and multiplexed detection, and for spatially distributed analyte detection in brain regions. The latter is particularly relevant for in vivo neurotransmitter assays, offering insights into the correlation between neural activity and neurotransmitter release, deepening our understanding of local neural dynamics.

Methods

CNT depositions

The smFET device utilized in this study incorporates single-walled CNTs deposited through two different methods: spin-casting and the chemical vapour deposition (CVD) (Extended Data Fig. 1). Both methods demonstrated comparable device performance, providing flexibility in the manufacturing set-up of the CNTs.

Spin-casting CNT FET fabrication

Commercially available CNTs suspended in an aqueous solution (250 μg ml⁻¹) (Super PureTubes, NanoIntegris) are randomly deposited on the silicon substrate by spin-coating using a previously established protocol³². After the CNT deposition, the bilayer resist is spun (DS-K101-312 antireflective coating at 5,000 r.p.m. for 1 min, followed by UV210-0.3 DUV photoresist at 2,500 r.p.m. for 1 min) and exposed using a DUV stepper. Titanium electrode pairs with source–drain separation of 0.5 μm are patterned. A total of 280 electrode pairs are arrayed in each reticle. Areas for platinum electrode bars, serving as pseudo-reference electrodes, are patterned on either side of the electrode array; 75 nm platinum is deposited in these areas. Gold bond pads are also patterned for device packaging. A 90 nm layer of gold is deposited on top of the titanium electrode arrays. This reticle is stepped across the 100 mm wafer, yielding 45 chips. The wafer is then exposed to O₂ plasma (50 W at 30 mT for 60 s), resulting in nanotubes being etched away from the regions outside the electrode gap. Subsequently, an additional photoresist covers the entire area of the wafer to protect the surface, and the wafer is cut into 45 chips. Individual chips are separated from the wafer, and the photoresist is stripped. Chips are then annealed in a vacuum at 425 °C for 15 min. The CNT FETs containing a single-bridging CNT are selected by inspecting the electrodes with scanning electron microscopy (SEM) and used in the experiments.

CVD-based CNT FET fabrication

CNTs are grown at predetermined locations in parallel by CVD. This process is enabled by patterning catalysts directly on thermally oxidized (285 nm SiO₂) silicon substrates with micrometre-scale precision using e-beam lithography (Nanobeam NB4). The growth of nanotubes is achieved by flowing a mixture of argon and hydrogen gas through ethanol and over the substrate at 890 °C (ref. 50). Positive-tone polymethyl methacrylate (Microchem) resists were assembled in a three-layer stack. The prepared photoresist trilayer was exposed to deep ultraviolet emission (Suss MA6 mask aligner). The initial metal (titanium) is deposited on the lithographic pattern of 79 parallel electrode pairs with a source–drain separation of 2.0 μm in each reticle (11 mm × 10 mm). The upper 30 nm platinum layer is applied on top of the wire bond pads and pseudo-reference electrode bars. CNT-transiting electrodes are passivated to protect from additional platinum overlay. After an overnight washing out of the platinum layer in acetone, the source–drain electrode gaps transiting the nanotubes are passivated again by patterning the isolated photoresist. After that, ectopic CNTs are eliminated by applying O₂ plasma. Subsequently, the protective passivation is removed with an overnight acetone strip at room temperature. The isolated and cleaned devices are then subjected to vacuum annealing at 425 °C for 15 min. The electrode pairs connected with a single CNT are selected by SEM imaging and used in the experiments.

SEM imaging

SEM images are taken with a Zeiss Sigma VP FESEM. Images are captured at all electrode sites to identify the number of contacted CNTs on an electrode pair.

Measurement set-up

The measurement systems are equipped with the previously developed homemade set-up^26,32. The measurement set-up consists of a polydimethylsiloxane (PDMS) microfluidic channel and a custom-made printed circuit board for the smFET data acquisition (Supplementary Fig. 2). The titanium electrode pairs with single bridging CNT and platinum pseudo-reference electrode bars on a chip are selected and connected to a land grid array package using an automated wire bonder (Kulicke & Soffa). The PDMS microfluidic channel was then stamped over the device array to allow liquid experiments. The channel was 7 mm long, 750 μm wide and approximately 500 μm deep. Two sterile tubing segments are connected to the channel inlet and outlet holes. A syringe pump connected to the outlet terminal withdraws fluid exiting the channel, thus controlling overflow rates. The circuit board contains 58 independent readouts that are simultaneously interrogated in real time. The circuitry for each readout incorporates tunable drain and source potentials. It comprises two gain stages: a front-end transimpedance amplification stage with a fixed resistive gain of 1 MΩ, followed by an inverting voltage amplifier with variable gain from 2× to 200×. Furthermore, each readout utilizes a second-order anti-aliasing filter topology, limiting the signal bandwidth to approximately 5.3 kHz. Readings from each device are sampled at a rate of 25 kilosamples per second. All measurements are performed at room temperature.

Transfer characterization in an aqueous solution (I_D–V_LG)

Transfer curves are measured in a PDMS microfluidic channel. The source–drain bias (V_DS) is chosen to be between 20 mV and 50 mV for each device to obtain the maximum source–drain current, I_max (I_D at V_LG = −0.5 V), reaching hundreds of nA. This initial source–drain bias is used for all measurements, including the 4-formylbenzene diazonium hexafluorophosphate (FBDP) functionalization experiment and the aptamer sensing experiment. For the transfer characterization, the liquid gate potential (V_LG) relative to the pseudo-reference platinum electrodes is tuned from −500 mV to 500 mV at intervals of 50 mV. The measurement time for the current is 1 s at each V_LG, and three successive I_D–V_LG sweeps are obtained to generate the average transfer curve.

Controlled single-point CNT functionalization chemistry

FBDP is synthesized using a previously established protocol³². FBDP is diluted in buffer solution (100 mM sodium phosphate buffer, pH 7.0), and a concentration between 1 μM and 100 μM was chosen to allow for real-time observations. The FBDP solution is applied to the device array through the microfluidic channel operated by a syringe pump. A constant source–drain voltage (V_DS) is applied during the experiment. The transfer curves are taken before and after the diazonium reaction. In addition, the time-domain I_D traces are monitored before, during and after the diazonium reaction to count and control the number of sp³ defect sites generated on the CNT sidewall. During the reaction, device conductance is continuously monitored while initially holding V_LG constant at −0.5 V. After that, the positive V_LG steps are applied until one discrete downward conductance step is detected. Then the V_LG is switched back to −0.5 V to halt further reactions. Device conductance is continuously monitored until the FBDP solution is thoroughly flushed from the device to ensure that the conductance remains unchanged without further defect formations.

Probe–aptamer conjugation

The aptamer probes specific to serotonin (5′-NH₂-CGA CTG GTA GGC AGA TAG GGG AAG CTG ATT CGA TGC GTG GGT CG-3′) and dopamine (5′-NH₂-CTC TCG GGA CGA CGC CAG TTT GAA GGT TCG TTC GCA GGT GTG GAG TGA CGT CGT CCC-3′) are synthesized with a 5′-amine termination²². Next, the aptamer probe is conjugated to the aldehyde end of the aryl groups on the CNT defect site by reducing Schiff’s base to a covalent bond. The reducing agent, sodium cyanoborohydride (NaBH₃CN, 200 μM), is added to the buffer solution containing 10 μM aptamer molecules. Then the solution is introduced to the functionalized device array at room temperature overnight. Next, PBS (pH 7.0) is used as the washing solution, which flows through the system at a rate of 25 μl min⁻¹ for 45 min to remove the weakly bound probe adsorbed on the CNT sidewall and microchannel walls.

Aptamer–neurotransmitter binding experiments

All experiments are performed in PBS buffer (pH 7.0) at room temperature. In addition, we performed control measurements in aCSF buffer (Supplementary Fig. 7). Neurotransmitter solutions (containing serotonin or dopamine) are freshly prepared with a concentration series between 0.1 nM and 1 μM. The time-domain I_D traces and transfer curves are recorded before, during and after the target introduction. The operating V_LG bias used in the I_D–t measurements is selected based on the transconductance characteristics of unfunctionalized CNT FETs. Therefore, these operating conditions varied between devices (Supplementary Table 1). First, the concentration series samples are introduced to the aptamer-functionalized device array. The devices are then incubated in the solution for 30 min before the measurement, allowing for sufficient diffusion of small molecules in the microfluidic channel (Extended Data Fig. 5). After the measurement, the existing sample solution is flushed away, and the microchannel is thoroughly rinsed with the buffer. Then, the following concentration series is introduced to the device array. The raw I_D–t data sampled at 25 kHz are filtered with a finite-impulse-response filter to 100 Hz with five times oversampling for the data analysis. The baseline drift and conductance fluctuation throughout the measurements are also corrected using the Whittaker baseline correction algorithm⁵¹ and iterative event finder⁵² using a custom Python toolchain (Supplementary Discussion).

State identification and dwell-time analysis

The preprocessed data are further analysed with a custom Python program to fit a two-state HMM using the Baum–Welch algorithm (Supplementary Discussion). Sections of the current trace are classified into high- and low-conductance states, and respective dwell times are extracted using the forward/backward maximum-likelihood algorithm. Dwell-time distributions for high and low states are transformed to log space⁵² and fitted to exponential distribution using a nonlinear least-squares algorithm. To estimate confidence intervals (0.05 and 0.95) associated with the fitted parameters, the fitting is conducted with a bootstrapping method with bootstrap samples (n = 2,000). The probability of occupying a low-current state (P_low) and a high-current state (P_high) and the respective confidence intervals (0.05 and 0.95) are estimated with bootstrapping (n = 2,000) based on the sum of individual state dwell times.

HMM analysis

In addition to the simple two-state HMM, more elaborate HMMs are fitted to the current traces of each device. The preprocessed data are fitted again using the Baum–Welch algorithm to various states of HMMs using 80% of the measurement data (Supplementary Discussion). The model fit is evaluated on the remaining 20% of data to calculate a BIC score for each model solution. The 4S-D state model gave the best overall BIC score across all devices, and concentrations above 100 pM serotonin are selected for any subsequent kinetic analysis of the CNT–aptamer system (Supplementary Fig. 10 and Supplementary Table 3). Rate constants based on the HMM are extracted from dwell-time distributions for each state transition fitted in semi-log scale using Python Lmfit module for a single-exponential distribution. Confidence values reported were extracted from the fitting routine. The individual HMM state occupations, P₀, P₁, P₂ and P₃, and the associated confidence intervals were calculated by bootstrapping (n = 2,000) the average dwell times in each state divided by the total trace time.

Statistics and reproducibility

All relevant information is stated in the figure captions. No statistical method was used to predetermine sample size.

Extended Data

Extended Data Fig. 1 |. SEM images and transfer characteristics of two types of CNT FET fabrications used in the study.

The number of CNTs crossing the source–drain electrode pair is counted on the SEM images. Conducting FET devices transiting a single CNT are screened by back-gate transfer characteristics using a semiconductor parameter analyser (Agilent 4155 C). The back-gate voltage (V_BG) is applied to the heavily doped p + + silicon substrate and swept from −10 to +10 V with a step size of 500 mV. Constant source–drain voltage (V_SD) of 100 mV is applied to the devices, and the current at each V_BG is measured. a) The representative SEM images of a CNT FET produced by a CVD method. CNT nucleation features (4 μm by 4 μm) containing iron nanoparticles are patterned on the top of silicon substrates by electron beam lithography. CNTs with controlled density and length are grown from the sites by the CVD method (left). The Ti electrode pairs with a 2 μm gap are deposited on the defined CNT growth sites (right). Each chip (11 mm by 10 mm) contains 79 pairs of source–drain electrical contacts. b) The two representative SEM images of CNT FETs fabricated by spin-casting. After spin-coating the CNT solution (250 μg/ml) on the four-inch wafer, Ti electrode pairs with source–drain separation of 0.5 μm are patterned on top. Each chip (11 mm by 10 mm) contains 280 pairs of source–drain electrical contacts. c) The representative transfer curves of 22 CNT FETs, including a single-crossing CNT produced by the CVD-grown method. d) The transfer curves of 26 single-crossing spin-cast CNT FETs. Devices that have a single CNT bridge and exceed threshold I_D values of 1 nA at −10 V of V_BG are used in the subsequent experiments.

Extended Data Fig. 2 |. Electrically controlled diazonium chemistry for single-point CNT functionalization.

a) The example of I_D–t trace (black line) displaying serial single sp³ defect generations by aryl-diazonium salt at a fixed liquid gate potential (V_DS = 30 mV, V_LG = −200 mV, red line). Discrete downward current steps are detected (Inset, red arrows), and the I_D level is rapidly reduced after introducing 4-formylbenzene diazonium hexafluorophosphate (FBDP) solution (shaded in blue). b) The representative I_D–t trace (from Device A, which is discussed in the main text) displays controlled diazonium chemistry (V_DS = 50 mV). At −500 mV of V_LG, the downward current steps are not observed in the FBDP solution. The discrete current steps are detected when the VLG is tuned to −300 mV. The two I_D steps cause a resistance change of 4.7 kΩ and 5.1 kΩ, respectively. c) The diazonium cation is reduced by electrons supplied from CNT; subsequently, aryl-radical is coupled with CNT. The V_LG is then turned down to −0.5 V to reduce the electron density in CNT, halting the reaction. d) Examples of transfer curves of Device A before (grey squares) and after the controlled diazonium reaction (red circles) and after DNA aptamer conjugation (blue triangles). All transfer curves are obtained in the phosphate buffer. Error bars represent a standard deviation of the mean current value for three measurements. The measurement V_LG point (100 mV) for temporal dynamic serotonin sensing is chosen from the curve. After the functionalization, the average conductance is decreased, indicating that permanent DNA-CNT couplings are created on the sp³ defect site on CNT.

Extended Data Fig. 3 |. The temporal electric signals originated from the single-molecule binding events observed on serotonin-specific aptamer-functionalized smFET devices.

a-f) 30-s-length I_D-t traces displaying single-molecule events (displayed in the increasing order of liquid gate potentials). The right panel shows the I_D distributions of each time trace, and two Gaussian fits with peak values μ₁ and μ₂. (a) Device A, V_LG = 200 mV, V_DS = 50 mV. (b) Device B, V_LG = 100 mV, V_DS = 50 mV. (c) Device C, V_LG = 300 mV, V_DS = 25 mV. (d) Device D, V_LG = 400 mV, V_DS = 50 mV. (e) Device E, V_LG = −200 mV, V_DS = 50 mV. (f) Device F, V_LG = 0 mV, V_DS = 200 mV. g) Measured SNR values for six devices $\left(SNR=\left({\mu }_1-{\mu }_2\right)/\sqrt{0.5\left({\sigma }_1^2+{\sigma }_2^2\right)}\right)$.

Extended Data Fig. 4 |. Concentration dependence for the fraction of time spent in the lower conductance state (P_LOW) of Device B (a) and Device F (b).

The plots of P_LOW are fitted into Langmuir isotherm function (black line). Data points are the mean probability of the low conductance state calculated from all dwell times by bootstrapping (N_boot = 2,000). Error bars represent the 90% confidence interval from bootstrapped mean value of P_LOW.

Extended Data Fig. 5 |. Histograms showing temporal conductance profiles of smFET device in response to the target and non-target introduction.

a) The experimental sequence of introducing serotonin (5-HT) and dopamine solutions to the device (Device D used for the experiments). b) Histograms of I_D distributions before target introduction. c-i) Histograms of I_D distributions displayed chronologically after each sample is introduced. 600-s of I_D-t trace was sampled at each recording point. 400 mV of V_LG was applied during the experiment. The data recording time after introducing blank PBS buffer (b), serotonin (c-f), and dopamine-diluted PBS solution (g-i) is indicated in each graph.

Extended Data Fig. 6 |. Control experiments on the aptamer-conjugated smFET devices.

I_D–t traces measured on serotonin-specific aptamer-functionalized CNT FET (Device B) measured in the presence of pure buffer, dopamine (100 nM), and serotonin solution (50 nM), respectively.

Extended Data Fig. 7 |. Control experiments on the unfunctionalized CNT FET.

a) I_D–t trace of an unmodified CNT FET (Device G), measured in presence of pure buffer solution. b) Response of the device to a 10 nM serotonin solution in the buffer. c) Device response at a higher concentration of 50 nM serotonin in the buffer. d) Graph showing the average I_D values measured over 10 minutes at varying serotonin concentrations displaying the extent of the non-specific charge adsorption. Standard deviations of the I_D measurements are represented by error bars.

Extended Data Fig. 8 |. Normalized dwell time distributions and concentration dependency of the high and low conductance states.

The semi-logarithmic dwell time histograms for the low conductance state (left column) and high conductance state (right column) at various serotonin concentrations are shown. The counts are presented per second to account for variations in measurement time between experiments. a, b) Normalized dwell time distributions for the low and high conductance states obtained from Device A (a) and Device B (b). The histograms are overlaid with single-exponential (red lines) and double-exponential decay (orange lines) functions. The double-exponential fits reveal two distinct populations characterized by fast and slow rates, represented by orange dotted lines. c) Concentration-dependent number of counts of smFET transitions for Device A (left) and Device B (right). The plot illustrates the relationship between the number of counts per second and the serotonin concentration, addressing the concentration dependency of the smFET transitions for each device.

Extended Data Fig. 9 |. The rate constants determined by HMM analysis.

a) The dwell time analysis with HMM showing single-exponential distribution (data from Device A, 50 nM of [5-HT]). b) The HMM includes eight interconversion rates. c) The two rate constants (on-rate: pink, off-rate: green) calculated from the HMM at each pathway are plotted as a function of serotonin concentrations. The smFET data of Device B are used. Data points are calculated from fitting dwell-time distributions for each state transition to a single-exponential distribution. The error bars indicate the one-sigma confidence interval of the model fit.

Data availability

Source data for all figures are provided with this paper. The measured raw data for all smFET devices are published on https://doi.org/10.5281/zenodo.10161590. Source data are provided with this paper.