Chiral discrimination of amino acids in meteorite and desert soil extracts via single-molecule nanogap conductance

Abstract

Life on Earth uses L-amino acids and D-sugars, whereas abiotic processes produce equal abundances of both enantiomers. Detecting enantiomeric excess among molecules of biological interest has therefore been proposed as a potential biosignature for life beyond Earth. However, robust electrical methods for determining the L/D ratios of amino acids and sugars at low concentrations, particularly in chemically complex samples, remain limited. Here we show that quantum tunneling currents electrically distinguish individual L- and D-amino acids and quantify their ratios at the single-molecule level without optical measurements or chiral recognition molecules. Individual enantiomers are distinguished with over 80% accuracy, and a blind test of 39 amino acids identifies both molecular identity and chirality with >50% accuracy, well above random chance. Mixtures containing four amino acids with different chiral compositions are quantified with >66% accuracy. Murchison meteorite and Atacama Desert extracts are analyzed using a targeted panel of 11 amino acids. Confidence-based filtering distinguishes high-confidence target assignments from low-confidence events, and the resulting compositions capture major compositional features observed by LC–MS while highlighting the effects of background and low-abundance signals. These results establish a targeted single-molecule electrical approach for chiral amino acid analysis and provide a foundation for future applications to study chemically complex samples, including those relevant to astrobiological research.

Chiral amino acids are important biomarkers in the search for life beyond Earth, but their analysis remains challenging. Here, the authors demonstrate single-molecule electrical identification of amino acid chirality and apply it to Murchison meteorite and Atacama Desert samples.

Introduction

Chiral molecules are chemically identical mirror images of each other that cannot be superimposed. Amino acids in biological systems are heavily biased toward their L-form^1–4, whereas sugars are heavily biased toward their D-form⁵. Because abiotic processes generally produce racemic mixtures, enantiomeric excess has been widely investigated as an indicator of biological and prebiotic processes. Accordingly, L-enantiomeric excess (L_ee) is routinely measured in extraterrestrial samples to help distinguish terrestrial contamination from indigenous organics. While recent work suggests possible abiotic routes to enantiomeric excess⁵, it has also been proposed as one potential biosignature for life beyond Earth^6,7. Robust analytical methods capable of quantitatively determining the chirality of trace amounts of amino acids and sugars are therefore of broad interest, including for the analysis of samples returned from space exploration missions.

Chiral isomers have identical chemical properties, so they generally cannot be distinguished using inherent differences in reactivity. There are two main methods for the chiral discrimination of samples: using polarized light to determine differences in optical properties and using molecular recognition to identify differences in conformation, including chirally-selective reactions⁸. Enantiomers can be physically separated by using a medium or sensor coupled to a chiral molecule. Recently, chiral identification has been performed via surface-enhanced Raman scattering (SERS)^9–11 and the interaction of chiral molecules with molecularly modified surfaces of bionanopores^12–14.

Chiral discrimination has attracted considerable interest because enantiomeric excess can provide important information about the origin and history of organic molecules and has been proposed as one potential biosignature⁵. While a threshold of >20% enantiomeric excess in multiple amino acids has been proposed as indicative of life⁷, earlier analyses of meteoritic amino acids reported L_ee > 30% for alanine, serine, leucine, aspartic acid, and glutamate, among others⁵, but these results are now generally considered to reflect terrestrial contamination. More recent NASA and JAXA asteroid sample-return missions have detected no significant enantiomeric excess (within experimental error), supporting racemic compositions of abiotic amino acids. Thus, reliable chiral analysis remains important for characterizing organic samples, particularly when interpreted together with complementary evidence.

Applications requiring compact, highly sensitive analytical systems—including future space exploration missions—place stringent demands on molecular detection technologies¹⁵. For example, measurements of trace organic molecules in extraterrestrial samples may require detection limits approaching ≈1 nM¹⁶, which is far lower than the amino acid abundance reported in meteorites, including those of Martian origin¹⁷. Direct electrical detection can simplify detector design compared to optical detection. Electrical systems do not require optics, can reduce payload size, and may reduce sensitivity to vibration. It is also generally desirable to avoid chemical or enzymatic processing steps, such as labeling reactions, to minimize the risk of false negatives and limit sample preparation complexity. Therefore, a measurement system capable of electrically detecting and identifying chiral molecules at low concentrations, ideally at the single-molecule level and without using recognition molecules, would be a valuable analytical platform for chemically complex samples and may prove useful for future astrobiological applications.

Recent developments in nanotechnology have made it possible to electrically detect and identify individual molecules of biological relevance^18–32. Here, we use an adjustable gap between two atomically pointed gold nanowires—formed with high stability and sub-nanometer precision using a mechanically controllable break-junction (MCBJ)—to monitor molecular identity. The electron tunneling current across the gap yields a current versus time plot, and a single molecule diffusing through the gap results in a characteristic waveform³¹. This approach directly counts the number of molecules diffusing through a gap, and has been used to identify and quantify nucleotides, polynucleotides, and amino acids^33–35. Developments in machine learning and artificial intelligence are enabling highly accurate measurements with electron tunneling nanogaps^36–38. If single-molecule counting can be combined with chiral identification, then the resulting approach will automatically determine enantiomeric ratios.

The identification of chirally distinct molecules requires a chirally defined detecting material—in this case the gold nanowire tips. Recent advances have enabled the development of chirality on solid surfaces^39–44, and crystal facets with a high mirror index enable chiral definition of gold nanoparticles^39–42. Further, the method of recognizing target chiral molecules with other chiral molecules has been extended from organic to inorganic materials. Atomically sharp solid structures, where high Miller indices tend to be formed, raise the possibility of identifying individual chiral molecules. The combination of atomically pointed gold electrodes, single-molecule counting, and artificial intelligence has enabled the identification of molecular chirality at the single-molecule level. Between amino acids and sugars, we focused on the chiral identification of amino acids, which are commonly subjected to single-molecule measurements^36–38.

Results

Single-molecule detection of L- and D-form amino acids

Single-molecule conductance measurements of amino acids were performed using gold nanogap electrodes formed at a mechanically controllable break junction (MCBJ) (Fig. 1a)^33,45. 0.1 μM aqueous amino acid solutions were measured at a voltage of 0.1 V DC. Because the protonation states and net charges of amino acids depend on the solution pH relative to their isoelectric points, all measurements were performed in 1× PBS at pH 7.4 to ensure consistent solution conditions. A SiO₂ layer was introduced around the gap junction to reduce electrical noise and define the nanochannel geometry. The nanogap was generated in situ by mechanically breaking the Au nanowire after device fabrication. As a result, gold is exposed at the fracture interfaces. The distance between the nanogap electrodes was controlled to be 0.56 and 0.58 nm during measurements. Nineteen of the 20 amino acids exist in both the L- and D-forms, with glycine being the exception due to its achirality. The total number of amino acids tested in this analysis was therefore (19 × 2 + glycine =) 39 (Table 1, Supplementary Note 1). In this experiment, amino acids interacted with the nanogap via diffusion.

Fig. 1. Schematics of chirality analysis for amino acid molecules with a nanogap device.

a As an amino acid transiently passes through the nanogap, it modulates the tunneling current and generates a characteristic current–time waveform. L- and D-enantiomers can produce distinguishable waveforms, potentially reflecting differences in transient molecule–junction interactions, molecular orientation, and local interfacial effects. b Single-molecule measurement system. The single-molecule measurement device is on the right side of the monitor. c Optical microscope image of the gold electrodes forming the nanogap. d Scanning electron microscope image of the nanogap between the gold electrodes. The images for c and d are representative of three independently fabricated devices.

Table 1.

Properties of amino acids

Amino acid	Molecular length (nm)	pI	td (ms)		Frequency (min-1)		F1-score
			L-form	D-form	L-form	D-form
Ala	67	6.00	3.58 ± 0.64	3.65 ± 0.63	197 ± 39.9	108 ± 16.4	0.814
Arg	148	10.8	3.69 ± 0.55	3 ± 0.49	160 ± 18.8	177 ± 10.9	0.955
Asn	96	5.41	4.2 ± 0.70	3.24 ± 0.43	84.1 ± 5.0	162 ± 25.5	0.860
Asp	91	2.77	4.56 ± 0.76	4.09 ± 0.57	40.1 ± 4.5	46.6 ± 10.4	0.919
Cys	86	5.07	3.34 ± 0.53	4.7 ± 0.81	133 ± 28.0	101 ± 13.0	0.857
Gln	114	5.65	4.31 ± 0.71	4.06 ± 1.14	58.6 ± 10.4	115 ± 13.6	0.784
Glu	109	3.22	3.43 ± 0.50	3.51 ± 0.50	131 ± 31.3	95.6 ± 6.5	0.876
His	118	7.59	4.49 ± 0.84	3.07 ± 0.46	111 ± 16.0	399 ± 56.9	0.901
Ile	124	6.02	2.96 ± 0.41	3.57 ± 0.60	443 ± 61.4	67.0 ± 14.8	0.974
Leu	124	5.98	3.92 ± 0.69	3.61 ± 0.54	61.1 ± 15.4	97.2 ± 9.6	0.890
Lys	135	9.75	5.28 ± 0.69	3.94 ± 0.65	46.1 ± 8.9	108 ± 7.6	0.792
Met	124	5.74	3.70 ± 0.60	3.42 ± 0.47	109 ± 7.2	105 ± 11.0	0.940
Phe	135	5.48	3.68 ± 0.67	3.81 ± 0.78	184 ± 8.9	102 ± 13.6	0.875
Pro	90	6.30	4.13 ± 0.82	3.48 ± 0.52	37.5 ± 3.9	87.9 ± 7.0	0.875
Ser	73	5.68	4.01 ± 0.67	3.6 ± 0.63	67.2 ± 4.1	159 ± 16.0	0.894
Thr	93	6.16	4.27 ± 0.69	4.41 ± 0.73	60.4 ± 12.9	68.7 ± 16.5	0.851
Trp	163	5.89	3.99 ± 0.68	3.59 ± 0.62	44.3 ± 8.3	175 ± 17.7	0.827
Tyr	141	5.66	4.02 ± 0.65	3.75 ± 0.84	207 ± 13.9	207 ± 23.3	0.869
Val	105	5.96	4.09 ± 0.67	5.54 ± 0.85	78.6 ± 5.4	63.4 ± 38.8	0.883
Gly	48	5.97	4.38 ± 0.89		82.7 ± 13.6		-

pI indicates the isoelectric point. Molecular length indicates the maximum interatomic distance within a molecule.

Measurements of aqueous solutions containing individual amino acid species showed spike-like signals of increased current against time (Fig. 2a, b, and Supplementary Figs. 1–5). Each spike-like signal can be characterized by the current height (I_p) and current duration time (t_d) (Fig. 2c). The current obtained is a tunneling current that flows through a single molecule present in the nanogap between the electrodes. The histograms of I_p for the L- and D-forms of tryptophan exhibit maxima at 35.7 and 36.9 pA, respectively (Fig. 2d). Of the 20 amino acids, tryptophan has the largest π-electron system. Because a π-electron system can easily transport electrons, tryptophan has the highest single-molecule conductance. The two histograms for chiral isoforms of tryptophan almost overlap. Histograms of t_d for the L- and D-forms of tryptophan with t_d peaks at 3.55 and 3.44 ms also exhibit a near overlap (Fig. 2g, h, Supplementary Fig. 6).

Fig. 2. Discrimination of L-tryptophan and D-tryptophan.

a Current–time profile for L-tryptophan. b Current–time profile for D-tryptophan. c Enlarged signal for L-tryptophan. I_p and t_d denote the peak current and signal duration, respectively. d Histogram of peak currents for L-tryptophan (orange) and D-tryptophan (green). e Correlation between signal frequency and t_d. Blue and red symbols represent L- and D-form amino acids, respectively. f Dependence of tryptophan signal frequency on concentration. Blue dotted lines indicate linear fits used to estimate the limit of detection (LOD) and limit of quantification (LOQ). Data are presented as mean ± s.e.m. from eight independent measurements (n = 8). Individual data points are shown as open circles, and squares with error bars represent the mean ± s.e.m. of eight independent measurements (n = 8). A broken y-axis is used to display zero-frequency measurements separately from the logarithmically scaled positive values. g, h Averaged conductance heatmaps constructed from all detected tunneling events of L-tryptophan (g) and D-tryptophan (h). Warmer colors indicate a higher event density, whereas cooler colors indicate a lower event density. Each heatmap was generated by normalizing the duration of each tunneling event from its start to its end, dividing the normalized signal duration into 12 equal segments, and plotting the average conductance normalized to the peak current of each event. Details of signal detection and feature extraction are described in the Methods section. i Confusion matrix for L-tryptophan and D-tryptophan. Values indicate the percentage of classified waveforms. The color scale represents the percentage of waveforms assigned to each class. Source data are provided as a Source Data file.

Amino acids with different isoelectric points have different net charges, molecular lengths, and molecular weights (Table 1). These physical quantities may affect their electron flow dynamics. In other words, these physical quantities may affect $t_d$ and signal frequency. The correlations among the isoelectric point, molecular length, molecular weight, $t_d$, and signal frequency were therefore investigated, and a clear negative correlation was observed between $t_d$ and signal frequency (Fig. 2e and Supplementary Fig. 7). To further interpret these observations, let us assume a model in which amino acids adsorbed on the nanogap electrode surface diffuse across the nanogap electrode surface to reach the sensing site. In this model, molecules with larger $t_d$, that is, those that are less likely to move on the electrode surface, are less likely to be observed. The experimental results are consistent with this model. These observations suggest that the local nanogap environment may contribute to chiral sensing. Possible contributions include local surface structure, molecular orientation, transient adsorption or coordination, and electrostatic interactions, although their relative contributions remain to be determined.

The dependence of tryptophan detection frequency on concentration was assessed to determine the limit of detection (LOD) of the single-molecule measurement system (Fig. 2f). The measurement time at each concentration was 15 min. The standard deviation of the signal frequency at each concentration was less than 0.27 min⁻¹. As the mean signal frequency of the blank was 2.1 ± 0.27 min⁻¹, the LOD and limit of quantification (LOQ) were determined to be 0.03 and 0.2 nM, respectively, from the calibration curve. At high concentrations (above 10 mM), the signal frequency no longer depended on the concentration of amino acids because the amino acids had reached saturation adsorption. A target detection limit of ≈1 nM has been proposed for life-detection measurements in planetary exploration¹⁶, and reported amino-acid abundances in some meteorites exceed this concentration¹⁷. The calibration-based LOD obtained here indicates that the nanogap system is sensitive under controlled measurement conditions. However, this value should not be interpreted as the practical identification or quantification limit in chemically complex environmental matrices, where process-derived background signals, nontarget molecular species, and confidence-based event rejection can substantially affect the number of retained target assignments. Establishing practical performance in such samples will require matrix-matched calibration and contamination-controlled validation experiments.

The L- and D-forms of tryptophan exhibited largely overlapping distributions in their Iₚ and t_d histograms, indicating that these parameters alone are insufficient to clearly distinguish between the two enantiomers (Fig. 2e, f).

However, this overlap suggests that additional waveform features should be considered. I_p and t_d do not capture all the information contained in a waveform. Complete density maps of the current-time waveforms for the L- and D-forms of tryptophan show significant differences, with the L-form waveform more spread out than the D-form waveform (Fig. 2g, h). Thus, the L- and D-forms of tryptophan yield distinct current–time waveforms. Across all amino acids, the standard deviations of the tunnel current distributions ranged from 16.9–26.1 for D-forms and 17.6–21.6 pA for L-forms, respectively. Although both distributions partially overlap, the two-dimensional density maps (Fig. 2g, h) show distinct regions of higher event probability for each enantiomer, reflecting characteristic differences in waveform spread and amplitude distribution. We conclude that waveform features enable discrimination between L- and D-forms of tryptophan.

The single-molecule conductance (I/V) is influenced by the Fermi level of the electrode, the highest occupied molecular orbital (HOMO) or lowest unoccupied molecular orbital (LUMO) energy level of the amino acid, and the interaction between the molecule and the junction^46,47. The HOMO and LUMO energies of L- and D-form amino acids are identical, and the nominal Fermi energies of the electrodes are expected to be equivalent. The observed waveform differences are therefore consistent with differences in transient molecule–junction interactions during transport through the nanogap. Possible contributions from local junction structure, molecular orientation, transient adsorption or coordination, and electrostatic interactions cannot be excluded. Further experimental and theoretical studies will be required to clarify the molecular and interfacial mechanisms underlying chiral discrimination.

Discrimination of L- and D-form amino acids

The L- and D-forms of tryptophan were measured independently, but we sought to determine whether individual traces could be correctly identified. A total of 25143 and 46129 current-time waveforms of the L- and D-form were extracted by our analysis software. In addition to I_p and t_d, a vector dividing each waveform into twelve equal parts in the time direction was used as a feature to characterize a given waveform (Supplementary Note 2). The measured data for the D and L forms of tryptophan constituted one data set each (Supplementary Fig. 9, Supplementary Table 1). To reduce the training bias, the data in the two data sets were undersampled, and the number of waveforms in each data set was standardized to the same number. The balanced dataset was randomly divided into training (80%) and held-out test (20%) datasets. Model training and validation were performed using 10-fold cross-validation on the training dataset, with 9 folds for training and the remaining fold for validation in each round. Classification performance was then evaluated on the held-out test dataset using the F₁-score and confusion matrix. L- and D-forms of tryptophan were successfully identified from individual waveforms with an F₁-score of 0.827 (Fig. 2i), substantially higher than the value of 0.5 expected for random binary classification.

Classifier performance is evaluated using the cross-validation and hold-out methods. In cross-validation, the dataset is repeatedly divided into training and testing subsets to assess model generalizability, whereas in the hold-out method, a fixed portion of the data is reserved for testing performance. The cross-validation method is commonly used to reduce training bias. However, in real-world use cases, training and test data are completely separate. For example, a classifier is trained on Earth-measured amino acid data, and a single-molecule measurement system implementing the trained classifier would then be transported to the Martian surface. Samples obtained from Mars regolith would be measured to determine which amino acids are present. Training and independent test data for the D- and L- forms of tryptophan measured using the different nanogap chips were prepared to determine the discrimination accuracy using the hold-out method, where the training and test data are completely separate (Supplementary Fig. 9). The numbers of D- and L-form waveforms in the training data were 7511 and 6158, respectively, whereas those in the test data were 5970 and 4889, respectively. Undersampling was performed in both the training and test processes. The hold-out method gave F₁−scores of 0.816 and 0.859 for the training and test processes, respectively (Supplementary Fig. 9). Thus, the cross-validation and hold-out methods gave similar discrimination accuracies.

The L- and D-forms of the other 18 amino acids were characterized using the same procedure as for tryptophan. The lowest F₁-score was 0.784 for glutamine, the highest F₁-score was 0.974 for isoleucine, and the average F₁-score was 0.876 ± 0.049 (mean ± standard deviation). (Fig. 3a, Table 1). These obtained F₁−scores indicate the accuracy of discriminating two molecules using a single current–time waveform. In a real use case, one solution is used to identify molecule A or B. The 5 min measurement of a single solution yields more than several hundred current–time waveforms (Table 1). If the identification accuracy with one waveform is $p,$ and $n$ waveforms are obtained, then the identification accuracy for one solution follows the binomial distribution ${\sum }_{n/2}^n{C}_t{p}^t{\left(1-p\right)}^{n-t}$. For $p$ = 0.9 and $n$ = 16, the identification accuracy is 99.9%.

Fig. 3. Chiral discrimination for each individual amino acid molecule.

a Confusion matrix for the L- and D-forms of nineteen amino acids. b Dependence of discrimination accuracy on amino acid species and nanogap distance. Large and small gaps were 0.56 and 0.54 nm, respectively. In the graph, orange and green indicate 0.54 nm and 0.56 nm gaps, respectively. c Correlations between discrimination accuracy and molecular length and between discrimination accuracy and molecular weight. $F_{\mathrm{small}}$ and $F_{\mathrm{large}}$ denote the F₁−scores at nanogap electrode distances of 0.54 and 0.56 nm, respectively. The color scale denotes the isoelectric points (pI) of amino acids. d Confusion matrix for L- and D-forms of 19 amino acids and glycine. The value in each matrix position indicates the percentage of waveforms. The color scale represents the percentage of waveforms assigned to each class. Source data are provided as a Source Data file.

The observed discrimination may depend on the interaction between the amino acid and the local nanogap environment during transient transport. We therefore examined the effect of interelectrode distance on classification accuracy. Changes in the interelectrode distance can alter the available molecular configurations, electronic coupling, local electric field, and the probability of transient interfacial interactions^46,47. Nineteen chiral amino acids were measured with a reduced interelectrode distance of 0.54 nm, instead of 0.56 nm. For 14 of the 19 amino acids, the accuracy of discriminating L and D forms using the 0.54 nm nanogap was higher than that using the 0.56 nm nanogap (Fig. 3b). The Thr and Pro discrimination accuracies were higher when the 0.56 nm nanogap was used. The discrimination accuracies for Val, Met, and IIe were similar for both nanogaps. The dependence of discrimination accuracy on interelectrode distance indicates that the local junction environment influences chiral classification. However, the present data do not resolve the relative contributions of electronic coupling, molecular orientation, steric confinement, transient adsorption or coordination, and electrostatic interactions.

The electronic interaction between the nanogap electrode and the amino acid showed only a weak correlation with molecular length and mass, suggesting that molecular size alone does not fully explain the observed variation in discrimination accuracy. However, the present measurements do not allow the contribution of molecular geometry to be separated quantitatively from local junction effects. The correlation coefficients between the discrimination accuracy and molecular length at interelectrode distances of 0.54 and 0.56 nm were 0.17 and −0.02, respectively, whereas those between the discrimination accuracy and molecular weight at interelectrode distances of 0.54 and 0.56 nm were 0.20 and 0.00, respectively (Fig. 3c). The isoelectric point did not correlate with the interelectrode distance, molecular size, or molecular weight. The discrimination accuracy at the interelectrode distance of 0.54 nm was positively correlated with the molecular length and molecular weight. Therefore, in a space with a more restricted orientation of amino acids, the larger the molecular length and molecular weight, the more susceptible the molecules are to orientation restrictions.

We next performed single-molecule identification of 39 individual amino acid species. We tested the ability of the nanogap to correctly identify L- and D-forms of 19 different amino acids, and glycine (Gly). Machine learning with the features and classifiers used in the identification of tryptophan showed F₁-score = 0.49 (Fig. 3d). This value is much higher than the F₁-score of 0.026 (1/39) expected for random detection. While the accuracy of identifying a single waveform for each amino acid may be limited, aggregating counts over multiple waveforms enables reliable discrimination. Therefore, each of the 39 amino acids tested can be identified with individual current-time waveforms.

Single-molecule analysis of mixed amino acid solutions

Single-molecule counting of chiral species makes it possible to determine the abundance ratios of different amino acid species in a sample. The dataset of single-molecule conductances of individual amino acids obtained above enables us to classify the current-time waveform obtained from the measurement of a mixed solution of several amino acids (Fig. 4a).

Fig. 4. Discrimination of individual amino acids in a mixed solution containing L-form and D-form species.

a Workflow for the identification and quantification of amino acids in mixed solutions using single-molecule electrical measurements. Signal events were extracted from current–time profiles, and a classifier was trained using datasets obtained from individually measured amino acids. The trained classifier was then applied to identify amino acid species in mixed samples, and the relative abundance of each amino acid was estimated by counting the classified events. b Current–time profile obtained from a mixed solution containing D-tryptophan (D-Trp), D-threonine (D-Thr), D-asparagine (D-Asn), L-methionine (L-Met), and glycine (Gly). Each amino acid was present at an equal molar ratio (20%). c Enlarged current–time profile corresponding to (b). d Confusion matrix for a mixed solution containing D-tyrosine (D-Tyr), D-leucine (D-Leu), D-proline (D-Pro), and L-phenylalanine (L-Phe). Values indicate the percentage of classified waveforms. The color scale represents the percentage of waveforms assigned to each class. e Quantitative analysis of the 1:1:1:1 D-Tyr:D-Leu:D-Pro:L-Phe mixture. f Confusion matrix for a mixed solution containing D-Tyr, L-histidine (L-His), L-Phe, and L-serine (L-Ser). Values indicate the percentage of classified waveforms. The color scale represents the percentage of waveforms assigned to each class. g Quantitative analysis of the 1:1:1:1 D-Tyr:L-His:L-Phe:L-Ser mixture. h Confusion matrix for a mixed solution containing D-histidine (D-His), L-His, L-alanine (L-Ala), L-glutamine (L-Gln), and L-Ser. Values indicate the percentage of classified waveforms. The color scale represents the percentage of waveforms assigned to each class. i Quantitative analysis of the 1:1:1:1:1 D-His:L-His:L-Ala:L-Gln:L-Ser mixture. j Quantitative analysis of the 3:1:1.5:1.5:1.5 D-His:L-His:L-Ala:L-Gln:L-Ser mixture. Source data are provided as a Source Data file.

Aqueous solutions with molar mixing ratios of D-Tyr:D-Leu:D-Pro:L-Phe = 1:1:1:1 were characterized as above. We obtained current-time waveforms showing the single-molecule conductance of individual amino acids (Fig. 4b, c). The machine learning classifiers were trained using 3800 extracted current-time waveforms. Four amino acids were identified with an F₁-score = 0.658 (Fig. 4d). The number of current-time waveforms identified as D-Tyr, D-Leu, D-Pro, and L-Phe were 10907, 6195, 9121, and 11781, respectively. The proportions of each amino acid waveform relative to all current-time waveforms were 28.7%, 16.3%, 24.0%, and 31.0% (Fig. 4e). The error frequencies, defined as |actual value − predicted value | /actual value, were 14.8%, 34.8%, 4.0%, and 24.0%, respectively (Table 2).

Table 2.

Quantitative analyses of mixed solutions

Sample name	F1-score	Molecule	Actual ratio (%)	Predicted ratio (%)
D-Tyr:D-Leu:D-Pro:L-Phe	0.658	D-Tyr	25	28.7
		D-Leu	25	16.3
		D-Pro	25	24.0
		L-Phe	25	31.0
D-Tyr:L-His:L-Phe:L-Ser	0.748	D-Tyr	25	24.9
		L-His	25	25.8
		L-Phe	25	14.8
		L-Ser	25	34.3
D-His:L-His:L-Ala:D-Gln:L-Ser	0.736	D-His	20	22.0
		L-His	20	22.9
		L-Ala	20	18.0
		D-Gln	20	15.0
		L-Ser	20	22.1
D-His:L-His:L-Ala:D-Gln:L-Ser	0.736	D-His	35.3	43.0
		L-His	11.8	15.4
		L-Ala	17.6	7.8
		D-Gln	17.6	13.4
		L-Ser	17.6	20.5
D-His:L-His:D-Ala:L-Ala:Gly:dGMP:dTMP	0.761	D-His	14.9	15.6
		L-His	14.9	10.5
		D-Ala	22.8	28.1
		L-Ala	22.8	18.3
		Gly	22.8	25.4
		dGMP	1.0	0.7
		dTMP	1.0	1.4

Aqueous mixtures of D-Tyr, L-His, L-Phe, and L-Ser in equimolar ratios were also measured by single-molecule counting to determine their ratios. The four amino acids were identified with F₁-score = 0.748 (Fig. 4f). The abundance ratios of D-Tyr, L-His, L-Phe, and L-Ser were estimated to be 24.9%, 25.8%, 14.8%, and 34.3%, respectively (Fig. 4g and Table 2), again very close to the known ratios.

The same approach on mixtures of five amino acids, including L- and D-form amino acids, resulted in an identification accuracy of 0.736 (Fig. 4h and Table 2). Aqueous solutions containing D-His, L-His, L-Ala, L-Gln, and L-Ser at a total amino acid concentration of 100 nM were prepared at mixing ratios of 1:1:1:1:1 and 3:1:1.5:1.5:1.5. The 1238 and 786 current–time waveforms obtained from each measurement were classified into D-His, L-His, L-Ala, L-Gln, and L-Ser via machine learning. The measured (classified) proportions were 1:1.0:0.8:0.8:1.0 (Fig. 4i) and 3:1.0:0.6:0.9:1.4 (Fig. 4j), respectively. These mixing ratios were chosen to test both equimolar and nonequimolar conditions, allowing evaluation of classification performance under different composition balances. The mixing-ratio error percentage, defined as the deviation between the actual and classified proportions of each amino acid, was within 25% for the equimolar (1:1:1:1:1) solution. However, in the 3:1:1.5:1.5:1.5 solution, L-Ala showed the largest deviation, with an error of 56%. This large error rate could be attributed to the small number of obtained current–time waveforms.

Samples of astrobiological interest may contain many different types of molecules, and we sought to begin testing the capacity of our approach to discriminate amino acid enantiomers in the context of more complex mixtures. Amino acids sourced from biology could, for example, also be rich in nucleic acids. We, therefore, considered a test case with nucleic acid monomer guanosine (dGMP) and thymine (dTMP) “contaminants.” We performed the single-molecule characterization of a mixture containing two kinds of nucleic acid monomers and five amino acids: dGMP, dTMP, glycine (Gly), and the L and D enantiomers of alanine (Ala) and histidine (His). The F₁−score for the seven species was 0.761 (Supplementary Fig. 10, Table 2). The mixing ratios were quantitatively analyzed by measuring mixed solutions with a D-His:L-His:D-Ala:L-Ala:Gly:dGMP:dTMP molar ratio of 15:15:23:23:23:1:1. The single-molecule measurement result was D-His:L-His:D-Ala:L-Ala:Gly:dGMP:dTMP = 16:10:28:18:25:1:1 (Supplementary Fig. 10 and Table 2). The error in the mixing ratio of each molecule was less than 33%. An actual sample is expected to contain more contaminants. In practical applications, especially for on-site analyses of environmental or extraterrestrial samples, the presence of various contaminants, such as inorganic ions, organic acids, and small peptides, is inevitable. To address this, we plan to enhance the classifier by incorporating measured data from such expected contaminants, aiming to achieve robust detection and reliable identification of amino acids under realistic sample conditions.

Although these experiments set a high standard for identifying chiral amino acids, there was nonetheless a difference between the actual abundance ratios and those measured from single-molecule counting. The driving forces for an amino acid molecule to approach, pass through, and then move away from the nanogap electrode are thermal motion and the electric field of the nanogap electrode^{48, 49}. The difference in the number of molecules detected in the experiments may partly reflect amino-acid-dependent differences in diffusion, interactions with the electrode surfaces, and detection probability at the nanogap.

Analysis of amino acids in meteorite and desert extracts

Given the promising results obtained with various molecular mixtures described above, we sought to determine whether our instrument and analysis pipeline could identify and quantify amino acids in natural samples of astrobiological relevance. We therefore conducted single-molecule tunneling-current measurements on amino-acid extracts from the Murchison meteorite and soil samples from two sites in the Atacama Desert, referred to as North of Antofagasta (NOA) and Diego de Almagro (DGA) (Supplementary Note 4). In parallel, we performed liquid chromatography–mass spectrometry (LC–MS), which serves as a reference method for chiral amino-acid characterization.

Figure 5a shows a comparison of the total amino acid abundance determined via LC–MS with the total number of tunneling-current events detected after signal acquisition via nanogap measurements. Total nanogap event counts include both events that were subsequently retained as high-confidence target assignments and those that did not meet the confidence criterion. Therefore, total event counts should not be interpreted as direct quantitative measures of total amino acid abundance. Murchison and DGA showed higher total event counts than the process blank, whereas NOA showed event counts closer to the blank level, indicating a larger relative contribution of process-derived background signals and/or nontarget matrix components under low-abundance conditions.

Fig. 5. Confidence-filtered targeted analysis of amino acids in extracts from the Murchison meteorite and Atacama Desert samples.

a Relationship between the total amino acid abundance determined by LC–MS and the total tunneling-current event counts detected after signal pickup by nanogap (NG) measurements for the Murchison meteorite extract, process blank, and the North of Antofagasta (NOA) and Diego de Almagro (DGA) Atacama Desert extracts. Each point represents one independently prepared extract analysed by both LC–MS and NG measurements (Murchison, n = 2; NOA, n = 3; DGA, n = 3; Blank, n = 1). Total NG event counts include both target-like events and non-target or low-confidence events. The dashed line indicates the least-squares linear regression. b Relative abundances of glycine, alanine, valine, serine, aspartic acid, and glutamic acid determined by LC–MS and confidence-filtered NG analysis for the Murchison, NOA, and DGA extracts. NG values were calculated using only high-confidence events assigned to a predefined 11-species target panel comprising glycine and the D- and L-enantiomers of alanine, valine, serine, aspartic acid, and glutamic acid. Events with a maximum predicted probability <0.5 were classified as unknown/rejected and excluded from the analysis. D- and L-enantiomer counts were combined for each amino acid, and the resulting six-group abundances were normalized to the total number of retained high-confidence target events. NG-derived values therefore represent the relative abundances within the predefined target panel rather than the overall molecular composition of the extracts. Data are presented as mean ± s.e.m. from three independent NG measurements (n = 3). c Comparison of D/L ratios determined by LC–MS and confidence-filtered NG analysis for alanine, valine, and aspartic acid in the Murchison and DGA extracts. The x-axis represents D/L ratios determined by LC–MS, and the y-axis represents D/L ratios determined by NG analysis using retained high-confidence assignments. LC–MS and NG measurements were performed independently on aliquots derived from the same extracts; therefore, individual measurements from the two methods were not paired. For Murchison, D/L ratios were determined from a single extract, with LC–MS and NG values each obtained from three repeated measurements of the same extract. For DGA, D/L ratios represent the mean values obtained from three independently prepared extracts (n = 3). Individual replicate values are shown alongside the mean for each amino acid and sample, and horizontal and vertical error bars indicate the s.d. for LC–MS and NG measurements, respectively. The dashed line indicates the line of identity (y = x). Source data are provided as a Source Data file.

For compound-specific analysis, we performed a targeted analysis using glycine and the D- and L-enantiomers of alanine, valine, serine, aspartic acid, and glutamic acid. These analytes were selected a priori based on previous Atacama soil studies and compatibility with the LC–MS workflow. Glycine was treated as a single achiral species, whereas the remaining five amino acids were resolved into D- and L-enantiomers, yielding an 11-species target panel.

The environmental-sample analysis was designed as targeted relative quantification within this predefined panel rather than comprehensive profiling of all amino acids or other organic constituents in the extracts. To reduce the forced assignment of low-confidence events, the maximum predicted probability across the 11 target classes was calculated for each feature-extracted event. Events with a maximum probability ≥ 0.5 were retained as high-confidence target events, whereas events below this threshold were classified as unknown/rejected events and excluded from the revised compositional and D/L-ratio analyses.

The proportions of high-confidence target events among feature-extracted events were 49.7% for Murchison, 28.3% for DGA, 25.4% for NOA, and 6.8% for the process blank. Accordingly, 93.2% of feature-extracted blank events were classified as unknown/rejected under the same criteria. These results show that the confidence-based filter removed most blank-derived events from the targeted environmental sample analysis.

The relative compositions of the six-amino-acid groups were recalculated using only retained high-confidence target events. For each chiral amino acid, D- and L-enantiomer counts were combined, and the resulting six-group counts were normalized to the total number of retained high-confidence target events.

In the Murchison sample, glycine remained the most abundant group in the confidence-filtered nanogap analysis (47.9%), followed by alanine (16.9%), glutamic acid (12.0%), and serine (11.7%) (Fig. 5b, Table 3). This composition retained the major glycine-, alanine-, and glutamic acid–rich patterns observed using LC–MS, although analyte-specific differences remained between the two methods, including higher nanogap-derived proportions of serine and valine compared to those detected via LC–MS.

Table 3.

Relative abundances (in %) of six amino acids in 6 M HCl-hydrolyzed hot-water extracts of the Murchison meteorite and Atacama Desert samples (DGA and NOA) measured via LC-FD–Q-ToF-MS and nanogap analysis^a

Sample	Murchison		NOA		DGA
Method	LC-MS	NG	LC-MS	NG	LC-MS	NG
Glycine	62.83 ± 0.69	47.9 ± 11.0	59.63 ± 7.49	56.4 ± 10.0	51.27 ± 1.75	52.6 ± 6.0
Alanine	15.28 ± 0.36	16.9 ± 7.0	8.43 ± 0.55	7.5 ± 7.0	20.85 ± 0.97	16.6 ± 1.0
Valine	2.55 ± 0.72	8.2 ± 2.0	1.23 ± 0.29	13.5 ± 5.0	9.43 ± 0.16	14.5 ± 3.0
Serine	0.34 ± 0.01	11.7 ± 1.0	11.74 ± 5.25	10.3 ± 1.0	1.69 ± 0.30	4.1 ± 1.0
Aspartic Acid	2.57 ± 0.05	3.3 ± 2.0	14.95 ± 2.13	12.2 ± 3.0	5.60 ± 1.94	12.1 ± 3.0
Glutamic acid	16.44 ± 0.30	12.0 ± 7.0	4.02 ± 1.29	0.0 ± 0.0	11.16 ± 0.86	0.0 ± 0.0

^aValues are means of three repeated measurements of a single extract for Murchison and three independently prepared extracts for NOA and DGA (n = 3). For each run r, relative abundances were computed as $p_{i,r}=\frac{A_{i,r}}{{\sum }_k{A}_{k,r}}.$The table reports ${\overline{p}}_i\pm \mathrm{SE}\left({\overline{p}}_i\right)$, where ${\overline{p}}_i$ is the across-run mean. Standard errors (SEs) were computed as $\frac{s}{\sqrt{n}}$ after per-run normalization, where s denotes the sample standard deviation across runs.

In the NOA and DGA samples, confidence filtering retained sample-specific target panel compositions but revealed analyte-dependent deviations from LC–MS results. In NOA, glycine, alanine, serine, and aspartic acid showed broadly comparable relative abundances between the two methods, whereas valine was higher and glutamic acid was not retained as a high-confidence nanogap assignment. In DGA, glycine and alanine showed similar relative contributions between the two methods, whereas valine, serine, and aspartic acid showed higher nanogap-derived proportions, and glutamic acid was not retained after confidence filtering. Thus, the confidence-filtered analysis supports targeted compositional comparison while indicating that quantitative agreement depends on both analyte identity and sample matrix.

Chiral analysis of environmental amino acid extracts

Chiral discrimination of alanine, valine, serine, aspartic acid, and glutamic acid was performed for amino-acid extracts from the Murchison meteorite and the NOA and DGA sites in the Atacama Desert. For the confidence-filtered analysis, D/L ratios were calculated using only retained high-confidence D- and L-enantiomer assignments and were compared with the corresponding values obtained via LC–MS.

Figure 5c shows a comparison of D/L ratios of alanine, valine, and aspartic acid in the Murchison and DGA extracts, for which ratios could be determined using both methods. Glutamic acid was excluded because its confidence-filtered nanogap-derived D/L ratio could not be obtained for the DGA sample. Serine was also excluded because the number of retained high-confidence assignments was insufficient to determine reliable D/L ratios under the confidence-filtered analysis.

The comparison showed analyte- and sample-dependent agreement rather than uniform quantitative correspondence across all amino acids. For the Murchison sample, the confidence-filtered nanogap-derived D/L ratio for aspartic acid was in good agreement with the corresponding LC–MS value, whereas differences remained for alanine and valine. The DGA sample likewise showed analyte-dependent discrepancies between the two methods. These results indicate that confidence filtering reduces contributions from ambiguous events while highlighting the current limitations of quantitative chiral analysis in chemically complex environmental matrices.

Accordingly, the confidence-filtered nanogap D/L ratios should be interpreted as high-confidence estimates within the predefined targeted amino-acid panel rather than as comprehensive chiral profiles of the environmental samples. Further validation using matrix-matched reference materials, increased event statistics, and the implementation of open-set recognition approaches will be required to establish practical quantitative performance for chemically heterogeneous environmental samples.

Discussion

We have demonstrated that a combination of single-molecule measurements using tunneling currents and machine learning can identify the 19 chiral amino acids and glycine. The single-molecule detection and discrimination of molecular chirality demonstrated here highlight a fundamental advance in chemical sensing. By distinguishing L- and D-forms of amino acids without optical or recognition-based methods, this approach establishes a direct electrical means of identifying molecular handedness at the quantum level^31,33. The resulting datasets of amino acid current–time waveforms may also contribute to future developments in peptide and protein sequencing technologies^33–35,38, but more importantly, they provide a foundation for probing molecular asymmetry—a key biosignature in astrobiology and life detection³¹.

Single-molecule chiral identification is not currently available using conventional mass spectrometry. The proposed nanogap method of chiral identification at the single-molecule level should be applicable to both large biomolecules and small molecules, and its principle is based on intermolecular interaction in which a chiral structure identifies the chiral molecule. The tip of a gold electrode with a high Miller index has been considered to have a chiral structure^41–43. However, the chiral structure of the tips of the two gold electrodes is not obvious. Gold electrodes are fabricated by breaking thin gold wires with an MCBJ. A system combining MCBJ and transmission electron microscopy may, in the future, be able to determine the chiral structure of the tips of the two gold electrodes.

Herein, we evaluated nanogap-based single-molecule analysis using amino-acid extracts from astrobiologically relevant samples. Confidence-based filtering enabled separation of high-confidence target assignments from low-confidence events within a predefined 11-species amino-acid panel. The process blank showed a low fraction of high-confidence target assignments, whereas the Murchison sample exhibited a substantially higher fraction of retained target assignments. The DGA and NOA samples showed intermediate target fractions, indicating that process-derived background signals and matrix-dependent non-target events contribute more strongly to the results obtained for some environmental extracts.

The confidence-filtered compositional and D/L ratio analyses captured selected major features observed via LC–MS, particularly for the Murchison sample, and showed analyte- and sample-dependent discrepancies. The present method should therefore be interpreted as a targeted, high-confidence identification approach rather than as comprehensive molecular profiling or absolute quantification of amino-acid mixtures in environmental samples. Future work should incorporate representative blank-derived and nontarget molecular datasets, matrix-matched calibration experiments, and open-set recognition approaches to improve discrimination between target analytes, background signals, and chemically unrepresented events.

Conventional LC–MS techniques remain important reference methods for terrestrial laboratory analysis. In contrast, the nanogap platform offers a compact, reagent-free, and potentially integrable approach for targeted molecular chirality detection. Its practical use in future planetary exploration will require further validation under contamination-controlled and matrix-matched conditions. Amino acids are relevant targets for astrobiology because they can persist for billions of years under cold and dry conditions, as demonstrated by their identification within primordial material, including meteorites. This stability makes them a promising target for characterization of potential abiotic and prebiotic environments, as well as for life detection. Earlier studies suggested that abiotic processes might result in enantiomeric excesses of L-amino acids, potentially influencing the preference of terrestrial life for the L-form. However, recent sample-return mission analyses (e.g., Ryugu and Bennu) have reported enantiomeric ratios consistent with racemic mixtures within analytical uncertainty, indicating that previously observed L-excesses in meteorites were likely due to terrestrial contamination⁵. Meteoritic amino acids tend to have higher HOMO-LUMO gaps than the set of amino acids used by terrestrial life⁵⁰ and this leads to higher conductance nanogap events for proteinogenic amino acids with their lower gaps⁵¹. Thus, the combination of estimated abundance, conductance, and enantiomeric excess may be especially informative in identifying traces of prebiotic chemistry or life beyond Earth, especially when biomass may be constrained by limitations in nutrients or free energy^8,52. Approaches for in situ extraction of amino acids include traditional acid hydrolysis, which can produce racemization of amino acids, subcritical water extraction⁵³, and solid phase extraction of powdered regolith (or meteorite) samples.

Methods

Materials

The 39 amino-acid standards used in this study comprised the L- and D-enantiomers of alanine, arginine, asparagine, aspartic acid, cysteine, glutamine, glutamic acid, histidine, isoleucine, leucine, lysine, methionine, phenylalanine, proline, serine, threonine, tryptophan, tyrosine, and valine, together with glycine. The standards were purchased from Tokyo Chemical Industry Co., Ltd. (TCI) and used without further purification. The stated purities were as follows: L-alanine, >99.0%; D-alanine, >98.0%; L-arginine, >98.0%; D-arginine, >98.0%; L-asparagine, >98.0%; D-asparagine, >98.0%; L-aspartic acid, >99.0%; D-aspartic acid, >98.0%; L-cysteine, >98.0%; D-cysteine, >98.0%; L-glutamine, >98.0%; D-glutamine, >98.0%; L-glutamic acid, >98.0%; D-glutamic acid, >98.0%; L-histidine, 97–99%; D-histidine, 97–99%; L-isoleucine, 97–99%; D-isoleucine, 97–99%; L-leucine, 97–99%; D-leucine, 97–99%; L-lysine, 97–99%; D-lysine, >95.0%; L-methionine, >99.0%; D-methionine, >99.0%; L-phenylalanine, >98.0%; D-phenylalanine, >98.0%; L-proline, >99.0%; D-proline, >98.0%; L-serine, >99.0%; D-serine, >98.5%; L-threonine, >99.0%; D-threonine, >99.0%; L-tryptophan, >98.5%; D-tryptophan, >98.0%; L-tyrosine, >98.5%; D-tyrosine, >98.0%; L-valine, >98.0%; D-valine, >98.0%; and glycine, >99.0%. Catalogue numbers and analytical specifications are provided in Supplementary Data 1.

Standards and reagents for environmental-sample analysis

All glassware, ceramics, and sample-handling tools used for environmental-sample preparation were rinsed with Milli-Q ultrapure water (18.2 MΩ cm, <3 ppb total organic carbon), wrapped in aluminum foil, and heated at 500 °C in air overnight. High-purity hydrochloric acid (Tamapure-AA-10-HCl) was used for acid-vapor hydrolysis, and AG50W-X8 cation-exchange resin (100–200 mesh, hydrogen form; Bio-Rad) was used for amino-acid purification. A 2 M NH₄OH solution was used for elution from the cation-exchange resin.

Sample collection

Atacama Desert samples were collected during a field expedition in February 2022 at two sites: Diego de Almagro (DGA; 26°23′11″S, 70°04′14″W) and North of Antofagasta (NOA; 22°47′46″S, 70°04′14″W), from depths of 5–10 cm below the surface, and stored in sterile Whirl-Pak® bags at ambient temperature. A 2 g interior fragment of the Murchison meteorite, free of fusion crust, was obtained from the Smithsonian National Museum of Natural History.

LC-FD/ToF-MS analysis

Amino acid abundances and enantiomeric ratios were determined by liquid chromatography with fluorescence detection and time-of-flight mass spectrometry (LC-FD/ToF-MS) using o-phthaldialdehyde/N-acetyl-L-cysteine (OPA/NAC) derivatization based on the procedure reported by Glavin et al.⁵⁴, with the experimental conditions used in the present study described below. The desalted, hydrolyzed, and non-hydrolyzed sample aliquots allocated for LC-FD/ToF-MS analysis were dried under vacuum. A 20-µL aliquot of each reconstituted sample was combined with 20 µL of sodium borate buffer (pH 9) and dried. The samples were then reconstituted in 20 µL of Milli-Q ultrapure water, mixed with 5 µL of 0.1 M OPA/NAC derivatization reagent, and allowed to react for 15 min at room temperature. The derivatization reaction was quenched with 75 µL of 0.1 M hydrazine.

The OPA/NAC derivatives were separated using a Waters ACQUITY UPLC coupled to a Waters Xevo G2-XS Q-ToF mass spectrometer operated in positive-ion mode. C2–C6 amino acids were chromatographically separated using a Waters BEH C18 column (2.1 mm × 50 mm, 1.7 µm particle size) and a Waters BEH phenyl column (2.1 mm × 150 mm, 1.7 µm particle size) connected in series. Both columns were maintained at 30 °C. Mobile phase A consisted of 50 mM ammonium formate containing 8% methanol (pH 8.0), and mobile phase B consisted of methanol. The flow rate was 150 µL min⁻¹. For C2–C6 amino acid separations, the gradient program was defined as time (min) (%B): 0 (0), 35 (55), and 45 (100). C5 amino-acid isomers and enantiomers were separated under the same chromatographic conditions using a modified gradient program: 0 (15), 25 (20), 25.06 (35), 44.5 (40), and 45 (100).

The Xevo G2-XS was operated in dual electrospray ionization (ESI) mode for lock-mass correction. The primary ESI source was operated with a capillary voltage of 3.0 kV, sampling cone voltage of 40 V, source temperature of 120 °C, desolvation gas (N₂) temperature of 350 °C, cone gas (N₂) flow rate of 50 L h⁻¹, and desolvation gas flow rate of 750 L h⁻¹. Leucine enkephalin was used as the lock-mass reference. The reference ESI source was operated with a sample infusion rate of 20 µL min⁻¹, sample fill volume of 250 µL, lockspray infusion rate of 10 µL min⁻¹, capillary voltage of 3.0 kV, reference cone voltage of 30 V, and collision energy of 6.0 V. The ToF analyzer was operated in Sensitivity mode, providing a full width at half maximum resolution of <22,000 based on the [M + H]⁺ ion of leucine enkephalin at m/z 556.2771.

Amino acid abundances and enantiomeric ratios were determined from the peak areas of the fluorescence chromatograms of the OPA/NAC derivatives (LC-FD; λₑₓ = 340 nm and λₑₘ = 450 nm) by comparison with amino acid standards analyzed under identical chromatographic conditions. Peak identities were confirmed by comparison of retention times with those of the corresponding standards and by accurate-mass measurements using a mass tolerance of ≤10 ppm.

Fabrication of a device for single-molecule detection

A 5 μm–thick polyimide film was formed on a 0.3 mm–thick silicon substrate. An electron beam resist was spin-coated on the polyimide film, and a fine metal line pattern was drawn through electron-beam lithography. Then, 5 nm Cr and 200 nm Au were deposited via RF sputtering. The metal wires were fabricated through the lift-off process. SiO₂ with a thickness of 300 nm was deposited via RF sputtering. After the spin-coating of the electron beam resist on the SiO₂ surface, the channel pattern was drawn via electron beam lithography. SiO₂ channels were formed through reactive ion etching using CHF₃ as the reaction gas. Finally, the polyimide directly under the metal wires was etched via reactive ion etching using oxygen as the reaction gas. The notch in the polyimide just below the metal wires makes it easy to break the metal wire^{33–35,37,38}. A nanochannel-integrated nanogap device was utilized. A cover made of polydimethylsiloxane (PDMS; Toray Dow Corning) was fused with the device substrate. The PDMS cover has a pre-formed microchannel with inlet and outlet ports to introduce the sample solution to the sensor’s nanochannel. Nanogap electrodes are fabricated in situ where they are measured. The polyimide, silicon, and silicon dioxide that constitute the nanogap electrode chip are heat-resistant up to over 300 °C. Thus, the nanogap electrode chip can be transported to its destination without requiring cooling, even inside a spacecraft, where temperatures could rise to ≈50 °C even in cold external conditions (e.g., upper range of vault temperature for Europa Lander¹⁶). The SiO₂ layer was deposited onto the fabricated device structure to reduce electrical noise and define the nanochannel geometry. The nanogap junction was subsequently formed in situ by mechanical rupture of the Au nanowire.

Single-molecule measurements

The nanogap electrode was formed using an MCBJ technique. When the piezo element is pushed up, the thin metal wire is gradually bent and broken, forming a nanogap electrode. When the piezo element is pulled down, the nanogap electrode is bonded. Repeating the breakage and bonding processes via up-and-down motion of the piezoelectric element gradually sharpens the electrode tip, eventually forming a single-atom contact. The formation of a gold monoatomic contact yields quantized conductance (77.5 μS), after which the single-atom contact is broken. After this breakage, a tunneling current flows between the nanogap electrodes. The tunneling current and the distance between electrodes ($d$) are related by $I=I_0\exp \left(-\beta d\right)$ ⁴⁵. The damping constant $\beta =-\frac{4\mathrm{\pi }\sqrt{2m\phi }}{\mathrm{h}}$ where m, φ, and h denote the electron mass, the work function of gold electrodes and the Planck constant. The distance between the nanogap electrodes is monitored through this tunneling current. If the distance deviates by ±0.0075 nm from the set distance, then feedback control is performed using a piezoelectric element to maintain the set distance between the electrodes. Consequently, the set distance between the nanogap electrodes is maintained at σ = ±0.0052 nm for more than 5 h (Supplementary Fig. 12).

For measurement, the operator places the solution to be measured into the nanogap chip, as shown in Fig. 1a. First, a nanogap with an interelectrode distance preset by the operator automatically forms. Then, the sample measurement starts automatically. During measurement, the distance between the nanogap electrodes is automatically feedback-controlled and held at the set value. Measurement is automatically terminated when the measurement time entered by the operator is reached. Each MCBJ device was used for about 1 h in total, consisting of four consecutive 15-min measurement periods.

The sample concentration was set to be 0.10 μM. The nanogap electrodes can be reused after being washed using ultrapure water (Supplementary Fig. 11). Pure amino acid data should be utilized as training and test data to reveal the principle accuracy of the single-molecule identification of amino acids. However, a substance may not be completely washed out of the nanogap device. Therefore, in this study, the nanogap electrodes were disposed of. Nonetheless, results confirmed that the reuse of the nanogap device did not significantly change the discrimination accuracy (Supplementary Fig. 11). Measurements were performed at interelectrode distances of 0.52, 0.54, 0.56, 0.58, and 0.60 nm, where 0.52 nm resulted in low discrimination accuracy. For 14 of the 19 amino acids, the discrimination accuracy increased when the interelectrode distance was reduced from 0.56–0.58 nm to 0.54 nm. However, the signal frequency at the 0.54 nm interelectrode distance was 6.5 min⁻¹, which was one-third of the signal frequency at the 0.56–0.58 nm interelectrode distance (18.4 min⁻¹). The signal frequency decreased to 4.3 min⁻¹ when the interelectrode distance exceeded 0.58 nm. Therefore, an interelectrode distance of 0.56–0.58 nm was used in this study. The current across the electrodes was amplified by a custom-built current amplifier and recorded at 10 kHz using a NI PXIe-4081 digital multimeter and NI PXI-5922 (National Instruments) under a DC bias voltage of 0.1 V^{33–35,37,38}.

Event detection

Signal regions were automatically extracted from the conductance–time profiles. The baseline and noise levels were determined from the preceding 10,000 data points. The noise level was defined as the mode of the standard deviations calculated from 200-data-point segments. A candidate tunneling event was detected when the current exceeded six times the standard deviation of the baseline (6σ). For each detected event, the event onset was determined retrospectively as the first preceding time point at which the current exceeded one standard deviation above the baseline (1σ). The event end was defined as the first subsequent time point at which the current fell below one standard deviation above the baseline (1σ). The interval between the event onset and end was defined as the event duration (t_d).

Machine learning

Machine-learning analysis was performed using Python 3.9.7⁵⁵, scikit-learn⁵⁵ version 0.24.2, and the XGBoost⁵⁶ classifier. Each pulse signal extracted from the current–time profiles was represented by a 15-dimensional feature vector comprising the peak current amplitude relative to the baseline (I_p), mean current amplitude (I_m), event duration (t_d), and 12 normalized shape features (T₁–T₁₂). To calculate the shape features, each event was divided into 12 equal-duration intervals along the time axis, and the mean current amplitude within each interval was normalized by I_p, yielding T₁–T₁₂. Events for which all 15 features could be extracted were retained for machine-learning analysis.

For the classification of amino-acid standards, random undersampling was applied to reduce class imbalance. The number of feature vectors for each molecular species was reduced to match that of the smallest class. The resulting balanced dataset was randomly divided into training (80%) and held-out test (20%) datasets. Model training and validation were performed using 10-fold cross-validation within the training dataset. Specifically, the training dataset was partitioned into ten subsets; in each cross-validation round, nine subsets were used for model training and the remaining subset was used for validation, with each subset serving once as the validation dataset. Classification performance was evaluated on the held-out test dataset using the F₁-score and confusion matrix.

For environmental sample analysis, a predefined 11-species target panel was used, comprising glycine and the D- and L-enantiomers of alanine, valine, serine, aspartic acid, and glutamic acid. For each feature-extracted event, the trained XGBoost classifier generated predicted probabilities for all 11 target classes. The highest predicted probability among the 11 classes was defined as the event-level classification confidence.

For the confidence-filtered environmental analysis, events with a maximum predicted probability ≥0.5 were retained as high-confidence target events. Events with a maximum predicted probability <0.5 were classified as unknown/rejected and excluded from the targeted compositional and D/L-ratio analyses. The same probability threshold was applied to the Murchison, NOA, DGA, and process blank datasets without sample-specific adjustment. The unknown/rejected category does not represent a chemically identified molecular class; rather, it includes events that did not meet the predefined confidence criterion for assignment to any of the 11 target species.

The total number of picked-up events was used for the comparison shown in Fig. 5a. The compositional analysis shown in Fig. 5b and Table 3 and the D/L-ratio analysis shown in Fig. 5c and Table 4 were calculated using only retained high-confidence target events. For each independent nanogap measurement, D- and L-enantiomer counts were combined to determine the relative abundance of each chiral amino acid, and the resulting abundances were normalized to the total number of retained high-confidence target events. Relative abundances are reported as the mean ± s.e.m. from three independent nanogap measurements (s.e.m. = s.d./√n). D/L ratios (Fig. 5c, Table 4) are reported as the mean ± s.d. from three independent nanogap measurements.

Table 4.

Confidence-filtered D/L ratios of five chiral amino acids in 6 M HCl-hydrolyzed hot-water extracts of the Murchison meteorite and Atacama Desert samples measured via LC-FD-Q-ToF-MS and nanogap analysis^a

Sample	Murchison		NOA		DGA
Method	LC-MS	NG	LC-MS	NG	LC-MS	NG
Alanine	1.66 ± 0.15	0.32 ± 0.32	n.d.	0.47 ± 0.40	0.19 ± 0.17	0.42 ± 0.46
Valine	0.35 ± 0.35	1.44 ± 0.27	n.d.	1.22 ± 0.47	0.02 ± 0.01	0.34 ± 0.31
Serine	n.d.	n.d.	n.d.	n.d.	n.d.	n.d.
Aspartic acid	0.91 ± 0.04	0.71 ± 0.41	n.d.	0.04 ± 0.41	0.31 ± 0.12	0.10 ± 0.05
Glutamic acid	0.95 ± 0.04	0.10 ± 0.42	n.d.	n.d.	0.14 ± 0.01	n.d.

^a Nanogap D/L ratios were calculated from retained high-confidence D- and L-enantiomer assignments within the predefined 11-species target panel. Events with a maximum predicted probability <0.5 were classified as unknown/rejected and excluded from the analysis. For Murchison, NG values are reported as the mean ± s.d. of three repeated measurements of a single extract. For NOA and DGA, NG values are reported as the mean ± s.d. obtained from three independently prepared extracts (n = 3). n.d., not determined because a reliable D/L ratio could not be calculated from the retained high-confidence events.

Quantum chemical calculations for measuring molecular length

Twenty amino acids were structurally optimized using quantum-chemical calculations implemented in the Gaussian 98 package⁵⁷. Structural optimization was performed using density functional theory (B3LYP) and 6-31 G(d, p) basis sets. The molecular length was defined as the longest interatomic distance in the molecule. The resulting optimized geometries were used to determine the longest interatomic distance of each amino acid, which was employed as the molecular length parameter in the correlation analysis shown in Table 1.

Amino acids extracted from astrobiologically relevant samples

Amino acids were extracted from the Murchison meteorite and from samples collected at two sites in the Atacama Desert, North of Antofagasta (NOA) and Diego de Almagro (DGA). For each extraction, ≈ 1.2 g of dried sample was mechanically homogenized and subjected to hot-water extraction in 2 mL of Milli-Q ultrapure water at 100 °C for 24 h. After centrifugation at 3000 × g for 3 min, the supernatant was collected. The residual sample was washed twice with 1 mL of Milli-Q ultrapure water, with centrifugation after each wash, resulting in a combined extract volume of ≈ 4 mL. The NOA and DGA extractions were performed independently in triplicate, whereas the Murchison meteorite was extracted once. Of the combined extract, 45% (1.8 mL) was dried under vacuum and subjected to vapor-phase acid hydrolysis with 6 M HCl at 150 °C for 3 h; the remaining 55% (2.2 mL) was reserved for other analyses not reported in this study. After hydrolysis, residual HCl was removed by centrifugal vacuum concentration, and the samples were reconstituted in 1 mL of Milli-Q ultrapure water. Amino acids were purified using AG50W-X8 cation-exchange resin (100–200 mesh, hydrogen form; Bio-Rad) and eluted with 7 mL of 2 M NH₄OH. The eluates were dried under vacuum and reconstituted in 300 µL of Milli-Q ultrapure water. The resulting samples were divided into aliquots for LC–MS analysis (10%, 30 µL), single-molecule nanogap analysis (65%, 195 µL), and additional analyses (25%, 75 µL) not reported in this study. For single-molecule nanogap analysis, the corresponding aliquots were reconstituted in Milli-Q ultrapure water to obtain amino-acid concentrations of 1–100 µM. Owing to the higher total amino-acid content, the Murchison extract was diluted to a final volume of 10 mL, whereas the NOA and DGA extracts were reconstituted in 1 mL. All single-molecule measurements were performed using freshly prepared 50-µL aliquots. A schematic overview of the sample-preparation workflow is provided in Supplementary Figs. 13–18.

Author contributions

M.T., M.T.Z., T.W., D.D., and C.E.C. conceptualized the study. T.O., Y.K., and T.T. curated the data. M.T., M.T.Z., T.W., D.D., and C.E.C. acquired funding and provided resources. T.O., Y.K., T.T., and T.Y. developed the software. J.L.R.-C. and H.L.M. performed amino-acid extractions and LC–MS analysis. M.T., M.T.Z., T.W., and C.E.C. supervised the study. D.D., M.T., T.W., C.E.C., J.L.R.-C., H.L.M., T.O., and Y.K. reviewed and edited the manuscript. All authors approved the final version of the manuscript.

Peer review

Peer review information

Nature Communications thanks Lingxin Chen, and the other, anonymous, reviewers for their contribution to the peer review of this work. A peer review file is available.

Funding

This research was partially supported by JST-CREST Grant Number JPMJCR1666, JST K Program Grant Number JPMJKP23H3, Grant-in-Aid for Scientific Research from the Ministry of Education, Culture, Sports, Science, and Technology (21H01741), a JSPS Invitational Fellowship for Research in Japan (DD), and NASA award 80NSSC22K0188.

Data availability

The data that support the findings of this study, including raw tunneling-current datasets, processed event-level feature matrices, sample metadata, and class labels, are available from Zenodo⁵⁸ and from the corresponding authors upon request. Source data are provided with this paper.

Code availability

Custom Python scripts used for event detection, feature extraction, and machine-learning analysis in this study are available from the corresponding authors upon request.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77947-6.