Time-resolved neurotransmitter detection in mouse brain tissue using an artificial intelligence-nanogap

The analysis of neurotransmitters in the brain helps to understand brain functions and diagnose Parkinson’s disease. Pharmacological inhibition experiments, electrophysiological measurement of action potentials, and mass analysers have been applied for this purpose; however, these techniques do not allow direct neurotransmitter detection with good temporal resolution by using nanometre-sized electrodes. Hence, we developed a method for direct observation of a single neurotransmitter molecule with a gap width of ≤ 1 nm and on the millisecond time scale. It consists of measuring the tunnelling current that flows through a single-molecule by using nanogap electrodes and machine learning analysis. Using this method, we identified dopamine, serotonin, and norepinephrine neurotransmitters with high accuracy at the single-molecule level. The analysis of the mouse striatum and cerebral cortex revealed the order of concentration of the three neurotransmitters. Our method will be developed to investigate the neurotransmitter distribution in the brain with good temporal resolution.


SI1. Schematic structure of MCBJ substrate
Fabrication process is described in method section in main manuscript.
A Au narrow wire is fabricated on elastic substrate. Polyimide insulating layer is removed by dry-etching. Narrowest part of the Au wire is free-standing structure. The substrate was bended by pushing using piezo from backside (see Figure 1d in main manuscript). The nanogap was formed by breaking Au wire by bending. The nanogap width is controlled by piezo displacement.
Estimation of gap width is described in next section in this supporting information.

SI2. Estimation of gap distance
The gap distance is estimated using by following current equation of direct tunneling current = exp (−  4  ℎ  √2  ).
Here, h, m,w, and l represents plank constant, electron mass, work function of gold electrode, gap distance. We used electron mass of 9.1 ×10 -31 kg as m, and work function of Au (111) The signals were analyzed by the conventional histogram-based method, where only single   features, such as the maximum and average currents, are analyzed to identify single-molecule signals. In this section, we only use the maximum current for signal discrimination. The current histograms per time unit are regarded as a probability density function for the molecules. Hence, the single-molecule signals are well-discriminated by comparing the detection rate in each current region, that is, by choosing the maximum probability densities among the target molecules when the overlap between the histograms of those molecules is small. The classification results obtained by the conventional histogram-based detection rate-comparison method are presented in Figure S4. Figure S4a shows the classification results for signals from pure solutions while Figure S4b shows the classification of signals obtained from mixtures.

SI5. Detail of Noise removal with PUC method
To remove noise signals, we perform Positive and Unlabelled data Classification (PUC).
Schematic image of PUC is shown in Figure S5a. PUC is appropriate algorithm for noise removal. Some signals are observed even in blank solution (See Figure S3, reference 28 in main text.). At first, to remove the blank signals, we performed PUC as shown in Figure S5b Figure S7.) The analysis was performed using Python 3.6 with the XGBoost library. We also analysed with random forest classifier from scikit-learn library version 0.21.1 instead of XGBoost classifier.
The classification result with random forest is represented in SI.8.

Figure S8
Relation between number of signals for training data and classification F-measure.

SI7. Improvement of classification accuracy by accumulation
The classification performance index (F-measure) for discrimination between the three neurotransmitters is 0.52, which is slightly higher than the value for random classification. This accuracy is not the accuracy determined by using multiple signals during application but only that for a single pulse. The classification accuracy can be improved by statistical analysis. In the method reported in this manuscript, each signal is classified one by one; the molecule is classified with majority vote of all signal classification results.
Here, we consider the relation between the classification accuracy and the number of signals n.
The prediction ratio for a single pulse of the true molecule p 1 is set to 0.5 while the prediction ratios of the other molecules p 2 and p 3 are set to 0.3 and 0.2, respectively. Then, the probability of accurate prediction by the majority vote P is described using the following equation: (1) where k i denotes the number of signals predicted as molecule I, S is the set of k values that satisfy: k 1 >k 2 +k 3 . The relation between P and n is shown in Figure S9. The accuracy determined by the majority vote is 80% for 20 signals, 90% for 40 signals, and 99% for 110 signals.
The relation between the classification accuracy and the number of signals for the ratio represented in Figure 2e of the manuscript is shown in Figure S9. The accuracies for all the three neurotransmitter molecules are improved by accumulation, as shown in Figure S10.

SI9. Origin of classification
Discrimination between two neurotransmitters were also performed ( Figure S12). All three case show accurate classification result. We compared serotonin and dopamine. The amino group is a typical anchoring group of single-molecule junctions and adsorbs onto gold electrodes via coordination bonds [39]. In absence of strong anchoring groups such as thiols, amino groups and π-conjugation planes act as anchoring groups [33, [40][41][42]. Given the junction structure of serotonin and dopamine, both these molecules form bonds via the amino groups with one electrode and via the π-conjugation planes with the other one; they have the same amino groups but different π-conjugation planes. Since the indole ring of serotonin is larger than the catechol ring of dopamine, serotonin consisted from larger π plane is deduced to form various structures in the junctions that originate larger current fluctuations. Dopamine and norepinephrine differ only by one hydroxyl group in the molecular structure. Despite this slight difference, the spectroscopic results in the gas phase indicate that the dopamine conformers are more abundant than the norepinephrine ones because of the intra-molecular interactions due to the hydroxyl groups [36, [43][44][45]. Conformational changes of the alkyl groups cause conductance changes [31,32], hence, the current fluctuation of norepinephrine is smaller than that of dopamine due to intra-molecular interactions. Although the current fluctuation factor was not directly used as a classification feature, the classification results suggest that the machine learning-based method classifies the molecules based on their behaviour in the nanogaps due to difference in the molecular structures. Analysis via machine learning could allow the observation of intramolecular interactions at the single-molecule level.

SI11. Analysis of mouse brain signals
The detail scheme of mouse brain analysis is represented in Figure S14. In this scheme, we perform PUC twice, first PUC is noise removal for the noise signals observed in blank solution due to migration of gold electrodes or contamination. Second PUC is extraction of target neurotransmitters from contamination. There are many other molecules in mouse brain. The signals from first PUC only contain neurotransmitters signals. The noise removed signals are trained as positive data, we obtain targeted-neurotransmitters signals from contamination. Then neurotransmitters signals in mouse brain were classified with supervised ML. Figure S14. Schematic flow for mouse brain analysis. In the first positive and unlabelled data (PU) classification run, the PU classifier was trained using the signals from the neurotransmitter solutions as unlabelled data and those from the blank measurements as positive data. Then, the signals from the neurotransmitter solutions were recognised again. The signals classified as negative were adopted as single neurotransmitter signals for the next training data. The second PU classification run was performed to remove the noise signals originated from contamination in the mouse brain; for the training data, the signals obtained from the brain measurements were treated as unlabelled and those from the first PU classification run were considered as positive for each neurotransmitter. The elimination of the predicted negative signals for all the three neurotransmitters provided the neurotransmitter signals in the brain. Current and dwell time histograms of mouse brain signals are shown in Figure S15. Compared striatum and cerebral cortex, striatum shows higher current signals caused by DA frequently.
time resemble each other. The neurotransmitters discrimination using current or dwell time histogram is difficult due to the similarity of histogram shapes. It suggests that ML-based analysis method is effective for neurotransmitters discrimination.