(Invited) Application of Machine Learning in Carbon Nanotube-Based Biosensors

Among electrochemical sensors, electrolyte-gated field-effect transistor (FET) with semiconducting single-walled carbon nanotubes (SWCNTs) is one of the most promising choices for the ultrasensitive biosensors. The biosensing is typically based on the recognition of analytes with capture probes (antibody, aptamer, receptor, enzyme, etc.) attached to SWCNTs. While the capture probes improve the chemical sensitivity and selectivity, the stability of the nanotube-based biosensors is limited by the activity of the capture probes. Another approach involves sensor arrays with different metal nanoparticles/self-assembled monolayers on the SWCNTs that act as nonspecific receptors. These sensor arrays can be trained with machine learning algorithms to distinguish analytes. When analytes bind to the sensor surface and interact with the semiconducting SWCNTs, they can cause modulations of the number of charge carriers in the FET channel, changing the device conductance, and eventually result in change in the FET characteristics, i.e., measured source-drain current vs. applied gate voltage. Rich information regarding the biorecognition process, sensing mechanism and sensing performances can be extracted from the changes in FET transfer characteristics. We proposed 11 features that can be selected to accurately describe the changes in FET transfer curves, including (1) relative change in transconductance, (2) threshold voltage (V th ) shift, (3, 10) relative change in conductance at ±0.6 V g , (4-9) change in overall conductance normalized to conductance at V th , and (11) the relative change in minimum conductance (see figure below). The transfer characteristics were extracted as distinct features for model training using established machine learning algorithms. We fabricated SWCNT-FET array functionalized with gold nanoparticles and different self-assembled monolayers (dodecanethiol and lipoic acid) for the sensing of nonmalignant and malignant cells, which were classified by linear discriminant analysis. [1] We have also demonstrated that the SWCNT-FET sensor array could be used for the screening of cell behaviors, and that live/dead mouse B16 melanoma cells could be successfully classified with machine learning. [2] [1] Silva, G. O.; Michael, Z. P.; Bian, L.; Shurin, G. V.; Mulato, M.; Shurin, M. R.; Star, A., Nanoelectronic Discrimination of Nonmalignant and Malignant Cells Using Nanotube Field-Effect Transistors. ACS Sens. 2017, 2 , 1128–1132. [2] Liu, Z. R.; Shurin, G. V.; Bian, L.; White, D. L.; Shurin, M. R.; Star, A., A Carbon Nanotube Sensor Array for the Label-Free Discrimination of Live and Dead Cells with Machine Learning. Anal. Chem. 2022 , 94 , 3565-3573. Figure 1