image: The multi-gate photoelectrochemical transistor can itself constitute an artificial neural network. Its operation is based on utilizing its synaptic memory to afford built-in VMM and light-tunable responsivity for weight updating.
Credit: ©Science China Press
Conventional electrochemical biosensing usually generates numerical values such as target concentrations, leaving the subsequent interpretation to users. Driven by the improved robustness and minimized information loss, recent artificial intelligence (AI)-aided electrochemical biosensing has achieved huge success in providing more interpretable outputs to identify specific physiological and pathological states. Nevertheless, most AI-aided biosensing systems still follow a separated workflow: biosensing, analog-to-digital conversion, data transfer and external decision-making. This increases system complexity, reduces working efficiency and may raise concerns about data security.
In-sensor computing offers a promising route to overcome these limitations by allowing sensors to process signals directly at the sensing terminal. This concept has been widely explored for physical signals such as light, pressure and electricity. Despite ongoing technical advancements, critical issues remain: (i) Existing solid-state systems can only perceive limited types of physical signals. (ii) In-sensor neural network necessitates an interconnected device array for the vector-matrix multiplication (VMM). Implementing in-sensor computing in biochemical systems still faces many inherent restrictions, e.g. lack of aqueous compatibility, device-to-device variability and crosstalk noise.
To address this challenge, a team led by Wei-Wei Zhao at Nanjing University realizes the electrochemical in-biosensing computing for the first time, which can simultaneously perform multi-target biosensing and artificial neural network classification. The key design lies in using the device’s multi-gate configuration and intrinsic synaptic memory to perform built-in VMM. On each photogate, concentration of a specific biochemical multiplies with light-tunable responsivities and then accumulates into a summed channel current as output, realizing the VMM. In this way, an in-situ single-layer artificial neural network classifier can be implemented, which can sense multiple targets, and based on this, generate reliable algorithmic decision-making. Using this device, a complete electrochemical in-biosensing computing system is constructed, which can realize:
In-situ training: Compared with ex-situ training based on fitted theoretical functions, the in-situ training strategy uses real-time experimental current to update the weights, allowing the device to better tolerate non-ideal behaviors and batch-to-batch variations. In classification tests using 30 randomly selected devices, the system achieves an overall accuracy of 93.33%, demonstrating the robustness of this method.
Real sample classification: By using nucleic-acid recognition strategies, three photogates are designed to individually respond to microRNA-21, microRNA-221 and microRNA-133a. The device successfully distinguishes microRNA expression patterns in different cell lysates with only one error among 30 cell lysate samples. The results can be directly displayed by LED lights within approximately 8.6 s, enabling rapid and visible decisions.
Zhao explained, “By making the electrochemical biosensing itself computing, we aim to revolutionize the workflow of conventional biosensing. We expected this work can offer a transformative paradigm to simpler, securer, cheaper, and more robust next-generation AI-driven electrochemical biosensing.”