Deep Neural Network-Assisted Terahertz Metasurface Sensors for the Detection of Lung Cancer Biomarkers

MicroRNAs (miRNAs) are essential for gene regulation, with abnormal levels leading to lung cancer. Specifically, miRNA-21, miRNA-92a, and miRNA-339-3p are key early cancer detection biomarkers. While traditional detection methods such as RT-PCR and qPCR are reliable, they are often complex and costly.

Terahertz (THz) sensing technology offers a non-ionizing, low-energy alternative for miRNA detection, with THz metasurfaces enhancing sensitivity through concentrated electromagnetic fields and high-Q Fano resonances. At the same time, deep neural networks (DNNs) provide a label-free method for accurately detecting miRNA concentrations and categories.

In this study, miRNA-21 was obtained from the Gemma gene, with concentrations ranging from 100 aM to 10 nM, while miRNA-92a and miRNA-339-3p were prepared at 1 pM. The THz metasurface biosensor was designed using CST Microwave Studio 2020, featuring a high-resistance silicon substrate, a silica layer, and a metal pattern.

An all-fiber THz time-domain spectroscopy system (TPF15K) with a spectral resolution of less than 3.5 GHz was utilized to generate a 1560 nm laser pulse. This pulse was split into pump and probe lights, where the pump light excited a THz wave that traversed the sensor, generating a photocurrent for data processing.

Experiments were conducted at normal temperature and pressure, with humidity maintained below 5%. A 20-µL miRNA solution was applied to the sensor, and ten readings were taken per scan, yielding 200 readings for each concentration. This process produced 1,800 samples for nine concentrations of miRNA-21 and 600 samples for three categories at 1 pM, alongside control experiments on a silicon substrate for comparable spectra.

Fast Fourier transform was performed using Python’s NumPy toolkit to normalize amplitudes and convert them to decibels (dB), generating frequency-domain spectra for miRNA-21 and other categories. The Savitzky-Golay filter was applied to enhance accuracy due to the minor spectral resonance changes observed in the sensor for clear classification.

The DNN model comprised six layers: two fully connected layers, a dropout layer, two additional fully connected layers, and an output layer. The model, implemented in Python 3.7.0 using Keras and TensorFlow, processed spectral data from 0.1 to 1.25 THz as a one-dimensional sequence of 1,150 frequency points, utilizing an 80%-20% dataset split for training and testing. After 500 epochs, training and test accuracies neared 1, with loss stabilizing at 0.6, indicating effective convergence without overfitting.

The metasurface sensor achieved a classification accuracy of 97.22% for miRNA-21, outperforming the 92.22% accuracy of silicon substrates. The sensor attained 95.00% accuracy for various miRNA categories compared to 74.16% for silicon substrates.

The DNN surpassed five other machine learning methods and demonstrated a lower limit of detection (LOD), highlighting its practicality for detecting miRNA-21 across various concentrations.

This study illustrates that a THz metasurface sensor integrated with a DNN effectively detects miRNAs associated with lung cancer, suggesting significant potential for real-time, nondestructive detection of nucleic acids and biomarkers in clinical applications. This approach offers a cost-effective, label-free solution for detecting low levels of miRNA without complex amplification.