Scholarlink

<p>Surface Plasmon Resonance (SPR) biosensing provides an effective platform to detect biomarkers in real-time; nevertheless, multiplexed quantification is difficult to achieve because of spectral overlap and nonlinear interactions. The present study suggests a calibration-free neural inversion scheme to accurately estimate biomarkers based on complex SPR spectra. A huge dataset of 25,000 spectra (400–800&#xa0;nm) was created under multiplexed conditions in realistic noise (20-40 dB SNR). Dimensionality reduction and preprocessing enhanced the quality of the signal and lowered the computational complexity. The developed deep learning model had a high predictive accuracy with an RMSE of 0.049 ng/mL, an R<sup>2</sup> of 0.987, and a MAPE of 3.8%. Preprocessing minimized the RMSE by 22%, and dimensionality reduction maintained more than 95% variance with insignificant performance deterioration. When the conditions were multiplexed, the RMSE was only slightly larger (0.035 to 0.053) with the increase in the number of biomarkers (2 to 5), but cross-talk was still low (less than 4%). The model was found to be resistant to noise, and RMSE increased from 0.041 to 0.058 as SNR decreased from 40 to 20 dB. In addition, it was more than 56% more accurate than traditional techniques and could infer in real-time (7.5 ms/sample). These findings show excellent possibilities for clinical diagnostics.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Calibration-free neural inversion of overlapping SPR spectral signatures for real-time multiplexed cancer biomarker quantification

  • 期刊论文

  • 发表日期：

    2026年7月30日

查看全文
  • D. Chandrakala,
  • Damodharan Palaniappan,
  • S. Thenappan,
  • P. Jayabharathi,
  • Palraj Karuppasamy,
  • R. Ramya

摘要

Surface Plasmon Resonance (SPR) biosensing provides an effective platform to detect biomarkers in real-time; nevertheless, multiplexed quantification is difficult to achieve because of spectral overlap and nonlinear interactions. The present study suggests a calibration-free neural inversion scheme to accurately estimate biomarkers based on complex SPR spectra. A huge dataset of 25,000 spectra (400–800 nm) was created under multiplexed conditions in realistic noise (20-40 dB SNR). Dimensionality reduction and preprocessing enhanced the quality of the signal and lowered the computational complexity. The developed deep learning model had a high predictive accuracy with an RMSE of 0.049 ng/mL, an R2 of 0.987, and a MAPE of 3.8%. Preprocessing minimized the RMSE by 22%, and dimensionality reduction maintained more than 95% variance with insignificant performance deterioration. When the conditions were multiplexed, the RMSE was only slightly larger (0.035 to 0.053) with the increase in the number of biomarkers (2 to 5), but cross-talk was still low (less than 4%). The model was found to be resistant to noise, and RMSE increased from 0.041 to 0.058 as SNR decreased from 40 to 20 dB. In addition, it was more than 56% more accurate than traditional techniques and could infer in real-time (7.5 ms/sample). These findings show excellent possibilities for clinical diagnostics.