New Publication | Superspectra-Guided SERS Enables Accurate Protein Quantification

We are excited to share our latest publication in ACS Omega, titled “Improving Protein Quantification with SERS Superspectra and Machine Learning.” 

Quantitative protein analysis using surface-enhanced Raman spectroscopy (SERS) has long been challenged by weak, orientation-dependent adsorption and highly heterogeneous spectral responses. Traditional peak-based calibration often fails because protein–surface interactions are nonlinear and substrate dependent. In this work, we introduce a superspectra-guided SERS framework that strategically integrates spectra from chemically complementary substrates to dramatically improve quantitative prediction accuracy.

🔬 What We Did

We fabricated silver nanorod (AgNR) SERS substrates and functionalized them with three distinct thiol monolayers—cysteamine (CM), cysteine (CN), and 6-mercapto-1-hexanol (MCH)—along with bare Ag nanorods as a neutral reference. These surface chemistries create different electrostatic and chemical interaction environments for protein adsorption. Using bovine serum albumin (BSA) as a model system, we collected SERS spectra across five orders of magnitude in concentration.

Instead of analyzing each substrate independently, we constructed “superspectra” by concatenating spectra acquired from different surface chemistries at the same protein concentration. These superspectra integrate complementary interaction information into a single high-dimensional feature vector. We then applied support vector regression (SVR) and random forest regression (RFR) to predict protein concentration from both single-substrate spectra and multisubstrate superspectra.

⚡ Key Results

Single-substrate spectra were insufficient for accurate quantification. Although characteristic BSA peaks were present, their intensities fluctuated irregularly with concentration due to competitive adsorption, surface-site heterogeneity, and orientation effects. Machine learning was therefore necessary to extract quantitative information from the full spectral pattern. Superspectra derived from chemically complementary surfaces dramatically improved prediction accuracy. The CM&CN combination reduced mean absolute error by more than fourfold compared with single-substrate models and achieved an R² value of 0.993. Adding the bare substrate to form the B&CM&CN combination further enhanced robustness. In contrast, including MCH often degraded performance because its negatively charged surface contributed weak or conflicting protein signals.

Across all models, random forest regression consistently outperformed support vector regression, demonstrating superior robustness in integrating chemically heterogeneous spectral inputs.

🌍 Why This Matters

This work establishes clear design principles for constructing effective SERS superspectra for protein quantification. Chemical complementarity between substrates is essential, and simply concatenating more spectra does not guarantee better performance. Substrate compatibility, rather than substrate count, determines predictive success.

More broadly, this study demonstrates how engineered surface chemistry combined with machine learning can expand the molecular information space of SERS. The superspectra framework provides a foundation for quantitative protein analysis, multianalyte detection, and future translation into complex biological fluids. It represents an important step toward AI-enabled, chemistry-informed SERS diagnostics.

👉 Read the full article:https://pubs.acs.org/doi/10.1021/acsomega.6c00157