🔬 physics

Label-free SERS Discrimination of Native Proline Hydroxylation at Single-molecule peptide by Deep Learning-assisted plasmonic nanopore

This paper presents a deep learning-assisted, label-free single-molecule SERS platform using a particle-in-pore nanopore sensor that successfully discriminates between hydroxylated and non-hydroxylated HIF peptide fragments by analyzing reproducible spectral changes and adsorption behaviors, achieving up to 89.74% classification accuracy.

Original authors: Yingqi Zhao, Kuo Zhan, Pei-Lin Xin, Yuge Liang, Enock Adjei Agyekum, Matti Putkonen, Shuai Li, Francesco De Angelis, Jianan Huang

Published 2026-05-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yingqi Zhao, Kuo Zhan, Pei-Lin Xin, Yuge Liang, Enock Adjei Agyekum, Matti Putkonen, Shuai Li, Francesco De Angelis, Jianan Huang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Finding a Tiny Change in a Giant Crowd

Imagine you have a massive library of books (proteins) that your body uses to function. Sometimes, a tiny editor makes a very small change to a single word in one of these books. In the world of biology, this is called a Post-Translational Modification (PTM). Specifically, this paper looks at a change where a "Proline" letter gets a tiny "Hydroxyl" sticker added to it.

This specific change is crucial because it acts like a switch for a protein called HIF-1α, which helps cells survive when oxygen is low (like in a tumor). Detecting this tiny "sticker" is incredibly hard because:

The sticker is microscopic.
The book (peptide) is short and floppy.
The "sticker" doesn't make a loud noise; it's a whisper in a hurricane.

The researchers built a high-tech "microscope" and a "smart brain" (AI) to hear that whisper.

The Tool: The "Plasmonic Nanopore" (The VIP Trap)

To see these tiny molecules, the scientists used a device they call a "Particle-in-Pore" sensor. Here is how it works, using an analogy:

The Gold Nanoparticle (The Passenger): Imagine the peptide (the protein fragment) is a passenger. It is glued to a tiny, shiny gold ball (a nanoparticle).
The Nanopore (The Tunnel): The device has a tiny tunnel (pore) made of gold, only 250 nanometers wide.
The Trap (The VIP Lounge): When the gold ball tries to swim through the tunnel, the scientists use electricity and a laser to gently push the ball against the side wall of the tunnel.
The Hotspot (The Spotlight): The space between the gold ball and the tunnel wall creates a super-intense "spotlight" (a plasmonic hotspot). When the laser hits this spot, the molecule inside glows with a unique fingerprint called SERS (Surface-Enhanced Raman Spectroscopy).

The Problem: Because the molecule is so small and flopping around, it changes its position constantly. Sometimes the "sticker" is in the spotlight; sometimes it's hidden. This makes the signal flicker and look different every time, like trying to recognize a friend in a crowd where they keep turning their head.

The Solution: The "Frequency Counter" and the "AI Detective"

Since the signal flickers, looking at just one snapshot doesn't work. The researchers used two clever tricks:

1. The Frequency Counter (POF Analysis)

Instead of trying to read one perfect picture, they took thousands of snapshots (over 6,800 for each test).

The Analogy: Imagine trying to identify a song by listening to a radio station that skips and changes volume. If you listen for an hour, you might not hear the whole song clearly, but you can count how often a specific drumbeat or guitar riff appears.
The Result: They counted how often specific "notes" (vibrations) appeared in their thousands of snapshots. They found that when the "sticker" (hydroxylation) was present, the "drumbeat" at 680 (from a specific part of the molecule) disappeared, and a new "guitar riff" at 712 appeared. This proved the sticker changed how the molecule sat on the gold ball.

2. The AI Detective (Deep Learning)

They fed all these thousands of flickering snapshots into a 1D-CNN (a type of Artificial Intelligence designed to recognize patterns).

The Analogy: Think of the AI as a detective who has seen thousands of photos of a suspect. Even if the suspect wears a hat in some photos and sunglasses in others, the AI learns the pattern of the face, not just the accessories.
The Result: The AI learned to distinguish between the "sticker" version and the "no-sticker" version, even though the signals were messy.
- For short peptides (7 and 9 letters long), the AI was about 73% to 78% accurate.
- For a longer peptide (15 letters long), the AI got 89% accurate.

The Surprise: The "Crowded Room" Effect

One of the most interesting findings happened with the longest peptide (15 letters).

The Observation: The researchers noticed that the "sticker" version of the long peptide left behind more of the original "glue" (citrate) on the gold ball than the non-sticker version did.
The Analogy: Imagine a dance floor (the gold ball).
- Without the sticker: The dancers (peptides) are very flexible and can squeeze onto the floor tightly, pushing all the old furniture (citrate) off the stage.
- With the sticker: The dancer is stiff or awkward. They can't squeeze in as well, so they leave some of the old furniture (citrate) on the dance floor.
Why it matters: The AI actually used this "leftover furniture" (the citrate signal) as a clue to tell the difference! The sticker changed the molecule's shape so much that it couldn't sit on the gold ball the same way, leaving a clear trace.

The Conclusion

This paper shows that by combining a super-sensitive trap (the nanopore), a frequency counter (to handle the flickering signals), and a smart AI (to find the patterns), scientists can detect a tiny chemical change (hydroxylation) on a single protein molecule.

They didn't just find the change; they discovered that the change also altered how the molecule "sat" on the sensor, which actually helped the AI find it even better. This proves that we can now "hear" the whispers of these tiny biological modifications, which is a big step forward for understanding how proteins work at the most basic level.

Technical Summary: Label-free SERS Discrimination of Native Proline Hydroxylation at Single-molecule Peptide by Deep Learning-assisted Plasmonic Nanopore

Problem Statement
Post-translational modifications (PTMs), specifically prolyl 4-hydroxylation of hypoxia-inducible factor-1α (HIF-1α), are critical for regulating cellular signaling and cancer progression. However, discriminating hydroxylated from non-hydroxylated peptide fragments at the single-molecule level remains a significant challenge. Existing methods face distinct limitations: mass spectrometry requires high molecular copy numbers ( $10^6$ – $10^8$ ) and suffers from false interpretations due to altered combinatorial rules; nanopore resistive pulse sensing provides indirect readouts encoded in current blockade amplitudes rather than direct chemical fingerprints; and conventional Surface-Enhanced Raman Spectroscopy (SERS) struggles with weak Raman cross-sections of hydroxylation, spectral overlap with the peptide backbone, and heterogeneity in molecular adsorption configurations and electromagnetic hotspots. Consequently, detecting subtle PTM signatures like hydroxylation in complex peptide systems is hindered by low reproducibility and difficulties in spectral interpretation.

Methodology
The authors developed a hybrid platform combining a "particle-in-pore" single-molecule SERS sensor with statistical spectral analysis and deep learning.

Particle-in-Pore SERS Platform: The system utilizes a gold nanohole array (250 nm diameter pores in a 100 nm gold film on a SiN membrane) integrated into a microfluidic device. Peptide-loaded 50 nm gold nanoparticles (AuNPs) are driven through the pores by electrophoresis and electroosmosis. A 785 nm laser exerts an optical force, trapping the AuNP against the pore sidewall. This configuration creates a confined plasmonic hotspot between the AuNP and the pore wall, enabling single-molecule sensitivity while minimizing backbone signal coverage.
Peptide Systems: Three pairs of HIF-1α fragment peptides containing the Pro-564 hydroxylation site were synthesized with lengths of 7, 9, and 15 amino acids (AAs). Non-hydroxylated (HIF7S, HIF9S, HIF15S) and hydroxylated (HIF7SPTM, HIF9SPTM, HIF15SPTM) variants were physically adsorbed onto citrate-capped AuNPs to form monolayers.
Data Acquisition and Processing: Time-series Raman spectra were collected during nanoparticle trapping events. Over 6,800 effective spectra per dataset were pre-processed (baseline correction, cosmic ray removal, normalization).
Statistical Analysis (POF): To address single-molecule spectral fluctuations, the authors employed Peak Occurrence Frequency (POF) analysis. This method aggregates large datasets to reconstruct representative spectral profiles, identifying reproducible spectral features based on the frequency of peak appearance at specific Raman shifts rather than absolute intensity.
Deep Learning Classification: A one-dimensional convolutional neural network (1D-CNN) was trained to classify hydroxylated vs. non-hydroxylated peptides using the full spectral data. Model interpretability was enhanced using Gradient-weighted Class Activation Mapping (Grad-CAM) to visualize which spectral regions drove the classification decisions.

Key Results

Spectral Signatures and Adsorption Changes: POF analysis revealed reproducible, hydroxylation-dependent spectral changes. For the 7AA and 9AA peptides, hydroxylation suppressed the Met $\nu$ (C–S) band at ~680 cm⁻¹ and enhanced bands near 712, 1016, 1121, and 1234 cm⁻¹. This suggests a shift in adsorption conformation from a Met-dominated orientation to alternative configurations involving Tyrosine (Tyr) ring interactions or peptide termini.
Discrimination Accuracy: The 1D-CNN achieved post-evaluation single-spectrum classification accuracies of 72.98% (7AA), 78.55% (9AA), and 89.74% (15AA). The Area Under the Curve (AUC) values were 0.80, 0.88, and 0.96, respectively. Notably, accuracy increased with peptide length, contrary to the expectation that longer, more complex peptides would be harder to distinguish.
Adsorption-Mediated Amplification in Longer Peptides: For the 15AA peptide pair, the hydroxylated variant (HIF15SPTM) exhibited a significantly higher POF for the citrate-associated band at 1073 cm⁻¹ (0.266) compared to the non-hydroxylated variant (0.033). This indicates that hydroxylation alters peptide conformation or surface affinity, reducing adsorption efficiency and leaving more citrate on the AuNP surface. This adsorption-mediated effect amplified the spectral differences, contributing to the superior discrimination accuracy (AUC > 0.96) for the 15AA pair.
Model Interpretability: Grad-CAM visualization confirmed that the CNN's sensitive spectral regions overlapped with the recurrent POF features identified as hydroxylation-dependent. High feature weights were observed at 680, 712, 1016, 1121, and 1234 cm⁻¹, as well as the citrate region (1073 cm⁻¹) for the 15AA peptides, validating the chemical relevance of the learned patterns.

Significance and Claims
The paper claims that the integration of particle-in-pore single-molecule SERS with POF analysis and deep learning provides a robust strategy for the ultrasensitive analysis of weak PTM signatures. Key contributions include:

Overcoming Spectral Heterogeneity: The study demonstrates that statistical reconstruction (POF) and deep learning can effectively extract discriminative patterns from the inherently heterogeneous and noisy data of single-molecule SERS.
Mechanistic Insight: The work reveals that discrimination is not solely based on intrinsic vibrational differences of the hydroxyl group but is significantly amplified by hydroxylation-induced changes in peptide adsorption behavior and surface interactions (e.g., citrate displacement).
Scalability: The method successfully discriminated PTMs in peptides of increasing structural complexity (up to 15 AA), suggesting that adsorption-mediated spectral changes may become more pronounced in larger systems, thereby aiding rather than hindering detection.
Label-free Detection: The approach offers a label-free alternative to mass spectrometry and antibody-based methods for detecting low-abundance PTMs, potentially enabling early disease diagnosis and the study of protein dynamics at the single-molecule level.

The authors conclude that this framework provides a promising pathway for analyzing subtle structural and chemical changes in biomolecules, bridging the gap between single-molecule sensing and complex biological interpretation.