Resolving Single-Peptide Phosphorylation Dynamics in Plasmonic Nanopores using Physics-Informed Bi-Path Model

This paper introduces a physics-informed deep learning framework that combines multiple-instance learning with spatiotemporal modeling to overcome signal stochasticity and background interference, enabling the robust, label-free detection of single-peptide phosphorylation events using particle-in-pore plasmonic nanopores.

The Gist

The Big Picture: The "Molecular Detective" Problem

Imagine you are a detective trying to identify a single, specific spy (a protein) in a crowded, noisy room. This spy is wearing a very subtle disguise: a tiny, almost invisible pin on their lapel (a phosphate group).

The problem is threefold:

1. The Spy is Shy: They only show up for a split second.
2. The Room is Noisy: There are other people (citrate molecules) wearing similar clothes, and the lighting is flickering.
3. The Camera is Blurry: You can only see a tiny part of the spy at a time, not their whole body.

Traditional methods are like trying to take a photo of the whole crowd at once. You get a blurry average where the spy's tiny pin gets lost in the background noise. This paper introduces a new way to catch that single spy and identify their pin with incredible precision.

---

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

The scientists built a tiny trap called a plasmonic nanopore.

• The Analogy: Imagine a tiny tunnel with a super-bright, focused flashlight (a "hotspot") shining inside it.
• How it works: When a molecule (the spy) drifts into this tunnel, it gets caught in the light. The light makes the molecule glow with a unique "fingerprint" called a Raman spectrum.
• The Catch: Because the molecule is moving and spinning, the light only hits a tiny piece of it at any given moment. It's like trying to identify a person by only seeing a flash of their shoe, then a flash of their sleeve, then a flash of their hat, all while the room is shaking.

The Challenge: The "Blinking" Signal

The signal from this flashlight is chaotic. It "blinks" on and off because the molecule is jiggling around (Brownian motion).

• The Noise: The gold nanoparticles used in the trap are held together by a sticky substance called citrate. This citrate also glows in the light, creating a background "hiss" that drowns out the spy's signal.
• The Similarity: The two spies the scientists are trying to tell apart are almost identical twins. One is a normal peptide (F-Ser), and the other is the same peptide but with a phosphate pin (F-pSer). They share 99% of their DNA. The only difference is that one tiny pin.

The Solution: The "Physics-Informed Bi-Path Model" (The Smart Brain)

Instead of trying to force a computer to look at every single blurry snapshot, the scientists built a special AI brain called a Physics-Informed Bi-Path Model. Think of this brain as having two specialized detectives working together:

1. The "Spotter" (Multiple Instance Learning)

• The Job: This detective doesn't care about every single frame of the video. They know that in a long, shaky video, only a few frames actually show the spy clearly.
• How it works: It scans the whole video and says, "Ignore the blurry frames where we just see the background noise. Focus only on the 5 seconds where the molecule actually looks like a molecule." It groups these good moments together to form a clear picture.
• The Analogy: It's like watching a security camera feed of a busy street. Instead of analyzing every second of the footage, the Spotter says, "Ignore the cars and the wind. Just show me the 3 seconds where the suspect walked by."

2. The "Storyteller" (Temporal Encoder)

• The Job: This detective looks at the order of events. How does the signal change over time?
• How it works: It uses advanced math (TCN and BiGRU) to understand the "story" of the molecule. Does the signal flicker in a pattern that looks like a phosphate group? Does it drift in a way that matches a specific amino acid?
• The Analogy: If the Spotter finds the suspect, the Storyteller checks the suspect's gait. "Even though we only saw a shoe, the way they walked into the light and out of it matches the gait of the phosphorylated spy, not the normal one."

3. The "Two-Stage Judge"

Once the Spotter and Storyteller agree, they pass the evidence to a Judge who works in two steps:

• Step 1: Is this a real molecule, or just background noise (citrate)?
• Step 2: If it is a real molecule, is it the normal twin or the phosphorylated twin?

Why This is a Game-Changer

• It's "Physics-Informed": The AI wasn't just fed random data; it was taught the rules of physics (how molecules move and how light interacts with them). This stops the AI from making "hallucinations" or guessing based on noise.
• It Handles Ambiguity: The AI is smart enough to say, "I'm not sure about this specific frame, but looking at the whole group of frames, I'm 99% sure this is the phosphorylated peptide."
• The Result: They achieved over 90% accuracy in distinguishing the two nearly identical peptides, even with the noisy background.

The Bottom Line

This paper is about teaching a computer to be a super-detective. By combining a clever light trap with an AI that understands both the shape of the data and the story of how it moves, the scientists can now spot tiny chemical changes (phosphorylation) on single molecules.

Why does this matter?
Protein phosphorylation is like a "light switch" in our bodies that controls how cells talk to each other. If this switch is broken, it can lead to diseases like cancer. Being able to see these switches on a single molecule, without needing millions of copies, opens the door to ultra-early disease detection and personalized medicine. It's like moving from needing a whole forest to find a single tree, to being able to find a single leaf on a single tree in a storm.

Technical Summary

1. Problem Statement

The paper addresses the critical challenge of detecting protein phosphorylation (a key post-translational modification or PTM) at the single-molecule level, specifically under conditions of low abundance and low stoichiometry. While Single-Molecule Surface-Enhanced Raman Spectroscopy (SM-SERS) using particle-in-pore plasmonic nanopores offers label-free, high-sensitivity detection, it faces three major analytical hurdles:

• Stochastic Signal Fluctuations: SM-SERS signals exhibit "blinking" and temporal fluctuations due to Brownian motion, molecular rotation, and transient binding within the plasmonic hotspot.
• Partial Excitation: The plasmonic hotspot is highly localized, exciting only 1–3 amino acid residues of a peptide at a time. This results in incomplete vibrational fingerprints, making it difficult to identify specific PTMs (like phosphorylation) that may be outside the current hotspot.
• Background Interference & Spectral Overlap: The stabilizing agent for gold nanoparticles (citrate) creates strong background signals that overlap with peptide spectra. Furthermore, phosphorylated peptides (F-pSer) and their unphosphorylated counterparts (F-Ser) share nearly identical backbones, differing only by a single phosphate group, leading to highly overlapping spectral features that are easily obscured by noise.

Traditional supervised learning methods fail here because they require pixel-level (frame-level) annotations, which are impossible to assign reliably to stochastic, fluctuating single-molecule time series.

2. Methodology

The authors propose a Physics-Informed Bi-Path Hierarchical Deep Learning Framework designed to decode complex SM-SERS dynamics. The methodology consists of four key stages:

A. Physics-Informed Data Segmentation

To address the stochastic nature of the data, the authors developed an adaptive segmentation algorithm based on Pearson correlation:

• Instead of treating the entire time series as a single label, the algorithm scans the time-series spectra to identify signal onset.
• Spectra are grouped into "bags" (segments) if their correlation ($r \geq 0.65$) with a reference spectrum ($S_{ref}$) is high.
• This mimics the physical residence time of a molecule in the hotspot (~2.5 seconds). It effectively separates distinct molecular states from idle noise and citrate interference without requiring manual frame-level labeling.

B. The Bi-Path Hierarchical Model

The core architecture is a dual-branch deep learning model that processes the segmented data:

1. Shared Spectral Encoder: A 1D Convolutional Neural Network (CNN) processes the input, which is engineered as a two-channel tensor (raw intensity + first derivative) to enhance subtle peak shifts and remove baseline drift. A channel-wise attention mechanism highlights discriminative Raman shifts.
2. Pathway 1: Spatial Aggregator (Multiple Instance Learning - MIL):
   • Utilizes Gated Attention to handle label ambiguity.
   • It autonomously identifies and aggregates the most informative "instances" (spectra) within a bag, filtering out idle states and noise, creating a robust bag-level representation.
3. Pathway 2: Temporal Encoder:
   • Comprises a Temporal Convolutional Network (TCN) with causal dilated convolutions and a Bidirectional Gated Recurrent Unit (BiGRU).
   • This pathway captures the temporal evolution of the signal, including short-term spectral bursts and long-range dependencies (blinking dynamics) that are crucial for distinguishing F-Ser from F-pSer.

C. Hierarchical Classification & Loss Function

• Two-Stage Classifier: The fused features from both paths are passed to a hierarchical classifier.
  • Stage 1: Distinguishes background (Citrate) from target analytes.
  • Stage 2: Distinguishes between the specific peptide variants (F-Ser vs. F-pSer).
• Physics-Informed Loss: The model is trained using a multi-objective loss function combining cross-entropy, stage-specific binary losses, and an MIL consistency loss.
• Peak-Sensitive Regularization (PSR): A regularization term is applied to the spectral attention mechanism to mathematically discourage the model from focusing on baseline noise, forcing it to attend to chemically meaningful Raman shifts.

D. Interpretability

The authors use 1D Integrated Gradients (1D-IG) to visualize feature attribution, ensuring the model's decisions are based on genuine molecular vibrations rather than statistical artifacts.

3. Key Contributions

• Novel Architecture: Introduction of a physics-informed, bi-path model combining MIL (for weak supervision) and temporal deep learning (TCN/BiGRU) specifically tailored for stochastic SM-SERS data.
• Adaptive Segmentation: A correlation-based segmentation method that aligns data processing with the physical diffusion dynamics of molecules in plasmonic hotspots.
• Robust PTM Detection: Successful differentiation of single-peptide phosphorylation (F-pSer) from its unphosphorylated form (F-Ser) despite strong citrate interference and partial excitation.
• Interpretability: Demonstrated that the deep learning model learns chemically valid features (e.g., specific phosphate vibrational modes) rather than noise, verified through Integrated Gradients.

4. Results

The model was tested on single-molecule SERS data of F-Ser (unphosphorylated), F-pSer (phosphorylated), and Citrate.

• Classification Accuracy:
  • Stage 1 (Citrate vs. Analyte): Achieved True Positive Rates of 87.1% (Citrate) and 91.4% (Analyte).
  • Stage 2 (F-Ser vs. F-pSer): Achieved accuracies of 93.2% (F-Ser) and 85.3% (F-pSer).
  • Overall Three-Class Accuracy: Citrate (87.1%), F-Ser (83.2%), F-pSer (85.3%).
• Performance Metrics:
  • AUC (Area Under Curve): Exceeded 0.95 for all classes.
  • AUPRC (Area Under Precision-Recall Curve): Maintained values above 91.6%, indicating high reliability in detecting rare binding events in imbalanced datasets.
• Physical Validation: The 1D-IG analysis confirmed that the model correctly identified key vibrational modes:
  • Citrate: Carboxylate bending/stretching (595, 673, 1015, 1345, 1550 $cm^{-1}$).
  • F-pSer: Phosphate group vibrations at 930, 950, and 1005 $cm^{-1}$, proving the model can isolate the specific PTM signature.

5. Significance

This work represents a significant advancement in ultrasensitive phosphoproteomics and single-molecule sensing:

• Overcoming Sensitivity Limits: It enables the detection of low-abundance PTMs without the need for fluorescent labeling or massive sample quantities required by mass spectrometry.
• Handling Stochasticity: By integrating physical constraints (diffusion, blinking) into the deep learning architecture, it solves the "label ambiguity" problem inherent in single-molecule flow-through experiments.
• Clinical Potential: The framework is presented as a robust paradigm for next-generation liquid biopsies, with potential applications in early-stage cancer detection and neurodegenerative biomarker analysis where subtle molecular changes must be detected in complex biological backgrounds.
• Generalizability: The core mechanism of decoding label-ambiguous temporal data can be adapted to other single-molecule biophysics and medical imaging modalities.