Learning Persistent Spatiotemporal Motifs Complex Molecular Interactions on Supported Lipid Bilayers

Zhen Yuan Yeo, Eun Ho Song, Jwa‐Min Nam, N. Duane Loh2025 · DOI

Abstract

This study introduces a framework for high-sensitivity analysis of molecular interactions on supported lipid bilayers (SLBs), enhancing traditional biosensing with unsupervised detection of persistent spatiotemporal motifs and temporal super-resolution. By leveraging rotationally invariant representations (RIZP) and t-SNE manifold learning, the framework autonomously quantifies complex interaction patterns, such as receptor-ligand binding. A pseudo-high-speed capture technique "stitches" sequential images to resolve transient molecular states that are otherwise lost in noise or limited by camera frame rates.

Introduction: The Challenge of Molecular "Dances"

Supported lipid bilayers (SLBs) serve as a critical platform for observing molecular interactions in biomimetic environments. However, traditional analysis often relies on rudimentary morphological features—average intensity, size, or eccentricity. While sufficient for simple systems, these metrics fail to capture the nuanced "dance" of multi-particle interactions. Transient states, such as the fleeting moments of DNA hybridization, often occur faster than standard imaging rates or are obscured by experimental noise, leaving a significant gap in our understanding of higher-order spatiotemporal dynamics.

Methodology: From Noisy Snapshots to Invariant Representations

The proposed pipeline addresses these limitations through a multi-step computational approach:

  1. Detection: Regions of interest (ROIs) are identified in dynamic movies using Trackpy.
  2. Rotational Invariance: To ensure the system recognizes interactions regardless of orientation, image patches are projected into Zernike Polynomial (ZP) bases. By utilizing the absolute values of these projections, the researchers create Rotationally Invariant ZP (RIZP) feature vectors.
  3. Manifold Learning: High-dimensional RIZP vectors are mapped onto a 2D manifold using t-SNE, clustering similar molecular configurations.
  4. Reconstruction: A modified Expectation-Maximization Compression (EMC) algorithm is used to reconstruct high-resolution, denoised images from the noisy input data.
Experimental dark-field imaging of DNA-tagged nanoparticles (A) and the Zernike Polynomial bases (B) used to generate rotationally invariant representations.

Results: Achieving Temporal Super-Resolution

The framework was validated using both simulated data and experimental dark-field movies of DNA-tagged nanoparticles.

Enhanced Signal-to-Noise Ratio

By averaging noisy patches within a cluster (vector quantization), the system significantly improves the signal-to-noise ratio (SNR). The EMC reconstruction successfully recovered particle structures from 500 randomly rotated, noisy inputs, demonstrating high fidelity in structural recovery.

Mapping and Ordering Interactions

The t-SNE embedding effectively separated distinct classes of particle configurations. A key innovation is the "pseudo-temporal evolution" achieved by applying a secondary t-SNE to class averages. This allows the researchers to order clusters chronologically, creating a "pseudo-high-speed" sequence that reveals the progression of binding events.

t-SNE manifold of class averages showing a clear trajectory, enabling the reconstruction of temporal sequences from transient snapshots.

Discussion and Future Outlook

By stripping away rotational variance, the model focuses exclusively on molecular configuration. This unsupervised approach allows for the discovery of persistent motifs without prior labeling. While computationally intensive, the technique is highly promising for high-throughput biosensing and drug discovery, where identifying subtle, fast-moving molecular dynamics at scale is essential. Future work will address the handling of extremely rare events and optimizing the pipeline for larger datasets.

Summary updated

Abstract

<p xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" class="first" dir="auto" id="d20116e98">This study introduces a framework for high-sensitivity analysis of molecular interactions on supported lipid bilayers (SLBs), enhancing traditional SLB biosensing applications with unsupervised detection of persistent spatio-temporal motifs and temporal super-resolution. Current SLB-based methods [ <a class="xref-link" href="#r1">1</a>] often rely on simple morphological features (e.g., average intensity, eccentricity, size, etc), which, while effective for limited experiments, have limited expressiveness when characterizing more complex spatiotemporal interactions between multiple particles. <p xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" dir="auto" id="d20116e103">Our proposed framework addresses this limitation by leveraging rotationally invariant representations developed for electron microscopy [ <a class="xref-link" href="#r2">2</a>] to autonomously quantify and classify reveal persistent, complex molecular interaction patterns, such as receptor-ligand binding or DNA hybridizations. The framework incorporates a pseudo-high-speed capture technique, which “stitches” sequential images to enhance the observation of transient molecular states and complex, higher-order interactions that would otherwise remain unresolved in real-time imaging. <p xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" dir="auto" id="d20116e108">Here is an overview of the framework. 1) Identify regions of interest (ROI) of a dynamic movie with many DNA-tagged particles interacting on an SLB (e.g., using Trackpy, a widely used particle tracking package). 2) Thereafter, we projected the image patch centered on each ROI (for each particle at each time) into its Zernike Polynomial (ZP) bases, separately for each RGB color channel. By taking the absolute values of these ZP projections, we transform each particle-centered patch into its Rotationally Invariant ZP (RIZP) representation [ <a class="xref-link" href="#r2">2</a>], ensuring orientation-independent descriptions of each particle. 3) Next, we reduce the dimensionality of the RIZP feature-vectors using t-SNE (t-distributed stochastic neighborhood embedding), which represents the wide variety of observed configurations on a lower-dimensional manifold, where clustering nearby patches reveals key patterns and features within the dataset.
BibTeX
@misc{yeo2025learning,
  author = {Zhen Yuan Yeo and Eun Ho Song and Jwa‐Min Nam and N. Duane Loh},
  title = {{Learning Persistent Spatiotemporal Motifs Complex Molecular Interactions on Supported Lipid Bilayers}},
  year = {2025},
  doi = {10.14293/apmc13-2025-0300},
}
Figure