Unveiling elemental fingerprints: A comparative study of clustering methods for multi-element nanoparticle data

Area of Science:

• Analytical Chemistry
• Materials Science
• Data Science

Background:

• Single particle-inductively coupled plasma-time of flight-mass spectrometry (SP-ICP-TOF-MS) generates complex multi-elemental nanoparticle composition data.
• Extracting meaningful information from large SP-ICP-TOF-MS datasets presents a significant analytical challenge.
• Hierarchical clustering (HC) is established for elemental fingerprinting but other methods remain unevaluated.

Purpose of the Study:

• To compare the effectiveness of hierarchical clustering (HC), spectral clustering, and t-distributed Stochastic Neighbor Embedding coupled with Density-Based Spatial Clustering of Applications with Noise (tSNE-DBSCAN) for analyzing SP-ICP-TOF-MS data.
• To evaluate clustering performance based on extracted cluster size and intra-cluster elemental composition similarity.
• To identify the optimal clustering approach for robust nanoparticle characterization using SP-ICP-TOF-MS.

Main Methods:

• Utilized SP-ICP-TOF-MS data comprising multi-elemental nanoparticle compositions.
• Applied and compared three distinct clustering algorithms: HC, spectral clustering, and tSNE-DBSCAN.
• Evaluated clustering outcomes by analyzing cluster size, nanoparticle homogeneity within clusters, and outlier sensitivity.

Main Results:

• Hierarchical clustering (HC) exhibited sensitivity to outliers, often failing to achieve optimal clustering solutions.
• Spectral clustering and tSNE-DBSCAN identified clusters missed by HC, revealing more intricate data structures.
• Spectral clustering outperformed both HC and tSNE-DBSCAN by consistently generating larger, more homogenous clusters of nanoparticles with similar elemental profiles.

Conclusions:

• While HC, spectral clustering, and tSNE-DBSCAN can all extract valuable information from SP-ICP-TOF-MS data, spectral clustering offers superior performance.
• Spectral clustering's ability to capture both global and local data structures makes it highly effective for nanoparticle elemental fingerprinting.
• The findings highlight spectral clustering as a robust and recommended method for analyzing complex SP-ICP-TOF-MS datasets.