Development of a brain tissue classifier using atmospheric solids analysis probe mass spectrometry

Thesis

Recent advances in compact, low-cost, and rapid analytical platforms have enabled direct measurement of molecular composition within complex biological materials. Clinical disciplines such as neurosurgery and neuropathology could benefit from analytical frameworks that combine fast chemical profiling with machine-learning-based interpretation. This thesis presents a brain tissue classification framework based on atmospheric solids analysis probe mass spectrometry (ASAP–MS) and supervised machine learning, and evaluates its potential to support molecular discrimination between brain tissue types in neurosurgical and neuropathological contexts.

To obtain high-quality data from biological samples, several sources of technical variability in ASAP–MS were systematically investigated. These included background contamination by residual calibrant, probe cooling after background acquisition, probe cleaning protocols, contamination from consumables, inter-user variability, and batch effects. For each factor, optimisation strategies were developed to improve data quality.

Beyond technical considerations, the influence of molecular class and sample composition on ASAP mass spectra was examined. Small polar metabolites in biological samples were found to undergo extensive in-source chemistry, generating dense and highly correlated spectral features. In contrast, lipid species from biological samples ionised more predictably and contributed largely additive signals, resulting in more stable and interpretable spectral patterns.

Building on this, the feasibility of applying ASAP–MS to formalin-fixed, paraffin-embedded (FFPE) brain tissue samples was evaluated by comparison with frozen brain tissue samples. Focused ultrasonication was shown to be an effective deparaffinisation approach. However, preservation method and fixation time strongly influenced spectral quality. Short-fixed FFPE samples retained more consistent molecular fingerprints than long-fixed samples, yet frozen brain tissue samples consistently outperformed FFPE brain tissue samples for ASAP–MS analysis.

Finally, an optimised protocol for analysing fresh brain tissue samples using ASAP–MS was developed. The resulting mass spectra were used to distinguish brain tumour from normal brain cortex. Robust performance was observed across iterative model updates and predictions on unseen samples. SHAP-based model interpretation indicated that classification was driven primarily by lipid-rich spectral regions. The best-performing model, a random forest with oversampling, achieved a sensitivity of 0.97, a specificity of 0.95, and a Cohen’s kappa of 0.90, indicating excellent agreement beyond chance. These results demonstrate that ASAP–MS, when combined with appropriate protocol optimisation and machine learning, has genuine potential as a rapid support tool for brain tissue classification.

Authors: Song, L

• DOI: 10.5287/ora-2n2xrbj60
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: brain biopsy; ambient ionisation mass spectrometry; atmospheric solids analysis probe; machine learning; brain tumour
• Subjects: Machine learning; Analytical biochemistry; Bioinformatics; Chemistry, Physical and theoretical; Neurosciences