Validation using sensitivity and target transform factor analyses of neural network models for classifying bacteria...

Area of Science:

• Computational biology
• Machine learning
• Analytical chemistry

Background:

• Neural networks offer potential for complex data analysis but lack transparency in input-output relationships.
• Classifying pathogenic bacteria using mass spectra requires robust and validated models.

Purpose of the Study:

• To develop and validate a Temperature Constrained Cascade Correlation Network (TCCCN) model for classifying five classes of pathogenic bacteria.
• To investigate the interpretability of TCCCN models by identifying key mass spectral features used in classification.

Main Methods:

• Utilized TCCCNs with a Latin-partition method for bacterial classification.
• Employed sensitivity analysis and Target Transformation Factor Analysis (TTFA) for model validation and feature identification.
• Acquired chemical ionization mass spectra from in situ thermal hydrolysis methylation of freeze-dried bacteria.

Main Results:

• A multiple-output TCCCN model achieved high classification accuracy (96 +/- 2%).
• Sensitivity analysis identified significant mass spectral peaks correlating with bacterial classes, many corresponding to known biomarkers.
• TTFA provided visual targets linking specific spectral peaks to bacterial identities.

Conclusions:

• TCCCNs provide a validated and interpretable approach for pathogenic bacteria classification from mass spectra.
• The study successfully divulged significant spectral peaks used by the neural network, enhancing model transparency.
• This work demonstrates the utility of TCCCNs and TTFA in identifying bacterial biomarkers from mass spectral data.