Resolution-Adaptive Binning Enhances Machine Learning Modeling by Interbatch and Multiplatform Orbitrap-Based Shotgun...

Area of Science:

• Analytical Chemistry
• Computational Biology
• Biomedical Science

Background:

• Machine learning (ML) on mass spectrometry (MS) data aids disease modeling but faces challenges with batch-specific models and limited generalizability.
• Traditional data integration methods struggle with mass-to-charge (m/z) features, hindering reliable data combination across different MS batches and platforms.
• The need for robust data integration strategies is critical for advancing ML applications in disease diagnostics using MS.

Purpose of the Study:

• To develop and validate a mass resolution-adaptive binning and integration strategy for MS data.
• To enhance the transferability and generalizability of ML models across diverse MS datasets and platforms.
• To improve disease detection accuracy by enabling reliable integration of multi-batch and multi-platform MS data.

Main Methods:

• A novel mass resolution-adaptive binning strategy was developed to handle m/z features across varying mass spectrometry resolutions.
• The strategy was tested on mixed standard solutions, recovering 88-99% of ground truth features in the low mass region.
• The approach was applied to a mouse model of hepatocellular carcinoma, integrating ambient MS imaging (MSI) and shotgun proteomics data.

Main Results:

• The proposed method demonstrated stable binning and integration across low, mid, and high mass regions, outperforming conventional techniques.
• Predictive models built using the integrated data showed superior performance compared to those using conventional methods.
• In a hepatocellular carcinoma mouse model, 10 generic metabolites were identified, leading to accurate disease detection via MSI (F1 score = 0.87) and direct infusion (recall = 0.89).

Conclusions:

• The mass resolution-adaptive binning and integration strategy effectively overcomes limitations of traditional methods for MS data integration.
• This novel approach significantly improves the generalizability of ML models for disease detection across various sample introduction methods.
• The strategy holds promise for advancing MS-based diagnostics by enabling more accurate and reliable integration of diverse datasets.