Variability analysis of LC-MS experimental factors and their impact on machine learning

Area of Science:

• Computational biology
• Analytical chemistry
• Bioinformatics

Background:

• Machine learning (ML), particularly deep learning (DL), is increasingly used in mass spectrometry (MS) for data analysis and prediction.
• Large datasets are crucial for training ML models, often sourced from public repositories.
• Variability in data acquisition, biological systems, and experimental designs across public MS datasets poses challenges for ML application.

Purpose of the Study:

• To systematically analyze sources of variability in public MS repositories.
• To evaluate the impact of these factors on ML model performance.
• To assess the effectiveness of transfer learning in MS data analysis.

Main Methods:

• Systematic analysis of public mass spectrometry datasets.
• Evaluation of ML model performance across diverse datasets.
• Application and assessment of transfer learning techniques.

Main Results:

• Homogeneity within projects significantly higher than between projects.
• Transferability of ML models is limited for datasets dissimilar to training data.
• Transfer learning improved model performance but not beyond non-pretrained models.

Conclusions:

• Dataset construction should prioritize similarity to future test cases due to limited transferability.
• While transfer learning offers some improvement, its benefit over non-pretrained models in this context is not substantial.