This groundbreaking study introduces a novel deep learning-based approach called Proformer for accurate classification of proteomic samples. Unlike conventional convolutional neural networks (CNNs), Proformer utilizes a transformer architecture that leverages a self-attention mechanism, enabling efficient learning of complex relationships within multi-dimensional proteomic data.

By integrating retention time, mass-to-charge, intensity, and ion mobility features from mass spectrometry measurements, Proformer demonstrates superior performance compared to CNNs.

The inclusion of ion mobility as a fourth feature dimension further enhances the model’s accuracy, underscoring the importance of leveraging advanced separation techniques. The study also highlights Proformer’s robustness against batch effects and its ability to classify samples even at the single-cell level, paving the way for broader applications in patient stratification and early disease detection. Additionally, the authors provide insights into the model’s decision-making process through interpretability methods like SHAP values, offering a comprehensive understanding of Proformer’s inner workings.

The IonOpticks Aurora Ultimate CSI Series, 25cm nanoflow UHPLC column enabled enhanced separation and contributed to the model’s impressive accuracy.

Pre – Published in BioRxiv, February 07, 2024.

Authors

Karl K. Krull, Arlene Kühn, Julia Höhn, Titus J. Brinker, Jeroen Krijgsveld

Title

Deep learning-based proteomics enables accurate classification of bulk and single-cell samples