A Transformer-Based Framework for Quality Control of Peptide Tandem Mass Spectra

In tandem mass spectrometry (MS/MS)-based proteomics, a significant portion of acquired spectra remains unidentified due to poor quality, which consumes excessive computational resources and increases false-positive rates during database searches. Traditional quality assessment methods rely on handcrafted features or classical machine learning that often generalize poorly across different instruments. While recent deep-learning approaches like SPEQ (spectrum quality) have introduced automation, their reliance on convolutional architectures and supervised learning limits their ability to capture global spectral dependencies and transfer across heterogeneous data sets. To address these limitations, we present a pretrained transformer framework that utilizes self-attention for automated MS/MS quality assessment. By leveraging self-supervised pretraining to learn robust, contextualized spectral representations, our model can capture global fragment relationships and ensure superior cross-instrument transferability. Our results demonstrate that it consistently outperforms existing models like SPEQ, offering higher accuracy and enhanced generalization across diverse data sets.