PrSMBooster: Improving the Accuracy of Top-Down Proteoform Characterization Using Deep Learning Rescoring Models

The aim of top-down mass spectrometry-based proteoform identification and characterization is to achieve optimal alignment between mass spectra and proteoforms. Consequently, the accuracy of identification results is crucial. Proteins with multiple primary structure alterations generate various proteoforms, leading to a combinatorial explosion due to their vast numbers. Furthermore, there is no gold set as a reference. So, enhancing the accuracy of identification results remains challenging. We propose a novel rescoring algorithm, PrSMBooster, which employs an ensemble approach. This approach utilizes non-deep models such as XGBoost, Decision Trees, and SVM as weak learners to extract latent features from proteoform spectrum matches. Ultimately, the deep learning model ResNeXt is used for final rescoring. We applied the PrSMBooster rescoring model to 47 independent cross-species datasets. Our comparison with the identification algorithm TopPIC demonstrates that PrSMBooster scores more accurately. In the vast majority of datasets, PrSM increases were observed at 1% FDR. Our findings indicate that PrSMBooster enhances scoring accuracy, reveals more identification results, and exhibits strong generalization capabilities.