Predicting peptide properties from mass spectrometry data using deep attention-based multitask network and...

Area of Science:

• Proteomics
• Computational Biology
• Machine Learning

Background:

• Database search algorithms in proteomics often use single properties like mass for peptide filtering.
• This single-property filtering can lead to the exclusion of valuable data, known as the 'streetlight' effect.
• There is a need for more comprehensive methods to filter candidate peptides effectively.

Purpose of the Study:

• To introduce ProteoRift, a novel deep learning network for predicting multiple peptide properties directly from spectra.
• To demonstrate ProteoRift's ability to significantly reduce the peptide search space.
• To develop uncertainty metrics for assessing data distribution and prediction confidence.

Main Methods:

• Developed ProteoRift, an attention and multitask deep network.
• Trained the network to predict peptide length, missed cleavages, and modification status from mass spectra.
• Formulated two uncertainty estimation metrics for data classification and high-scoring spectra prediction.

Main Results:

• ProteoRift achieved up to 97% accuracy in predicting peptide properties.
• The method reduced the search space by over 90%.
• The end-to-end pipeline demonstrated 8x to 12x speedups with comparable peptide deduction accuracy.
• Uncertainty metrics showed high performance in distinguishing data types (ROC-AUC 0.99) and predicting correct peptides (ROC-AUC 0.94).

Conclusions:

• ProteoRift offers a powerful deep learning approach for accelerating proteomic data analysis.
• The tool effectively mitigates the 'streetlight' effect by considering multiple peptide properties.
• Integrated uncertainty metrics enhance the reliability and interpretability of proteomic search results.