Retention Time Prediction Using Neural Networks Increases Identifications in Crosslinking Mass Spectrometry

Abstract:

Crosslinking mass spectrometry (Crosslinking MS) has developed into a robust technique that is increasingly used to investigate the interactomes of organelles and cells. However, the incomplete and noisy information contained in mass spectra limits the numbers of protein-protein interactions (PPIs) that can be confidently identified. Here, we successfully leveraged chromatographic retention time (RT) information to aid the identification of crosslinked peptides from mass spectra. Our Siamese machine learning model xiRT achieved highly accurate RT predictions of crosslinked peptides in a multi-dimensional separation of crosslinked E. coli lysate. We combined strong cation exchange (SCX), hydrophilic strong anion exchange (hSAX) and reversed-phase (RP) chromatography and reached  0.94 in RP and a margin of error of 1 fraction for hSAX in 94%, and SCX in 85% of the predictions. Importantly, supplementing the search engine score with retention time features led to a 1.4-fold increase in PPIs at a 1% false discovery rate. We also demonstrate the value of this approach for the more routine analysis of a crosslinked multiprotein complexes. An increase of 1.7-fold in heteromeric crosslinked residue-pairs was achieved at 1% residue-pair FDR for Fanconi anaemia monoubiquitin ligase complex, solely using reversed-phase RT. Retention times are a powerful complement to mass spectrometric information to increase the sensitivity of Crosslinking MS analyses.

Conclusion:

Using a Siamese network architecture, we succeeded in bringing RT prediction into the Crosslinking MS field, independent of separation setup and search software. Our open source application xiRT introduces the concept of multi-task learning to achieve multi-dimensional chromatographic retention time prediction, and may use any peptide sequence-dependent measure including for example collision cross section or isoelectric point. The black-box character of the neural network was reduced by means of interpretable machine learning that revealed individual amino acid contributions towards the separation behavior. The RT predictions – even when using only the RP dimension – complement mass spectrometric information to enhance the identification of heteromeric crosslinks in multiprotein complex and proteome-wide studies. Overfitting does not account for this gain as known false target matches from an entrapment database did not increase. Leveraging additional information sources may help to address the mass-spectrometric identification challenge of heteromeric crosslinks.