Deep learning from harmonized peptide libraries enables retention time prediction of diverse post translational modifications

Abstract

In proteomics experiments, peptide retention time (RT) is an orthogonal property to fragmentation when assessing detection confidence. Advances in deep learning enable accurate RT prediction for any peptide from sequence alone, including those yet to be experimentally observed. Here we present Chronologer, an open-source software tool for rapid and accurate peptide RT prediction. Using new approaches to harmonize and false-discovery correct across independently collected datasets, Chronologer is built on a massive database with >2.2 million peptides including 10 common post-translational modification (PTM) types. By linking knowledge learned across diverse peptide chemistries, Chronologer predicts RTs with less than two-thirds the error of other deep learning tools. We show how RT for rare PTMs, such as OGlcNAc, can be learned with high accuracy using as few as 10-100 example peptides in newly harmonized datasets. This iteratively updatable workflow enables Chronologer to comprehensively predict RTs for PTM-marked peptides across entire proteomes.

Machine learning strategies to tackle data challenges in mass spectrometry-based proteomics

“…Chronologer [50] is a residual convolutional neural network developed for peptide RT prediction. To train their prediction tool, the authors undertook significant efforts in composing a large-scale, comprehensive peptide RT dataset by aggregating and harmonizing 11 public datasets.…”

“…To gain insight into the current state of peptide property prediction datasets, we evaluated the performance of a transformer neural network encoder [56] in predicting the RT and CCS of a mixture of modified and unmodified peptides that were not previously seen. For RT prediction, we used a curated version of the Chronologer dataset [50], containing 2.1 million peptides and 12 PTMs. The dataset compiled by Meier et al [11], featuring ~714,000 peptides and 3 PTMs, was used for CCS prediction.…”

“…We investigated the effects of self-supervised pretraining, using both additional related data and the same dataset, as well as multi-task learning on peptide property prediction performance. For the following experiments, we used a combination of the RT dataset, which is a curated version of the dataset composed by Chronologer [50], and the dataset composed by Meier et al [11] for CCS values.…”

Machine learning‐based peptide‐spectrum match rescoring opens up the immunopeptidome

“…In this way, model performance can be optimized for specific datasets, even when only a limited amount of training data is available. Transfer learning is used by several deep learning tools [34, 44, 57] and may help to create models better suited for different scenarios, such as finetuning them for different fragmentation mechanisms, instrument platforms, and even lab‐specific data properties [58].…”

“…However, it is important to note that currently most PTMs found with open modification searching might not be supported by the underlying predictors. In this case, transfer learning can be used [57], as well as emerging algorithmic solutions that can predict peptide properties for unseen PTMs [34, 68]. Additionally, even though several approaches have been proposed to control the FDR in open modification searching [69], further research is necessary to investigate how the FDR behaves and to ensure that PSM rescoring does not introduce biases.…”