Evaluation of machine learning models for proteoform retention and migration time prediction in top-down mass spectrometry

Wenrong Chen, Elijah N. McCool, Liangliang Sun, Yong Zang, Xia Ning, Xiaowen LiubioRxiv 2021 · DOI

TL;DR

This paper tackles the challenge of accurately predicting proteoform retention and migration times in top-down mass spectrometry, a crucial step for improving proteoform identification. They found that the Prosit model, originally developed for bottom-up proteomics, performs exceptionally well in this top-down context, achieving high accuracy for both retention and migration time predictions.

What Problem Does This Paper Address?

Top-down mass spectrometry (MS) is a powerful technique for analyzing intact proteins (proteoforms), enabling their identification, characterization, and quantification. However, accurately identifying these proteoforms can be challenging. One way to improve identification accuracy is by predicting their retention time (in liquid chromatography, LC) or migration time (in capillary electrophoresis, CZE). These predicted times can act as an additional filter, helping to distinguish correct proteoform identifications from incorrect ones. While significant progress has been made in predicting peptide retention times in bottom-up MS (where proteins are fragmented into smaller peptides), there's a notable gap in methods for predicting these times for intact proteoforms in top-down MS. This is partly due to the inherent complexity of longer proteoforms and the scarcity of high-quality training data.

Methods

The researchers evaluated six different machine learning models for predicting proteoform retention and migration times. These included:

  • GPTime: A model that uses a large set of features related to amino acid properties and sequence.
  • Fully Connected Neural Network (FNN): A standard neural network architecture.
  • DeepRT+: A convolutional neural network (CNN)-based model.
  • Prosit: A recurrent neural network (RNN)-based model, originally designed for peptide spectrum prediction.
  • DeepDIA: A hybrid CNN-RNN model.
  • Semi-empirical model: A model based on fundamental physicochemical properties like molecular mass and charge.

For migration time prediction in CZE, they also developed a method to remove batch effects, which are variations introduced by different experimental runs. This involved converting migration times to electrophoretic mobility, correcting for batch effects using linear regression, and then converting back to migration times.

A key aspect of their work was also exploring transfer learning. This involved pre-training models on large datasets from bottom-up MS (peptides) and then fine-tuning them on the top-down MS data.

Datasets

The study utilized two primary datasets:

  • Ovarian Tumor RPLC-MS/MS data (LC-OT): This dataset was used for training and evaluating retention time prediction models. It contained proteoforms identified from ovarian tumor samples.
  • SW480 Colon Cancer CZE-MS/MS data (CZE-SW480): This dataset was used for training and evaluating migration time prediction models. It was generated from SW480 colon cancer cells and involved prefractionation using size exclusion chromatography (SEC).

The researchers also leveraged publicly available datasets for transfer learning, specifically a large bottom-up MS dataset of peptides from human cell lines and tissues, and a bottom-up CZE-MS/MS dataset from HeLa cells.

Key Insights and Ideas

  • Prosit's surprising success: The Prosit model, originally designed for a different task (peptide spectrum prediction) and a different MS approach (bottom-up), proved to be highly effective for proteoform retention and migration time prediction in top-down MS. This suggests that its underlying architecture, particularly its ability to process sequential data, is well-suited for this problem.
  • Transfer learning is beneficial, but not universally: Pretraining models on bottom-up peptide data significantly improved the accuracy of some models (like DeepRT+) for top-down predictions, especially when top-down training data was limited. However, it didn't consistently boost the performance of already high-performing models like Prosit, suggesting that the benefit of transfer learning depends on the model architecture and the nature of the transferable information.
  • Batch effect correction is critical for CZE: Variations between CZE runs can significantly impact migration time predictions. The study demonstrated that batch effect correction is an indispensable step to achieve accurate migration time predictions in CZE-MS.
  • Migration time prediction is generally easier than retention time: Across most models, migration time prediction achieved higher accuracy than retention time prediction. This might be due to the more direct relationship between molecular properties (mass, charge) and migration time in CZE compared to the complex interplay of factors influencing retention in RPLC.

Key Results

  • Retention Time Prediction: The Prosit model achieved the highest accuracy for proteoform retention time prediction, with an R² of 0.906.
  • Migration Time Prediction: The FNN model achieved the highest accuracy for proteoform migration time prediction, with an R² of 0.959. The Prosit model also performed very well, with an R² of 0.943.
  • Batch Effect Correction: After batch effect correction, the R² for migration time prediction on the CZE-SW480 data improved from 0.613 to 0.915.
  • Transfer Learning Impact: For retention time prediction on the LC-OT dataset, transfer learning improved the Prosit model's R² from 0.860 to 0.914. For migration time prediction on the CZE-SW480 dataset, transfer learning improved DeepRT+'s R² from 0.750 to 0.888.

Limitations and Future Work

  • Limited Top-Down Data: The authors acknowledge that the scarcity of large, high-quality training datasets for top-down proteoform prediction remains a significant challenge. This limitation likely contributed to the lower performance of some complex deep learning models.
  • Model Suitability for Long Sequences: Complex deep learning models might struggle with the longer sequences of proteoforms compared to peptides, impacting their prediction accuracy.
  • PTMs and Batch Effects: Post-translational modifications (PTMs) and batch effects are identified as compounding factors that complicate proteoform migration/retention time prediction. These are highlighted as key areas for future research.
  • Experiment Setting Variability: Predicting retention/migration times across experiments with different settings (e.g., different LC columns or CZE conditions) is a challenging problem that requires further investigation.
  • Need for Larger Datasets: The authors emphasize the need for larger datasets to further test and improve machine learning models for proteoform retention/migration time prediction.

Summary updated

Abstract

Abstract Reversed-phase liquid chromatography (RPLC) and capillary zone electrophoresis (CZE) are two popular proteoform separation methods in mass spectrometry (MS)-based top-down proteomics. The prediction of proteoform retention time in RPLC and migration time in CZE provides additional information that can increase the accuracy of proteoform identification and quantification. Whereas existing methods for retention and migration time prediction are mainly focused on peptides in bottom-up MS, there is still a lack of methods for the problem in top-down MS. We systematically evaluated 6 models for proteoform retention and/or migration time prediction in top-down MS and showed that the Prosit model achieved a high accuracy ( R 2 > 0.91) for proteoform retention time prediction and that the Prosit model and a fully connected neural network model obtained a high accuracy ( R 2 > 0.94) for proteoform migration time prediction.
BibTeX
@misc{chen2021evaluation,
  author = {Wenrong Chen and Elijah N. McCool and Liangliang Sun and Yong Zang and Xia Ning and Xiaowen Liu},
  title = {{Evaluation of machine learning models for proteoform retention and migration time prediction in top-down mass spectrometry}},
  year = {2021},
  doi = {10.1101/2021.10.31.466700},
  howpublished = {bioRxiv},
}
Figure