TL;DR
This paper tackles the challenge of accurately predicting proteoform retention times (RTs) in liquid chromatography and migration times (MTs) in capillary electrophoresis for top-down mass spectrometry. By evaluating various machine learning models, including a novel transfer learning approach, they demonstrate that deep learning models, particularly GRU-based architectures, can achieve high prediction accuracy, significantly aiding proteoform identification.
What Problem Does This Paper Address?
Top-down mass spectrometry (MS) is crucial for analyzing intact proteins (proteoforms) and understanding complex post-translational modifications. However, accurately identifying these proteoforms can be challenging. Retention time (RT) in liquid chromatography (LC) and migration time (MT) in capillary electrophoresis (CE) are key experimental parameters that can help confirm proteoform identities. While methods for predicting peptide RT/MT in bottom-up MS exist, there's a significant gap in similar predictive tools for the larger, intact proteoforms analyzed in top-down MS. This lack of predictive power hinders accurate proteoform identification and quantification, especially when spectral data is limited.
Methods
The researchers systematically evaluated eight machine learning models for proteoform RT prediction and five for MT prediction. These included traditional methods like logistic regression (LR), support vector regression (SVR), and random forest regression (RFR), alongside established deep learning architectures such as convolutional neural networks (CNNs), recurrent neural networks (RNNs, specifically GRUs and LSTMs), and fully connected neural networks (FNNs). Notably, they also implemented and assessed a transfer learning strategy. This involved pre-training models on large datasets of peptides from bottom-up MS experiments and then retraining them on smaller proteoform datasets from top-down MS. This approach aims to leverage knowledge learned from abundant peptide data to improve predictions for sparser proteoform data.
Datasets
The study utilized two primary top-down MS datasets:
- RPLC-MS/MS data: From ovarian tumor samples (referred to as LC-ONE and LC-TEN).
- CZE-MS/MS data: From SW480 and SW620 colon cancer cells (referred to as CZE-ONE and CZE-ALL).
For transfer learning, they also incorporated large bottom-up MS datasets:
- LC-PEPTIDE: A dataset of 146,587 unique peptides identified from human cell lines and tissues.
- CZE-PEPTIDE: A dataset of 4234 unique peptides identified from HeLa cells.
Key Insights and Ideas
- Deep Learning Dominance: Advanced deep learning models, particularly GRU-based architectures, consistently outperformed traditional machine learning methods for both RT and MT prediction.
- Transfer Learning's Power: Transfer learning proved highly effective, especially when top-down proteoform datasets were limited. Pretraining on large peptide datasets significantly boosted the prediction accuracy of proteoform models.
- RT Prediction is Tricky: MT prediction generally achieved higher accuracy than RT prediction across most models, suggesting that RT prediction is inherently more complex, possibly due to the greater number of factors influencing chromatographic separation.
- Calibration is Crucial: For CZE-MS, variations between runs significantly impacted experimental MTs. Calibration of these variations was shown to be an indispensable step for achieving high prediction accuracy.
- Model Architecture Matters: The GRU + FNN model, with its design for sequential data and attention mechanisms, showed remarkable robustness and high accuracy even with limited proteoform data, suggesting its suitability for handling longer proteoform sequences.
Key Results
- Proteoform RT Prediction: The GRU + FNN model with transfer learning achieved a high accuracy of R = 0.978 on the LC-TEN dataset.
- Proteoform MT Prediction: The GRU + FNN and a simple FNN model achieved high accuracies of R = 0.982 and R = 0.981, respectively, on the CZE-ONE and CZE-ALL datasets.
- Transfer Learning Impact: For the GRU + FNN model, transfer learning improved RT prediction accuracy from R = 0.929 to R = 0.978 on the LC-TEN test set.
- Filtering Benefits: Incorporating RT/MT prediction for filtering identified proteoforms led to an increase in proteoform identifications (e.g., 5.9% for LC-ONE data) by helping to remove false positives.
Limitations and Future Work
- Data Scarcity: A major limitation is the lack of large, comprehensive top-down proteoform datasets, which hinders the training of highly complex models. The authors suggest combining data from multiple species or experimental settings.
- Modified Proteoforms: The current models primarily focus on unmodified proteoforms. Predicting RT/MT for proteoforms with post-translational modifications (PTMs) remains a significant challenge due to even sparser data.
- Method Transferability: Applying models trained on one dataset to data generated with different MS settings can be problematic, requiring recalibration or retraining.
- Future Directions: The authors plan to generate larger training datasets, further improve prediction accuracy, and extend their work to predict RT/MT for modified proteoforms. They also aim to investigate how to better transfer models across different experimental settings.
Summary updated