Retrieve and Revise: Improving Peptide Identification with Similar Mass Spectra

Zhengcong FeiAAAI 2021 · DOI

Introduction

In the field of proteomics, tandem mass spectrometry (MS/MS) serves as the primary tool for identifying protein sequences. The core challenge lies in accurately mapping a mass spectrum—a complex plot of ion intensities—to its corresponding peptide sequence. Traditional deep learning approaches typically employ an encoder-decoder architecture, where a Convolutional Neural Network (CNN) encodes the spectrum and a Recurrent Neural Network (RNN) generates the sequence amino acid by amino acid.

However, these "one-pass" decoding methods suffer from two critical limitations. First, they lack global context; predicting the initial amino acids occurs without information about the remainder of the sequence. Second, they are highly sensitive to noise. Because MS/MS data is inherently noisy, an error in the early stages of decoding can propagate through the entire sequence, leading to incorrect identifications.

The Retrieve and Revise Framework

To address these challenges, we introduce a "Retrieve and Revise" framework. This approach shifts the paradigm from generating sequences from scratch to editing a high-quality reference.

The Retrieve and Revise architecture combines a similar spectrum retriever with a sequence reviser and RL-based optimization.

The framework consists of four primary components:

  1. Similar Spectrum Retriever: For any input spectrum $S$, the system queries a database to find the most similar spectrum $S'$ using Pearson and Spearman correlation. This retrieves both the reference spectrum and its known peptide sequence $P'$.
  2. Difference Information Encoding: Rather than processing the input in isolation, the model focuses on the differences between $S$ and $S'$. A dual-attention mechanism identifies these discrepancies to create a "difference vector."
  3. Peptide Sequence Reviser: An LSTM-based decoder is conditioned on both the reference peptide $P'$ (processed via a BiLSTM) and the difference vector. This allows the model to "edit" the reference sequence to match the new input, leveraging the reference as a global prior.
  4. Disturb-based RL Optimization: To enhance robustness against noise, the model is trained using Reinforcement Learning. By intentionally perturbing the difference vector during training and rewarding the model for correct identification despite these disturbances, the attention mechanism is forced to focus on high-signal features.

Experimental Results

The model was evaluated across five diverse datasets, including cross-species (Mouse) and cross-lab (Human, E. coli, Yeast) validations. The "Retrieve and Revise" approach consistently outperformed both traditional statistical tools and modern deep learning baselines.

Comparison of Area Under Curve (AUC) across four major datasets shows our model consistently outperforming state-of-the-art baselines.

Key performance metrics include:

  • Superior Accuracy: On the Mann-Mouse-QEHF dataset, the model achieved a Peptide Recall (PR) of 0.293, surpassing the previous state-of-the-art DeepTag (0.289) and nearly doubling the performance of the industry-standard PEAKS (0.145).
  • Generalizability: The model maintained its performance edge across different organisms and laboratory setups, demonstrating that the logic of comparative revision is broadly applicable.
  • Precision-Recall Balance: The framework showed a 3.1% absolute improvement in PR score over the best competitor, indicating higher reliability in identifying entire peptide sequences correctly.

Conclusion

The "Retrieve and Revise" paradigm offers a significant advancement in peptide identification by treating the problem as a comparative task rather than a generative one. By utilizing existing databases as a source of global context, the model effectively mitigates the impact of experimental noise and the limitations of one-pass decoding. Future work will focus on optimizing the computational overhead of database retrieval and expanding the framework to handle a wider array of post-translational modifications.

Summary updated

BibTeX
@inproceedings{fei2021retrieve,
  author = {Zhengcong Fei},
  title = {{Retrieve and Revise: Improving Peptide Identification with Similar Mass Spectra}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i17.17734},
}
Figure