π -PrimeNovo: An Accurate and Efficient Non-Autoregressive Deep Learning Model for De Novo Peptide Sequencing

Xiang Zhang, Tianze Ling, Zhi Jin, Sheng Xu, Zhiqiang Gao, Boyan Sun, Zijie Qiu, Nanqing Dong, Guangshuai Wang, Guibin Wang, Leyuan Li, Muhammad Abdul-Mageed, Laks V. S. Lakshmanan, Wanli Ouyang, Cheng Chang, Siqi Sun

TL;DR

This paper introduces π-PrimeNovo, a novel non-autoregressive deep learning model for de novo peptide sequencing. By breaking away from traditional autoregressive methods, π-PrimeNovo achieves significantly higher accuracy and dramatically faster inference speeds, making it a powerful tool for complex proteomic analyses like metaproteomics and PTM discovery.

What Problem Does This Paper Address?

Identifying proteins is a cornerstone of biological research, and a key step is determining the amino acid sequence of peptides derived from them. While database searching is common, it fails when dealing with novel or uncharacterized proteins. This is where de novo peptide sequencing comes in, aiming to deduce sequences directly from mass spectrometry data. Existing deep learning models for this task predominantly use an autoregressive approach, where each amino acid is predicted based on the previous ones. This can lead to error propagation, where a mistake early on cascades and reduces overall accuracy. Furthermore, these sequential methods, often combined with beam search, are computationally slow, hindering their application in large-scale analyses like metaproteomics.

Methods

π-PrimeNovo tackles these challenges with a novel non-autoregressive Transformer architecture. Unlike autoregressive models, it predicts all amino acids simultaneously, allowing for a more holistic understanding of the peptide sequence and avoiding error cascading. A key innovation is the Precise Mass Control (PMC) unit, which works in tandem with the non-autoregressive framework to ensure the generated peptide sequence precisely matches the measured precursor mass from the mass spectrometer. This is achieved through a dynamic programming approach that considers both predicted probabilities and mass constraints. The model is trained using Connectionist Temporal Classification (CTC) loss, which helps maintain global sequence coherence. For computational efficiency, a CUDA-accelerated algorithm is employed for the PMC unit.

Datasets

The study utilizes several datasets:

  • MassIVE-KB: A large-scale spectral library used for training the core π-PrimeNovo model.
  • Nine-species benchmark dataset: A standard dataset for evaluating de novo sequencing models, used for both training (with cross-validation) and testing.
  • Revised nine-species benchmark dataset: A more recent and larger version of the nine-species dataset, used for performance evaluation.
  • ProteomeTools (PT) dataset: A collection of synthetic human peptides, used for fine-tuning and evaluating adaptability.
  • HCC dataset: A real-world dataset from early-stage human hepatocellular carcinoma patients, used for evaluating generalizability.
  • IgG1-Human-HC dataset: A dataset of human antibody sequences processed with various enzymes, used to assess generalization on novel sequences.
  • Three-species dataset: Comprising samples from Arabidopsis thaliana, Caenorhabditis elegans, and Escherichia coli, used for comparative analysis.
  • PXD019483 dataset: Used for comparing performance with PepNet.
  • 21PTMs dataset: A benchmark dataset containing a wide variety of post-translational modifications (PTMs).
  • 2020-Cell-LUAD dataset: A phosphoproteomics dataset from human lung adenocarcinoma, used for evaluating PTM detection capabilities, particularly on non-enriched samples.

Key Insights and Ideas

  • Non-autoregressive is the future: The paper strongly advocates for non-autoregressive models in peptide sequencing, demonstrating that simultaneous prediction can lead to both higher accuracy and speed.
  • Mass matters: The PMC unit is crucial. It's not enough to predict the sequence; it must match the physical properties (mass) of the peptide, and this integration is key to π-PrimeNovo's success.
  • Robustness is key: π-PrimeNovo shows remarkable resilience to common challenges in MS/MS data, such as missing peaks and varying peptide lengths, outperforming existing methods across these variations.
  • PTMs are no longer a bottleneck: The model's ability to accurately predict peptides with PTMs, even in non-enriched datasets, opens doors for discovering low-abundance modifications.

Key Results

  • Accuracy: π-PrimeNovo achieves an average peptide recall of 64% on the nine-species benchmark dataset, a 10% improvement over the previous state-of-the-art (Casanovo V2). On a revised benchmark, it reached 75%.
  • Speed: The model is up to 69x faster than state-of-the-art autoregressive methods, making it practical for large-scale metaproteomic analyses.
  • Metaproteomics: It significantly enhances the identification of taxon-unique peptides, aiding in microbial species identification.
  • PTM Detection: Demonstrates powerful capability in accurately mining phosphopeptides from non-enriched datasets, offering a new solution for low-abundance PTM discovery.
  • Generalization: Shows strong performance across diverse datasets and conditions, including zero-shot scenarios.

Limitations and Future Work

  • Computational Cost of PMC: While CUDA acceleration is used, the dynamic programming for PMC can still be computationally intensive, especially with very high precision requirements.
  • Fine-tuning Data Requirements: The paper notes that significant improvements in fine-tuning performance are observed after incorporating a substantial amount of new data (around 10,000 points), suggesting that data availability is still a factor.
  • Catastrophic Forgetting: While addressed by mixing training data, the phenomenon of forgetting original data distributions during fine-tuning is acknowledged.

Future work could involve further optimizing the PMC unit, exploring more advanced attention mechanisms, and applying the model to even more complex and challenging proteomic datasets.

Summary updated

Abstract

Abstract Peptide sequencing via tandem mass spectrometry (MS/MS) is fundamental in proteomics data analysis, playing a pivotal role in unraveling the complex world of proteins within biological systems. In contrast to conventional database searching methods, deep learning models excel in de novo sequencing peptides absent from existing databases, thereby facilitating the identification and analysis of novel peptide sequences. Current deep learning models for peptide sequencing predominantly use an autoregressive generation approach, where early errors can cascade, largely affecting overall sequence accuracy. And the usage of sequential decoding algorithms such as beam search suffers from the low inference speed. To address this, we introduce π -PrimeNovo, a non-autoregressive Transformer-based deep learning model designed to perform accurate and efficient de novo peptide sequencing. With the proposed novel architecture, π -PrimeNovo achieves significantly higher accuracy and up to 69x faster sequencing compared to the state-of-the-art methods. This remarkable speed makes it highly suitable for computation-extensive peptide sequencing tasks such as metaproteomic research, where π -PrimeNovo efficiently identifies the microbial species-specific peptides. Moreover, π -PrimeNovo has been demonstrated to have a powerful capability in accurately mining phosphopeptides in a non-enriched phosphoproteomic dataset, showing an alternative solution to detect low-abundance post-translational modifications (PTMs). We suggest that this work not only advances the development of peptide sequencing techniques but also introduces a transformative computational model with wide-range implications for biological research.
BibTeX
@inproceedings{zhang2024i,
  author = {Xiang Zhang and Tianze Ling and Zhi Jin and Sheng Xu and Zhiqiang Gao and Boyan Sun and Zijie Qiu and Nanqing Dong and Guangshuai Wang and Guibin Wang and Leyuan Li and Muhammad Abdul-Mageed and Laks V. S. Lakshmanan and Wanli Ouyang and Cheng Chang and Siqi Sun},
  title = {{<i>π</i> -PrimeNovo: An Accurate and Efficient Non-Autoregressive Deep Learning Model for De Novo Peptide Sequencing}},
  booktitle = {Nature Communications},
  year = {2024},
  url = {https://openreview.net/forum?id=AW0DaPH5aw},
}
Figure