PEAKS is well-known and considered the gold standard for peptide de novo sequencing. To advance de novo sequencing even further, BSI was first to introduce deep learning technology into peptide de novo sequencing of mass spectrometry data and reported a significant improvement in accuracy [1,2]. The speed and accuracy of this technology naturally draws interest to integrate our solution, DeepNovo, into PEAKS.

As a key technology for finding novel peptides from mass spectrometry (MS) data, DeepNovo provides an advanced solution and will push research like antibody sequencing and neoantigen discovery further.

Key Features:

  • NEW DeepNovo – deep learning-based de novo sequencing to increase amino acid and peptide level accuracy
  • Harness GPUs and accelerate de novo sequencing
  • Improvements of peptide de novo sequencing observed across different MS technologies

DeepNovo outperforms our previous PEAKS de novo algorithm by 20.5% accuracy at the amino acid level and 16.6% at the peptide level, when tested with Bruker timsTOF SCP data [3] (Table 1). Moreover, significant improvements with ThermoFisher Orbitrap data [4] was observed, with an accuracy increase of 5.5% and 10.6% at the amino acid and peptide levels, respectively (Table 2).

Accuracy
ToolsAmino AcidPeptide
PEAKS de novo60.9%34.6%
DeepNovo81.4%51.2%
Table 1. timsTOF Accuracy of PEAKS de novo and DeepNovo predictions of PSMs at the amino acid and peptide levels.
Accuracy
ToolsAmino AcidPeptide
PEAKS de novo72.0%42.5%
DeepNovo77.5%53.1%
Table 2. Orbitrap Accuracy of PEAKS de novo and DeepNovo predictions of PSMs at the amino acid and peptide levels.

From testing, we compared PEAKS DB, PEAKS de novo, and DeepNovo peptides. We observed an overlap between peptide sequences, as well as the overlap of the amino acids predicted by PEAKS de novo and DeepNovo. PEAKS de novo and DeepNovo shared a large amount of amino acid assignments (up to 64%). More importantly, compared to PEAKS de novo, DeepNovo shared significantly more peptides with PEAKS DB.

Figure 1. Venn diagrams showing an overlap of PEAKS DB, PEAKS de novo and DeepNovo peptides (left) and an overlap of amino acids predicted by PEAKS de novo and DeepNovo (right).
Figure 2. The accuracy-versus-score distributions of PEAKS de novo and DeepNovo peptides from timsTOF SCP data.
Figure 3. The accuracy-versus-score distributions of PEAKS de novo and DeepNovo peptides from Thermo Orbitrap data.

To establish a new peptide scoring threshold with DeepNovo, we compared the amino acid accuracy vs. score distributions between PEAKS de novo and DeepNovo results using timsTOF and Orbitrap data.  In both cases, DeepNovo achieves ~95% accuracy at the amino acid level with a score cutoff of 55, whereas at the same score cutoff PEAKS de novo peptides have ~80% accuracy in amino acid predictions. These results demonstrate the enhanced accuracy in peptide sequencing when using deep learning technology. Consistent with this, our experience with many large public datasets and internal data indicates that ~95% accuracy in peptide amino acid sequencing requires minimum scores of 80 and 55 when using PEAKS de novo and DeepNovo, respectively (for timsTOF data).

References & Resources

References

  1. Tran, N. H., et al. (2017). De novo peptide sequencing by deep learning. Proceedings of the National Academy of Sciences of the United States of America, 114(31), 8247–8252. https://doi.org/10.1073/pnas.1705691114
  2. Tran, N. H., et al. (2019). Deep learning enables de novo peptide sequencing from data-independent-acquisition mass spectrometry. Nature methods, 16(1), 63–66. https://doi.org/10.1038/s41592-018-0260-3
  3. Phulphagar KM, et al. (2023). Sensitive, high-throughput HLA-I and HLA-II immunopeptidomics using parallel accumulation-serial fragmentation mass spectrometry. bioRxiv. 532106. doi: 10.1101/2023.03.10.532106
  4. Tretter C, et al. (2023). Proteogenomic analysis reveals RNA as a source for tumor-agnostic neoantigen identification. Nat Commun. 14(1):4632. doi: 10.1038/s41467-023-39570-7

Resources