Personalized de novo peptide sequencing to identify non-canonical HLA peptides from individual-patient immunopeptidome

Background

De novo peptide sequencing provides a direct and unbiased solution to identify non-canonical HLA peptides from mass spectrometry (MS) based immunopeptidomics. However, de novo peptides may often contain sequencing errors while neoantigen applications ultimately require completely accurate peptide sequences. Recent studies have developed new de novo sequencing methods and dedicated models that were optimized specifically for HLA data. Here we go one step further and propose a personalized deep learning model to learn from the immunopeptidome of each individual patient and then to predict non-canonical HLA peptides of that patient.

Methods

Given an MS-based immunopeptidomics of an individual patient, PEAKS DB search was performed to identify canonical peptide-spectrum matches (PSMs) at false discovery rate (FDR) of 1%. A universal GraphNovo model was applied to perform de novo sequencing and predict 100 candidate peptides per spectrum. The model was further fine-tuned on the PSMs identified by PEAKS DB to obtain a personalized de novo sequencing model. The personalized GraphNovo model was finally combined with other retention time and binding prediction models to select the best peptide out of 100 candidates per spectrum. Thus, the final peptide was strongly supported by multiple dimensions of evidence, including its spectrum, retention time, binding motif, and the personal HLA profile of the patient.

Results

We evaluated our personalized de novo sequencing approach on the datasets of ten cancer patients from three previously published studies. We also compared our personalized models to the universal GraphNovo model and to PEAKS De novo algorithm. The evaluation results show that our personalized models improved the peptide accuracy by 5-10% and increased the number of identified HLA peptides by 20-30%. Furthermore, the identified peptides also exhibited high binding levels and accurate binding motifs, especially at the second amino acid position, which often cause de novo sequencing errors due to the missing fragmentation problem of HLA peptides.

Conclusions

Personalized de novo peptide sequencing models represent the most advantageous and ideal solution to identify non-canonical HLA peptides of each individual patient.