MSRescue improves De Novo candidate selection

Background: Successful immunotherapy against cancer necessitates the precise characterization of human leukocyte antigen class I (HLA-I) binding peptides to identify suitable targets. Liquid chromatography-tandem mass spectrometry (LC-MS) based immunopeptidomics offers high throughput and accuracy in identifying HLA-I binding peptides from canonical protein degradation products, but still lacks streamlined procedures for non-canonical antigen identification. Previously, we developed the MARS algorithm that proposes to incorporate HLA binding prediction and retention time differences to improve de-novo candidate selection, which improved identification of non-canonical neoantigens from aberrant or unknown RNA translations. Here, we propose an alternative strategy, which incorporates sequence clustering for improved identification of peptide sequences.

Method: We trained our previously published neural network model MSRescue using sequence clusters from database search-identified peptides with low false discovery rates (FDR) to predict HLA binding ranks. We used the pancreatic immunopeptidome of Human HLA Atlas UDN02 with 3017 unique peptides, to predict the de novo candidates of the lung immunopeptidome from the same subject using Peaks X de novo, Peaks 11 DeepNovo, and NovoR candidates. Fractional ranks (F-Ranks) of true candidates were used to measure overall improvement, and full sequence recall (FSR) was used to measure performance of a simple identification scheme by picking top candidates per spectrum.

Results: We constructed a set of ground truth spectra using Peaks X and Peaks 11 database searches against the SwissProt human proteome, resulting in 7097 unique peptides from 27561 spectra. By allowing up to 20 candidates per spectrum, DeepNovo reported 2765 peptides as the first candidate from 6007 spectra containing the correct sequence in the candidate list, Peaks X de novo recovered 3792 peptides / 7899 spectra, and NovoR identified 5502 peptides / 12514 spectra. Both our original MARS method and our new approach significantly improved F-Ranks of true candidates when applied to Peaks and NovoR candidates, but the FSR improvements were not significant.

Conclusion: Applying MSRescue improves de novo candidate ranking, and enables post-processing of de novo candidate data without prior knowledge on HLA haplotype. In contrast to our previously published MARS method, this technique can be applied in samples where HLA typing is not possible or permitted, facilitating the exploration of non-canonical neoantigens.