Discovering and Validating Neoantigens by Mass Spectrometry-based Immunopeptidomics and Deep Learning

Abstract

SummaryHere we propose a personalized machine learning approach to predict the collective response of a patient’s CD8+ T cells by modeling the positive and negative selection processes, i.e. the central tolerance of T cells. In particular, for each individual patient, we collected his/her HLA-I self peptides derived from mass spectrometry-based immunopeptidomics as negative selection, and allele-matched immunogenic T cell epitopes from the Immune Epitope Database as positive selection. The negative and positive peptides were used to train a binary classification model, which was then applied to predict the immunogenicity of candidate neoantigens of that patient. Evaluation results on three cancer patients and one mouse cancer cell line and showed that our personalized models achieved an average accuracy of 79% and outperformed existing immunogenicity prediction tools. Furthermore, the models were able to rank neoantigens that elicited CD8+ T-cell responses within the top 15% of candidate peptides identified from the patients, thus reducing the time and costs of complicated in vitro validations.

A transformer model forde novosequencing of data-independent acquisition mass spectrometry data

“…We anticipate a number of future directions for work to further improve Cascadia on DIA data. First, Cascadia can be specifically fine-tuned for specific applications of interest, for example by training on a dataset of MHC bound peptides to improve immunoproteomics analysis [28] or non-tryptic data to ameliorate the tryptic bias of the model [29]. Furthermore, there are many successful ideas proposed by de novo methods in the DDA setting which may also be beneficial in the DIA setting, including the addition of additional auxiliary training tasks [4,9], alternate decoding strategies [6,15], and post-processing algorithms to refine predictions [19,30].…”

Deep learning methods for de novo peptide sequencing

“…The DeepNovo method, along with its successor method, PointNovo, has been incorporated into the commercial software PEAKS, and most of these 16 applications make use of that software. Among all 27 studies, the most common application is detection of neoantigens and noncanonical antigens [77,79,86,87,89,91,100,103], followed by antibody sequencing [84,88,99 Peptide assignments for the MS/MS spectrum from the nine-species benchmark data with universal spectrum identifier [104] mzspec:PXD003868:BY 04 heat 3:scan:28038. This spectrum is annotated as peptide "VFAIAN[+0.984]AFAK" in the original version of this dataset [1] but was reannotated as "VFAIANAFAK" recently [65].…”

“…Evaluating the tools on human and mouse antibody data, the authors concluded that Casanovo and PointNovo show improved peptide recall across different enzymes and datasets compared with competing methods. Second, Tran et al [76] have evaluated PEAKS, PointNovo, Casanovo, and GraphNovo on five datasets: human tryptic data, Method Year Application Ind DeepNovo 2020 Detection of neoantigens [77] 2021 De novo sequencing in metaproteomics [78] ✓ 2021 Detection of neoantigens [79] ✓ 2021 Detection of junction peptides [80] ✓ 2021 Detection of shell proteins [81] ✓ 2023 Detection of neuropeptides [82] ✓ 2023 Detection of proteasome-generated spliced and non-spliced peptides [83] ✓ 2023 Antibody sequencing [84] ✓ 2023 Detection of short peptides [85] ✓ 2023 Noncanonical antigen detection [86] ✓ 2023 Detection of neoantigens [87] 2024 Antibody sequencing [88] ✓ 2024 Detection of neoantigens [89] 2024 Detection of venom proteins [ ✓ 2022 Detection of neoantigens [103] ✓ Table 3: Applications of deep learning de novo sequencing methods. The final column ("Ind") indicates whether the application was published independently of the original authors of the work.…”