Bogaert, C. | Search bioRxiv

bioRxiv2022

Improving T-cell mediated immunogenic epitope identification via machine learning: the neoIM model

The identification of immunogenic peptides that will elicit a CD8+ T cell-specific immune response is a critical step for various immunotherapeutic strategies such as cancer vaccines. Significant research effort has been directed towards predicting whether a peptide is presented on class I major histocompatibility complex (MHC I) molecules. However, only a small fraction of the peptides predicted to bind to MHC I turn out to be immunogenic. Prediction of immunogenicity, i.e. the likelihood for CD8+ T cells to recognize and react to a peptide presented on MHC I, is of high interest to reduce validation costs, de-risk clinical studies and increase therapeutic efficacy especially in a personalized setting where in vitro immunogenicity pre-screening is not possible. To address this, we present neoIM, a random forest classifier specifically trained to classify short peptides as immunogenic or non-immunogenic. This first-in-class algorithm was trained using a positive dataset of more than 8000 non-self immunogenic peptide sequences, and a negative dataset consisting of MHC I-presented peptides with one or two mismatches to the human proteome for a closer resemblance to a background of mutated but non-immunogenic peptides. Peptide features were constructed by performing principal component analysis on amino acid physicochemical properties and stringing together the values of the ten main principal components for each amino acid in the peptide, combined with a set of peptide-wide properties. The neoIM algorithm outperforms the currently publicly available methods and is able to predict peptide immunogenicity with high accuracy (AUC=0.88). neoIM is MHC-allele agnostic, and in vitro validation through ELISPOT experiments on 33 cancer-derived neoantigens have confirmed its predictive power, showing that 71% of all immunogenic peptides are contained within the top 30% of neoIM predictions and all immunogenic peptides were included when selecting the top 55% of peptides with the highest neoIM score. Finally, neoIM results can help to better predict the response to checkpoint inhibition therapy, especially in low TMB tumors, by focusing on the number of immunogenic variants in a tumor. Overall, neoIM enables significantly improved identification of immunogenic peptides allowing the development of more potent vaccines and providing new insights into the characteristics of immunogenic peptides.

bioinformatics↗

bioRxiv2022

neoMS: Attention-based Prediction of MHC-I Epitope Presentation

Personalised immunotherapy aims to (re-)activate the immune system of a given patient against its tumour. It relies extensively on the ability of tumour-derived neoantigens to trigger a T-cell immune reaction able to recognise and kill the tumour cells expressing them. Since only peptides presented on the cell surface can be immunogenic, the prediction of neoantigen presentation is a crucial step of any discovery pipeline. Limiting neoantigen presentation to MHC binding fails to take into account all other steps of the presentation machinery and therefore to assess the true potential clinical benefit of a given epitope. Indeed, research has uncovered that merely 5% of predicted tumour-derived MHC-bound peptides is actually presented on the cell surface, demonstrating that affinity-based approaches fall short from isolating truly actionable neoantigens. Here, we present neoMS, a MHC-I presentation prediction algorithm leveraging mass spectrometry-derived MHC ligandomic data to better isolate presented antigens from potentially very large sets. The neoMS model is a transformer-based, peptide-sequence-to-HLA-sequence neural network algorithm, trained on 386,647 epitopes detected in the ligandomes of 92 HLA-monoallelic datasets and 66 patient-derived HLA-multiallelic datasets. It leverages attention mechanisms in which the most relevant parts of both putative epitope and HLA alleles are isolated. This results in a positive predictive value of 0.61 at a recall of 40% on its patient-derived test dataset, considerably outperforming current alternatives. Predictions made by neoMS correlate with peptide identification confidence in mass spectrometry experiments and reliably identify binding motif preferences of individual HLA alleles thereby further consolidating the biological relevance of the model. Additionally, neoMS displays extrapolation capabilities, showing good predictive power for presentation by HLA alleles not present in its training dataset. Finally, it was found that neoMS results can help refine predictions of response to immune checkpoint inhibitor treatment in certain cancer indications. Taken together, these results establish neoMS as a considerable step forward in high-specificity isolation of clinically actionable antigens for immunotherapies.

bioinformatics↗