Background T cell-based immunotherapy is unquestionably a promising strategy against cancer. Regardless, the efficacy of T-cell therapies depends on mitigating off-target effects during the treatment. Off-target toxicity is associated with T-cell cross-reactivity (CR), i.e., therapeutic T-cell receptors (TCR) recognizing undesirable targets, leading to an auto-immune response. In this context, predicting CR risk remains an unsolved challenge. Here we present Crossdome, a tool that performs peptide screening and predicts cross-reactivity risk based on multi-omics data.

Methods Crossdome was developed by leveraging data from peptide-loaded Human Leukocyte Antigen (pHLA) complexes and TCR molecules, including biochemical properties, immunopeptidomics data, and three-dimensional fingerprints/interactions. Together these data were used i) to predict peptides that are biochemically similar to desirable targets, i.e., putative off-target candidates; and ii) to highlight TCRpHLA interactions associated with shared hotspots among the target and other peptides-HLA. Furthermore, we integrate functional data to evaluate immunogenic potential and expression associated with the putative candidates. As proof-of-principle, we benchmark our approach using well-known cross-reactivity cases, such as MAGEA3, NY-ESO-1, AFP, and TMEM161A.

Results Biochemical properties (BP) can be applied to calculate relatedness between peptides. The validated CR cases were found among the best-scored hits into thousand putative candidates (p-value ≤ 0.01). Particularly, on the MAGEA3 screening, the known cross-reactive peptide ESDPIVAQY (TITIN) was listed as a best-scored candidate, position 20th out of ~25.000 eluted peptides. Other cross-reactivity peptides were also enriched among the highly similar candidates, including MAGEA3 paralogs and synthetic peptides from yeast-displayed experiments (p-value ≤ 0.05). To strengthen our predictions, we implemented a penalty system based on the MAGEA3-specific TCR fingerprint. The hotspot positions were uncovered by molecular dynamics. Positions 1, 4, 5, 7, and 8 were used to recalculate the relatedness scores among the target and candidates. The combination between BP and fingerprints improved the CR predictions to 83%. Finally, we focus on functional characterization among the cross-reactivity candidates associated with the MAGE3A peptide. CR candidates have shown a large variability related to immunogenic prediction and expression pattern, spanning from dangerous to harmless profiles. On the dangerous profile, TITIN and LCAT have shown a high immunogenic level and expression-biased to heart and lung organs, respectively. Curiously, YTDPVGVLY (LCAT) peptide was not reported in previous studies.

Conclusions Crossdome is an interactive R package to predict cross-reactivity risk among peptide databases. Additionally, our method can integrate TCR repertoire data to improve the predictions. Currently, we are focusing our efforts on experimental validation and new antigens discovery.