TCR-based cancer immunotherapy (e.g., TILs, engineered TCR-T cells, TCR mimic antibodies) is emerging as a promising alternative or a complementary approach to surgery, radiotherapy, and chemotherapy (Hunter 2017). However, before it is fully embraced as a form of cancer treatment, several shortcomings need to be addressed, including the risks of adverse therapy effects caused by off-target toxicity (García-Fernández et al. 2021). To help mitigate this problem, we have introduced ARDitox—a novel computational method that can identify off-target epitopes even when they differ significantly from the targeted epitope presented on a given pMHC.

So far, very few computational algorithms for predicting off-target binding in TCR-based cellular cancer immunotherapies have been proposed. Similarly to ARDitox, Expitope (Jaravine et al. 2017) and iVax (Moise et al. 2013) approaches start with querying the human proteome for peptides homologous to the target with a predefined number of mismatches allowed. However, iVax does not consider the presentation of recognized putative OTEs in the process of ranking them, while Expitope estimates peptide presentation on HLA by proxy using a combination of the proteasomal cleavage probability and transport associated with antigen processing, as well as HLA binding. ARDitox is therefore the first method for off-target epitope detection that leverages HLA presentation probabilities and an understanding of epitope residue orientation. Overall, ARDitox is a novel approach to identify putative OTEs with a unique pipeline that includes an in-house AI-trained pHLA presentation model, a unique scoring function focused on physico-chemical properties of the TCR-facing residues, and an extended search of peptides derived from frequent mutations.

We tested ARDitox on four clinically validated cases of TCRs targeting TAA-derived epitopes, one TCR in a preclinical stage and one virus epitope, as well as on two datasets: i) TAA vs Virus epitopes and ii) TAA vs frameshift-derived epitopes. ARDitox correctly identified the OTEs that caused the immunotoxicity in the treated patients (Table 2). Furthermore, in Case 3, where no OTEs were experimentally observed, ARDitox did identify one putative cross-reactive epitope with a safety score < 3. While this finding is interesting from a methodological standpoint, given the lack of observed off-target cross-reactivity in patients, this result should be considered a potential false positive result of our in silico screen. However, it is also critical to consider the expression profile of the predicted off-target protein. If this protein is expressed at negligible levels or in a clinically inaccessible location, this could explain the lack of observed patient cross-reactivity, even if the epitope is predicted to bind. This highlights the need for further validation of our findings using wet-lab methods and also points to areas where our technique may require further optimization to reduce the number of false positive outcomes. In case 5, we have successfully identified an OTE that might lead to autoimmunity as a result of molecular mimicry after EBV infection, showing ARDitox’s potential usefulness in the development of vaccines, while taking into account molecular mimicry. The fact that the observed clinical cross-reactivity did not meet our 3.0 cutoff is a valuable finding. It likely highlights a scenario where individual clonotypes within the polyclonal mixture may have had weak off-target scores, but their combined, cumulative effect in vivo led to a clinical outcome. Lastly, in case 6, we experimentally identified the ADH1A epitope. This off-target would not be detected in mouse models due to the lack of presentation of the orthologous Adh1-derived epitope (PLDPLITHV) (Fig. 3E). Importantly, ADH1A had a safety score close to 5, which corresponded to a weak binding of the TCR. Early identification of this OTE is valuable as additional safety measures can be considered to ensure that activation of the TCR-T cell against ADH1A and ADH1B will not occur during clinical trials.

In assessing the risk of a given TCR therapy, ARDitox serves as a tool that provides qualitative guidelines aimed at giving a comprehensive view of potential threats, rather than predicting clinical outcomes. To this end, we recommend a detailed analysis of three key variables: i) The total number of potential off-target epitopes (OTEs), which provides a general idea of the scale of the potential issue, ii) The number of OTEs with a safety score < 3, which helps identify the most probable, high-risk cases, iii) The expression levels of OTEs with a safety score < 3, which is crucial for assessing potential clinical consequences.

Furthermore, to ensure maximum safety, we highly recommend that all putative OTEs with a score < 5 be subjected to in vitro verification, as some weak TCR off-target interactions (e.g., cases 5 and 6), especially if highly expressed in crucial organs, might have negative consequences for the patient's health. It should be stressed that the importance of the variables mentioned above should not be neglected, as exemplified by TTN’s OTE in use case 2. In this case, the number of all putative OTEs was considered moderate. Based on this variable, the TCR against this target epitope seemed to be very promising; however, if the cross-reactivity of each OTE with a safety score < 3 would be checked experimentally because low safety scores indicate that the TCR may bind to both the target and the OTE similarly. Lastly, checking the expression status of each putative OTE with a safety score < 5 should indicate which tissue types are of particular interest for the experimental verification. This would have been important in case 2, as during the preclinical in vitro studies, no toxicity towards the tested heart muscle cell line was detected, as the TTN protein is expressed only in contracting cardiac myocytes (Coles et al. 2020). Identifying TTN as OTE upfront could have enforced the addition of appropriate cell lines to the test panel. One potential shortcoming of our model is that, currently, for some less frequent HLA-types, incorrect amino acids may be scored as the ones facing the TCR. However, this problem is minor as it occurs only for HLA types that have a generally low frequency. Furthermore, it can be mitigated as more data regarding TCR-faced amino acid positions for rare HLAs becomes available through various experimental and computational methods. These include experimental techniques such as X-ray crystallography, X-scan, Cryo-EM, and peptide elution data, as well as AI tools like Boltz-2 for (Passaro, et al. 2025) generating structural data. Another limitation is that a fraction of the OTEs identified by ARDitox are likely false positives, including some with safety scores below 3. This is an inherent trade-off in a tool designed to prioritize potentially dangerous off-targets for subsequent in vitro validation, rather than to replace experimental testing. Importantly, our case studies demonstrate that ARDitox tends to assign high priority rankings to known, biologically relevant off-targets, supporting its utility as an effective pre-screening tool. Additionally, we acknowledge that a formal assessment of the false negative rate is essential for a safety-focused pipeline. This remains an important direction for future work to better define model sensitivity. A key consideration of our homology-based approach is its inability to capture off-targets arising from TCR promiscuity, where recognition occurs despite little or no sequence similarity. As the basic version of ARDitox relies on homology searches, such interactions may be missed. However, by incorporating empirical TCR binding motif data (e.g., from alanine- or X-scans) or structural modeling, the ARDitox pipeline can be enhanced to make TCR-specific off-target predictions. This functionality positions ARDitox as a unique tool that complements existing methods by bridging the gap between initial target selection and final therapeutic TCR validation.

In order to assess the effectiveness of the proteome search for OTEs and the proposed scoring methodology, we compared the analysis performed on TAAs and viral epitopes. The higher number of OTEs detected by ARDitox from TAA-derived peptides compared to viral epitopes and frameshift peptides is consistent with the hypothesis that, due to their self-origin, TAAs are more likely to share homology with the human proteome. These results support our recommendation for interpreting ARDitox output. When assessing the risk of a target causing off-target toxicity, the primary focus should be on verifying the number of putative OTEs with a safety score below 3.

Lastly, we wanted to check whether frameshift mutations are promising targets for immunotherapeutic strategies, since they give rise to multiple, out-of-frame, random protein products that should not map to the reference proteome (Spaanderman et al. 2021). Furthermore, frameshift-derived epitopes usually do not share functional domains with other genes found in the human genome, and, as such, the general number of putative OTEs should be both lower and with higher safety scores. To verify this, we used a database composed of 16 frameshift-derived neoepitopes that were predicted as presented by our presentation model (Chen, et al. 2022; Vita et al. 2019). When compared to 16 TAA epitopes, the number of OTEs from frameshift neoepitopes was 10 × lower (frameshift OTEs: 336 vs TAA OTEs: 3911). Furthermore, none of the putative frameshift OTEs had a safety score < 3 (Fig. 4B). As such, the results from ARDitox suggest that, provided nonsense-mediated decay does not occur, frameshift neoepitopes may be promising alternatives to TAA epitopes in TCR therapies.

In conclusion, we have developed a method for the identification of off-target toxicity that can be successfully applied in the development of cellular immunotherapies. Our tool, ARDitox, takes into account: (i) peptide processing, (ii) pHLA binding, (iii) pHLA presentation probability, (iv) determination and similarity of TCR-faced amino acids, (v) frequent variants as a source of off-target epitopes, and (vi) mRNA and protein expression levels. Most importantly, the application of our platform, ARDitox, to process data from several use case studies allowed efficient identification of OTEs, which proves its applicability in the development of TCR-based cancer immunotherapies.