Beyond sequence similarity: ML-powered identification of pHLA off-targets for TCR-mimic antibodies using high throughput binding kinetics

Abstract

T-cell receptor mimic (TCRm) antibodies are an emerging class of tumor-targeting agents used in advanced immunotherapies such as bispecific T-cell engagers and CAR-T cells. Unlike conventional antibodies, TCRms are designed to recognize peptide - human leukocyte antigen (pHLA) complexes that present intracellular tumor-derived peptides on the cell surface. Due to the typically low surface abundance and high sequence similarity of pHLAs, TCRms require high affinity and exceptional specificity to avoid off-target toxicity. Conventional methods for off-target identification such as sequence similarity searches, motif-based screening, and structural modeling focus on the peptide and are limited in detecting cross-reactive peptides with little or no sequence homology to the target. To address this gap, we developed EpiPredict, a TCRm-specific machine learning framework trained on high-throughput kinetic off-target screening data. EpiPredict learns an antibody-specific mapping from peptide sequence to binding strength, enabling prediction of interactions with unmeasured pHLA sequences, including sequence-dissimilar peptides. We applied EpiPredict to two distinct TCRms targeting the cancer-testis antigen MAGE-A4. The model successfully predicted multiple off-targets with minimal sequence similarity to the intended epitope, many of which were experimentally validated via T2 cell binding assays. These findings establish EpiPredict as a valuable tool for lead optimization of TCRms, enabling the identification of antibody-specific off-targets beyond the scope of traditional peptide-centric methods and supporting the preclinical de-risking of TCRm-based therapies.

Keywords: BLI; Binding kinetics; HLA; MAGE-A4; MHC; SCORE; T2-binding; TCR-like antibody; X-scan; machine learning; off-target toxicity; pHLA; pMHC; peptide-MHC complex; preclinical de-risking.

Figure 1.

EpiPredict core workflow. (A) A dataset of 530 peptides (339 EpiTox-predicted off-targets and 190 X-scan variants) were selected. (B) Peptides were formatted as pHLA-A *02:01 microarrays and profiled using SCORE for each antibody. (C) Binding data was used to training machine learning models for each antibody. (D) 316,855 peptides from the human proteome likely to be presented by HLA-A *02:01 (NetMHCpan rank $\le$2) were fed to the models. (E) High-probability hits prioritized for confirmatory in in-vitro assays.

Figure 2.

Kinetic profiling of two TCRms, antibody A (A, C, E) and antibody B (B, D, E). (A, B) Bio-layer interferometry (BLI) binding kinetics of TCRms with target MAGE-A4. Association and dissociation phases are shown, with 1:1 global fit curve overlaid (solid black lines). (C, D) Heatmap of $K_D$ ratio values from positional peptide library scanning (X-scan) of the MAGE-A4 wild-type peptide (GVYDGREHTV). Higher log ${ }_2$($K_D$ WT/$K_D$ mutant) ratios, indicative of stronger binding, are shown in warmer colors; lower ratios are shown in cooler colors. White tiles denote wild-type residues, grey tiles indicate positions where $K_D$ values could not be determined consistent with abolished binding, and black tiles represent positions where peptide loading could not be confirmed by ${\beta }_2$ M staining. (E) log ${ }_2$($K_D$ WT/$K_D$ mutant) ratios of peptides with more than one mutation that showed binding to either antibody A (blue) or antibody B (green). In this case, no bar is equal to no binding observed.

Figure 3.

EpiPredict machine learning model training and predictions. (A) Model architecture and training scheme. (B) & (C) Predicted antibody binding probabilities across sequence dissimilarity levels. Scatter plot of individual peptide-antibody pairs showing EpiPredict-predicted binding probabilities (y-axis) as a function of sequence dissimilarity to the MAGE-A4 decapeptide, quantified by Hamming distance (x-axis; 1–10 substitutions). Each point represents a unique peptide differing by the indicated number of amino acids. Peptides with predicted binding probability $>$0.5 were selected for experimental validation. The distribution demonstrates that predicted binders occur even among peptides with high sequence dissimilarity.

Figure 4.

Validation of predicted off-target peptide binding in SCORE microarrays and the T2 cell assay. (A) log2 binding ratios of SCORE and T2 binding experiments for antibody A. (B) log2 binding ratios of SCORE and T2 binding experiments for antibody B. No bars is equal to no binding detected. The more negative the values are, the weaker the binding. Positive values would have meant a stronger binding against the off-target peptide compared to the target. (C) $\mathit{E}{C}_{50}$ dose – response titration of antibody a and antibody B on T2 cells loaded with the target peptide. (D) $\mathit{E}{C}_{50}$ dose – response titration of antibody a and antibody B on T2 cells loaded with peptide PP5.

Figure 5.

Structural basis of antibody-specific binding to predicted off-target peptides. (A) TCRm-MAGE-A4 peptide interfaces for antibody A as modelled by chai-1 (F2). (B) TCRm-MAGE-A4 peptide interfaces for antibody B as modelled by chai-1 (F2). Portions of the antibody and HLA-A *02:01 complex are omitted for clarity. Hydrogen bonds are shown as red clouds, ionic contacts as purple clouds. (C) Structural alignment of the MAGE-A4 (green) with PP5 (red) within the HLA-A *02:01 binding groove. (D) Structural alignment of the MAGE-A4 (green) with PP9 peptides (red) within the HLA-A *02:01 binding groove. Antibody and HLA complex are omitted for clarity. (E) Antibody A – PP5 interfaces. (F) Antibody B – PP5 interfaces. Hydrogen bonds are depicted as red clouds and ionic contacts as purple clouds. (G) Structural alignment of MAGE-A4 (green) with PP25 (red) in the HLA-A *02:01 groove. (H) Structural alignment of MAGE-A4 (green) with PP14 (red) in the HLA-A *02:01 groove. Antibody and HLA complex are omitted for clarity.