Predicting Antigen-Specific T Cell Immunity Against Wilms Tumor 1 in Hematologic Cancer

Wilms tumor 1 (WT1) is a tumor-associated antigen expressed in both solid tumors and hematological cancers. In myeloid malignancies, such as acute myeloid leukemia (AML), chronic myeloid leukemia (CML), and myelodysplastic syndrome (MDS), WT1 is overexpressed and relates to worse prognosis. Clinical trials utilizing modified T cell receptors (TCRs) targeting WT1 are already ongoing. Endogenous T cell responses against WT1 have not been analyzed at scale, due to lack of suitable tools. Therefore, we aimed to develop a new computational tool that could detect WT1-specific T cell responses from TCR sequencing data.

To generate a prediction tool and queryable database of CD8+ TCRs we first enriched WT1-specific TCRs by using peptide pulsing of peripheral blood mononuclear cell samples from 6 AML patients and 6 healthy with HLA-A*02:01 restricted WT1 segments RMFPNAPYL (RMF; WT1_126–134) or VLDFAPPGA (VLD; WT1_37-45). Initial peripheral blood peptide stimulation was followed by a secondary peptide-loaded CD14+ monocyte stimulation, both accompanied with cytokine supplementation. After a 2-week pulsing protocol, WT1 specific T cells were sorted utilizing RMF and VLD specific multimers after which samples were sequenced with single cell RNA+TCR⍺β sequencing (scRNA+TCRseq). Overall, we generated 853 RMF-specific and 3378 VLD-specific AML and healthy subject-derived TCR⍺β sequences. ScRNA+TCRseq analysis showed that WT1 specific T cells were predominantly central memory and effector memory phenotype in both healthy donors and AML patients, regardless of which epitope was recognized.

This data was then utilized for the TCRGP machine learning method for predicting TCR recognition from unseen samples. To select TCRs for training the TCRGP models for VLD and RMF, we removed 33 TCRs that appeared in both VLD and RMF pulsing experimental datasets. In addition, we utilized recently published epitope-specific TCRs, 1261 for VLD and 101 for RMF, and applied GLIPH2 to the combined set of VLD- or RMF-specific TCRs to form representative repertoires of TCR that had motifs enriched in these repertoires. For training the models against each epitope, we used epitope-specific and control TCRs in a 1:10 ratio to factor in the larger number and variety of non-specific TCRs. The control TCRs included TCRs from the VDJdb database that are specific to other epitopes or previously published non-specific TCR repertoire data.We obtained the area under receiving operator characteristic curve (AUROC) values of 0.74 and 0.75 for VLD and RMF epitopes, respectively, allowing relatively reliable identification of TCRs recognizing these epitopes. In addition to WT1, we also trained TCRGP to multiple other epitopes, such as common viral epitopes from COVID-19 (SARS-CoV-2),cytomegalovirus (CMV), Epstein-Barr virus (EBV), and epitopes related to melanoma. These models were created similarly using TCRs with confidence score ≥ 1 from VDJdb.

To analyze the antigen specific TCRs in primary patient samples, we have sequenced samples from 21 AML, 26 CML and 25 MDS patients with bulk TCRβ sequencing, and 12 AML patients and 6 CML patients with scRNA+TCRseq. In addition, we utilized a healthy cohort of 40 donors. Our initial findings suggest that on average the frequency of RMF specific T cells is higher within AML and CML patients than in healthy subjects, mean frequency 0.89% in healthy, 1.81% in AML (p=0.02), and 1.10% in CML (p=0.006). Some patients had significantly higher frequency of these TCRs (up to 14%). With scRNA+TCRseq data of AML and CML patients we were able to study the phenotypes of the RMF specific T cells and examine how their frequency relates to the WT1 expression of their leukemic cells. With the 4 AML patients with HLA-A*02, on average the phenotypes of RMF-specific CD8+ T cells were 46% effector, 45% effector memory, 7% progenitor exhausted, and 2% naïve. The 2 patients with most RMF-specific T cells (12%) had the highest and lowest expression of WT1 in their myeloid bone marrow cells.

In conclusion, the presented peptide pulsing protocol enabled us to create WT1-specific TCRs and train a TCRGP computational tool to recognize antigen specific T cells from primary patient samples. The tool could be utilized to identify patients possibly benefiting from novel immunotherapies targeting WT1 and for monitoring leukemia antigen specific immune responses during treatments and allogenic hematopoietic stem cell transplantation.