Acute myeloid leukemia (AML) is a life-threatening hematologic malignancy characterized by the accumulation of highly proliferative and poorly differentiated myeloid cells in the blood and the bone marrow (1). It commonly affects older people with a median age at diagnosis of 68 years and a five year overall survival rate lower than 30% (1). Allogeneic hematopoietic stem cell transplant (allo-HSCT) remains the most effective therapeutic strategy for high-risk AML patients likely due to the graft versus leukemia (GvL) effect, in which donor T cells recognize and target leukemic clones (2, 3). In addition to driving disease development and progression, mutations in the leukemic cells (1, 4) present ideal leukemia specific neoantigens when presented by the major histocompatibility complexes (MHCs) and then recognized by the T cell receptors (TCR) triggering the GvL effect (1, 5). A unique TCR signature is generated by a complex rearrangement process of an array of variable (V), diversity (D), and joining (J) exons to select one of each that will recombine into a functional receptor with complementary determining regions 3 (CDR3) responsible for the interaction between the TCR and the peptide-MHC (6–8). The generated TCR repertoire undergoes dynamic changes with the onset and progression of disease and the course of treatment (9). Specific features of the TCR repertoire such as the TCR diversity and clonal expansion were found to correlate with a patients' clinical characteristics, disease status and clinical outcome (10–15).

In AML, an increase in TCR expanded clones following PD-1 blockade therapy was observed in responders’ patients (11). Furthermore, TCR diversity was lower in patients with graft versus host disease (GVHD) and in relapsed patients among those who received allo-HSCT (14). Diversity of the TCR repertoires undergoes significant skewing post-allo-HSCT with an increased clonal expansion compared with healthy subjects (16). Considering the clinical and biological relevance of the TCR repertoire, computational methods to extract the TCR sequences from widely available public RNA-seq data have increasingly become available (17, 18). Among these computational tools, TRUST was recently developed and applied to analyze public TCR/BCR repertoire from The Cancer Genome Atlas (TCGA) cancer data and in pediatric and adult AML (10, 19). Infants with AML had significantly higher TCR CDR3s per kilo TCR reads (CPK) compared with children or adult patients (10). The study also reported an association between CBFB-MYH11 mutation and lower TCRB CPK value in pediatric patients with AML (10).

Here we applied two computational methods (the updated version of TRUST (TRUST4) and MIXCR on the LAML-TCGA RNA-seq data to profile TCR repertoires across a cohort of 151 patients with AML to investigate the TCR repertoire in peripheral blood (PB) samples in relationship with patients’ clinical and molecular characteristics to better understand the role of TCR repertoire and its impact on disease progression and outcome. Our study serves as an independent unbiased approach to apply two commonly used computational tools to analyze TCR clones from RNA-seq dataset in AML. This research provides valuable insights into the identified TCR clones from RNA-seq data and calls for the need of refinement and optimization of computational methodologies in this evolving field.

Characteristics of the TCR repertoire such as diversity indices or clonal expansion and/or composition reflect the antigen specificity and T cell dynamic changes. Highly expanded clones within tumors may contribute significantly to the antitumor response as previously observed in different types of cancer (29). Several tools that leverage the RNA-seq data to extract TCR repertoire data have been developed (22, 23). RNA-seq data are capable of profiling dominant clonotypes with high frequencies but will unlikely capture the diversity of the repertoire or identify the less abundant clones. Li et al. developed TRUST (19), which then was employed by the group on both pediatric and adult AML samples to characterize T and B cell receptors (TCR and BCR) repertoires (10). The study showed lower number of normalized TCR CDR3 counts in infants (0-3 years old) and children (3-20 years old) when compared with healthy children, and in adults (over 20 years old) with AML when compared with healthy adults. Although the study has explored the association between TCR repertoire characteristics and major common AML mutations and clinical outcome, a comprehensive analysis of how TCR repertoire features correlate with both clinical and molecular characteristics of patients with AML remains to be fully elucidated. Here we applied an updated version of TRUST (TRUST4), previously found to have significantly improved computational efficiency and the number of recovered CDR3s (30) along with another recently developed tool (MIXCR) to compare their output and comprehensively examine the TCR repertoires in the adult AML cohort and how their features are associated with patients’ molecular and clinical characteristics based on available RNA-seq database.

To ensure robustness of the analysis we applied two of the most commonly used computational tools to extract TCR information from RNA-seq data – TRUST4 and MIXCR. These tools were previously benchmarked with other computational tools to extract TCR information from RNA-seq data originating from the same samples focusing on the characterization of T-cell rich and T cell poor tissues. Both methods were found effective in capturing estimated repertoire diversity in T-cell rich tissues particularly in samples with a limited number of hyperexpanded clones (31). However, the exact accuracy of the alignment from TRUST4 and MIXCR remains unknown due to the lack of a control framework of known TCR composition both quantitatively and qualitatively for direct comparison and assessment. This highlights the need to establish rigorous controls that can allow for identification of potential disparities in alignment accuracy to allow for instilling confidence in the reliability of T-cell receptor analysis output. The average numbers of TCRA and TCRB reads from RNA-seq were higher in TRUST4 output compared with MIXCR output. Both tools have resulted in satisfactory coverage of TRAV and TRBV genes capturing 65-75.5% of the V and 76-86% of the J known genes. While TRUST4 has a similar number of TCRA and TCRB reads on average, MIXCR has a higher number of reads for TCRA than TCRB and less coverage across TRBV segment usage. Mapped reads obtained from TRUST4 and MIXCR were used to estimate the frequency of different TRAV and TRBV genes across a cohort of patients with AML. Among the top 10 V genes identified in a previous study on 29 cancer types from the TCGA data (19), TRUST4 identified six TRAV and six TRBV and MIXCR identified five of TRAV and seven of TRBV genes (19) in the AML cohort. Small overlap between clonotypes identified by TRUST4 and MIXCR suggest differences in the performance of the two computational tools and highlights that study design needs to be carefully considered when applying these tools on RNA-seq data. The alignment algorithm is significantly different between TRUST4 and MIXCR and detailed explanation of each alignment method is reported in detail by their developers (17, 22, 23). Differences in V gene assignment algorithms and CDR3 amino acid sequence start points can lead to discrepancies in identified CDR3 sequences. Additionally, variations in alignment methods and contig generation can impact the number of aligned reads per patient. To ensure confidence in the associations between TCR features and molecular/clinical characteristics, we utilized the two computational tools. While both tools showed similar patterns, only approximately 30% of the identified CDR3 sequences were shared between them. This discrepancy makes interpreting data containing specific sequences and quantitative information of certain clones challenging, as the tools assemble different clonotypes.

We observed a slightly higher normalized number of unique clones in female patients compared with males, an observation that was seen in prior studies (19, 32). We also found that clinically relevant characteristics such as WBC, BM blast percentage and PB blast percentage negatively correlate with TCRA and TCRB unique clone counts and therefore patients with higher than median white blood cell count, percentage of bone marrow blasts and percentage of PB blasts have been found to have lower normalized number of unique clones. One possible explanation for this association is that leukemic blasts may outcompete normal T cells, leading to a reduction in the number of unique T cell clones in the sample. Similarly, a higher WBC count may indicate a more aggressive disease with a higher burden of leukemic blasts, which in turn may result in a lower number of T cells and therefore fewer unique T cell clones. Skewed TCR repertoire may be related to a response to myeloblast microenvironment, however it was previously suggested that while T cells may recognize cancer cells, effective antitumor response is impaired.

Previous studies found a strong positive association between the normalized number of unique clones and the number of nonsynonymous mutations in 29 cancer types from TCGA including cancers with high mutational burden (19). This is possible due to higher number of neoantigens being presented for the T cells (19). We did not observe such association when the number of the most common AML mutations and unique TCRA or TCRB clones were considered. Previous study using TCGA dataset investigated five genes of high clinical significance (FLT3, NPM1, KIT, CEBPA and WT1), along with three gene fusions (RUNX1-RUNX1T1, CBFB-MYH11, and PML-RARA). Only CBFB-MYH11 was found to be associated with significantly lower TCRB CPK value in pediatric patients with AML, similar but not significant trend was found in infant and adult patients with AML (10). We found that certain mutations, such as FLT3 and NPM1 were associated with a decreased normalized number of unique T cell clones while TP53 was associated with increased normalized number of unique TCR clones. However, these associations were lost when considering the WBC, BM blasts and PB blasts. This is likely due to the fact that these mutations are known to be associated with these clinical features (33–38). Mutations in WT1 however were associated with the TCR repertoire when adjusting by other clinical characteristics. It was previously shown that WT1-specific T cells can be generated to target leukemia cells with limited effect on normal progenitor cells (39). Similarly, WT1-specific CD8 T cells were found to be induced in azacitadine and donor lymphocyte infusion treatment contributing to graft versus leukemia effect in MDS and AML patients (40). The presence of WT1-specific CD8 T cells was previously suggested to contribute to maintenance of complete remission (41). Vaccination with peptides derived from WT-1 induce immune responses resulting in an improved clinical outcome (42).

Specific features of the TCR repertoire such as the TCR diversity and clonal expansion were found to correlate with patients’ clinical characteristics, disease status and clinical outcome in various cancers (10, 43–45). Diversity of the TCR repertoire was found to be inversely correlated with age in patients with breast cancer (46). Lower diversity indices were observed in advanced disease stages (46) and upon anti-CTLA-4 therapy in melanoma patients (47). Higher usage of TRBV20.1 was reported in the HER2- patients during complete pathological remission while TRBV30 was higher in HER2+ patients responding to trastuzumab (46). Tumor-infiltrating T cell repertoire was more diverse compared with normal pancreatic tissue in pancreatic ductal adenocarcinoma suggesting recognition of tumor derived antigens (48). Higher TCR diversity was found to be correlated with better progression free interval in a few cancer types (49).

Our analysis found no statistically significant difference in overall survival between patients with a lower versus higher than median normalized number of unique clones for both TCRA and TCRB. However, we observed that patients that had CANTGELFF within the identified TCRB clones by TRUST4 tended to have better but not statistically significant overall survival compared with those without such CDR3. According to the TCRdb database the CANTGELFF was previously found in patients with breast cancer, urothelial cancer, healthy subjects and in patients with CMV infection among others (50). However, this finding was not corroborated by the results obtained from MIXCR. Additionally, MIXCR failed to detect public clonotypes shared by more than 3 patients. Consequently, we believe that while both MIXCR and TRUST4 are suitable for analyzing associations between TCR diversity estimates and molecular/clinical patient characteristics, they may not be the most appropriate tools when the analysis involves the identification and quantification of specific CDR3 sequences. To validate our findings, confirmation through TCR targeted sequencing in an independent AML dataset and functional studies are necessary. The observed discrepancy underscores the need for appropriate benchmarking of current algorithms against a known ground truth - a library of TCR receptors with known composition. Regrettably, such a control library is not yet available, hindering the use of RNA-seq data for both quantitative and qualitative analyses in this context.

While the focus of our study was on the association between TCR repertoire and clinical and molecular features in AML patients, the identification of viral epitopes in our analysis is an interesting finding that may have implications for the understanding of the role of viral epitopes in AML patients. Association of CMV infection with AML relapse after allogeneic hematopoietic stem cell transplant (HCT) has been previously identified along with an increase of multifunctional CMV-specific T cells (51, 52). Also, CMV reactivation is a frequent complication after HCT. Even though previous studies have not found association of EBV with the pathogenesis of AML, it has been suggested that it is a secondary event related to compromised immune systems in AML (53). No significant associations were found for HCV, however, it was found to be increased in 8% of patients with AML (54). It is important to note that our epitope analysis approach has some limitations, including the potential for false positives and the inability to distinguish between TCR clones that are specific to the viral epitope versus those that are cross-reactive with other antigens. For example, we identified SARS-CoV-2 related epitopes even though the patient samples are pre-pandemic, however, it is now known that cross-reactive epitopes were present in individuals before the SARS-CoV-2 pandemic (55). It was also previously reported that T cells could cross-react with both CMV and SARS-CoV-2 in SARS-CoV-2 unexposed individuals. Functional studies have shown non-naive spike and non-spike specific T cells in the blood of pre-pandemic individuals with several HLA backgrounds (56–58). Interpretation of antigen-specificity predictions indicating virus-related targets is based on clustering analysis to identify similar specific TCR sequence in a reference dataset. While this database lacks AML-specific TCRs, RNA-seq is only able to capture the most abundant clones due to its shallow nature indicating that these clones were expanded within these patients; therefore it would be suggested that virus-specific memory T cells were leveraged for anti-AML response through T-cell cross-reactivity. Virus specific T cells were previously shown to populate repertoire of human healthy tissues and tumors where they can be leveraged to trigger cytotoxicity in the tumor (59, 60). To confirm whether the virus-specific memory T cells identified in AML patients present an anti-AML response through T-cell cross reactivity versus just background dominance of those TCR clones, functional studies and deep TCR sequencing would be needed. Also, further validation studies would be needed to confirm the clinical relevance of the identified viral epitopes in AML patients.

One limitation of our study is the reliance on bulk RNA-seq data, which unfortunately lacks information on the pairing between TCRA and TCRB. Consequently, this hinders our ability to fully predict the functional receptor responsible for antigen binding. Although single-cell sequencing is the most informative and suitable approach for confidently pairing TCRA and TCRB, previous research has demonstrated that the beta chain of the CDR3 sequence holds the most significant influence on antigen specificity (61).

Public TCRs were previously described in infectious diseases, cancer and autoimmune disease and were found to associate with clinical outcome (62–67). Consistently with previous studies we found that public CDR3s have shorter amino acid length (68). It was previously demonstrated that shorter CDR3 length is related to smaller number of N-insertions including CDR3s specific to viral epitopes such as CMV and EBV (69). In our study we identified CANTGELFF CDR3 which resulted from the recombination of either TRBV12-3, TRBV2, TRBV28, TRBV30, TRBV3-1, TRBV5-5, TRBV5-7, TRBV6-6, TRBV7-2 or TRBV9 with TRBJ2-2 in patients with a better overall survival. However, this finding was not reproduced using MIXCR and can possibly be a false-positive. Further analysis recognizing public TCRs in patients with AML might be beneficial to recognize leukemia specific TCRs.

AML is characterized by increased prevalence of immunosuppressive phenotype leading to impaired immune response (70–74). Single cell RNA-seq analysis suggested altered T regulatory cells as feature of newly diagnosed AML patients (11). In agreement with the previous studies we observed strong positive correlation between foxp3 expression and the number of unique TCRA and TCRB. Studies on AML animal models suggest impaired T cell function with features of exhaustion or antigen-specific T cell tolerance (75–77). Anergy is then developed through lack of co-stimulatory signals with excessive stimulation of the TCR and high levels of inhibitory signals such as PD-1, CTLA4, TIGIT, LAG3 (78). Consistently we found that inhibitory signals such as TIGIT and CTLA-4 were found to be positively correlated with unique number of TCR clones. Interestingly PDCD1 was found not to be associated with TCR repertoire. Further high positive correlation of GZMB, but not PRF1 with the number of unique TCRA and TCRB was observed suggesting association between cytotoxic T cells and TCR repertoire. Observed correlations suggest more exhausted T cell profile in characterized patients with AML consistently with previous research (11). Although exhausted CD8 T cells can still identify antigens through TCR, they do not effectively respond to the antigen (79). Lack of correlation with perforin and correlations with co-inhibitory receptors but not activation markers further supports dysfunctional immune microenvironment within patients with AML.

Several limitations of this study are present. AML is a heterogeneous hematological malignancy; the relatively small sample size might not fully reflect all aspects of this disease. Secondly, RNA-seq data provides a shallow estimation of the TCR repertoire, as very limited number of clones and low number of reads corresponding to the most abundant clones, and the broader TCR repertoire cannot be captured. Low number of reads does not allow for calculation of features of TCR repertoire such as diversity indices or detection of rare clonotypes therefore we used normalized unique clone count as an approximate metric for TCR diversity. While we analyzed public clonotypes there is a chance we missed some individuals also sharing the same clonotypes shared in this population due to low coverage. Associations made between TCR repertoire features and clinical and molecular characteristics are mainly based on abundant clones which are more expanded in comparison to not detected rare clonotypes; therefore full representation of the repertoire is not accessible. If certain antigen specific clonotypes related to certain mutations are less abundant they might not be detected here which would affect the mutational data analysis. Furthermore 447 out of 2820 TCRA clones and 830 out of 1603 TCRB clones MIXCR were assigned to multiple TRAV or TRBV genes possibly due to the short read length. The small percentage of overlap between clones identified by TRUST4 and MIXCR further highlights the need for benchmarking studies that validate these computational methods. Furthermore, although we identified top 10 TRAV and TRBV exons, only single cell sequencing would be able to determine which clones are paired to form a functional TCR receptor consisting of alpha and beta chains.

Despite the limitations of low coverage associated with the use of RNA-seq data to extract TCR sequence information and the small sample size, we were able to report statistically significant associations between the TCR repertoire and clinical and molecular features in patients with AML. Our findings highlight the importance of analyzing the TCR repertoire in AML to better understand the adaptive immune response against leukemic cells and the impact of the disease on the immune repertoire and the later on the clinical outcome.

Our study revealed unexpectedly poor reproducibility of detected CDR3 clones between the two software packages. This finding is important because it highlights the need for caution when interpreting results obtained using different software tools such as those focusing on the specific expanded clones or the V segment utilizations. Nevertheless, for the type of correlative analyses we performed, the two packages performed very similarly and showed the same patterns of associations observed between the TCR repertoire and clinical and molecular features were consistent. This consistency supports the robustness of our findings and suggests that the associations we reported are likely to be biologically relevant.

While the utilization of RNA-seq data enables the identification of highly abundant clonotypes in patient samples, it is important to acknowledge the limitations associated with the low coverage of this approach. The study’s findings further highlight the need for TCR-Seq analyses to provide the depth needed to characterize the TCR repertoire in patients with AML accurately. Further, the poor reproducibility in terms of identified CDR3 clones between MIXCR and TRUST4 highlight the need of development of a control of known composition to serve as a biological calibrator for accurate qualitative and quantitative TCR analyses. Future studies using TCR-seq analyses could provide a more comprehensive understanding of the immune response to AML and the potential for TCR-based therapies to improve patient outcomes.

Publicly available datasets were analyzed in this study. This data can be found here: https://portal.gdc.cancer.gov/projects/TCGA-LAML.

Ethical approval was not required for the studies involving humans because this study was conducted on publicly available dataset. The studies were conducted in accordance with the local legislation and institutional requirements. Written informed consent to participate in this study was not required from the participants or the participants’ legal guardians/next of kin in accordance with the national legislation and the institutional requirements.

MP: data analysis, validation and visualization, writing-original draft, reviewing-editing, methodology. MT: data curation, software, formal analysis, validation, methodology, writing the original draft. Y-CW: formal analysis, writing-original draft, review, and editing. AK: data curation, software, formal analysis, validation, methodology, writing-review editing. ML and JM: methodology, writing-review editing, and provided statistical input. HA: conceptualization, resources, supervision, funding, validation, writing original draft, project administration, writing-review, and editing. All authors contributed to the article and approved the submitted version.