Document resource
Background Determining the antigen specificities of T-cell responses remains a central challenge in immuno-oncology. While single-cell atlases have advanced understanding of the tumor immune microenvironment, the antigenic targets of tumor-infiltrating T cells—particularly viral, neoantigen, and tumor-associated antigens—are largely unresolved. To address this, we developed two resources: the Cancer Immunotherapy T Cell Atlas (CITA), the first pan-cancer single-cell atlas linking transcriptomes, TCR sequences, and antigen specificity (TAS) with standardized metadata, and the T Cell Antigen Specificity Database (TAS-db), a harmonized TCR-antigen specificity data model, aggregated from public databases and studies.Methods We generated RNA-TCR embeddings in CITA using mvTCR 1 trained on single-cell RNA/TCR data. TCR sequences were standardized with IMGT2 and mapped to the TAS-db3 4 using PRISM, a novel protein language model and probabilistic framework. Antigen specificity dynamics were inferred using Slingshot.9 HLA alleles were inferred from VJCDR3α/β using HLAGuesser.10 Results We developed the CITA, a pan cancer atlas of 1,238,142 T cells from 21 studies, including 755,261 with TCRs across 302 donors, 30 cancer types, 19 tissues, and 4 sequencing platforms ( figure 1A). Our annotation framework identified 14 CD4+ and 12 CD8+ T-cell phenotypes (figure 1B).11–23 Multimodal analysis revealed lesion-specific differences in T-cell phenotype frequencies and TCR clonality. CD4+ cytotoxic and late-effector CD8+ T cells were enriched in tumors, while resident memory CD8+ T cells localized to tumor and normal adjacent tissue. Naive CD4+ T cells were broadly distributed, including peripheral blood. CD8+ exhibited greater clonality than CD4+ T cells, with maximal expansion in tumors (figure 1E). TCR diversity varied by tumor types (figure 1F), likely reflecting differential antigenic landscapes. Joint embeddings uncovered associations between inferred HLA alleles and TCR features, suggesting TCR data may inform HLA inference (figure 1G).We also developed the TAS-db, a harmonized resource of 164,680 TCR-antigen pairs and constructed a protein language embedding for antigen specificity annotation (figure 2A-E) While exact CDR3αβ matching identified few antigen-specific T cells in CITA, our model yielded many high-confidence predictions (>0.9 likelihood), including neo- and tumor-associated antigen (TAA) targets for CD4+ and CD8+ T cells (figure 2F-M). Viral-specific CD8+ T cells exhibited memory-like states, while neoantigen- and TAA-specific cells displayed distinct activation and exhaustion phenotypes (figure 2N-O), revealing shared antigen-specific T-cell phenotypes in cancer.Conclusions With CITA, TAS-db and PRISM, we established a scalable framework for mapping T-cell specificity and phenotypes across cancers, enabling high-resolution analysis of antigen-driven transcriptional programs to accelerate discovery for next-generation precision immunotherapies.Acknowledgements We thank SITC Sparkathon for supporting this work. R.Z. is supported by the Parker Institute for Cancer Immunotherapy Bridge Fellows Award. R.Z. acknowledges funding from the NCI SPORE (P50-CA192937) and the Leukemia & Lymphoma Society and receives grant support from Bristol Myers Squibb and AstraZeneca. D.V. is supported by the University of California Cancer Research Coordinating Committee. V.D.J. was supported, in part, by the University of California Hellman Fellowship.References Drost F, An Y, Bonafonte-Pardàs I, et al. Multi-modal generative modeling for joint analysis of single-cell T cell receptor and gene expression data. Nat Commun. 2024;15:5577.Lefranc MP, Giudicelli V, Ginestoux C, et al. IMGT®, the international ImMunoGeneTics information system®. Nucleic Acids Res. 2009;37:D1006-D1012.Shugay M, Bagaev DV, Zvyagin IV, et al. VDJdb: a curated database of T-cell receptor sequences with known antigen specificity. Nucleic Acids Res. 2018;46:D419-D427.Gowthaman R, Pierce BG. TCR3d: the T cell receptor structural repertoire database. Bioinformatics. 2019;35:5323–5325.Vita R, Mahajan S, Overton JA, et al. The immune epitope database (IEDB): 2018 update. Nucleic Acids Res. 2019;47:D339-D343.Koşaloğlu-Yalçın Z, Blazeska N, Vita R, et al. The cancer epitope database and analysis resource (CEDAR). Nucleic Acids Res. 2023;51:D845-D852.Tickotsky N, Sagiv T, Prilusky J, et al. McPAS-TCR: a manually curated catalogue of pathology-associated T cell receptor sequences. Bioinformatics. 2017;33:2924–2929.Lee Y, Freitag R, Ganesan R, et al. PATCRdb: database of TCRs from data mining patent documents. medRxiv. 2023;2023–01.Street K, Risso D, Fletcher RB, et al. Slingshot: cell lineage and pseudotime inference for single-cell transcriptomics. BMC Genomics. 2018;19:477.Ruiz Ortega M, Pogorelyy MV, Minervina AA, et al. Learning predictive signatures of HLA type from T-cell repertoires. PLoS Comput Biol. 2025;21:e1012724.Michel M, Demel C, Zacher B, et al. TT-seq captures enhancer landscapes immediately after T-cell stimulation. Mol Syst Biol. 2017;13:920.Matsuzawa Y, Oshima S, Takahara M, et al. TNFAIP3 promotes survival of CD4 T cells by restricting MTOR and promoting autophagy. Autophagy. 2015;11:1052–1062.Kortekaas KE, Santegoets SJ, Sturm G, et al. CD39 identifies the CD4+ tumor-specific T-cell population in human cancer. Cancer Immunol Res. 2020;8:1311–1321.Clarke J, Panwar B, Madrigal A, et al. Single-cell transcriptomic analysis of tissue-resident memory T cells in human lung cancer. J Exp Med. 2019;216:2128–2149.Salmon AJ, Shavkunov AS, Miao Q, et al. BHLHE40 regulates the T-cell effector function required for tumor microenvironment remodeling and immune checkpoint therapy efficacy. Cancer Immunol Res. 2022;10:597–611.Gingras S, Pelletier S, Boyd K, et al. Characterization of a family of novel cysteine-serine-rich nuclear proteins (CSRNP). PLoS One. 2007;2:e808.Sener Z, Cederkvist FH, Volchenkov R, et al. T helper cell activation and expansion is sensitive to glutaminase inhibition under both hypoxic and normoxic conditions. PLoS One. 2016;11:e0160291.Bourque J, Kousnetsov R, Hawiger D. Roles of hopx in the differentiation and functions of immune cells. Eur J Cell Biol. 2022;101:151242.Crawford A, Angelosanto JM, Kao C, et al. Molecular and transcriptional basis of CD4+ T cell dysfunction during chronic infection. Immunity. 2014;40:289–302.Zheng C, Fass JN, Shih YP, et al. Transcriptomic profiles of neoantigen-reactive T cells in human gastrointestinal cancers. Cancer Cell. 2022;40:410–423.Opejin A, Surnov A, Misulovin Z, et al. A two-step process of effector programming governs CD4+ T cell fate determination induced by antigenic activation in the steady state. Cell Rep. 2020;33(8):108356.Toribio-Fernández R, Zorita V, Rocha-Perugini V, et al. Lamin A/C augments Th1 differentiation and response against vaccinia virus and Leishmania major. Cell Death Dis. 2018;9(1):9.Michel M, Demel C, Zacher B, et al. TT-seq captures enhancer landscapes immediately after T-cell stimulation. Mol Syst Biol. 2017;13(3):920.Abstract 1112 Figure 1Cancer Immunotherapy T Cell Atlas (CITA) data layers and TAS (T Cell Antigen Specificity) database. (A) Schematic of the Cancer Immunotherapy T Cell Atlas (CITA) data layers and TAS (T Cell Antigen Specificity) database collections. (B) We identified diverse CD8+ and CD4+ T cell phenotypes reflecting known differentiation states, with tumor-enriched terminal effectors and exhausted cells, and blood-enriched naïve and early memory cells. CD4+ subsets include N/CM, Tregs, and MEM, with further specialization in Tfh-like, Th1-like, and stress-associated subsets. (C) Cellular fractions of CD4+ and CD8+ T cell subsets are plotted for each of six lesion types, with individual dots representing donor samples. Accompanying pie charts depict the overall distribution of lesion origins in the cohort: N (normal), LN (lymph node), NAT (normal adjacent tumor), MET (metastasis), PB (peripheral blood), and PE (pleural effusion). The plots summarize how each T cell subset’s abundance varies across these lesion groups. (D) Summary statistics of TCR repertoire metrics in CITA. Clonality was computed based on Shannon entropy, with clonotypes defined by the V and J genes in the CDR3α and CDR3β regions. CD4+ T cells displayed a greater number of unique clonotypes compared to CD8+ T cells. Boxplot and stripplot show clonality distribution along the VJCDR3αβ axis for CD4+ and CD8+ T cells, highlighting a broader distribution and higher clonality in the CD8+ population (Mann-Whitney U test, p < 2.085e-11). (E) Clonality across six lesion types (N, LN, PB, MET, NAT, and T), showing significant variability. Tumor (T) lesions exhibit significantly higher TCR clonality than other tissue types (p-adjusted < 0.05). (F) Shannon diversity across cancer types, with marker size indicating clonotype richness (i.e., number of unique clonotypes). (G) CD4+, CD8+ UMAP of joint RNA and TCR multimodal embedding. UMAP projection of joint RNA and TCR multimodal embedding from the mvTCR model showing inferred MHC class II alleles for CD4+ T cells and MHC class I alleles for CD8+ T cells, predicted using HLA Guesser. Only associations with p < 0.05 are displayedAbstract 1112 Figure 2Antigen and TCR features of clonotypes in the T Cell Antigen Specificity. (A) Pie charts show the distribution of antigen categories, (B) epitopes, and (C) antigen protein symbols associated with the top expanded clonotypes, defined by VJCDR3β in TAS-db. The majority of clonotypes are specific to viral (VIR) antigens, followed by those recognizing oncogenic viral (ONCV) and bacterial (BAC) antigens. Tumor-associated antigens (TAA) and neoantigens (NEO) account for 2.8% and 1.3% of clonotypes, respectively. Antigen categories include VIR (viral), ONCV (oncoviral), BAC (bacterial), NEO (necantigen), UN (unclassified), PAT (patent), AUTO (autoimmune), TAA (tumor-associated antigen), ALLG (allergen), CUN (cancer-unclassified), PAR (parasite), SYN (synthetic peptide). FUNG (fungal), and ALLO (allogeneic). (D) 1-SNE projection of the top 10,000 TCRB clonotypes in TAS based on ESM protein language model embeddings, colored by CDR3β-TRBV gene usage. (E) Bar plot showing the distribution of unique TCRβ clonotypes and peptides across antigen protein categories in the TAS database. (F), (H) CITA RNA-based UMAPs of CD8 and CD4+ T cells, colored by P(antigen | TCR_CITA)-the computed likelihood that a CITA TCR is assigned to a specific antigen protein based on sequence similarity, using TAS as a reference through a protein language model. Only cells with P(antigen | TCR CITA) > 0.9 are shown. (G), (1) UMAPs of CD8 and CD4 cells (P(antigen | TCR_CITA) > 0.9), colored by predicted antigen category. (J), (L) CD8’ and CD4 CITA TCRS mapped to TAS antigens via exact matches of VJCDR3β sequences. (K), (M) UMAPs of CD8+ and CD4+ cells (P(antigen | TCR_CITA) > 0.9), colored by predicted antigen protein. (N) Antigen-specific cluster dynamics inference map of CITA CD8+ T cells. UMAP represents CD8+ T cells with VIR or TAA/NEO specificity, with TCRs annotated by TAS reference mapping and trajectories inferred using the Slingshot method, colored by manual RNA-based annotation. For VIR-specific clusters, four lineages were identified: L1: Pre-mem → Antiviral → Cytotoxic; L2: Pre-mem → Antiviral → EM; L3:Pre-mem → Late Effector → RM/EM; L4: Pre-mem → Late Effector → STR. (O) For TAA/NEO-specific clusters, the following lineages were observed: L1: Pre-mem → Late Effector → EM; L2: Pre-mem → Antiviral → ISG; L3: Pre-mem → Late Effector → RM; L4: Pre-mem → Antiviral → Cytotoxic