The detection of circulating tumor cells (CTCs) through liquid biopsy offers a non-invasive approach for accurately monitoring cancer dissemination and evaluating therapeutic efficiency. However, their rarity and heterogeneity limit conventional tumor antigen labelling-based methods in identifying and tracing CTCs. Here, we developed a novel metric, termed chromatin unwinding state (CUS), which leverages activated transcriptional regions related to cell-identity processes from single-cell transcriptomic data while overcoming technical variances. Using CUS features, we trained attention-based neural network models, panCTC, to in situ identify and lineage trace rare single CTCs directly from 5 mL of peripheral blood mononuclear cells scRNA-seq without enrichment. We benchmarked panCTC on various in silico-simulated, public, and in-house sequenced data, demonstrating its robustness across sample types and platforms. PanCTC could provide real-time scRNA-seq profiles of fresh CTCs, supporting early cancer detection and targeted anti-metastatic therapy. Circulating tumor cells (CTCs) are key metastatic precursors, yet their extreme rarity and molecular heterogeneity—including downregulation of epithelial markers (e.g., EpCAM⁻/CK⁻) 1, acquisition of hybrid phenotypes (e.g., EpCAM⁺CD68⁺) 2, or taking up platelets 3—severely challenge their detection. Current marker-dependent platforms (e.g., CellSearch® or CTC-iChip) 4, 5 often miss these aggressive CTC subsets, compromising their clinical translation. Current single-cell analysis tools, including general-purpose annotation tools (e.g., scATOMIC 6 and CopyKAT 7) and large-scale pretrained foundation models (e.g., scGPT 8 and scFoundation 9), are not designed for CTC identification. Although dedicated tools such as CTCTracer 10 and iCTC 11 have been developed, they are trained on limited CTC datasets and exhibit poor predictive power. The marked phenotypic plasticity of CTCs further complicates the reliable determination of their cellular origin. Therefore, the core challenge in accurately identifying and tracing CTCs from scRNA-seq data lies in the ability to extract both pan-cancer conserved features and disease-specific heterogeneous signatures from tumor transcriptomic data. To address these limitations, we introduced the concept of “chromatin unwinding state” (CUS) derived from scRNA-seq data. In a given cell, certain chromatin regions adopt an unwound architecture, creating a permissive nanoenvironment that facilitates transcriptional activity by modulating molecular accessibility and reaction energetics 12. These chromatin regions span sub-megabase domains within chromosomal territories 13. CUS captures transcriptional activation patterns linked to cell identity and biological state, which we hypothesized are conserved from primary tumors to CTCs and can serve as robust, marker-independent signatures. Here, we present panCTC, a deep-learning algorithm that extracts pan-cancer conserved CUS features to identify and trace CTCs directly from scRNA-seq data of peripheral blood mononuclear cells (PBMCs), without prior enrichment or labeling. PanCTC detects classical and non-classical CTCs at single-cell resolution, offering a powerful new approach to advance liquid biopsy and precision oncology. We propose the CUS as a cell-intrinsic metric for quantifying local transcriptional activity and present the panCTC framework. This framework uses CUS features derived from primary tumors to identify CTCs and trace their tissue of origin. We hypothesize that (1) CUS profiles reflect lineage-inherited transcriptional activity states, and (2) CTCs retain CUS features of their corresponding primary tumors (Figure 1A). Accordingly, the panCTC framework first converts the single-cell unique molecular identifier (UMI) count matrix into a CUS matrix. The CUS value is computed by statistically comparing the expression level of a genomic window (e.g., spanning 10 consecutive genes) against a cell-specific expression baseline, without relying on population-level references (Figure 1B, Text S1, Figure S1A). Subsequently, an attention-based network performs a two-tier classification: first distinguishing CTCs from other cells, then classifying CTCs into one of 12 candidate primary tumor types (Text S1, Figure S1B,C). For model training and validation, we integrated over 660,000 cells from 42 datasets, including primary and metastatic tumor cells from 12 cancer types, PBMCs from healthy donors and cancer patients, and previously published CTC scRNA-seq data (Figure 1C, Tables S1–S5). To validate the biological relevance of CUS, we first assessed transcriptomic profiles via CUS and measured its correlation with chromatin accessibility (ATAC-seq peaks) (Figure S2A,B). Across six major cell types, CUS showed a substantially stronger association with promoter/enhancer ATAC peaks than gene expression alone (Figure 1D). In chromatin regions where CUS and ATAC peaks were highly correlated in kidney cancer cells, we identified pro-tumor transcription factors such as BACH1 and ZNF12, which are not readily discernible from expression data alone (Figure S2C). As such, CUS serves as a transcriptomic proxy for chromatin accessibility. CUS also effectively captured cell-type-specific signatures. Canonical gene markers were enriched within corresponding CUS regions (e.g., the immune-cell-associated CUS-255 region contains CD45, and the epithelial-cancer-associated CUS-373 region contains EPCAM) (Figure S3A). Gene sets derived from immune- and tumor-specific CUS were enriched for cell type-specific functional pathways, respectively (Figure S3B, Tables S6–S7). Moreover, CUS preserved cell identity while minimizing technical noise across scRNA-seq platforms (10X Genomics, Smart-seq, DNBelab), outperforming raw UMI counts in terms of data consistency across platforms (p < 0.001, Kruskal–Wallis test) (Figure S4). These findings demonstrate that high-CUS regions are associated with multi-gene chromosomal contexts correlated with chromatin states, regulatory factors, and cellular functions, providing a stable new metric for cell identification. PanCTC comprises two AI models trained on CUS matrices from scRNA-seq data of primary tumor tissues. Both models were subsequently validated using computationally generated pseudo-PBMC datasets containing unseen cancer cells, including previously confirmed CTCs by established microfluidic isolation methods, primary and metastatic tumor cells, all with ground-truth annotations (Figure S5A). PanCTC's Model 1, for CTCs identification, achieved exceptional performance (sensitivity: 0.962 ± 0.079; specificity: 1.0 ± 0) across 1136 unseen datasets (Figure 1E, Figure S5B–D). It surpassed six state-of-the-art methods (scATOMIC 6, CopyKAT 7, scGPT 8, scFoundation 9, CTCTracer 10, and iCTC 11) in accuracy (Figure 1F, Figure S5F–H) and maintained the lowest false-positive rate of 0.001 ± 0.001 in 21 healthy-donor PBMC samples (Figure S5E). PanCTC's Model 2, for twelve-class cancer classification, far exceeded the random baseline (0.083) on 260 independent datasets, achieving near-perfect accuracy on primary cancer cells (1 ± 0.002), high accuracy on metastatic cells (0.923 ± 0.067), and moderate but highly variable accuracy on experimentally isolated CTCs (0.529 ± 0.335) (Figure 1G, Figure S6A–F). It demonstrated significantly superior prediction accuracy compared to existing methods (scATOMIC 6, scGPT 8, scFoundation 9, and CTCTracer 10) (Figure 1H, Figure S6G–I). Finally, on unseen scRNA-seq data of real PBMCs from eight cancer types, panCTC was benchmarked against high-confidence CTCs. Given the distinct strengths of various existing deep learning methods for CTC identification and the lack of an established gold standard, we defined a set of high-confidence CTCs derived from the intersection of at least three technical approaches. PanCTC exhibited the most stringent CTC detection rate (0.63%, panCTC-identified CTCs in all PBMCs) and the highest overlap with high-confidence CTCs (59.5%, high-confidence CTCs in panCTC-identified CTCs) (Figure 2A). It achieved patient-level single-cell detection rates 0.102%–1.449% (CTCs in PBMCs of one patient's blood sample) versus <0.5% (for current methods) and attained patient-level sensitivity 100% versus 83.3%–97% (Figure 2B). These results establish panCTC as a novel platform leveraging conserved CUS features of primary tumor cells for CTCs detection and lineage tracing from millions of PBMCs. We further characterized CTCs identified by panCTC using copy number variations (CNVs) and somatic mutational analyses between CTCs and patient-matched primary tumors. In pancreatic ductal adenocarcinoma (PDAC), CTCs exhibited common CNV patterns shared with specific tumor subclones (Figure S7). These shared CNVs involved genes related to the cytoskeleton, adhesion, and epithelial-mesenchymal transition (e.g., S100A8, VIM, and COL1A1). In cervical cancer (CC), however, CNV analysis revealed only partial conservation at certain chromosomes between CTCs and primary tumors, likely due to evolutionary divergence between CTCs and primary tumor cells (Figure S8). Subsequent analysis of somatic mutations in colorectal and breast cancer patients confirmed shared mutations in cancer-related genes (including CDKN2C, LMNA, MUC1, PDE4DIP, HRAS, B2M, COL1A1, and PIK3CA) between CTCs and patient-matched primary tumor cells (Figure S9). Together, these results indicate that CNV and mutation analyses support the malignancy nature of at least a fraction of CTCs identified by panCTC. PanCTC enables comprehensive single-cell characterization of CTCs across 10 cancer types. Transcriptomic analysis revealed cancer-type-specific patterns: PDAC and breast cancer (BC) CTCs highly expressed epithelial (EPCAM, KRTs) and stemness (SOX4, SOX9) markers, while CTCs from cervical, lung, and gastric cancers displayed elevated proliferation (DUSP1, BTG2), migration (S100A8, MMP9, VEGFA), platelet interaction (TUBB1, PPBP), and oncogenesis (JUN, FOS) (Figure 2C, Figure S10A). EMT status varied, with PDAC/BC CTCs showing hybrid epithelial-mesenchymal phenotypes and others being predominantly mesenchymal (Figure S10B–C). CTCs were categorized into EPCAM+ and EPCAM– subpopulations with distinct marker expression and functional profiles. Most EPCAM+ CTCs were predominantly from PDAC and BC, whereas EPCAM– CTC-like cells predominated in liver, lung, and other cancers, highly expressing non-typical CTC markers (C1QA, PPBP) 1-3 (Figure 2D–E, Figure S11). Ligand-receptor analysis further elucidated potential mechanisms mediating CTC intravasation, survival, extravasation, and colonizing metastases. For instance, ITGB-ICAM is involved in migration and CTC cluster formation 14, MHC-I and LGALS9-CD44 may mediate immune evasion 15, CLEC-KLRB1 may protect CTCs from NK cell-mediated lysis 16, 17, CD99-CD99 and MIF-CD74 can promote extravasation 18 (Figure 2F, Figure S12). Together, these findings demonstrate that panCTC can uncover the molecular heterogeneity and microenvironmental interactions of CTCs within the blood compartment. We further evaluated the clinical relevance of CTCs identified by panCTC. Analysis of PBMC scRNA-seq data from NSCLC, BC, and GC patients with tumor stage revealed CTCs molecular features associated with disease progression. Metastatic patients exhibited significantly more CTCs (n = 525, detection rate 3.6%) than non-metastatic patients (n = 111, detection rate 1.1%) (Figure S13A, Table S8). CTCs were further subclassified into EPCAM+ and EPCAM– subsets (Figure S13B). EPCAM+ CTCs were mainly detected in metastatic BC patients, while EPCAM– CTC-like cells were present across all patients (Figure S13C). Metastasis-derived EPCAM– CTC-like cells showed upregulation of adhesion, migration, and fusion markers (S100A8, S100P, MMP9) and enrichment in migration and potassium channel pathways (Figure S13D–E). Cell–cell communication analysis revealed EPCAM–-specific signals involving innate immunity (GZMA, RETN), platelet adhesion (PF4), and epithelial proliferation (GRN) (Figure S13F). In an ovarian cancer (OVC) cohort 19, all panCTC-detected CTCs were EPCAM– CTC-like cells and shared expression of markers (e.g., VIM and S100A4) with metastatic cells from matched primary tumors, lymph nodes, and ascites (Figure 2G). High expression of these markers in bulk RNA datasets correlated with significantly reduced overall survival (Figure 2H, Figure S13G), supporting their potential as non-invasive prognostic biomarkers. Therefore, panCTC enables molecular characterization of individual CTCs and may provide non-invasive biomarkers with broad clinical implications. The in situ detection of CTCs within PBMC scRNA-seq profiles offers considerable potential for cancer screening and metastasis research. In this study, we developed panCTC, an AI method leveraging CUS, genomic regions of high transcriptional activity inferred from scRNA-seq, to accurately identify and trace CTCs. PanCTC outperformed existing methods (scATOMIC 6, scGPT 8, CTCTracer 10, etc.) in challenging scenarios, such as samples with extremely low CTC counts or EPCAM– cases, or early-stage disease. Its superior reliability stems from a design specialized for CTCs, utilizing robust CUS features and training expanded with millions of primary tumor cell data to overcome CTC-sample limitations (Table S9). Compared to current clinical liquid biopsy approaches (e.g., CellSearch 4 and ctDNA assays 20, Table S10), panCTC requires no physical enrichment and achieves higher single CTC detection sensitivity. It can trace CTC tissue origin using only peripheral blood, complementing imaging-based diagnostics. Furthermore, panCTC enables functional characterization of CTCs, revealing insights into mechanisms like cluster formation, immune evasion 15, and extravasation 18 under native conditions. Limitations include restricted cancer-type coverage due to scRNA-seq data availability, scarcity of annotated CTC datasets, and computational simplifications in CUS estimation. Expanding datasets, in vitro and animal models, and refining algorithms will enhance future performance. In conclusion, panCTC provides a deep learning framework for the in situ identification and lineage tracing of CTCs directly from blood, independent of enrichment. It detects rare single CTCs and traces their origin using conserved CUS patterns, offering a powerful, minimally invasive platform for advancing metastasis research and liquid biopsy applications. Bin Ye: Conceptualization; methodology; data curation; software; investigation; writing—original draft; writing—review and editing; visualization. Zhen Wang: Methodology; software; data curation; investigation; writing—review and editing; visualization. Yinqi Bai: Methodology; software; data curation; investigation; writing—review and editing. Xu Zhang: Methodology; software; data curation; investigation; writing—review and editing. Rui Zhang: Methodology; software; data curation; investigation; writing—review and editing. Jingru Lian: Data curation; investigation; methodology; writing—review and editing; resources. Xuefei Liu: Methodology; data curation; investigation; writing—review and editing; resources. Yan Zhang: Investigation; formal analysis; data curation; writing—review and editing; resources. Zhiyuan Xu: Investigation; data curation; formal analysis; writing—review and editing; resources; funding acquisition. Li Yang: Investigation; formal analysis; data curation; writing—review and editing; resources. Haiman Jin: Investigation; writing—review and editing; formal analysis; data curation; resources. Fang Chen: Investigation; writing—review and editing; formal analysis; data curation; resources. Zhihao Xie: Formal analysis; investigation; data curation; resources; writing—review and editing. Ping Zhou: Investigation; writing—review and editing; formal analysis; data curation; resources; funding acquisition. Jun Tan: Investigation; writing—review and editing; formal analysis; data curation; resources. Shan Zeng: Investigation; writing—review and editing; formal analysis; data curation; resources. Changzheng Du: Investigation; writing—review and editing; formal analysis; data curation; resources. Yang Min: Investigation; formal analysis; data curation; writing—review and editing; resources. Xiaomin Ni: Data curation; resources; formal analysis; writing—review and editing; investigation. Jingxian Duan: Investigation; writing—review and editing; formal analysis; resources. Zhicheng Li: Investigation; writing—review and editing; formal analysis; resources. Hui Yang: Investigation; writing—review and editing; formal analysis; resources. Yunpeng Cai: Investigation; writing—review and editing; formal analysis; resources. Hongyan Wu: Investigation; writing—review and editing; formal analysis; resources. Catherine C. Liu: Investigation; writing—review and editing; formal analysis; resources. Jing Cai: Investigation; writing—review and editing; formal analysis; resources. Yi Lu: Funding acquisition; investigation; writing—review and editing; resources; formal analysis. Jian Zhang: Investigation; funding acquisition; writing—review and editing; formal analysis; resources. Hao Wu: Investigation; writing—review and editing; formal analysis; resources. Hairong Zheng: Resources; writing—review and editing. Longqi Liu: Writing—review and editing; resources. Xin Hong: Methodology; data curation; writing—review and editing; funding acquisition; investigation; resources; project administration; validation; supervision. Hao Yu: Project administration; resources; supervision; data curation; validation; writing—review and editing; writing—original draft; conceptualization; methodology; software; investigation; funding acquisition; visualization. All authors have read the final manuscript and approved it for publication. We wish to acknowledge all authors who deposited their data with open availability to the public databases. We thank Wenjun Ding (College of Life Sciences, University of Chinese Academy of Sciences, China), Xinyuan Guan and Wei Dai (Department of Clinical Oncology, University of Hong Kong, Hong Kong, China), Li Fu (Medical Research Center, Shenzhen University, China) for technical assistance. We also thank the Biological cloud computing platform at the Institute of Advanced Computing and Digital Engineering, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences for model computing. This work was supported by grants from the Science and Technology Project of Shenzhen (JCYJ20220530154407017, JCYJ20240813155824032), Guangdong provincial funding and awards (2025A1515011714), Science and Technology special fund of Hainan Province (ZDYF2024SHFZ045), Guangdong Province Overseas Postdoctoral Talent Support Project (T2250610233), and National Natural Science Foundation of China (82173336, 82573274). We apologize for not being able to cite additional work owing to space limitations. The authors declare no conflicts of interest. All participants provided written informed consent, and the research has been approved by the Medical Ethics Committee of the University of Hong Kong-Shenzhen Hospital (No. 2019049-001) and the First Affiliated Hospital of Hainan Medical University (No. 2021126). The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions. Due to ethical and legal restrictions, individual-level data of patients cannot be made publicly available. Data are available from the corresponding author (email protected) upon reasonable request and subject to local rules and regulations. This includes submitting a proposal to the management team, where upon approval, analysis needs to be done on a local server with protected access, complying with GDPR regulations. Code of panCTC is available at https://github.com/SiatBioInf/panCTC. Supplementary materials (methods, text, Chinese and may be in the or Science Figure of the panCTC framework for CTCs detection from scRNA-seq data of PBMCs. Figure between chromatin accessibility and CUS across cell types. Figure CUS cell-type-specific transcriptional signatures. Figure CUS while signatures. Figure of panCTC Model for CTC identification. Figure of panCTC Model for tracing the tissue origin of CTCs. Figure of PDAC primary cancer cells and CNV between CTCs and primary Figure of sequenced colorectal cancer samples and CNV between CTCs and primary Figure mutations shared between sequenced CTCs and matched primary cancer Figure Biological of panCTC-identified CTCs are with established CTC Figure profiles and enrichment of EPCAM+ EPCAM– CTC-like Figure Cell–cell interactions between CTCs and immune cells in PBMCs. Figure Clinical relevance of in CTCs. Table RNA data of primary tumor and in this Table data of 5 types of immune cells from healthy PBMC and in this Table RNA data of CTCs and for in this Table data of metastasis tumor and in this Table and public data of cancer PBMCs and healthy and their in this Table chromatin unwinding of immune and Table chromatin unwinding of cancer types. Table stage of public PBMC samples to in this Table of panCTC versus existing deep learning for CTC detection. Table of panCTC versus with current clinical approaches. The is not for the or of supporting by the than be to the corresponding author for the
Ye et al. (Mon,) studied this question.