
==== Front
Brief Bioinform
Brief Bioinform
bib
Briefings in Bioinformatics
1467-5463
1477-4054
Oxford University Press

10.1093/bib/bbae028
bbae028
Review
AcademicSubjects/SCI01060
New techniques to identify the tissue of origin for cancer of unknown primary in the era of precision medicine: progress and challenges
https://orcid.org/0009-0002-1561-0031
Ma Wenyuan Department of Medical Oncology, The First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, China

Wu Hui Department of Medical Oncology, The First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, China

Chen Yiran Department of Surgical Oncology, The First Affiliated Hospital, College of Medicine, Zhejiang University, Hangzhou, Zhejiang, China

Xu Hongxia Zhejiang University-University of Edinburgh Institute (ZJU-UoE Institute), Zhejiang University School of Medicine, Zhejiang University, Haining, China

Jiang Junjie Department of Gastroenterology, Affiliated Hangzhou First People's Hospital, Zhejiang University School of Medicine, Hangzhou, China

Du Bang Real Doctor AI Research Centre, School of Medicine, Zhejiang University, Hangzhou 310058, China

Wan Mingyu Department of Medical Oncology, The First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, China

Ma Xiaolu Department of Medical Oncology, The First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, China

Chen Xiaoyu Department of Medical Oncology, The First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, China

Lin Lili Department of Nuclear Medicine, The First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, China

Su Xinhui Department of Nuclear Medicine, The First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, China

Bao Xuanwen Department of Medical Oncology, The First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, China

https://orcid.org/0000-0003-2720-724X
Shen Yifei Department of Laboratory Medicine, the First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, China

Xu Nong Department of Medical Oncology, The First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, China

Ruan Jian Department of Medical Oncology, The First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, China

Jiang Haiping Department of Medical Oncology, The First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, China

https://orcid.org/0000-0002-9670-9246
Ding Yongfeng Department of Medical Oncology, The First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, China

Corresponding authors. Yongfeng Ding, Department of Medical Oncology, the First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou 310006, China. Tel.: +86-0571-87235841. E-mail: dingyongfeng@zju.edu.cn; Haiping Jiang, Department of Medical Oncology, the First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou 310006, China. Tel.: +86-0571-87235841. E-mail: jianghaiping@zju.edu.cn
3 2024
11 2 2024
11 2 2024
25 2 bbae02816 7 2023
10 12 2023
11 1 2024
© The Author(s) 2024. Published by Oxford University Press.
2024
https://creativecommons.org/licenses/by-nc/4.0/ This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (https://creativecommons.org/licenses/by-nc/4.0/), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited. For commercial re-use, please contact journals.permissions@oup.com

Abstract

Despite a standardized diagnostic examination, cancer of unknown primary (CUP) is a rare metastatic malignancy with an unidentified tissue of origin (TOO). Patients diagnosed with CUP are typically treated with empiric chemotherapy, although their prognosis is worse than those with metastatic cancer of a known origin. TOO identification of CUP has been employed in precision medicine, and subsequent site-specific therapy is clinically helpful. For example, molecular profiling, including genomic profiling, gene expression profiling, epigenetics and proteins, has facilitated TOO identification. Moreover, machine learning has improved identification accuracy, and non-invasive methods, such as liquid biopsy and image omics, are gaining momentum. However, the heterogeneity in prediction accuracy, sample requirements and technical fundamentals among the various techniques is noteworthy. Accordingly, we systematically reviewed the development and limitations of novel TOO identification methods, compared their pros and cons and assessed their potential clinical usefulness. Our study may help patients shift from empirical to customized care and improve their prognoses.

cancer of unknown primary (CUP)
tissue of origin (TOO)
gene expression profiling
precision medicine
liquid biopsy
image omics
Natural Science Foundation of Zhejiang Province 10.13039/501100004731 LQ20H160043 Clinical Research Fund Project of Zhejiang Medical Association 2019ZYC-A83 National Natural Science Foundation of China 10.13039/501100001809 82303950
==== Body
pmcINTRODUCTION

Despite standardized diagnostic workups, cancers of unknown primary (CUPs) comprise a heterogeneous collection of metastatic tumors with unknown primary tumors [1]. Epidemiologically, CUPs are estimated to account for 2–5% of all diagnosed tumors worldwide [2–4]. Indeed, most patients (80–90%) with CUPs fall into an unfavorable subset, with median overall survival (OS) durations ranging from 3 to 11 months and a 1-year OS of 25–40% [5–7]. Clinically aggressive, early-spreading and unpredictable metastases define these tumors [2, 4, 8]. Due to the lack of standard treatment, most CUP patients receive empirical chemotherapy, including platinum–taxane regimens [1, 9].

Patients diagnosed with CUP have a worse prognosis than those with metastatic cancer of known origin [10, 11]. This suggests that tissue of origin (TOO) identification and subsequent site-specific therapy for CUP may enhance survival and prognosis. Data from various clinical trials suggest the possibility of this hypothesis [12–15]. Varadhachary et al. [14] were the pioneers in demonstrating, using a small sample of patients with CUP associated with a colon-cancer profile (CCP-CUP), and found that patients with CCP-CUP derive substantial benefits from using specific treatments developed for colon cancer. Similarly, Hainsworth et al.’s [15] assay-directed site-specific therapy yielded a median survival time of 12.5 months, surpassing outcomes associated with empiric CUP regimens. Furthermore, regarding whether patients with CUP benefit from site-specific therapy, Ding et al. [16] conducted a meta-analysis and concluded that identifying the TOO and administration of site-directed therapy is effective, specifically for CUP patients with responsive tumor types. Accordingly, identifying the TOO of CUP is critical for optimizing and pinpointing treatment. The meta-analysis by Ding et al. [16] also showed that improving the accuracy of TOO identification could significantly improve patient prognosis. From a psychological point of view, CUP patients have more psychological perplexity and stress than those with recognized TOO [17–21], which concerns both clinicians and patients. Therefore, it is of great clinical significance to develop TOO identification techniques for patients diagnosed with CUP.

A comprehensive diagnostic approach for CUP typically includes physical examination, medical history review, hematology assessment, endoscopy, imaging studies and pathological analysis, as depicted in Figure 1 [1, 2, 7]. However, these methods may not consistently identify the TOO in all suspected CUP patients. For example, conventional imaging techniques exhibit a TOO detection rate of only 20–27%, while positron emission tomography (PET) improves this rate slightly to 37% [22, 23]. Immunohistochemical analysis, although necessary, is a labor-intensive and often inefficient method for TOO identification in malignant tumors, with a notable 27% of cases remaining undetermined. The concordance rate between pathological and clinical diagnoses was only 59% [24, 25]. Notably, multiple immunostainings consume a limited amount of the tumor tissue. The limitations of traditional clinical methods underscore the need for innovative and effective TOO identification techniques.

Figure 1 Diagnostic methods recommended for the anticipatory diagnosis of CUP patients. The clinical evaluation of CUP begins with a thorough tumor history, family history and physical examination. This is followed by analysis, including a basic hematologic examination; CT and PET scans of the chest, abdomen and pelvis; and determination of tumor biomarkers. Endoscopy, like gastrointestinal endoscopy and laryngoscopy, can not only visualize the location of the tumor but also provide the tissue samples needed for pathological examination, and various immunohistochemical combinations can play a role in identifying the tumor category. If the location of the primary tumor cannot be determined, the diagnosis of CUP remains. αSMA, α-smooth muscle actin; AFP, α-fetoprotein; CA, cancer antigen; CD, cluster of differentiation; CK, cytokeratin; CLA, cutaneous lymphocyte-associated antigen; CT, computed tomography; CUP, cancer of unknown primary; hCG, human chorionic gonadotropin; HMB45, human melanoma black 45; myoD1, myoblast determination protein 1; PET, positron emission tomography; PSA, prostate-specific antigen; S100, calcium-binding protein G. Potential cancer type designation is determined by marker positivity unless otherwise noted (Figure 1 belongs to the Introduction section).

Identifying TOO in metastatic tumors is the cornerstone of clinical work in oncology. However, a notable challenge emerges in a small subset of cases, particularly within poorly differentiated carcinomas and squamous cell carcinomas, where the diagnostic process is significantly hampered by the absence of specific site-specific immunohistochemical markers [24]. New TOO detection techniques for CUP are being proposed and validated as technology develops (Figure 2). On the one hand, the prediction accuracy, sample requirements and technical principles of the different identification techniques show significant diversity. Contrarily, a comprehensive evaluation comparing the merits, demerits and future prospects of these methods is yet to be undertaken. Thus, we systematically assessed emerging TOO detection methods for CUP to determine their clinical utility.

Figure 2 New techniques developed at different molecular levels for the detection of TOO in CUP are emerging. CUP, cancer of unknown primary; RT-PCR, reverse transcription-polymerase chain reaction; TOO, tissue of origin (Figure 2 belongs to the New techniques of identifying TOO of CUP section).

MAIN TEXT

Methods of literature search and criteria for article selection

A systematic literature search was performed using PubMed, Web of Science, Embase, Cochrane Library and ClinicalTrials.gov from 1 January 2000 to 1 May 2023, with English language restrictions. Conference abstracts from the American Society of Clinical Oncology (ASCO) and European Society of Medical Oncology (ESMO) meetings were also included. Search terms were as follows: [(cancer* OR carcinom* OR neoplas* OR malignan*) AND (‘unknown primary’ OR ‘occult primary’ OR ‘primary metastatic’) AND (origin* OR type) AND (trace* OR infer* OR classif* OR identif* OR predict*) AND (accuracy OR sensitivity OR specificity)]. Only studies conducted on CUP patients were included. The primary lesion was identified using non-routine clinical diagnostic methods, and the research was limited only to human model. Case reports, editorials and commentaries were excluded (Supplementary Figure 1).

A total of 14 369 potentially eligible studies were initially identified from the systematic literature search, as shown in Supplementary Figure 1. After removing the duplicates from the different databases (n = 7249), irrelevant studies (n = 7018) were excluded by title and abstract screening. A total of 102 studies were assessed for eligibility. Eight articles were excluded because they did not introduce the technique to identify TOO, 16 articles were excluded because they did not focus on patients with CUP, 20 articles were excluded because accuracy data were not available and 15 articles were excluded because their sample size was less than 30 cases. Due to the limited number of comparative studies available, one comparative study conducted by Chen et al. was still included in our research, despite its small sample size (Supplementary Figure 1).

New techniques of identifying TOO of CUP

Based on genomic profiling

The TOO of CUP can be identified at the deoxyribonucleic acid (DNA) level. DNA copy number variations (CNVs) [26, 27], somatic and germline mutation [28–30], expression quantitative trait loci (eQTL) [31] and single-nucleotide polymorphisms (SNPs) [32, 33] have been used to identify TOO in tumor tissues (Table 1).

Table 1 Comparison of approaches used in the identification of the site of origin in CUP

		Year published	Authors	Sample size	Source	Type	Materials	NT	Accuracy	External	NE	CUP	AE	Prospective	Therapy	Time	Details of the method	
Genomic profiling	
	1	2013	Felix Dietlein et al. [32]	905	TCGA, CCLE	FF, FFPE	DNA	23	79.00%	Yes	431	No	71.00%	No	Yes	NA	Single-nucleotide polymorphism (SNP) + second generation sequencing	
	2	2015	Andrea Marion Marquard et al. [30]	2820	COSMIC	NA	DNA	28	87.60%	Yes	993	No	85.00%	No	No	NA	Somatic mutation, copy number profiles + random forest (RF)	
	3	2020	Bingsheng He et al. [33]	4909	ICGC	FFPE	DNA	13	88.22%	No				No	No	NA	SNP + RF	
	4	2020	Ying Liang et al. [26]	3480	TCGA	FFPE	DNA	6	83.31%	Yes	NA	No	79.00%	No	No	NA	Copy number alteration (CNA) + RF	
	5	2020	Xiaojun Liu et al. [29]	3374	ICGC	FFPE	DNA	13	81.00%	No				No	No	NA	Somatic mutation + RF	
	6	2020	Yulin Zhang et al. [27]	4566	TCGA	FFPE	DNA	10	89.13%	Yes	1262	No	74.21%	No	No	NA	Copy number variations (CNVs) + eXtreme Gradient Boosting (XGBoost)	
	7	2022	Yongchang Miao1 et al. [31]	Over 7000	TCGA	NA	DNA	20	96.00%	Yes	87	No	90.00%	No	No	NA	Expression quantitative trait loci (eQTL) + XGBoost	
	8	2022	Luan Nguyen et al. [28]	6756	PCAWG	FFPE	DNA	35	90.00%	Yes	141	No	58.00%	No	No	NA	Genome-wide mutation features + RF	
Gene expression profiling (GEP)	
	9	2005	Richard W. Tothill et al. [42]	229	TCGA	FFPE	RNA	14	89.00%	Yes	13	Yes	84.60%	No	No	NA	Quantitative PCR, microarray + support vector machine (SVM)	
	10	2007	Gauri R. Varadhachary et al. [37]	104	Self-collected cases	FFPE	RNA	13	61.00%	No				Yes	Yes	NA	10 genes Quantitative reverse-transcription PCR (qRT-PCR)	
	11	2008	Nitzan Rosenfeld et al. [68]	253	Self-collected cases	FF, FFPE	RNA	22	>90.00%	Yes	83	No	89.00%	No	No	NA	MicroRNA array	
	12	2009	Agendia BV et al. [43]	633	Self-collected cases	FF, FFPE	RNA	30	81.00%	Yes	229	No	82.00%	Yes	No	NA	Microarray	
	13	2010	Rosetta Genomics Ltd et al. [63]	356	Self-collected cases	FFPE	RNA	25	90.00%	Yes	204	No	85.00%	Yes	No	NA	MicroRNA array + qRT-PCR	
	14	2013	Rolf Søkilde et al. [65]	208	Self-collected cases	FFPE	RNA	15	85.00%	Yes	48	No	88.00%	Yes	No	NA	MicroRNA array + microarray profiling	
	15	2013	John D. Hainsworth et al. [15]	252	Self-collected cases	FFPE	RNA	NA	98.00%	No				Yes	Yes	2 to 3 weeks	CancerTYPE ID (92-gene qRT-PCR)	
	16	2015	Richard W. Tothill et al. [48]	450	Self-collected cases	FFPE	RNA	18	82.00%	Yes	58	No	78.00%	Yes	No	NA	Microarray	
	17	2017	Yuanyuan Li et al. [59]	9096	TCGA	FFPE	RNA	31	95.60%	No				No	No	NA	RNA sequencing	
	18	2017	Marcos Tadeu dos Santos et al. [41]	4429	Self-collected cases	FFPE	RNA	25	86.60%	Yes	102	Yes	83.80%	No	No	2 weeks	95-gene qRT-PCR	
	19	2019	Mengyao Li et al. [60]	1007	GEO	FFPE	RNA	1	99.36%	Yes	24	Yes	100.00%	Yes	Yes	NA	Relative gene expression orderings (REOs)	
	20	2020	Qing Ye et al. [39]	609	Self-collected cases	FFPE	RNA	21	89.80%	Yes	141	No	71.60%	Yes	No	NA	90-gene qRT-PCR	
	21	2020	Yue Zhao et al. [56]	18,217	TCGA, ICGC	FFPE	RNA	32	98.54%	Yes	486	Yes	96.70%	No	No	NA	RNA sequencing +1D-Inception	
	22	2021	Sijie Chen et al. [45]	5759	TCGA	NA	RNA	21	96.38%	Yes	42	No	83.30%	No	No	NA	Microarray + XGBoost	
	23	2021	Noemi Laprovitera et al. [64]	159	Self-collected cases	FFPE	RNA	17	95.00%	Yes	53	No	90.00%	Yes	No	2 days	MicroRNA array + prediction analysis of microarrays, the least absolute shrinkage and selection operator	
	24	2021	Julien Vibert et al. [54]	20,918	TCGA, GTEX, HPA	FF	RNA	39	96.26%	Yes	48	Yes	79.00%	Yes	Yes	NA	RNA sequencing + RF, K-nearest neighbors (KNN)	
	25	2021	Ruixi Li et al. [52]	7713	TCGA	FFPE	RNA	20	96.10%	Yes	79	No	83.50%	No	No	NA	RNA sequencing + gradient boosting (GBDT)	
	26	2021	Yifei Shen et al. [53]	10,553	TCGA	FFPE	RNA	24	93.00%	No				Yes	No	NA	RNA sequencing + rank-based majority vote algorithm	
	27	2022	James Hong et al. [55]	1528	TCGA	FFPE	RNA	14	97.00%	No				No	No	NA	RNA sequencing + deep learning	
	28	2022	Qingfeng Lu et al. [44]	5708	TCGA	FFPE	RNA	15	96.90%	No				No	No	NA	Microarray + XGBoost	
	29	2022	Wei Sun et al. [40]	1417	Self-collected cases	FFPE	RNA	21	94.40%	No				Yes	No	NA	90-gene qRT-PCR	
	30	2022	Weiqin Jiang et al. [58]	17	TCGA, self-collected cases	FFPE	RNA	4	94.40%	Yes	7	Yes	100.00%	No	No	NA	RNA sequencing + naive Bayes algorithm	
	31	2023	Jackson Michuda et al. [57]	52,936	TCGA	FFPE	RNA	68	91.10%	Yes	1708	No	80.00%	No	No	NA	RNA sequencing + multinomial logistic regression classifier with L2 regularization	
Epigenetics	
	32	2016	Sebastian Moran et al. [12]	7691	TCGA	FFPE	DNA	38	99.60%	Yes	216	No	87.00%	Yes	Yes	NA	Microarray DNA methylation signatures	
	33	2022	Agustin F. Fernandez et al. [72]	42	Self-collected cases	NA	DNA	19	69.00%	No				No	No	NA	DNA methylation fingerprint—1505 CpG sites	
	34	2023	Ze Zhang et al. [73]	7735	TCGA, GEO	FFPE	DNA	30	99.00%	Yes	1775	Yes	93.00%	No	No	NA	DNA methylation + multilayer perceptron	
Multi-omics	
	35	2016	Wei Tang et al. [96]	Over 5000	TCGA	FFPE	RNA	14	87.78%	No				No	No	NA	miRNA array expression	
							DNA		96.43%					No	No	NA	DNA methylation profiles + maximum relevance maximum distance	
							DNA		97.06%					No	No	NA	DNA methylation profiles + principal component analysis	
	36	2020	Binsheng He et al. [101]	7008	TCGA, GEO	FFPE	RNA	20	89.28%	Yes	19	No	NA	No	No	NA	Gene expression + multi-classifier RF	
							DNA		53.51%				NA	No	No	NA	Gene mutation + multi-classifier RF	
							RNA + DNA		95.77%				89.47%	No	No	NA	Gene expression + gene mutation + multi-classifier RF	
	37	2021	Haiyan Liu et al. [102]	7244	TCGA	FFPE	RNA	21	94.63%	No				No	No	NA	Gene expression	
							DNA		43.33%					No	No	NA	Somatic mutation	
							DNA		86.13%					No	No	NA	DNA methylation	
							RNA + DNA		94.02%					No	No	NA	DNA methylation + gene expression + somatic mutation	
	38	2022	Kaiyan Chen et al. [103]	16	Self-collected cases	FFPE	DNA	1	43.30%	No				No	No	NA	Genomic profiling	
							DNA		100.00%					No	No	NA	DNA methylation	
							DNA		100.00%					No	No	NA	Genomic profiling + DNA methylation	
Liquid biopsy	
	39	2015	Ronald Lebofsky et al. [76]	34	Self-collected cases	Blood samples	DNA	18	97.00%	No				Yes	No	NA	Cell-free tumor DNA	
	40	2015	Myron G. Best et al. [80]	283	Self-collected cases	Blood samples	RNA	6	71.00%	No				Yes	No	NA	RNA sequencing of tumor-educated platelets	
	41	2020	Ayuko Hoshino et al. [79]	34	Self-collected cases	FFPE	Protein	4	100.00%	Yes	9	No	100.00%	No	No	NA	Extracellular vesicle (EV) and particle Biomarkers	
				37	Self-collected cases	Blood samples	Protein	5	100.00%	Yes	10	No	100.00%	No	No	NA	EV and particle biomarkers	
Others	
	42	2022	Enrico Moiso et al. [94]	11,744	TCGA, MOCA	FFPE	RNA	33	97.40%	Yes	52	Yes	95.00%	Yes	Yes	NA	Developmental deconvolution	
	43	2021	Ming Y. Lu et al. [95]	32,537	Self-collected cases	Whole-slide images	Images	18	83.00%	Yes	317	Yes	61.00%	No	No	NA	Pathomics + CNN encoder, attention-based multiple-instance learning algorithm	
Predictive accuracy was generally determined by comparison with designations made using gold-standard clinicopathological criteria. Accuracy, accuracy of internal data; AE, accuracy of external data; CUP, presence or absence of CUP in the external data; External, presence or absence of validation of external data; Materials, materials used in the study; Source, source of selected samples; Type, type of tissue; NT, number of tumor types; NE, presence or absence of number of external data; Prospective, whether prospective or not; Therapy, whether site-specific therapy received or not; Time, time required to identify primary; Year, year of publication. CCLE, Cancer Cell Line Encyclopedia; CNN, convolutional neural network; COSMIC, the Catalogue of Somatic Mutations in Cancer; CUP, cancer of unknown primary site; FF, fresh frozen; FFPE, formalin-fixed paraffin-embedded; GEO, Gene Expression Omnibus; GTEx, Genotype-Tissue Expression; HPA, Human Protein Atlas; ICGC, International Cancer Genome Consortium; miRNA, microRNA; MOCA, Mouse Organogenesis Cell Atlas; NA, not available; PCAWG, Pan-Cancer Analysis of Whole Genomes; PCR, polymerase chain reaction; TCGA, The Cancer Genome Atlas; TOO, tissue of origin.

CNV, a genetic marker of the genome, is a variation of DNA fragments ranging in size from 1 kb to 3 Mb. CNVs are critical in affecting gene function through gene dosage, breakage, fusion and position effects and have a strong association with tumors [34, 35]. The machine learning (ML) model utilized genomic data to identify TOO of CUP. One noteworthy tool in this context is CNAOrigin, developed by Liang et al., which harnesses a convolutional neural network (CNN) model. After subjecting the model to rigorous 10-fold cross-validations, CNAOrigin demonstrated an impressive predictive accuracy of 83.81% on internal datasets and 79% on independent datasets [26]. Another useful DNA-level method for TOO identification is eQTL, which explains the association between SNPs and gene expression levels. A recent study by Miao et al. explicitly integrated eQTL into the eXtreme Gradient Boosting (XGBoost) classification model. This integration yielded a remarkable prediction accuracy of over 96% in 10-fold cross-validation using The Cancer Genome Atlas (TCGA) data [31]. Combining multiple DNA-level methods may help detect the TOO of CUP more accurately. For instance, Marquard et al. [30] performed a comprehensive study in which an approach using only point mutation had an accuracy of only 69% and an approach that integrated point mutation and CNV significantly improved accuracy to 85%.

Based on gene expression profiling

Metastatic tumors may retain gene expression patterns from cell-type-specific tumors [36]. Therefore, gene expression profiling (GEP) is vital for TOO detection. Currently, several GEP methods (Table 1), such as reverse transcription-polymerase chain reaction (RT-PCR) [15, 37–41], microarrays [42–51], second-generation sequencing of ribonucleic acid (RNA) [38, 52–59] and relative gene expression orderings (REOs) [60], are available to aid in the search for the TOO of CUP.

RT-PCR was one of the pioneering methods used for discrimination and is still utilized today. Several commercial platforms, such as CancerType ID (a 92-gene RT-PCR-based cancer classifier), have been developed. However, RT-PCR is limited compared to microarrays and second-generation sequencing. Microarrays and second-generation sequencing offer the advantage of identifying a broader spectrum of tumor types, assessing older samples with preservation periods extending up to a decade and facilitating targeted therapies [15, 39]. Our research group extracted approximately 1000 signature molecules from the TCGA and RNA sequencing of clinical samples from our institution to create the Bayes algorithm for tissue origin diagnosis (TOD-Bayes algorithm) to diagnose the TOO of hepatobiliary pancreatic malignancies [58]. The accuracy rate of our internal data exceeded 95%, and the external validation corroborated an accuracy rate of 94.4% [58]. Sample REO is stable, which minimizes the impact of experimental batch, data conversion, RNA degradation and tumor tissue sampling site randomization [60, 61]. For example, Li et al. used five gene pairs as markers to predict the TOO of metastatic colorectal cancer (CRC), achieving accuracy rates of 99.36% and 100% for internal and external data, respectively [60].

MicroRNA (miRNA) is a non-coding family of 22-nucleotide single-stranded RNA molecules encoded by endogenous genes [62]. MiRNAs are persistent and resistant to ribonuclease (RNase) degradation in compromised clinical samples, making miRNA array a reliable TOO detection technique [63–68]. Laprovitera and colleagues used 89 miRNAs to deduce the TOO of CUPs. The miRNA expression was evaluated in 159 samples using digital droplet PCR and the least absolute shrinkage and selection operator (LASSO) model combined with the predictive analysis of microarrays (PAMR) nearest shrinkage center of mass method. This integrated approach yielded an internal data accuracy of 95% and increased OS in CUP patients [64]. This study highlights the potential utility of miRNA array in identifying the TOO of CUP.

Based on epigenetics

Epigenetics mechanisms, including DNA methylation, histone modification and chromosomal remodeling, regulate gene expression independently of changes in the DNA sequence [69, 70]. Studies have shown that CUP is characterized by a substantial overall loss of DNA methylation, resulting in a decrease in 5-methylcytosine levels ranging from 20% to 60%, making DNA methylation an ideal biomarker for identifying the TOO of CUP [71–73]. Recent research has used improved DNA methylation platforms to detect 10 481 tumor samples with 99.6% specificity and 97.7% sensitivity using approximately 450 000 CpG sites in the human genome [12].

Based on proteins

Several proteomic methods are available, including tandem mass tagging/isobaric tags for relative and absolute quantification (TMT/iTRAQ) and data-independent acquisition/sequential window acquisition of all theoretical fragment ions (DIA/SWATH) [24]. Nonetheless, no research currently employs rigorous proteomic techniques to identify the TOO of CUP. Hasegawa et al. conducted a retrospective analysis of 90 patients with an unfavorable subset of CUP using a combination of immunohistochemical markers to identify TOO. Fifty-six patients (62.2%) with predicted TOO using this technique received site-specific therapy and had a median OS of 20.3 months, a significant improvement in survival compared to 10.7 months for patients receiving empiric chemotherapy [13]. Although this may not meet the criteria of a ‘stringent’ proteomic technique, it nonetheless underscores the considerable prospective utility of proteomics.

Based on liquid biopsy

Liquid biopsy, a non-invasive methodology, has the potential to revolutionize the diagnosis, treatment and prognosis of CUP [74, 75]. Key biomarkers employed in this approach encompass circulating tumor cells (CTCs), circulating tumor DNA (ctDNA) [76–78], extracellular vesicles (EVs) [79], peripheral blood circulating RNA and tumor-educated platelets (TEPs) [80] (Table 1, Figure 3).

Figure 3 Comparison between liquid biopsy and tissue biopsy for the detection of TOO in CUP. ARMS, amplification refractory mutation system; CUP, cancer of unknown primary; NGS, next-generation sequencing; PCR, polymerase chain reaction; TOO, tissue of origin (Figure 3 belongs to the Based on liquid biopsy section).

CTCs, originating from the primary tumor and circulating within the bloodstream [81], contrast with ctDNA, which comprises DNA fragments shed by cancer cells through apoptosis or necrosis [82]. CTCs and ctDNA can reveal cancer genetic and phenotypic traits and predict TOO [77, 82–85]. Lebofsky and colleagues reported a remarkable 97% concordance between ctDNA analysis and the accurate identification of TOO across 34 patients encompassing 18 distinct tumor types [76]. Nonetheless, it is important to note that the effectiveness may be somewhat limited in detecting early-stage tumors or in older patients [86] because of the diverse metastatic nature of the tumor [74, 75, 87].

EVs refer to a heterogeneous population of small, membrane-bound vesicles found in various body fluids, which contain diverse biomolecules [88]. These EVs can be categorized into subgroups such as exosomes, endosomes, microbubbles and apoptotic bodies are EV subgroups, distinguished by their size and morphological characteristics [88–91]. Hoshino et al. employed EV protein patterns derived from tumor tissue and plasma to differentiate melanoma, colorectal, pancreatic and lung cancers, with a 100% accuracy rate. Moreover, the study showed that the specificity of the EV-based diagnostic method remained consistent across different stages of cancer and could even detect cancers in their early stages [79].

TEPs are important in the systemic and local responses to tumor growth, thereby altering their RNA profile. Best et al. determined the diagnostic potential of TEPs by mRNA sequencing of 283 platelet samples. The TOO was accurately identified in 71% of cases across six different tumor types [80]. These findings suggest that blood platelets are a valuable platform for detecting TOO of CUP.

Circulating RNA in the peripheral blood has the potential to aid in the diagnosis and treatment of CUP [92], although its utilization in research remains limited. In a recent study by Yao et al., the authors demonstrated the potential of this approach by analyzing miRNA profiles from plasma samples obtained from individuals with gastric cancer and non-cancer patients using two independent gene expression synthesis datasets. Three miRNAs-hsa-miR-320a, 1260b and 6515-5p have demonstrated exceptional specificity in distinguishing primary gastric tumors [93]. However, further research is needed to determine the efficacy of this method for CUP patients.

Based on other techniques

In addition to the above techniques for TOO detection, tumor developmental atlases and image omics show considerable potential. Moiso’s team has constructed a comprehensive human tumor development atlas by analyzing and comparing single-cell data from TCGA tumor samples with the Mouse Organogenesis Cell Atlas (MOCA). The atlas aims to establish correlations between cancer biology and development. The team used a developmental multilayer perceptron (D-MLP) classifier constructed from this atlas, which showed remarkable accuracy of 0.974 in identifying TOO [94]. Image omics could also determine the TOO of CUP. Lu et al. developed an artificial intelligence–based (AI-based) pathology training model capable of simultaneously predicting the metastatic status and identifying the origin of 18 different tumor types. On the known primary tumor test set, the model demonstrated outstanding performance, achieving a maximum level 1 accuracy of 0.83 and a level 3 accuracy of 0.96. On the external test set, it also achieved the highest levels 1 and 3 accuracy of 0.80 and 0.93, respectively [95].

ML for identifying TOO of CUP

The basic flowchart of building a TOO classifier with an ML algorithm

The process of developing a classifier for TOO utilizing an ML algorithm entails the following steps (Figure 4): initially, the training set must be created by obtaining multimodal data either from public databases like TCGA and International Cancer Genome Consortium (ICGC) or through self-collection of the data. The data collected may include genomic profiling, gene expression profiling and proteins data from tumor tissue, CTCs and EVs data from plasma and CT images or pathological images. Using genomic profiling data as an example, bioinformatics and ML algorithms are applied to score and rank the most relevant genes for creating tumor–gene associations and constructing TOO classifiers. Several ML algorithms to identify the TOO of CUP have been applied in this context [28, 29, 33, 42, 44, 45, 52–54, 57, 58, 64, 73, 96] (Supplementary Figure 3 and Table 2). These associations are subsequently assessed through independent validation sets, and the classifier’s efficacy is further verified with challenging clinical cases. Finally, the classifier can calculate the ‘tissue origin score’ when applied to CUP patients and then choose the tissue source with the highest score for site-specific therapy [95, 97, 98].

Figure 4 The basic flowchart of building a TOO classifier with an ML algorithm. (A) Multimodal data used to identify the TOO of CUP. (B) The ML algorithm for identifying TOO of CUP. (C) The Construction and Application of TOO Classifier. CTCs, circulating tumor cells; ctDNA, circulating tumor DNA; CUP, cancer of unknown primary; EVs, extracellular vesicles; ICGC, International Cancer Genome Consortium; TCGA, The Cancer Genome Atlas; TEPs, tumor-educated platelets; TOO, tissue of origin (Figure 4 belongs to the The basic flowchart of building a TOO classifier with an ML algorithm section).

Table 2 Comparison of ML algorithms used in the identification of the TOO in CUP

Machine learning algorithms [111]	Strength	Weaknesses	Range of application	
Stepwise additive logistic regression [130]	1. Automatic feature selection: It can automatically select features that have significant predictive power for the response variable, simplifying the model.	1. Possibility for suboptimal models: Stepwise selection or elimination of features may not always find the best model.	1. Suitable for classification problems, especially when interpretability is crucial.	
2. Model interpretability: By reducing unnecessary features, a more interpretable model can be obtained.	2. Instability in selection: Small changes in the data can result in significant variations in the selected feature set.	2. Applicable when the dataset has numerous features.	
3. Control overfitting.	3. High computational cost of iterative calculations.	3. Can serve as a preliminary feature selection stage in a multifaceted modeling procedure.	
Random forest (RF) [131, 132]	1. High predictive accuracy.	1. Lack of interpretability.	1. Suitable for both classification and regression problems.	
2. Robust to overfitting: RF is less prone to overfitting due to the ensemble averaging effect.	2. Computational complexity: It can be computationally expensive.	2. Applicable to datasets with a mixture of numerical and categorical features.	
3. Feature importance estimation: RF can provide information about the relative importance of different features.	3. Bias toward features with more categories: Potentially leading to biased feature importance rankings.	3. Not suitable when interpretability is a primary concern or when computational resources are limited.	
4. Robust to outliers and missing data.			
Support vector machine (SVM) [133, 134]	1. Effective in high-dimensional spaces.	1. Computationally intensive and time-consuming.	1. Effective in complex classification problems with datasets exhibiting complex distributions or clear boundaries.	
2. Robust to overfitting: SVM uses a regularization parameter to control overfitting.	2. Requires feature scaling: SVM is sensitive to the scale of input features and often requires feature standardization.	2. Effective when dealing with various features, especially when the number of features exceeds samples.	
3. Versatility in kernel selection: Different kernel functions can handle non-linear relationships between features.	3. Lack of interpretability.		
4. Effective in small sample sizes.	4. No direct probability estimation: Computing the probability of instances belonging to a class requires additional steps.		
K-nearest neighbors classifier (KNN) [135]	1. Simplicity: Easy to understand and implement.	1. Computational complexity.	1. Small dataset scenarios.	
2. Non-parametric: KNN makes no assumptions about the underlying data distribution.	2. Sensitivity to feature scaling.	2. Non-linear dataset scenarios.	
3. No training phase: New data points can be classified immediately.	3. Lack of robustness to noisy data.	3. Choosing an appropriate distance metric is crucial for accurate classification.	
4. Interpretable results: KNN provides a transparent decision-making process.	4. Boundedness of dimensionality: KNN performance deteriorates as the number of dimensions increases, due to the sparsity of data in high-dimensional spaces.		
eXtreme Gradient Boosting (XGBoost) [136]	1. High performance.	1. Complexity: It requires careful tuning of hyperparameters to achieve optimal performance.	1. Large dataset scenarios, various classification and regression problems.	
2. Flexibility: It can handle various types of data, including numerical and categorical features.	2. Computationally expensive.	2. More suitable for structured data tasks.	
3. Handling missing values: It can reduce the need for extensive preprocessing.	3. Lack of interpretability.		
4. Cross-validation: Allow cross-validation to easily obtain the optimal number of boosting iterations.			
(Continued)

Table 2 Continued

Machine learning algorithms [111]	Strength	Weaknesses	Range of application	
	1. Dimensionality reduction.	1. Loss of interpretability.	1. It can be used for data dimensionality reduction, visualization and preprocessing.	
Principal component analysis (PCA) [137]	2. Feature extraction: PCA can extract a smaller set of features (principal components) that capture the maximum variance in the data.	2. Assumption of linearity: If the underlying data have complex non-linear relationships, PCA may not capture the most important features accurately.	2. It captures the maximum variance by searching for the principal components of the data, thereby simplifying the data structure.	
	3. Noise reduction: PCA can help remove noise or irrelevant features from the dataset by focusing on the components with the highest variance.	3. Sensitive to outliers.		
	4. Data visualization: PCA can be used to visualize high-dimensional data in lower-dimensional spaces.	4. Boundedness of dimensionality: Its performance deteriorates as the number of dimensions increases.		
Naive Bayes algorithm [138]	1. Simplicity and efficiency.	1. Strong independence assumption: In real-world scenarios, features may have dependencies, leading to suboptimal performance.	Naive Bayes is a simple probability-based classifier that is particularly suitable for high-dimensional data and text classification tasks.	
2. Scalability: It performs well with a small amount of training data, real-time or streaming data.	2. Limited expressiveness: Due to its simplicity, it may struggle with capturing complex relationships.	
3. Robust to irrelevant features: It assumes that features are conditionally independent given the class label, making it robust to irrelevant features and helping avoid overfitting.	3. Data scarcity issue: When data are scarce, it may result in poor performance and unreliable predictions.	
4. Interpretability.	4. Sensitive to feature distributions.	
Majority vote algorithm [53]	1. Simplicity.	1. Increase computational burden.	The algorithm is an integrated technology that can combine the prediction results of multiple models to achieve better performance and stability.	
2. Reduction of bias: By combining the predictions of multiple classifiers, it can improve the overall accuracy and robustness of the ensemble.	2. Not always providing improvements: If some of the models have poor performance, integration may not provide any benefits.	
3. Stability: It is less sensitive to small changes in the training data.	3. Not applicable to all problems.	
4. Interpretability.	4. Limited decision boundaries: It may struggle to capture complex or non-linear relationships.	
Convolutional neural network (CNN) [139, 140]	1. Effective feature extraction.	1. Computationally expensive.	1. Suitable for image-related image classification, object detection and semantic segmentation tasks.	
2. Spatial invariance: CNNs are able to capture spatial relationships in data, making them robust to translations, rotations and scale variations.	2. Large memory footprint.	2. Suitable for scenarios with other spatially structured data and a large number of training samples.	
3. Parameter sharing: CNNs utilize weight sharing across different spatial locations, reducing the number of parameters.	3. Limited interpretability.		
	4. Data requirements: CNNs typically require a large amount of labeled training data to generalize well.		

The algorithm underlying these experimental techniques for identifying TOO of CUP

Our literature review indicates a current concentration on applying supervised learning algorithms [27, 29, 30, 42, 54], with limited exploration of unsupervised learning methods [96]. As representatives of supervised learning algorithms, the random forest (RF) model and the XGBoost model are frequently applied algorithms for identifying TOO of CUP [29–31]. The algorithms possess high accuracy, incorporating strategies for handling missing feature data and thus provide an advantage in processing DNA- and RNA-related information (Supplementary Figure 3). Among the reviewed studies, only one article utilized the principal component analysis (PCA) algorithm within the unsupervised learning realm [96]. Notably, there is a conspicuous absence of discussions regarding the application of reinforcement learning algorithms, highlighting a research gap at the algorithmic level.

The feature extraction process is a crucial preliminary step in model construction, involving selecting a subset of the most relevant features from the original set. Different types of data can adopt different types of feature selection strategies. For text data, simple statistical methods like the Pearson correlation algorithm employed by Zhang et al. [27] and Hoshino et al. [79] can filter features. Yet, complex gene interactions challenge traditional methods assuming feature independence. Many studies have employed decision tree models (such as RF) to address feature selection [28, 30, 33]. In addition, Laprovitera et al. [64] used the LASSO algorithm, and Jiang et al. [58] used correlation-based feature selection (CFS), considering the correlation between the target variable and features. Tang et al. [96] proposed a two-tier feature selection strategy, with the first tier based on miRNA differential expression and DNA differential methylation analysis and the second tier mainly employing mathematical algorithms like the PCA algorithm. Traditional ML algorithms, which often rely on a limited set of genes or characteristics, may be constrained in their capacity to discern numerous cancer subtypes. To overcome this limitation, deep learning (DL) algorithms that use various image features to achieve higher accuracy rates have been introduced [99]. For image data, the CNN is a practical feature extraction method, with Lu et al. [95] segmenting images on this basis and extracting local feature descriptors to learn essential features in the images.

Selecting the appropriate algorithm poses a challenge due to significant variations in their advantages, limitations and application areas (Supplementary Figure 3 and Table 2). Using DL algorithmic models is a necessary approach when working with image data. The most elementary of these models is the CNN model. DL models, including Transformer and ResNet, can be employed depending on the objective, such as image detection, classification or segmentation [97]. For textual or sequential data, such as DNA, RNA and proteins, employing XGBoost and lightGBM classification models can produce the desired outcomes.

DISCUSSION

Progress over the past two decades

Over the past two decades, methods for TOO identification have changed drastically (Figure 5). Firstly, the broadening of research perspectives: whereas in earlier years, researchers focused on traditional DNA and mRNA levels, the focus has shifted to the novel, multifunctional analytes such as non-coding RNAs [63–68], proteins [13], epigenetic markers [12, 71–73], ctDNA [76–78] and EVs [79]. Another hallmark is the merging of multiple perspectives and unique insights [96, 99–101]. Some studies including eQTL [31] and REO [60] now analyze multiple genes simultaneously, increasing dimensionality of research. Secondly, the expansion of research tools: with the rapid changes in relevant technologies, the tools used by researchers have changed from PCR to second-generation sequencing [32, 52–56, 59] and tumor developmental atlas [94], thus achieving greater efficiency and accuracy. Thirdly, the expansion of materials used: research has expanded beyond traditional tumor tissue. Liquid biopsy techniques have enabled the shift toward plasma samples [76, 79, 80], whereas image omics have also empowered pathomics to discover the TOO of CUP [95]. Fourthly, expanding the scope of research: anticipated tumor diversity is expanding, and researchers are entering previously unreachable areas (Supplementary Figure 2A). Fifthly, advances in accuracy: accuracy rates have increased from an average of about 80% to nearly 100% in 20 years (Supplementary Figure 2B and C). This progress is, in part, attributed to the proliferation of ML, fostering the growth of bioinformatics (Supplementary Figure 2D) and enabling the analysis of extensive biological datasets, holding significant promise [102–104].

Figure 5 Progress for the detection of TOO in CUP over the past two decades. CUP, cancer of unknown primary; PCR, polymerase chain reaction; TOO, tissue of origin (Figure 5 belongs to the Discussion section).

Challenges of the current studies on experiential methods

We identified some challenges in the current study by summarizing all the studies. The following issues may need to be addressed to improve TOO detection in CUP prognosis. Many techniques used to identify TOO are highly accurate, but whether this ‘digital’ accuracy can be translated into clinical benefits remains to be discussed. Despite a predictive rate of 78.6% for TOO, site-specific therapy based on microarray analysis did not significantly enhance 1-year survival, according to a study by Hidetoshi Hayashi and colleagues [51].

Compared to empirical chemotherapy, identifying TOO and pursuing organ-specific treatment will inevitably result in some time delay. John D. Hainsworth et al. found that TOO takes 2–3 weeks to identify [15], which may not be feasible for CUP patients with short OS. In this regard, reducing the time delay is critical, and the time spent searching for TOO should be a vital criterion for evaluating the technology’s efficacy. However, despite its clinical importance [7, 105], few studies have reported the TOO identification time [15, 64]. This requires researchers’ attention and effort in the future.

Moreover, in clinical practice, tumor tissue from patients with CUP is limited (coarse needle aspiration/biopsy specimens) and can only meet the needs of routine immunohistochemistry in the clinic. Conducting TOO-related tests may require a second tissue biopsy due to insufficient samples, a procedure patients often avoid due to its inherent risks [74, 75, 87, 106, 107]. This emphasizes the need for non-intrusive methods.

Notably, there are few comparative studies on different techniques [96, 99–101, 108]. Atara Posner et al. used DNA features to identify TOO in 51 out of 61 CUP cases, with GEP proving useful in only 21 cases. Their study concluded that DNA mutation analysis outperformed GEP in TOO identification [108]. The authors also found that GEP had lower classification accuracy for cholangiocarcinoma because its transcriptional profile resembles that of pancreatic or upper gastrointestinal tract tumors [56]. On the other hand, DNA mutation profiling is particularly useful because some gene mutations (alterations in IDH1, FGFR2 and BAP1) are highly enriched and have diagnostic significance [108, 109]. While Wei Tang et al. applied miRNA expression and DNA methylation profiles to identify the TOO, the overall accuracy was 87.78% based on the miRNA dataset and 97.06% based on DNA methylation [96]. Haiyan Liu et al. discovered that DNA methylation, GEP and somatic mutation data were best classified by GEP (mean accuracy 94.63%) and worst classified by somatic mutation (mean accuracy 43.33%) [100]. However, no consensus has been regarding the superiority of the different techniques.

Challenges of the current studies on computational methods

In addition to experimental technical obstacles, computational methods have significant pitfalls.

The first statistical challenge, known as the ‘curse of dimensionality’, is a common concern among bioinformatics experts [110]. This term refers to the overfitting problem caused by the excessive number of features, resulting in poor predictive performance on validation sets [111]. Due to the high dimensionality of omics data, the problem of dimensional curse is particularly prominent. Moreover, Chen et al. [110] pointed out the presence of feature redundancy or lack of correlation, introducing noise in high-dimensional space and making it more challenging for models to extract genuine signals. One solution discussed in the paper by Michuda et al. [57] is using regularization indicators to penalize prediction models with too many features, thus supporting simpler models with a relatively minor set of features. Simultaneously, it is necessary to divide the dataset into three subsets: the training, model selection and test set. The model selection set is to identify models with optimal generalization performance. However, new challenges arise, such as the current high cost of next-generation sequencing for liquid biopsy samples, leading to datasets often insufficient for three subsets [76, 79, 80, 90].

Though promising, developing models that integrate prior biological knowledge (e.g. known gene regulatory pathways for specific types of tumors) has not been extensively explored due to the limited availability of such prior knowledge [110]. Selecting the most appropriate features from a multitude of features may also alleviate the issue of the curse of dimensionality. Despite these efforts, organically selecting features from multimodal data and enhancing the interpretability of selected algorithms remain significant challenges in the future.

Future perspective

Different research methods are complex, and each has its own advantages and disadvantages. However, with improvement of medical quality, simple, applicable and accurate research may be the future direction [112, 113]. Liquid biopsy, a safer, cost-effective and less invasive alternative, has emerged as a novel diagnostic, predictive and prognostic window for CUP. Blood is widely believed to be a reservoir for tumor cells in vivo. Thus, liquid biopsy can potentially reduce the sampling bias of tissue biopsy and ultimately provide greater predictive accuracy. Liquid biopsy evaluates prognosis, disease load, risk of recurrence, therapeutic alternatives and dynamic mutational processes [74, 87, 114, 115]. Although the concordance between tissue and liquid biopsy in CUP patients has not been fully evaluated [116–119], the performance of liquid biopsy is a promising direction for predicting TOO in CUP (Figure 3). Besides the above-mentioned materials for CUP, future studies can be conducted on peripheral blood circulating RNA [93, 120] and circulating tumor vascular endothelial cells (cTECs) [121].

Lu et al. demonstrated the enormous benefit of pathomics in identifying TOO of CUP [95], and another potential area is radiomics [122–125]. Conventional tumor evaluation through radiography relies primarily on qualitative features, also known as ‘semantic’ features, like tumor density, enhancement pattern, intratumoral cellular and acellular composition, regularity of tumor margins and anatomical relationships with surrounding tissues [102, 126]. Radiomics allows radiographic images to be quantified according to their shape, size and texture patterns [103, 127, 128]. In cases of extremely high accuracy and integration of multiple data sources, CT- or PET-based imaging may be desirable.

As shown above, comparing various strategies is difficult owing to the significant variability of the tumor types selected in each study, the diverse model development methods and the limited data of the selected samples. Perhaps studies are also needed to compare the ability of different biomarkers under the same conditions, including the same dataset, preprocessing scheme and classification algorithm.

Multi-omics is still intriguing as sequencing costs decrease and technology advances, but its potential to enhance prediction accuracy requires further investigation. Haiyan Liu et al. downloaded GEP, somatic mutation and DNA methylation data of 7224 samples from TCGA and generated seven different feature matrices through various combinations. They found that the best accuracy was 94.63% for the single method and 94.02% after combination, revealing that simply combining multiple biomarkers did not do much to improve prediction accuracy [100]. In contrast, He et al. employed the RF model and integrated gene mutations and expression (TOOme) to infer tumor TOO, which differs from Liu. Their approach yielded higher accuracy (95.77%) compared to using somatic mutations (53.51%) or GEP data alone (89.28%) [99]. While these findings indicate potential, it’s clear that simply stacking multi-omics data is insufficient [129]. A more integrated approach using ML models is likely necessary. Accordingly, further studies are required to determine which omics approaches work best and how to combine them to predict TOO.

CONCLUSIONS

In the era of precision medicine, the endless stream of new technologies has led to rapid advances in TOO identification of CUP: accuracy rates are increasing by leaps and bounds; molecular profiling, including techniques based on genomic profiling, gene expression profiling and epigenetics, is flourishing; and ML is rising. Liquid biopsy and image omics enable non-invasive methods for TOO detection. However, it remains to be confirmed whether the current technological advances have improved patient prognosis. Large-scale clinical studies, multi-institutional collaborations and a unified standard database may need more work.

Key Points

Tissue of origin (TOO) identification for cancer of unknown primary and subsequent site-specific therapy can improve prognosis of CUP.

Of these, techniques to identify the TOO are the critical part.

We systematically review the development and limitations of novel TOO identification methods, compare their pros and cons and assess their potential clinical usefulness in the future.

Supplementary Material

Supplementary_Materials_1_bbae028 Click here for additional data file.

FUNDING

This study is kindly supported by the Natural Science Foundation of Zhejiang (LQ20H160043); Clinical Research Fund Project of Zhejiang Medical Association (2019ZYC-A83); and National Natural Science Foundation of China (82303950).

DATA AVAILABILITY

The data underlying this article are available in the article and in its online supplementary material.

Author Biographies

Wenyuan Ma is a student in the Department of Medical Oncology, The First Affiliated Hospital, Zhejiang University School of Medicine. His research interests include data integration analysis and cancer of unknown primary.

Hui Wu is a student of the Department of Medical Oncology, The First Affiliated Hospital, Zhejiang University School of Medicine. Her research interests include gastric cancer and immunotherapy.

Yiran Chen is a research fellow of the Department of Surgical Oncology, The First Affiliated Hospital, Zhejiang University School of Medicine. Her research interests include bioinformatics and phenotyping and treatment of gastric cancer.

Hongxia Xu is a research fellow of the Zhejiang University-University of Edinburgh Institute (ZJU-UoE Institute), Zhejiang University School of Medicine. Her research interests are in the field of bioinformatics and cancer-related research.

Junjie Jiang is a physician in the Department of Gastroenterology, Affiliated Hangzhou First People’s Hospital, Zhejiang University School of Medicine. His research interests include bioinformatics and gastrointestinal neoplasms.

Bang Du is a researcher at the Real Doctor AI Research Center, School of Medicine, Zhejiang University. Her research interests include medical and artificial intelligence applications.

Mingyu Wan is a student of the Department of Medical Oncology, The First Affiliated Hospital, School of Medicine, Zhejiang University. Her research interests include bioinformatics and cancer-related research.

Xiaolu Ma is a student in the Department of Medical Oncology, The First Affiliated Hospital, Zhejiang University School of Medicine. Her research interests include bioinformatics and cancer-related research.

Xiaoyu Chen is a student of the Department of Medical Oncology, The First Affiliated Hospital, Zhejiang University School of Medicine. Her research interests include machine learning with applications in bioinformatics.

Lili Lin is a doctor of the Department of Nuclear Medicine, The First Affiliated Hospital, Zhejiang University School of Medicine. Her research interests include the development and biological application of molecular imaging probes.

Xinhui Su is a professor in the Department of Nuclear Medicine, The First Affiliated Hospital, Zhejiang University School of Medicine. His laboratory focuses on the basic and clinical study of PET/CT imaging and the synthesis.

Xuanwen Bao is a researcher in the Department of Medical Oncology, The First Affiliated Hospital, Zhejiang University School of Medicine. His laboratory focuses on exploring tumor microenvironment and tumor immune heterogeneity through computational biology and bioinformatics.

Yifei Shen is a researcher in the Department of Laboratory Medicine, The First Affiliated Hospital, Zhejiang University School of Medicine. His laboratory focuses on the use of machine learning and data mining techniques for early screening and prognostic assessment of diseases using multidimensional omics data.

Nong Xu is a professor in the Department of Medical Oncology, The First Affiliated Hospital, Zhejiang University School of Medicine. His laboratory focuses on the mechanism and medical treatment of gastrointestinal tumor and lung cancer.

Jian Ruan is a professor in the Department of Medical Oncology, The First Affiliated Hospital, Zhejiang University School of Medicine. His laboratory focuses on the mechanism and treatment of lung cancer.

Haiping Jiang is a professor in the Department of Medical Oncology, The First Affiliated Hospital, Zhejiang University School of Medicine. Her laboratory focuses on bioinformatics, therapeutic and molecular mechanism research of gastric cancer.

Yongfeng Ding is a professor in the Department of Medical Oncology, The First Affiliated Hospital, Zhejiang University School of Medicine. His laboratory focuses on image omics, therapeutic and molecular mechanism exploration of gastric cancer and CUP.
==== Refs
References

1. Fizazi  K, Greco  FA, Pavlidis  N, et al.  Cancers of unknown primary site: ESMO clinical practice guidelines for diagnosis, treatment and follow-up. Ann Oncol  2015;26 (Suppl 5 ):v133–8.26314775
2. Lee  MS, Sanoff  HK. Cancer of unknown primary. BMJ  2020;371 :m4050.33288500
3. Siegel  RL, Miller  KD, Wagle  NS, Jemal  A. Cancer statistics, 2023. CA Cancer J Clin  2023;73 :17–48.36633525
4. Pavlidis  N, Pentheroudakis  G. Cancer of unknown primary site. Lancet Lond Engl  2012;379 :1428–35.
5. Huebner  G, Link  H, Kohne  CH, et al.  Paclitaxel and carboplatin vs gemcitabine and vinorelbine in patients with adeno- or undifferentiated carcinoma of unknown primary: a randomised prospective phase II trial. Br J Cancer  2009;100 :44–9.19066607
6. Hess  KR, Abbruzzese  MC, Lenzi  R, et al.  Classification and regression tree analysis of 1000 consecutive patients with unknown primary carcinoma. Clin Cancer Res  1999;5 :3403–10.10589751
7. Rassy  E, Pavlidis  N. Progress in refining the clinical management of cancer of unknown primary in the molecular era. Nat Rev Clin Oncol  2020;17 :541–54.32350398
8. Varghese  AM, Arora  A, Capanu  M, et al.  Clinical and molecular characterization of patients with cancer of unknown primary in the modern era. Ann Oncol  2017;28 :3015–21.29045506
9. Kato  S, Alsafar  A, Walavalkar  V, et al.  Cancer of unknown primary in the molecular era. Trends Cancer  2021;7 :465–77.33516660
10. Kim  CS, Hannouf  MB, Sarma  S, et al.  Survival outcome differences based on treatments used and knowledge of the primary tumour site for patients with cancer of unknown and known primary in Ontario. Curr Oncol Tor Ont  2018;25 :307–16.
11. Rassy  E, Parent  P, Lefort  F, et al.  New rising entities in cancer of unknown primary: is there a real therapeutic benefit?  Crit Rev Oncol Hematol  2020;147 :102882.32106012
12. Moran  S, Martínez-Cardús  A, Sayols  S, et al.  Epigenetic profiling to classify cancer of unknown primary: a multicentre, retrospective analysis. Lancet Oncol  2016;17 :1386–95.27575023
13. Hasegawa  H, Ando  M, Yatabe  Y, et al.  Site-specific chemotherapy based on predicted primary site by pathological profile for carcinoma of unknown primary site. Clin Oncol R Coll Radiol G B  2018;30 :667–73.
14. Varadhachary  GR, Raber  MN, Matamoros  A, Abbruzzese  JL. Carcinoma of unknown primary with a colon-cancer profile-changing paradigm and emerging definitions. Lancet Oncol  2008;9 :596–9.18510991
15. Hainsworth  JD, Rubin  MS, Spigel  DR, et al.  Molecular gene expression profiling to predict the tissue of origin and direct site-specific therapy in patients with carcinoma of unknown primary site: a prospective trial of the Sarah Cannon research institute. J Clin Oncol  2013;31 :217–23.23032625
16. Ding  Y, Jiang  J, Xu  J, et al.  Site-specific therapy in cancers of unknown primary site: a systematic review and meta-analysis. ESMO Open  2022;7 :100407.35248824
17. Wolyniec  K, O’Callaghan  C, Fisher  K, et al.  A qualitative study of patients with cancer of unknown primary: perceptions of communication, understanding of diagnosis and genomic testing, and information needs. Psychooncology  2023;32 :589–96.36690922
18. Hyphantis  T, Papadimitriou  I, Petrakis  D, et al.  Psychiatric manifestations, personality traits and health-related quality of life in cancer of unknown primary site. Psychooncology  2013;22 :2009–15.23359412
19. Wagland  R, Bracher  M, Drosdowsky  A, et al.  Differences in experiences of care between patients diagnosed with metastatic cancer of known and unknown primaries: mixed-method findings from the 2013 cancer patient experience survey in England. BMJ Open  2017;7 :e017881.
20. Wolyniec  K, Sharp  J, Fisher  K, et al.  Psychological distress, understanding of cancer and illness uncertainty in patients with cancer of unknown primary. Psychooncology  2022;31 :1869–76.35765251
21. Richardson  A, Wagland  R, Foster  R, et al.  Uncertainty and anxiety in the cancer of unknown primary patient journey: a multiperspective qualitative study. BMJ Support Palliat Care  2015;5 :366–72.
22. Kwee  TC, Kwee  RM. Combined FDG-PET/CT for the detection of unknown primary tumors: systematic review and meta-analysis. Eur Radiol  2009;19 :731–44.18925401
23. Ambrosini  V, Nanni  C, Rubello  D, et al.  18F-FDG PET/CT in the assessment of carcinoma of unknown primary origin. Radiol Med (Torino)  2006;111 :1146–55.17171520
24. Selves  J, Long-Mira  E, Mathieu  M-C, et al.  Immunohistochemistry for diagnosis of metastatic carcinomas of unknown primary site. Cancer  2018;10 :108.
25. Mokhtari  M, Safavi  D, Soleimani  N, et al.  Carcinoma of unknown primary origin: application of immunohistochemistry with emphasis to different cytokeratin 7 and 20 staining patterns. Appl Immunohistochem Mol Morphol  2022;30 :623–34.36036642
26. Liang  Y, Wang  H, Yang  J, et al.  A deep learning framework to predict tumor tissue-of-origin based on copy number alteration. Front Bioeng Biotechnol  2020;8 :701.32850687
27. Zhang  Y, Feng  T, Wang  S, et al.  A novel XGBoost method to identify cancer tissue-of-origin based on copy number variations. Front Genet  2020;11 :585029.33329723
28. Nguyen  L, Van Hoeck  A, Cuppen  E. Machine learning-based tissue of origin classification for cancer of unknown primary diagnostics using genome-wide mutation features. Nat Commun  2022;13 :4013.35817764
29. Liu  X, Li  L, Peng  L, et al.  Predicting cancer tissue-of-origin by a machine learning method using DNA somatic mutation data. Front Genet  2020;11 :674.32760423
30. Marquard  AM, Birkbak  NJ, Thomas  CE, et al.  TumorTracer: a method to identify the tissue of origin from the somatic mutations of a tumor specimen. BMC Med Genomics  2015;8 :58.26429708
31. Miao  Y, Zhang  X, Chen  S, et al.  Identifying cancer tissue-of-origin by a novel machine learning method based on expression quantitative trait loci. Front Oncol  2022;12 :946552.36016607
32. Dietlein  F, Eschner  W. Inferring primary tumor sites from mutation spectra: a meta-analysis of histology-specific aberrations in cancer-derived cell lines. Hum Mol Genet  2014;23 :1527–37.24163242
33. He  B, Dai  C, Lang  J, et al.  A machine learning framework to trace tumor tissue-of-origin of 13 types of cancer based on DNA somatic mutation. Biochim Biophys Acta Mol Basis Dis  2020;1866 :165916.32771416
34. Poduri  A, Evrony  GD, Cai  X, Walsh  CA. Somatic mutation, genomic variation, and neurological disease. Science  2013;341 :1237758.23828942
35. Redon  R, Ishikawa  S, Fitch  KR, et al.  Global variation in copy number in the human genome. Nature  2006;444 :444–54.17122850
36. Dermawan  JK, Rubin  BP. The role of molecular profiling in the diagnosis and management of metastatic undifferentiated cancer of unknown primary: molecular profiling of metastatic cancer of unknown primary. Semin Diagn Pathol  2021;38 :193–8.33309276
37. Varadhachary  GR, Talantov  D, Raber  MN, et al.  Molecular profiling of carcinoma of unknown primary and correlation with clinical evaluation. J Clin Oncol  2008;26 :4442–8.18802157
38. Greco  FA, Lennington  WJ, Spigel  DR, Hainsworth  JD. Molecular profiling diagnosis in unknown primary cancer: accuracy and ability to complement standard pathology. J Natl Cancer Inst  2013;105 :782–90.23641043
39. Ye  Q, Wang  Q, Qi  P, et al.  Development and clinical validation of a 90-gene expression assay for identifying tumor tissue origin. J Mol Diagn  2020;22 :1139–50.32610162
40. Sun  W, Wu  W, Wang  Q, et al.  Clinical validation of a 90-gene expression test for tumor tissue of origin diagnosis: a large-scale multicenter study of 1417 patients. J Transl Med  2022;20 :114.35255924
41. Santos  MTD, de  Souza  BF, Cárcano  FM, et al.  An integrated tool for determining the primary origin site of metastatic tumours. J Clin Pathol  2018;71 :584–93.29248889
42. Tothill  RW, Kowalczyk  A, Rischin  D, et al.  An expression-based site of origin diagnostic method designed for clinical application to cancer of unknown origin. Cancer Res  2005;65 :4031–40.15899792
43. van  Laar  RK, Ma  X-J, de  Jong  D, et al.  Implementation of a novel microarray-based diagnostic test for cancer of unknown primary. Int J Cancer  2009;125 :1390–7.19536816
44. Lu  Q, Chen  F, Li  Q, et al.  A machine learning method to trace cancer primary lesion using microarray-based gene expression data. Front Oncol  2022;12 :832567.35530331
45. Chen  S, Zhou  W, Tu  J, et al.  A novel XGBoost method to infer the primary lesion of 20 solid tumor types from gene expression data. Front Genet  2021;12 :632761.33613644
46. Monzon  FA, Medeiros  F, Lyons-Weiler  M, Henner  WD. Identification of tissue of origin in carcinoma of unknown primary with a microarray-based gene expression test. Diagn Pathol  2010;5 :3.20205775
47. Ades  F, de  Azambuja  E, Daugaard  G, et al.  Comparison of a gene expression profiling strategy to standard clinical work-up for determination of tumour origin in cancer of unknown primary (CUP). J Chemother  2013;25 :239–46.23906077
48. Tothill  RW, Shi  F, Paiman  L, et al.  Development and validation of a gene expression tumour classifier for cancer of unknown primary. Pathology (Phila)  2015;47 :7–12.
49. Staub  E, Buhr  H-J, Gröne  J. Predicting the site of origin of tumors by a gene expression signature derived from normal tissues. Oncogene  2010;29 :4485–92.20514016
50. Ojala  KA, Kilpinen  SK, Kallioniemi  OP. Classification of unknown primary tumors with a data-driven method based on a large microarray reference database. Genome Med  2011;3 :63.21955394
51. Hayashi  H, Kurata  T, Takiguchi  Y, et al.  Randomized phase II trial comparing site-specific treatment based on gene expression profiling with carboplatin and paclitaxel for patients with cancer of unknown primary site. J Clin Oncol  2019;37 :570–9.30653423
52. Li  R, Liao  B, Wang  B, et al.  Identification of tumor tissue of origin with RNA-Seq data and using gradient boosting strategy. Biomed Res Int  2021;2021 :6653793.33681364
53. Shen  Y, Chu  Q, Yin  X, et al.  TOD-CUP: a gene expression rank-based majority vote algorithm for tissue origin diagnosis of cancers of unknown primary. Brief Bioinform  2021;22 :2106–18.32266390
54. Vibert  J, Pierron  G, Benoist  C, et al.  Identification of tissue of origin and guided therapeutic applications in cancers of unknown primary using deep learning and RNA sequencing (TransCUPtomics). J Mol Diagn  2021;23 :1380–92.34325056
55. Hong  J, Hachem  LD, Fehlings  MG. A deep learning model to classify neoplastic state and tissue origin from transcriptomic data. Sci Rep  2022;12 :9669.35690622
56. Zhao  Y, Pan  Z, Namburi  S, et al.  CUP-AI-dx: a tool for inferring cancer tissue of origin and molecular subtype using RNA gene-expression data and artificial intelligence. EBioMedicine  2020;61 :103030.33039710
57. Michuda  J, Breschi  A, Kapilivsky  J, et al.  Validation of a transcriptome-based assay for classifying cancers of unknown primary origin. Mol Diagn Ther  2023;27 :499–511.37099070
58. Jiang  W, Shen  Y, Ding  Y, et al.  A naive Bayes algorithm for tissue origin diagnosis (TOD-Bayes) of synchronous multifocal tumors in the hepatobiliary and pancreatic system. Int J Cancer  2018;142 :357–68.28921531
59. Li  Y, Kang  K, Krahn  JM, et al.  A comprehensive genomic pan-cancer classification using the cancer genome atlas gene expression data. BMC Genomics  2017;18 :508.28673244
60. Li  M, Li  H, Hong  G, et al.  Identifying primary site of lung-limited cancer of unknown primary based on relative gene expression orderings. BMC Cancer  2019;19 :67.30642283
61. Berry  S, Pelkmans  L. Mechanisms of cellular mRNA transcript homeostasis. Trends Cell Biol  2022;32 :655–68.35660047
62. He  B, Zhao  Z, Cai  Q, et al.  miRNA-based biomarkers, therapies, and resistance in cancer. Int J Biol Sci  2020;16 :2628–47.32792861
63. Rosenwald  S, Gilad  S, Benjamin  S, et al.  Validation of a microRNA-based qRT-PCR test for accurate identification of tumor tissue origin. Mod Pathol  2010;23 :814–23.20348879
64. Laprovitera  N, Riefolo  M, Porcellini  E, et al.  MicroRNA expression profiling with a droplet digital PCR assay enables molecular diagnosis and prognosis of cancers of unknown primary. Mol Oncol  2021;15 :2732–51.34075699
65. Søkilde  R, Vincent  M, Møller  AK, et al.  Efficient identification of miRNAs for classification of tumor origin. J Mol Diagn  2014;16 :106–15.24211363
66. Ferracin  M, Pedriali  M, Veronese  A, et al.  MicroRNA profiling for the identification of cancers with unknown primary tissue-of-origin. J Pathol  2011;225 :43–53.21630269
67. Varadhachary  GR, Spector  Y, Abbruzzese  JL, et al.  Prospective gene signature study using microRNA to identify the tissue of origin in patients with carcinoma of unknown primary. Clin Cancer Res  2011;17 :4063–70.21531815
68. Rosenfeld  N, Aharonov  R, Meiri  E, et al.  MicroRNAs accurately identify cancer tissue origin. Nat Biotechnol  2008;26 :462–9.18362881
69. Casado-Pelaez  M, Bueno-Costa  A, Esteller  M. Single cell cancer epigenetics. Trends Cancer  2022;8 :820–38.35821003
70. Dawson  MA, Kouzarides  T. Cancer epigenetics: from mechanism to therapy. Cell  2012;150 :12–27.22770212
71. Moran  S, Martinez-Cardús  A, Boussios  S, Esteller  M. Precision medicine based on epigenomics: the paradigm of carcinoma of unknown primary. Nat Rev Clin Oncol  2017;14 :682–94.28675165
72. Fernandez  AF, Assenov  Y, Martin-Subero  JI, et al.  A DNA methylation fingerprint of 1628 human samples. Genome Res  2012;22 :407–19.21613409
73. Zhang  Z, Lu  Y, Vosoughi  S, et al.  HiTAIC: hierarchical tumor artificial intelligence classifier traces tissue of origin and tumor type in primary and metastasized tumors using DNA methylation. NAR Cancer  2023;5 :zcad017.37089814
74. Ignatiadis  M, Sledge  GW, Jeffrey  SS. Liquid biopsy enters the clinic—implementation issues and future challenges. Nat Rev Clin Oncol  2021;18 :297–312.33473219
75. Li  W, Liu  J-B, Hou  L-K, et al.  Liquid biopsy in lung cancer: significance in diagnostics, prediction, and treatment monitoring. Mol Cancer  2022;21 :25.35057806
76. Lebofsky  R, Decraene  C, Bernard  V, et al.  Circulating tumor DNA as a non-invasive substitute to metastasis biopsy for tumor genotyping and personalized medicine in a prospective trial across all tumor types. Mol Oncol  2015;9 :783–90.25579085
77. Lu  S-H, Tsai  W-S, Chang  Y-H, et al.  Identifying cancer origin using circulating tumor cells. Cancer Biol Ther  2016;17 :430–8.26828696
78. Laprovitera  N, Salamon  I, Gelsomino  F, et al.  Genetic characterization of cancer of unknown primary using liquid biopsy approaches. Front Cell Dev Biol  2021;9 :666156.34178989
79. Hoshino  A, Kim  HS, Bojmar  L, et al.  Extracellular vesicle and particle biomarkers define multiple human cancers. Cell  2020;182 :1044–1061.e18.32795414
80. Best  MG, Sol  N, Kooi  I, et al.  RNA-Seq of tumor-educated platelets enables blood-based Pan-cancer, multiclass, and molecular pathway cancer diagnostics. Cancer Cell  2015;28 :666–76.26525104
81. Batth  IS, Mitra  A, Rood  S, et al.  CTC analysis: an update on technological progress. Transl Res J Lab Clin Med  2019;212 :14–25.
82. Cheng  ML, Pectasides  E, Hanna  GJ, et al.  Circulating tumor DNA in advanced solid tumors: clinical relevance and future directions. CA Cancer J Clin  2021;71 :176–90.33165928
83. Dang  DK, Park  BH. Circulating tumor DNA: current challenges for clinical utility. J Clin Invest  2022;132 :e154941.35703177
84. Pellini  B, Chaudhuri  AA. Circulating tumor DNA minimal residual disease detection of non-small-cell lung cancer treated with curative intent. J Clin Oncol  2022;40 :567–75.34985936
85. Schuster  E, Taftaf  R, Reduzzi  C, et al.  Better together: circulating tumor cell clustering in metastatic cancer. Trends Cancer  2021;7 :1020–32.34481763
86. Hemminki  K, Pavlidis  N, Tsilidis  KK, et al.  Age-dependent metastatic spread and survival: cancer of unknown primary as a model. Sci Rep  2016;6 :23725.27009354
87. von  Felden  J, Garcia-Lezana  T, Schulze  K, et al.  Liquid biopsy in the clinical management of hepatocellular carcinoma. Gut  2020;69 :2025–34.32883873
88. EL Andaloussi  S, Mäger  I, Breakefield  XO, et al.  Extracellular vesicles: biology and emerging therapeutic opportunities. Nat Rev Drug Discov  2013;12 :347–57.23584393
89. Cheng  L, Hill  AF. Therapeutically harnessing extracellular vesicles. Nat Rev Drug Discov  2022;21 :379–99.35236964
90. Nikanjam  M, Kato  S, Kurzrock  R. Liquid biopsy: current technology and clinical applications. J. Hematol. Oncol. J Hematol Oncol  2022;15 :131.36096847
91. Yu  W, Hurley  J, Roberts  D, et al.  Exosome-based liquid biopsies in cancer: opportunities and challenges. Ann Oncol  2021;32 :466–77.33548389
92. Mugoni  V, Ciani  Y, Nardella  C, Demichelis  F. Circulating RNAs in prostate cancer patients. Cancer Lett  2022;524 :57–69.34656688
93. Yao  Y, Ding  Y, Bai  Y, et al.  Identification of serum circulating MicroRNAs as novel diagnostic biomarkers of gastric cancer. Front Genet  2020;11 :591515.33597967
94. Moiso  E, Farahani  A, Marble  HD, et al.  Developmental deconvolution for classification of cancer origin. Cancer Discov  2022;12 :2566–85.36041084
95. Lu  MY, Chen  TY, Williamson  DFK, et al.  AI-based pathology predicts origins for cancers of unknown primary. Nature  2021;594 :106–10.33953404
96. Tang  W, Wan  S, Yang  Z, et al.  Tumor origin detection with tissue-specific miRNA and DNA methylation markers. Bioinforma Oxf Engl  2018;34 :398–406.
97. Kleppe  A, Skrede  O-J, De Raedt  S, et al.  Designing deep learning studies in cancer diagnostics. Nat Rev Cancer  2021;21 :199–211.33514930
98. Howard  J . Algorithms and the future of work. Am J Ind Med  2022;65 :943–52.36128686
99. He  B, Lang  J, Wang  B, et al.  TOOme: a novel computational framework to infer cancer tissue-of-origin by integrating both gene mutation and expression. Front Bioeng Biotechnol  2020;8 :394.32509741
100. Liu  H, Qiu  C, Wang  B, et al.  Evaluating DNA methylation, gene expression, somatic mutation, and their combinations in inferring tumor tissue-of-origin. Front Cell Dev Biol  2021;9 :619330.34012960
101. Chen  K, Zhang  F, Yu  X, et al.  A molecular approach integrating genomic and DNA methylation profiling for tissue of origin identification in lung-specific cancer of unknown primary. J Transl Med  2022;20 :158.35382836
102. Huang  S, Yang  J, Fong  S, Zhao  Q. Artificial intelligence in cancer diagnosis and prognosis: opportunities and challenges. Cancer Lett  2020;471 :61–71.31830558
103. Bi  WL, Hosny  A, Schabath  MB, et al.  Artificial intelligence in cancer imaging: clinical challenges and applications. CA Cancer J Clin  2019;69 :127–57.30720861
104. Jiang  Y, Yang  M, Wang  S, et al.  Emerging role of deep learning-based artificial intelligence in tumor pathology. Cancer Commun Lond Engl  2020;40 :154–66.
105. Pentheroudakis  G, Briasoulis  E, Pavlidis  N. Cancer of unknown primary site: missing primary or missing biology?  Oncologist  2007;12 :418–25.17470684
106. Brezgyte  G, Shah  V, Jach  D, Crnogorac-Jurcevic  T. Non-invasive biomarkers for earlier detection of pancreatic cancer-a comprehensive review. Cancer  2021;13 :2722.
107. Fitzgerald  RC, Antoniou  AC, Fruk  L, Rosenfeld  N. The future of early cancer detection. Nat Med  2022;28 :666–77.35440720
108. Posner  A, Prall  OW, Sivakumaran  T, et al.  A comparison of DNA sequencing and gene expression profiling to assist tissue of origin diagnosis in cancer of unknown primary. J Pathol  2023;259 :81–92.36287571
109. Arai  Y, Totoki  Y, Hosoda  F, et al.  Fibroblast growth factor receptor 2 tyrosine kinase fusions define a unique molecular subtype of cholangiocarcinoma. Hepatol Baltim Md  2014;59 :1427–34.
110. Chen  P-HC, Liu  Y, Peng  L. How to develop machine learning models for healthcare. Nat Mater  2019;18 :410–4.31000806
111. Greener  JG, Kandathil  SM, Moffat  L, Jones  DT. A guide to machine learning for biologists. Nat Rev Mol Cell Biol  2022;23 :40–55.34518686
112. Olivier  T, Fernandez  E, Labidi-Galy  I, et al.  Redefining cancer of unknown primary: is precision medicine really shifting the paradigm?  Cancer Treat Rev  2021;97 :102204.33866225
113. Rassy  E, Labaki  C, Chebel  R, et al.  Systematic review of the CUP trials characteristics and perspectives for next-generation studies. Cancer Treat Rev  2022;107 :102407.35569387
114. Ye  Q, Ling  S, Zheng  S, Xu  X. Liquid biopsy in hepatocellular carcinoma: circulating tumor cells and circulating tumor DNA. Mol Cancer  2019;18 :114.31269959
115. Kato  S, Krishnamurthy  N, Banks  KC, et al.  Utility of genomic analysis in circulating tumor DNA from patients with carcinoma of unknown primary. Cancer Res  2017;77 :4238–46.28642281
116. Chae  YK, Davis  AA, Carneiro  BA, et al.  Concordance between genomic alterations assessed by next-generation sequencing in tumor tissue or circulating cell-free DNA. Oncotarget  2016;7 :65364–73.27588476
117. Chae  YK, Davis  AA, Jain  S, et al.  Concordance of genomic alterations by next-generation sequencing in tumor tissue versus circulating tumor DNA in breast cancer. Mol Cancer Ther  2017;16 :1412–20.28446639
118. Zugazagoitia  J, Ramos  I, Trigo  JM, et al.  Clinical utility of plasma-based digital next-generation sequencing in patients with advance-stage lung adenocarcinomas with insufficient tumor samples for tissue genotyping. Ann Oncol  2019;30 :290–6.30535340
119. Schwaederlé  MC, Patel  SP, Husain  H, et al.  Utility of genomic assessment of blood-derived circulating tumor DNA (ctDNA) in patients with advanced lung adenocarcinoma. Clin Cancer Res  2017;23 :5101–11.28539465
120. Hamam  R, Hamam  D, Alsaleh  KA, et al.  Circulating microRNAs in breast cancer: novel diagnostic and prognostic biomarkers. Cell Death Dis  2017;8 :e3045.28880270
121. Lin  PP . Aneuploid circulating tumor-derived endothelial cell (CTEC): a novel versatile player in tumor neovascularization and cancer metastasis. Cell  2020;9 :1539.
122. Lambin  P, Leijenaar  RTH, Deist  TM, et al.  Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol  2017;14 :749–62.28975929
123. Mayerhoefer  ME, Materka  A, Langs  G, et al.  Introduction to Radiomics. J Nucl Med  2020;61 :488–95.32060219
124. Bera  K, Braman  N, Gupta  A, et al.  Predicting cancer outcomes with radiomics and artificial intelligence in radiology. Nat Rev Clin Oncol  2022;19 :132–46.34663898
125. Guiot  J, Vaidyanathan  A, Deprez  L, et al.  A review in radiomics: making personalized medicine a reality via routine imaging. Med Res Rev  2022;42 :426–40.34309893
126. Aerts  HJWL, Velazquez  ER, Leijenaar  RTH, et al.  Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nat Commun  2014;5 :4006.24892406
127. Hosny  A, Parmar  C, Quackenbush  J, et al.  Artificial intelligence in radiology. Nat Rev Cancer  2018;18 :500–10.29777175
128. Huynh  E, Hosny  A, Guthier  C, et al.  Artificial intelligence in radiation oncology. Nat Rev Clin Oncol  2020;17 :771–81.32843739
129. Zheng  B, Fang  L. Spatially resolved transcriptomics provide a new method for cancer research. J Exp Clin Cancer Res  2022;41 :179.35590346
130. Padoa-Schioppa  C . Logistic analysis of choice data: a primer. Neuron  2022;110 :1615–30.35334232
131. Amaratunga  D, Cabrera  J, Lee  Y-S. Enriched random forests. Bioinforma Oxf Engl  2008;24 :2010–4.
132. Chai  Z, Zhao  C. Multiclass oblique random forests with dual-incremental learning capacity. IEEE Trans Neural Netw Learn Syst  2020;31 :5192–203.31995504
133. Nedaie  A, Najafi  AA. Support vector machine with Dirichlet feature mapping. Neural Netw  2018;98 :87–101.29223012
134. Wang  H, Shao  Y, Zhou  S, et al.  Support vector machine classifier via L0/1 soft-margin loss. IEEE Trans Pattern Anal Mach Intell  2022;44 :7253–65.34166185
135. Shi  J, Chen  X, Xie  Y, et al.  Delicately reinforced k -nearest neighbor classifier combined with expert knowledge applied to abnormity forecast in electrolytic cell. IEEE Trans Neural Netw Learn Syst  2023;1–11.
136. Li  F, Zuo  Y, Lin  H, Wu  J. Boostxml: gradient boosting for extreme multilabel text classification with tail labels. IEEE Trans Neural Netw Learn Syst  2023;1–14.
137. Mi  J-X, Zhang  Y-N, Lai  Z, et al.  Principal component analysis based on nuclear norm minimization. Neural Netw  2019;118 :1–16.31228720
138. Ding  X, Zhang  H, Ma  C, et al.  User identification across multiple social networks based on naive Bayes model. IEEE Trans Neural Netw Learn Syst  2022;1–12.
139. Sarıgül  M, Ozyildirim  BM, Avci  M. Differential convolutional neural network. Neural Netw  2019;116 :279–87.31125914
140. Anwar  SM, Majid  M, Qayyum  A, et al.  Medical image analysis using convolutional neural networks: a review. J Med Syst  2018;42 :226.30298337
