As shown in Supplementary Figure 1 , a total of 37 patients clinically diagnosed as carcinoma of the unknown primary were retrospectively enrolled from multiple hospitals across China according to the NCCN guidelines. The carcinoma of unknown primary was defined as a histologically proven metastatic malignant tumor whose primary site cannot be identified during pretreatment evaluation. Patient follow-up has also been performed to confirm that the primary tumors remained unknown until the time of this study. Two patients were excluded due to the lack of evaluable tumor tissue and plasma or the difference in the targeted sequencing panel used. Finally, 26 tumor tissue and 22 plasma samples from a total of 35 CUP patients were subjected for further analysis. The tumor and plasma samples were all collected after treatment. Written consent was collected from each patient. The NGS tests were performed in a centralized clinical testing center (Nanjing Geneseeq Technology Inc.) with a target gene panel of 416 genes according to protocols reviewed and approved by the ethical committee of each participating hospital. 5 to 10 ml of peripheral blood was collected from each patient in EDTA-coated tubes (BD Biosciences, San Jose, CA, USA). Plasma was extracted within 2 h of blood collection and shipped to the central testing laboratory within 48 h. Formalin-fixed paraffin-embedded (FFPE) tumor tissue blocks/sections or fresh tumor tissues were obtained from the hospitals, and with confirmation by the pathologists from the centralized clinical testing center for diagnosis and tumor purity.

The DNA extraction, quantification, and library preparation were performed as previously described (9). In brief, FFPE samples were de-paraffinized with xylene, and DNA was extracted using the QIAamp DNA FFPE Tissue Kit (Qiagen, Hilden, Germany) according to the manufacturer’s protocols. Genomic DNA from fresh tumor tissue was extracted using the DNeasy Blood & Tissue Kit (Qiagen) according to the manufacturer’s protocols. Peripheral blood samples were centrifuged at 1,800g for 10 min. Then, the plasma was isolated for extraction of cfDNA and the genomic DNA of white blood cells in sediments served as normal control. The Circulating Nucleic Acid Kit (Qiagen, Germany) was used to purify cfDNA from plasma. The genomic DNA from white blood cells was extracted using DNeasy Blood and Tissue Kit (Qiagen). Genomic DNA was qualified using a NanoDrop 2000 (Thermo Fisher Scientific, Waltham, MA), and cfDNA fragment distribution was analyzed on a 2100 Bioanalyzer using the High Sensitivity DNA Kit (Agilent Technologies, Santa Clara, CA). All DNA was quantified using the dsDNA HS assay kit on a Qubit 3.0 fluorometer (Life Technologies, Carlsbad, CA, USA) according to the manufacturer’s recommendations. Sequencing libraries were prepared using the KAPA HyperPrep Kit (KAPA Biosystems, Woburn, MA, USA) with an optimized manufacturer’s protocol and sequenced as previously described (9). The mean sequencing depths of tumor tissue and plasma were 500× and 3,000×, respectively.

Sequencing data were processed as previously described (9). In brief, the data were first demultiplexed and subjected to FASTQ file quality control to remove low-quality data or N bases. Qualified reads were mapped to the reference human genome hg19 using the Burrows–Wheeler Aligner, and the Genome Analysis Toolkit (GATK 3.4.0) was employed to apply the local realignment around indels and base quality score recalibration. Picard was used to remove PCR duplicates. VarScan2 was employed for the detection of single-nucleotide variations (SNVs) and insertion/deletion mutations. SNVs were filtered out if the mutant allele frequency (MAF) was less than 1% for tumor tissue and 0.3% for plasma samples. Common SNVs were excluded if they were present in >1% of the population in the 1000 Genomes Project or the Exome Aggregation Consortium (ExAC) 65,000 exomes database. The resulting mutation list was further filtered by an in-house list of recurrent artifacts based on a normal pool of whole blood samples. Parallel sequencing of matched white blood cells from each patient was performed to further remove sequencing artifacts, germline variants, and clonal hematopoiesis. The copy number alterations were analyzed as previously described (10, 11). The tumor purities were first estimated using ABSOLUTE (12). Somatic CN alteration events were assigned based on sample-ploidy values calculated in the FACETS algorithm (13). Loss of heterozygosity (LOH) was also calculated using FACETS and determined using the minor copy number estimates of each segment for genes in the targeted panel. The minor copy number is by definition 0 in an LOH event (14, 15). Structural variants were detected using FACTERA with default parameters (16). The fusion reads were further manually reviewed and confirmed on Integrative Genomics Viewer (IGV).

The microsatellite instability (MSI) was determined by evaluating 52 embedded mononucleotide repeats with a minimum of 15-bp repeats that were included in the sequencing panel. The baseline length distribution of each repeat was determined from a pool of microsatellite-stable samples. A sample was identified as MSI if more than 40% of the qualified sites displayed instability. Tumor mutational burden (TMB, mutation per megabase) was determined based on the number of somatic base substitutions and indels in the targeted regions of the gene panel covering 0.85 Mb of coding genomes, excluding known driver mutations as they are overrepresented in the panel. Samples within the highest mutation-load tertile (top 33.3%) were classified as having high TMB, which was equal to ≥20 mutations/Mb in TMB, as previously described (17). LOH-high was also defined as the samples within the highest LOH tertile (top 33.3%), which was equal to ≥33% in LOH.

Homologous recombination (HR) pathway genes were defined according to previous publications including the genomic analyses of ovarian carcinoma published by the Cancer Genome Atlas Research Network (TCGA) and clinical trials of the PARP inhibitor in different cancer types (NCT04042831, NCT03207347, and NCT03209401) (18–21). The HRR genes included BRCA1/2, PALB2, CHEK2, ARID1A, NBN, BARD1, BRIP1, RAD50, RAD51C, RAD51D, RAD54L, MRE11, the set of Fanconi anemia genes (FANCA, FANCC, FANCD2, FANCE, FANCF, FANCG, FANCL, FANCM), C11orf30 (EMSY), the DNA damage sensing genes ATM and ATR, and PTEN.

This CUP cohort included 24 males (69%) and 11 females (31%) with a median age at diagnosis of 62 years, ranging from 42 to 82 years. Of the 35 CUP cases included in this retrospective study, 42% (11 out of 35) was adenocarcinoma (ACUP) and 23% (8 out of 35) was squamous cell cancer (SCUP). Two cases were identified as poorly differentiated CUP. Histology type was unknown in 40% (14 out of 35) of cases. The distribution of tumor location is shown in Table 1 , and the detailed tumor metastasis status is shown in Supplementary Table 1 . The numbers of organ metastatic lesions were classified as follows: lymph node metastasis (10, 29%), single-organ metastasis (8, 23%), organ and lymph node metastasis (3, 8%), multiple-organ metastasis (3, 8%), and unknown metastasis (11, 31%). In patients with single-organ metastasis, tumor was found in the lung (2), liver (2), bone (2), retina (1), and abdomen (1). In patients with organ and lymph node metastasis, organ metastasis was detected in the bone, tongue, bone, and adrenal gland, respectively. In the three patients with multiple organ metastasis, two patients with both lung and bone metastasis showed metastasis in intestine and gall bladder respectively. The other patient had metastasis in the liver and gut.

CUP has been shown associated with the occurrence of lung, kidney, and colorectal cancers in families (22). Here eleven patients had cancer occurrence in the family, including three families with esophageal cancers, two families with lung cancers, and two with colon cancers ( Table 1 and Supplementary Table 1 ). Both the sister and brother of patient 30 were diagnosed with lung cancer while the brother of patient 15 was diagnosed with lung cancer. The brother of patient 10 and the mother of patient 07 were diagnosed with colon cancer. However, we did not find any potentially deleterious germline variant that should have prompted patients to undergo germline testing nor had they had undergone germline testing before the study.

In this CUP cohort, tissue samples were available in 26 patients and plasma samples were available in 22 patients. Tumor tissues were from metastatic tumors including lymph nodes tumors (13), lung tumors (4), liver tumors (2), bone (2), retina (1), and abdomen (1). The genomic landscape of CUP patients in tumor tissue and plasma is shown in Figure 1 and Supplementary Figure 2 , respectively. The detailed mutation list is shown in Supplementary Table 2 . Commonly altered genes found in both tissue and plasma samples included TP53 (77% of tissue, 64% of plasma), KRAS (31% of tissue, 36% of plasma), and SMAD4 (23% of tissue, 14% of plasma), similar to reported CUP cohorts (23). MYC amplification was the most frequently observed copy number variation (CNV) which accounted for 23% (6 out of 26) of total cases. Other frequently observed CNVs included CDKN2A deletion (15%) and ZNF217 amplification (15%). EML4-ALK fusion was detected in one patient.

Based on the ten canonical pathways profiled by The Cancer Genome Atlas (TCGA) (24), we looked into the frequently altered pathways in CUP. As shown in Figure 2 , the TP53 signaling pathway (23/26, 88%) was the most frequently altered canonical pathway in CUP. RTK-RAS (20/26, 77%) and PI3K (15/26, 58%) signaling pathways were also found to be in tumor tissues of more than half of CUP patients. Intriguingly, 58% (15/26) of the CUP cases harbored alterations in homologous recombination pathways ( Figure 3 and Supplementary Table 3 ).

We further compared the altered pathways in CUP with three published cancer cohorts (one stage II–IV esophageal squamous cell cancer cohort/ESCC, one metastatic lung cancer cohort, one stage II–III colorectal cancer cohort) (25–27) ( Supplementary Figure 3 ). The tumor tissues of the ESCC and lung cancer cohorts were sequenced with the same 416 gene panels while the colorectal cancer cohort was sequenced with a 425-gene panel covering the 416 genes. Lung cancer and colorectal cancer have been reported with a high ratio of RTK-RAS pathway gene alterations while the ESCC cohort and colorectal cancer have a high ratio of TP53 pathway alterations according to other publications (28–31). The HRR pathway has been reported to be enriched in esophageal squamous cell cancer (32). In this study, TP53 and RTK-RAS pathways were the top two pathways altered in CUP. The ratio of altered RTK-RAS in lung cancer was significantly higher than that in ESCC and colorectal cancer (p < 0.001). The difference in the RTK-RAS ratio between CUP and lung cancer was insignificant due to the limited cohort size. The ratio of TP53 pathway alterations was insignificant among the four groups but with a trend of decrease in lung cancer. Interestingly, the CUP cohort showed a higher ratio of patients with HRR pathway gene mutations than the other three cohorts (p < 0.05).

The median tumor mutational burden (TMB) of the 26 CUP cases with tumor tissue was 12.5 mutation/Mb. The differences in TMB were significant between patients with HRR wild-type and mutated HRR (p = 0.013, mean TMB HRR WT vs. mutated: 9.9 vs. 23.9 mutation/Mb) Figure 2B ). The microsatellite instability (MSI) analysis revealed that the 26 tumor specimens were all microsatellite stable (MSS). Genomic loss of heterozygosity (LOH) was also assessed, showing a median LOH of 28.0%. Although there was no significant difference in the LOH status between HRR WT and mutated tumors, 67% (10/15) of HRR-mutated tumors displayed a LOH higher than 20% and 40% (6/15) had a LOH over 33% ( Supplementary Table 3 ).

According to the OncoKB database, clinically actionable mutations identified in this cohort were KRAS G12C(3), PIK3CA E545K(2), PIK3CA E542K(1), BRAF V600L(1), EML4-ALK(1), and ALK L1196M(1) ( Table 2 ), accounting for 23% (8 out of 35) of total cases. 78% of the actionable mutations were detected in both plasma and tissue samples from the same patient. The available treatment history of patients is shown in Figure 3B . Platinum-based chemotherapy was applied in eight cases without intervention of tyrosine kinase inhibitor (TKI) therapy.

Four patients with metastatic squamous cell cancer (SCC) in lymph nodes (P05, P07, P08, P12) progressed on platinum-based chemotherapy within 10 months of treatment ( Figure 3B ). Although three of four patients displayed HRR mutated and LOH-high, the PARP inhibitor may not be a good option for these patients. Meanwhile, PIK3CA E545K mutation was identified in tumor tissue and plasma of P05 and 12, and no other known driver mutation was identified in both cases ( Table 2 ). Target therapy with PI3K inhibitors such as buparlisib, copanlisib, and taselisib might be a better treatment option, especially for patient 05 with an allele frequency (AF) of 39.25% detected in tissue. Patient 26 with metastatic adenocarcinoma in lymph node was treated with pemetrexed disodium plus oxaliplatin for 6 months and achieved stable disease (SD). KRAS G12C was found at an AF of 77.75% in tissue and 58.96% in plasma, indicating that the KRAS G12C inhibitor such as AMG-510 could be beneficial for this patient. P17 and P23 were HRR-mutated and LOH-high adenocarcinoma. Carboplatin or oxaliplatin-based chemotherapy was administrated for more than 5 months. PARP inhibitors may be used for these patients if they reach a platinum-free interval of 6 months or longer.