The definition of rare tumors is primarily based on Wang et al.^[9,16], who defined rare tumors in China as cancer types with an annual incidence rate of fewer than 2.5 per 100,000. To update the list of rare tumors, our expert group added several new entities, including Mesothelioma^[17-20], Thymic epithelial tumors^[21], and NUT carcinoma^[22], and additional cancer types from the Rare tumor list of China^[23]. The final list comprises 137 cancer types, which are classified into 15 tumor systems according to their primary anatomical sites and/or histopathological features [Supplementary Table 1].

Genomic and clinical data were obtained from cBioPortal public studies, including The Cancer Genome Atlas (TCGA) cohorts [Supplementary Table 2]. Inclusion criteria were as follows: (1) the cancer type met the specified definition of rare tumors; (2) genomic testing of at least 50 genes required; and (3) molecular profiling data included at least one qualified genomic data type, i.e. somatic mutations, copy number alterations, structural variations/gene fusions, tumor mutation burden (TMB), or microsatellite instability (MSI). Patients diagnosed with multiple types of primary cancer were excluded. Studies focusing on specific gene mutations (such as basket trials) and non-real-world studies (including cell lines, patient-derived xenografts, and organoids) were also excluded.

For each eligible patient, demographic characteristics (including age and gender), cancer type and histology, specimen type, genomic alteration profile (including somatic mutations, copy number alterations, and gene fusions), TMB, and MSI status were extracted. TMB was calculated as the total number of somatic non-synonymous mutations per megabase (Mb) of sequenced genome. If multiple TMB values were recorded for a patient, the median value was used as the individual’s TMB. High TMB was defined as ≥ 10 mutations/Mb. MSI status was determined using standard genomic instability metrics.

Actionability and resistance were queried using the CCBM knowledgebase (version 30 June, 2026). A case is considered “Actionable” if the patient’s genomic alterations and/or immune biomarkers match any intervention(s) in the knowledgebase, and the drug targets are related to the signaling pathways of the detected biomarkers. If matching interventions are found only for cancer types different from the patient’s diagnosis, the case is classified as “Off label”. If no matching interventions are found, the case is classified as “Non-actionable”. Only oncogenic mutations listed in the COSMIC database were considered.

Evidence levels are allocated as follows: Level 1 includes biomarker-driven indications approved by regulatory authorities (i.e. NMPA and FDA); Level 2 includes biomarker-driven indications recommended by national guidelines (i.e. CSCO, CACA, and NCCN); Level 3 includes biomarker-guided interventions supported by clinical trials with more than 10 cases; and Level 4 includes biomarker-guided interventions supported by case reports or by preclinical studies.

In the actionability analysis, single and compound biomarker(s) were evaluated separately, as some compound biomarkers have distinct medical interpretations. For example, third- and fourth-generation EGFR-TKIs are recommended for non-small cell lung cancer patients with EGFR T790M and C797S mutations, respectively, but when EGFR T790M and C797S (trans) co-mutated, combination therapy with first- and third-generation EGFR-TKIs is recommended instead^[24-26]. The analysis also considered the hierarchical relationships between cancer types and genomic alterations. For instance, interventions matching broader categories such as “Solid tumors” are also defined as “Actionable”. Similarly, interventions for broader genomic alterations, such as EGFR exon 20 oncogenic mutations and EGFR PACC mutations, are considered “Actionable” for specific mutations like EGFR C797S. This approach helps avoid gaps in interpretation due to tumor-agnostic agents or common biomarkers.

The genomic landscape was conducted at the sample level and the actionability analyses were conducted at the patient level. Descriptive statistics were used to summarize patient characteristics. Median values with ranges represented continuous variables, while frequencies with percentages represented categorical variables. Bioinformatics analysis comprised coordinating clinical annotations, integrating and aggregating genomic mutation profiles, classifying TMB and MSI status, and identifying genes with recurrent alterations in each tumor classification system. Heatmaps plotted by ComplexHeatmap (Version 2.26.1) presented genomic discoveries at the sample level. Bar charts displayed the top mutated genes for each classification system. All analyses were performed using R software (Version 4.5.1) and Python (Version 3.11.3). Visualization was performed using basic R functions.

The final analysis included 44,790 samples from 42,569 rare tumor patients. Supplementary Table 3 summarizes the demographic characteristics of the study cohort. There was a total of 137 cancer types, spanning 15 classification systems [Supplementary Table 1]. The top five cancer types were glioblastoma (n = 3,439), cancer of unknown primary (n = 2,927), neuroblastoma (n = 2,753), leiomyosarcoma (n = 2,520), and angiosarcoma (n = 1,798).

The overall mutational profile of rare tumors and the mutational profiles of each tumor classification system are shown in Supplementary Figures 1 and 2, respectively. In our cohort, the most frequently mutated genes among all rare cancer types were TP53 (16.5%), CDKN2A (5.0%), ATRX (4.6%), RB1 (4.2%), and DNMT3A (4.1%). Among them, TP53, NF1, PTEN, IDH1, and PIK3CA were the most common potential actionable targets.

The top mutated genes in rare tumor patients stratified by tumor classification system were shown in Figure 1. Tumors from different classification systems possessed unique mutational characteristics, with GATA3, NOTCH1, BAP1, MEN1, ARID1A, and GTF2I being the most prevalent mutated genes in Breast, Head and neck/Eye, Mesothelioma, Neuroendocrine system, Reproductive system, and Thymus tumors, respectively. TP53 was the most prevalent mutated gene in the rest of the tumor classification systems.

The median TMB of rare tumors was 0.30 mutations/Mb (range: 0-510.47 mutations/Mb). Only 4.3% (1,046/24,354) and 0.9% (122/13,013) patients were classified as TMB-high and MSI-high, respectively. Figure 2 shows the prevalence of actionable immune biomarkers (i.e., TMB-high and MSI-high) in rare tumor patients stratified by tumor classification system. The distribution of TMB and MSI status varied substantially across tumor classification systems, with Skin tumors (28.2%, 256/909) and Endocrine system tumors (1.9%, 6/317) exhibiting the highest rates of TMB-high and MSI-high, respectively.

Figure 3 presents the proportion of rare tumor patients with actionable biomarkers identified, stratified by tumor classification system. In total, 24.7% (10,518/42,569) of patients harbored at least one potential actionable biomarker. When analysis was limited to biomarkers supported by high evidence levels (Level 1-2), the proportion decreased to 8.7% (n = 3,713). The actionable rates based on clinical trial/biological evidence and off-label indications were 5.4% (n = 2,292) and 10.6% (n = 4,513) for, respectively.

Among tumor classification systems, Lung tumors (55.0%, 72/131) exhibited the highest actionable rate, followed by Reproductive system (54.6%, 179/328), Digestive system (39.9%, 1,677/4,203), Skin (35.5%, 323/909), and Urinary system (32.6%, 182/559) tumors. With the exception of NUT carcinoma, which only included one case, Thymus tumors (8.7%, 30/343) demonstrated the lowest actionable rate. Supplementary Table 4 provides a summary of the top mutated actionable genes and their corresponding possible therapeutic agents.

Additionally, several copy number alterations that may predict the expression of targetable proteins were identified. These alterations included: CD274 amplification (n = 50, predicting PD-L1 overexpression), KIT amplification (n = 24, predicting CD117 overexpression), ERBB2 amplification (n = 127, predicting HER2 overexpression), PTEN loss (n = 7, predicting PTEN deficiency), BAP1 loss (n = 38, predicting BAP1 deficiency), SMARCB1 loss (n = 23, predicting INI-1 deficiency), CCND1 amplification (n = 6, predicting Cyclin D1 overexpression), FGFR1 amplification (n = 1, predicting FGFR1 overexpression), and HGF amplification (n = 1, predicting HGF overexpression). It should be noted that these findings require validation through protein-based testing prior to informing treatment decisions.

Detailed case analyses revealed several complex situations that require special consideration. A total of 1,602 (3.8%) patients harbored two or more potential actionable biomarkers with the same evidence level. In such cases, it is important to consider combination therapy and prioritize treatment options. For example, two biliary tract cancer patients (Study ID: biliary_tract_msk_2026, Patient ID: P-0003289 & P-0011240) simultaneously detected BRCA1/2 oncogenic mutation and ERBB2 amplification, and one diffuse astrocytoma patient (Study ID: glioma_mskcc_2019, Patient ID: P-0019039) simultaneously detected BARF fusion and IDH1 R132H mutation. To better interpretate these complicated results, molecular tumor board is recommended to tailor the treatment decision based on comprehensive information, including but not limited to multi-omics molecular test results, medical history, personal/family history, physical condition, and drug availability.

In addition, 10,299 (24.2%) patients harbored alterations related to targeted therapy resistance. The most frequently observed resistance-associated alterations included TP53 (15.5%, n = 6,590, possibly resistant to ALK-TKI, Pemigatinib, and etc.), PTEN (3.0%, n = 1,267, possibly resistant to Adagrasib, Lorlatinib, and etc.), KRAS (2.6%, n = 1,114, possibly resistant to ALK-TKI, Cetuximab, and etc.), NF1 (2.0%, n = 854, possibly resistant to Adagrasib, HER2-TKI, and etc.), and MDM2 (1.5%, n = 648, possibly resistant to PD-(L)1 inhibitors). Notably, 59.5% (6,254/10,518) of patients with actionable biomarker(s) identified also exhibited co-occurring resistance alteration(s), which may limit the efficacy of targeted therapy. For example, an inflammatory myofibroblastic tumor patient (Study ID: sarcoma_mskcc_2022, Patient ID: P-0023268) had concurrent ALK fusion and TP53 oncogenic mutation, indicating that ALK-TKI may be resistant; and a spindle cell carcinoma of lung patient (Study ID: msk_chord_2024, Patient ID: P-0011125) had concurrent KRAS G12C and PTEN oncogenic mutation, indicating that Adagrasib may be resistant.