我们已经匹配来自两个独立队列的癌症患者的肿瘤 - 正常的WGS数据：Hartwig和PCAWG。补充注释1中描述了Harartwig和PCAWG队列收集和处理和处理与HARTWIG SOMATIC PIPLINE的PCAWG样本重新分析的全面文档的详细描述。

　　.VCF文件中的每个突变均由紫色给出亚克隆的可能性。遵循紫色指南，我们认为亚克隆分数等于或高于0.8的突变为亚克隆和低于0.8阈值的突变为克隆。对于每个样本，我们通过将克隆突变的数量除以总突变负担（包括SBS，多核苷酸变体和ID）来计算克隆突变的平均比例。最后，对于每种癌症类型，我们都使用Mann -Whitney检验来评估原发性肿瘤和转移性肿瘤之间克隆性差异的重要性。使用Benjamini – Hochberg程序调整了P值，以根据错误发现率（FDR）。调整后的p< 0.05 was deemed as significant.

　　In addition, we leveraged biopsy site data in patient reports to further investigate differences in metastatic tumour clonality according to the metastatic biopsy site (see also Supplementary Note 2). If the metastatic biopsy site was in the same organ or tissue as the primary tumour, we considered them as ‘local’, whereas if the metastatic biopsy site was reported in the lymphoid system or other organs or tissues, they were dubbed as ‘lymph’ and ‘distant’, respectively. Cancer types for which there was a minimum of five samples available for each of the biopsy groups were selected and Mann–Whitney test was used to compare the clonality between the biopsy groups.

　　Chromosome arm level and genome ploidy was estimated as previously described20.

　　First, for each chromosome arm, tumour purity and ploidy-adjusted copy number (CN) segments (as determined by PURPLE) were rounded to the nearest integer. Second, arm coverage of each integer CN was calculated as the fraction of chromosome arm bases with the specific CN divided by the chromosome arm length (for example, 60% of all chromosome 5p segments have a CN of 2, 30% have a CN of 1 and 10% have a CN of 3). We defined the arm-level ploidy level as the CN with the highest coverage across the whole arm (in the example above it would be 2). Third, we computed the most recurrent chromosome arm ploidy levels across all chromosome arms per sample (that is, observed genome ploidy).

　　Next, we estimated the true genome ploidy by taking WGD status (given by PURPLE) into account. If a sample did not undergo WGD, its total expected genome ploidy was deemed to be 2n. If a sample did undergo WGD and its observed genome ploidy was less than six, the estimated genome ploidy was deemed to be 4n, and 8n if the observed genome ploidy was six or more. An observed genome CN of more than eight was not found in our dataset.

　　Then, for each chromosome arm in each sample, we defined the normalized arm ploidy as the difference between the arm-level ploidy level and the expected genome ploidy. The resulting value was classified as 1 for differences higher than or equal to 1 (representing arm gains), as −1 for differences lower than or equal to −1 (representing arm losses) or as 0 (no difference). Normalized arm ploidy values were averaged across all samples from a cancer type in a cohort-specific manner (that is, separating primary and metastatic samples). A Mann–Whitney test was performed per cancer type and chromosome arm to assess the mean difference in arm gains or losses at the cancer-type level. The resulting P value was FDR adjusted across all arms per cancer type. Finally, q < 0.01 and a normalized arm ploidy difference higher than 0.25 was deemed to be significant.

　　To compare the differences in aneuploidy scores and the LOH proportions in each group, a Mann–Whitney test was performed per cancer type. The aneuploidy score represents the number of arms per tumour sample that deviate from the estimated genome ploidy as previously described20. The LOH score of a given sample represents the sum of all LOH regions divided by the GRCh37 total genome length. A genomic region is defined as LOH when the minor allele CN < 0.25 and major allele CN ≥ 0.8.

　　To compare the fraction of samples with a driver mutation in TP53 as well as the fraction of WGD samples per cohort, a Fisher’s exact test was performed per cancer type. Any TP53 driver alteration (non-synonymous mutation, biallelic deletion and homozygous disruption) was considered in the analysis. Multiple driver mutations per sample in a single gene were considered as one driver event. WGD was defined as present if the sample had more than 10 autosomes with an estimated chromosome CN of more than 1.5. P values were FDR corrected across all cancer types. A q < 0.01 was deemed to be significant for all statistical tests.

　　The number of somatic mutations falling into the 96 SBS, 78 DBS and 83 ID contexts (as described in the COSMIC catalogue51; https://cancer.sanger.ac.uk/signatures/) was determined using the R package mutSigExtractor (https://github.com/UMCUGenetics/mutSigExtractor, v1.23).

　　SigProfilerExtractor (v1.1.1) was then used (with default settings) to extract a maximum of 21 SBS, 8 DBS and 10 ID de novo mutational signatures. This was performed separately for each of the 20 tissue types that had at least 30 patients in the entire dataset (aggregating primary and metastatic samples; see Supplementary Table 3). Tissue types with less than 30 patients as well as patients with metastatic tumours with unknown primary location type were combined into an additional ‘Other’ group, resulting in a total of 21 tissue-type groups for signature extraction. To select the optimum rank (that is, the eventual number of signatures) for each tissue type and mutation type, we manually inspected the average stability and mean sample cosine similarity plot output by SigProfilerExtractor. This resulted in 440 de novo signature profiles extracted across the 21 tissue-type groups (Supplementary Table 3). Least squares fitting was then performed (using the fitToSignatures() function from mutSigExtractor) to determine the per-sample contributions to each tissue-type-specific de novo signature.

　　The extracted de novo mutational signatures with high cosine similarity (≥0.85) to any reference COSMIC mutational signatures with known cancer-type associations51 were labelled accordingly (288 de novo signatures matched to 57 COSMIC reference signatures).

　　For the remaining 152 unlabelled de novo signatures, we reasoned that there could be one or more signatures from one cancer type that is highly similar to those found in other tissue types, and that these probably represent the same underlying mutational process. We therefore performed clustering to group likely equivalent signatures. Specifically, the following steps were performed:

　　For certain de novo signature clusters, we could manually assign the potential aetiology based on their resemblance to signatures with known aetiology described in COSMIC51, Kucab et al.26 and Signal (access date 1 February 2023)52. Some clusters were an aggregate of two known signatures, such as SBS_denovo_clust_2, which was a combination of SBS2 and SBS13, both linked to APOBEC mutagenesis. Other clusters had characteristic peaks of known signatures, such as DBS_denovo_clust_4, which resembled DBS5 based on having distinctive CT>AA and CT>交流峰。最后，将DBS_DENOVO_CLUST_1注释为可疑的极点突变和MMRD，因为该簇的贡献很高（超过150个突变）的样品通常是微卫星（MSI）或极点突变的。同样，将DBS_DENOVO_CLUST_2注释为可疑的MMRD作为Aetiology，因为该群集的贡献很高（超过250个突变）的样品都是MSI。有关所有手动分配的Aetiologies的列表，请参见补充表3。

　　在应用病因分配后，从头提取产生了69个SB，13个DB和18个ID代表性突变特征（补充表3）。其中大多数（69个SBS中的42个，13个DBS中的7个和18个ID中的8个）映射到人类癌症中描述的突变特征上35,53。

　　然后，我们比较了原发性肿瘤和转移性肿瘤之间每个突变过程的活性（即，突变的数量）。对于每个样本，我们首先将相同突变类型（即SBS，DBS或ID）的特征的贡献概括为具有相同的病因，此后称为“病因学贡献”。根据癌症类型和鉴于病因，我们进行了两侧Mann – Whitney测试，以确定原发性和转移性肿瘤的病因学贡献是否存在显着差异。根据癌症类型和每种突变类型，我们使用p. adjust（）碱基函数来使用Holm的方法进行多次测试校正。接下来，我们在贡献中添加了1个假数（避免除以0），并计算了中值的贡献log2折叠变化，即log2（（转移性肿瘤中的中位数贡献 + 1）/（原发性肿瘤中的中位数贡献 + 1））。我们认为，当Q时，原发性肿瘤和转移性肿瘤之间的病因贡献显着差异< 0.05, and log2 fold change ≥ 0.4 or log2 fold change ≤ −0.4 (= ± ×1.4).

　　Relative aetiology contribution was calculated by dividing aetiology contribution by the total contribution of the respective mutation type (that is, SBS, DBS or ID). To determine the significant difference in relative aetiology contribution, we performed two-sided Mann–Whitney tests as described above. We also calculated the median difference in contribution (that is, median relative contribution in metastatic tumours − median relative contribution in primary tumours). We considered the relative aetiology contribution between primary and metastatic tumours to be significantly different when q < 0.05 and median difference was 0.01 or more.

　　We also determined whether there was an increase in the number of samples with high aetiology contribution (that is, hypermutators) in metastatic versus primary cohorts. For each signature, a sample was considered a hypermutator if the aetiology contribution was 10,000 or more for SBS signatures, 500 or more for DBS signatures or 1,000 or more for ID signatures. For each cancer type, for each aetiology, we performed pairwise testing only for cases in which there were five or more hypermutator samples for either metastatic or primary tumours. Each pairwise test involved calculating P values using two-sided Fisher’s exact tests, and effect sizes by multiplying Cramer’s V by the sign of the log2(odds ratio) to calculate a signed Cramer’s V value that ranges from −1 to +1 (indicating enrichment in primary or metastatic, respectively). We then used the p.adjust() base R function to perform multiple testing correction using Bonferroni’s method.

　　To count the SBS1 mutations, we relied on the definition from ref. 54 that is based on the characteristic peaks of the COSMIC SBS1 signature profile: single-base CpG > TpG mutations in NpCpG context. To ensure that these counts and the downstream analyses are not affected by differential APOBEC exposure in primary and metastatic cohorts, we excluded CpG > TPG在TPCPG中，这也是宇宙SBS2签名轮廓中的特征峰。此外，对于皮肤黑色素瘤，排除了与SBS7A重叠的[C/T] PCPG中的CpG> TPG。为了获得SBS5和SBS40计数，我们依靠它们从本研究中执行的突变签名分析得出的暴露（如上所述）。

　　为了评估SBS1负担与患者年龄之间的相关性，在活检时，我们进行了癌症类型和队列特异性线性回归（即对原发性和转移性肿瘤样本的单独回归）。为了避免由高温肿瘤引起的虚假作用，排除了TMB大于30,000的样品以及SBS1负担大于5,000的样品。

　　对于每种癌症类型和队列，我们​​通过随机选择75％的可用样品来计算100个独立的线性回归。我们选择了中位线性回归（基于回归斜率）作为代表性回归进行进一步分析。同样，置信区间是从计算回归的第一和第99个百分点中得出的。

　　要评估原发性和转移性代表性线性回归之间差异的重要性（以下称为简单性的线性回归），我们首先过滤了未能显示出未能显示出SBS1负担和年龄在原发性和转移性肿瘤中活检中的正相关趋势的癌症类型（Pearson的折叠率和metakeft of Metseft of Primental and Metsatigation and Metsatigation Aspressication Aspression and 0.1）。接下来，对于每种选定的癌症类型，我们计算了原代和转移性SBS1突变计数的回归残差，在这两种情况下都是基本线性回归。然后，使用Mann -Whitney测试比较了主要和转移性残留分布，以评估显着性。Mann -Whitney P类型< 0.01 were deemed as significant. Finally, to ensure that the differences were uniform across different age ranges (that is, not driven by a small subset of patients), we only considered significant cancer types in which the metastatic linear regression intercept was higher than the primary intercept.

　　SBS5/SBS40 correlations were computed following the same procedure and using the sum of SBS5 and SBS40 exposures for each tumour sample. If none of the mutations were attributed to SBS5/SBS40 mutational signatures, the aggregated value was set to zero. In the ploidy-corrected analyses, we divided the SBS1 mutation counts (and SBS5/SBS40 mutation counts for the SBS5/SBS40 ploidy-corrected regression, respectively) by the PURPLE-estimated tumour genome ploidy.

　　For each cancer type, the mean fold change (fc) was defined as where MPredi and PPredi are the estimated number of SBS1 mutations for a given age ith according to the metastatic and primary linear regressions, respectively. Similarly, the mean estimated SBS1 burden difference (SBS1diff) was defined as: .

　　SBS1 individual mutations were identified as described in the previous section. For SBS5 and SBS40 mutations, we used a maximum likelihood approach to assign individual mutations to the SBS5 and SBS40 mutational signatures in a cancer-type-specific manner.

　　For every SBS1 (and SBS5/SBS40 mutation), we then assign the clonality according to the PURPLE subclonal likelihood estimation, in which only mutations with subclonal (SUBCL) likelihood ≥ 0.8 were considered as such (see above).

　　For each tumour sample, the SBS1 clonality ratio (or respectively SBS5/SBS40 clonality ratio) was defined as the ratio between the proportion of clonal SBS1 mutations () divided by the total proportion of clonal alterations in the sample ().

　　We computed for each primary cancer type the average number of SBS1 per year as the number of SBS1 mutations divided by the age of the patient at biopsy (only considering primary samples and excluding hypermutated samples as described above). We then used a Spearman’s correlation to assess its association with the estimated mean SBS1 mutation rate fold change in metastatic tumours (see above). In addition, to exclude potential biases in our primary cohort, we repeated the same analysis relying on an independent measurement of primary cancer SBS1 yearly accumulation. Specifically, we used the best-estimated accumulation of SBS1 per year from ref. 30 (Supplementary Table 6) and regressed it to the fold-change estimates for the matching cancer types present in both datasets.

　　LINX55 chains one or more SVs and classifies these SV clusters into various event types (‘ResolvedType’). We defined deletions and duplications as clusters with a ResolvedType of ‘DEL’ or ‘DUP’ whose start and end breakpoints are on the same chromosome (that is, intrachromosomal). Deletions and duplications were split into those less than 10 kb and 10 kb or more in length (small and large, respectively), based on observing bimodal distributions in these lengths across cancer types (Extended Data Fig. 5b). We defined complex SVs as clusters with a ‘COMPLEX’ ResolvedType, an inversion ResolvedType (including: INV, FB_INV_PAIR, RECIP_INV, RECIP_INV_DEL_DUP and RECIP_INV_DUPS) or a translocation ResolvedType (including: RECIP_TRANS, RECIP_TRANS_DEL_DUP, RECIP_TRANS_DUPS, UNBAL_TRANS and UNBAL_TRANS_TI). Complex SVs were split into those with less than 20 and 20 or more SVs (small and large, respectively), based on observing similar unimodal distributions in the number SVs across cancer types whose tail begins at approximately 20 breakpoints (Extended Data Fig. 5b). Finally, we defined long interspersed nuclear element insertions (LINEs) as clusters with a ResolvedType of ‘LINE’. For each sample, we counted the occurrence (that is, SV burden) of each of the seven SV types described above. In addition, we determined the total SV burden by summing counts of the SV types.

　　We then compared the SV-type burden between primary versus metastatic tumours as shown in Fig. 3a. First, we performed two-sided Mann–Whitney tests per SV type and per cancer type to determine whether there was a statistically significant difference in SV-type burden between primary versus metastatic. The Bonferroni method was used for multiple testing correction on the P values from the Mann–Whitney tests (to obtain q values). Next, we calculated relative enrichment as follows: log10(median SV-type burden in metastatic tumours + 1) − log10(median SV-type burden in primary tumours + 1); and calculated fold change as follows: (median SV-type burden in metastatic tumours + 1) / (median SV-type burden in primary tumours + 1). When calculating relative enrichment and fold change, the pseudocount of 1 was added to avoid the log(0) and divide by zero errors, respectively. Fold changes are displayed with a ‘>’在图3a中，当原发性肿瘤的SV负担为0时（即，如果没有伪金，则会出现鸿沟0）。我们认为，当Q <0.05和折叠变化≥1.2或折叠变化≤0.8时

　　为了确定可以解释SV负担增加的功能，我们将SV负担与各种肿瘤基因组特征相关联。其中包括：（1）基因组倍（由紫色确定）；（2）同源重组缺陷（由Chord33确定）和MSI（由紫色）状态；（3）在345个与癌症相关的基因（不包括经常受CN蚀变影响的脆弱的位点基因5）中存在突变5），此后称为“基因状态”；（4）治疗史，包括存在放射疗法，存在79种不同的癌症疗法之一以及接受的治疗总数。所有主要样品和所有没有治疗信息的转移样品都没有治疗。基因组倍性和收到的治疗总数是数字特征，而其余的则是布尔（即是真或错误）特征。总共有429个功能。

　　SV型负担转换为Log10（SV型负担+1），并使用多元线性回归模型（LMS）与429个功能相关。这是针对七种SV类型的每一种以及每种癌症类型（或亚型）分别执行的。在SV主要分析（图4B – F）中，有23种癌症类型，总共161种（23种癌症类型×7 SV类型）LM模型。

　　每个LM模型（即每SV类型和癌症类型）都涉及对（1）转移和主要样品（主要 +转移性）的三个独立LMS培训，（2）仅HARTWIG样品（仅转移），以及（3）仅PCAWG样品（仅主要）。这样做是为了滤除特征之间的相关性和仅由于原发性和转移性肿瘤之间特征值的差异而增加的SV型负担。然后，我们需要与初级 +转移性LM中SV型负担正相关的特征，以独立显示仅转移或仅一级LMS中的相同关联。只有独立表现出与SV负担正相关的基因组特征被进一步视为重要（即在Lollipop图中表示）。

　　三个LMS中的每一个都经过如下培训：

　　对于每个LM分析，我们使用以下过滤标准来识别与SV型负担增加相关的功能：

　　最后，为了确定与原发性肿瘤（反之亦然）相比，在转移性肿瘤中富集了哪些特征（与增加的SV型负担相关），我们计算了Cliff Delta的数字特征和Boolean特征的Cramer的V。悬崖的三角洲的范围为-1至+1，其中-1代表原发性肿瘤的完全富集，而+1代表转移性肿瘤的完全富集。Cramer的V仅在0到1（1代表原发性肿瘤或转移性肿瘤中的富集）范围内，对数（优​​势比）的符号被分配为Cramer V值的符号，因此其范围为-1至+1。与原发性癌症相比，效果大小超过0的特征被认为可以解释转移性癌症的SV负担增加。

　　我们依靠紫色和linx55构建的患者特异性癌症驱动器和融合目录。仅保留驾驶员可能性超过0.5的驾驶员。对先前在文献中报道的人过滤融合驱动器。同样，由于低易近度区域中的高负数突变，我们手动策划了驱动程序列表并删除了SMAD3热点突变。最终的驱动程序目录总共包含453个驱动基因，最终融合目录中包含554个融合。

　　为了比较原发性和转移性肿瘤中驱动因素的数量，然后我们将融合与LINX驱动器变体相结合，以计算特定于患者的驾驶员事件。涉及相同驱动基因但驱动器类型的驱动因素被认为是单个驱动程序（例如，在同一样本中，TP53突变和TP53删除被视为一个驱动程序事件）。进行了癌症特异性的Mann-Whitney测试，以评估原发性肿瘤和转移性肿瘤之间的差异。调整后的Q <0.01被认为是显着的。

　　为了评估驾驶员的富集，从驱动程序目录中构建了一个应变矩阵，其中包含每个驱动器类型（即删除，放大或突变）和每个队列中的癌症类型的驱动器突变频率（转移性和主要）。为融合构建了第二个应变矩阵。部分扩增被认为是扩增的，而同源的破坏被认为是缺失。对这些应急矩阵进行过滤的基因过滤，这些基因显示了主要或转移队列中五个突变样品的最小频率。然后，对每个基因，癌症类型和突变类型进行了双面Fisher的精确测试，并针对每种癌症类型的FDR调整了P值。Cramer的V和优势比用作效应尺寸度量。调整后的P <0.01被认为是显着的。

　　为了确定每个样本中观察到的可操作变体量，我们将SNPEFF（v5.1）56注释的变体与从三个不同数据库（Oncokb57，civic58和cgi59）衍生的变体进行了比较，这些变异是根据常见的临床证据级别（https://civic.redcivic.readtecs.iococs.ioc.ioce.ioces.io/nevelle and andle/nmodef）/late/latlate/lattefter/latemnlem andlatek/modlate/m，先前描述的5。在我们的研究中，我们仅考虑了A和B水平的证据，这些证据代表了已批准用于治疗的变体，目前正在分别在后期临床试验中进行评估。当癌症类型与批准或正在研究的癌症类型相匹配的癌症类型时，确定一种变体是“标签上的”，否则“标签”。仅考虑敏感类别的可起作用变体（即包含变体的肿瘤对某种治疗敏感）。由于它们倾向于掩盖其他变体，尤其是在标签外类别中，因此未评估样品级别可起作用的变体，例如TMB高/低或MSI状态。此外，出于相同的原因，在此分析中未考虑野生型可起到的变体。由于缺乏可用数据，因此未考虑与基因表达或甲基化有关的变体。此外，我们发现源自白血病的可起作变体与数据集中的实体瘤有很大不同，这就是为什么我们将它们排除在内。为了分析带有治疗性可行变体的样品比例，我们认为，按照a ON/off-Blabel的订单为b/off标签的订单，每个样本都保留了每个样本的最高证据水平。为了评估全球和在转移性肿瘤中A上标签水平的可行变体的富集， Fisher的精确测试在泛滥的范围内和每种癌症类型进行。调整后的P <0.05被认为是显着的。频率的折叠变化仅针对具有全球显着差异的癌症类型显示。

　　为了确定哪些变体对观察到的显着频率差异有最大的贡献，使用Fisher的每种癌症类型和等级水平的Fisher精确测试测试了单个可起作用的变体在转移性肿瘤中的富集。P值是根据癌症调整FDR的，Q <0.05被认为是显着的。在扩展数据中，只有来自具有全球显着差异的癌症类型的可操作变体（请参见上文），并且在初级或转移性队列中以最小频率为5％，并且显示了它们之间的最小频率差异5％。但是，作为补充表7的一部分，所有筛选变体的差异都可以使用。

　　我们的目的是指出可能导致转移队列中某些癌症治疗的反应可能导致的驱动因素。因此，我们设计了一项测试，该测试鉴定了与来自相同癌症类型的未经治疗的患者相比，用特定治疗类型治疗的患者富含的驱动因素改变（有关工作流程的说明）。

　　根据其作用机理对处理进行分组，以便将具有共同作用机理的多种药物分组为机械治疗类别（例如，顺铂，奥沙利铂和卡泊肽分组为铂）。我们通过在活检前根据治疗记录对患者进行治疗注释的患者来创建323个治疗和癌症类别。如果一名患者接受多种治疗或同时组合多种药物，则可能参与多组。在分析中，进一步考虑了至少10例患者的92例和癌症类别。

　　因此，对于每种癌症类型（或亚型，在乳房和结直肠癌的情况下）和治疗组，我们执行了以下步骤：

　　TEDS的完整目录及其突变频率可以在补充表8中找到。

　　我们使用带有默认参数的DNDSCV（V0.0.1）60来鉴定编码突变中的癌症驱动基因。全局Q <0.1用作显着性的阈值。从DNDSCV输出中提取每个驱动基因的突变频率。我们将突变频率定义为具有非同义突变的样品数量。

　　我们使用ActivedRiverWGS61（v1.1.2，默认参数）在基因组的五个调节区域中识别非编码驱动器元素，包括3'未翻译区域（UTRS），5'UTRS，5'UTRS，长的非编码RNA，近距离RNA，近端启动子和夹具站点。对于每个元素类别，我们从Ensembl V101提取了基因组坐标。每个监管区域都进行了独立测试。为了选择明显的命中，我们对调整后的P值（FDR <0.1）和最少三个突变样品过滤了。我们将突变频率定义为治疗组中每个显着突变元件的突变样品数量。

　　我们使用以下设置运行了92个治疗和癌症组中的每一个（参考文献62）（v2.0.23）：

　　然后使用Ensembl V101的坐标用功能元件对焦点峰（Q≤0.1和<1 mb）进行注释。如上所述，使用Fisher的精确测试评估了每个基因上处理和未经处理的队列之间的频率差异。为此，我们首先计算了每个样品中每个基因的焦点扩增和深层耗竭状态。当基因的倍性水平比其全基因组平均倍性水平高（通过紫色测量）高2.5时，将一个基因放大，并且当基因倍虫水平低于0.3时被删除（即深层缺失）。我们观察到，大多数峰包含多个重要基因候选物（在多个校正Q <0.05之后），因此我们保留了最紧密地位于峰峰的基因，该基因是跨处理样品中最显着富集的区域。接下来，我们还发现，未经处理的对照组中，每种癌症类型的多个治疗组之间的复发峰不存在，因为大多数Hartwig样品都接受了多种治疗类型。因此，我们将峰与重叠的范围合并，每种癌症类型的每个基因组区域都会产生一个峰值。对于每个倒塌的峰，我们选择了显示峰峰值附近基因的最低Q值的处理类型。分别处理缺失和扩增峰。

　　为了估计TEDS对转移队列中每个样品总驱动程序总数的贡献，我们在癌症特异性，基因特异性和驱动器特异性的方式中排除了从驱动器突变目录（请参阅上面的“驱动器变化”）中的任何TED。

　　有关研究设计的更多信息可在与本文有关的自然投资组合报告摘要中获得。