We acquired the tumor and matched normal tissues from 132 lung cancer patients at the Samsung Medical Center in Seoul. This study was approved by Samsung Medical Center institutional review board (IRB 2018-03-110), and informed consent was obtained from each patient. Tumor tissues were transplanted into the NSG mouse, NOD.Cg-Prkdc^scidIl2rg^tm1Wjl/SzJ (Stock No. 005557) [11], purchased from the Jackson Laboratory to establish the PDX mouse models. All tumor tissues of the patient and PDX were prepared for formalin-fixed paraffin-embedded, and a pathologist examined histopathology of tissue through hematoxylin and eosin staining. Whole exome sequencing (WES) was performed using the Illumina TruSeq Exome kit and HiSeq 2500 sequencing platform (Illumina, San Diego, CA, USA). The WES data for patient tumor, patient normal, and PDX tumor tissues were deposited at the Korean Bioinformation Center (KOBIC) (ID 10050154).

Preprocessing steps consist of adapter trimming, quality control, and filtering mouse reads. First, adapter sequences were removed and sequence reads whose quality score < 33 in more than 50% of bases were discarded using fastx-toolkit (ver. 0.0.14). After the trimming process, we removed the single reads and kept the paired-end reads only using cmpfastq perl program (http://compbio.brc.iop.kcl.ac.uk/software/cmpfastq.php). For WES data from PDX mouse tissues, we applied an additional step to filter out mouse-originated reads using Xenome software (ver. 1.0.1) [12] with the reference genomes of human (UCSC hg19 in https://genome.ucsc.edu/) and mouse (UCSC mm10). We kept the human-specific reads only for subsequent analyses.

Resulting reads were mapped to the human reference genome (UCSC hg19) using Burrows-Wheeler Alignment (BWA)-MEM alignment tool [13] with default options. After mapping, reads were sorted by samtools version 1.8 [14]. We performed Genome Analysis Toolkit (GATK, v4.0.7.0) [15] AddOrReplaceReadGroups command for adding read group information, MarkDuplicates command for removing duplicated reads, BaseRecalibrator and ApplyBQSR commands for correcting realignment. Data processing summary statistics are given in Supplementary Table 1. Then, somatic single nucleotide variations (SNVs) and insertions/deletions were called using the GATK4-Mutect2 [16] pipeline. Filter-based annotation in ANNOVAR [17] was used for variant annotations. In addition, we calculated the copy number alterations using EXCAVATOR2 (v1.1.2) [18]. All statistical analysis and visualizations were performed using R version 3.6.1

We analyzed 132 lung cancer patients whose tumor and matched normal tissues were available and the PDX tumor samples were successfully harvested. The pathophysiological information of patients is summarized in Table 1. Briefly, we had 54 adenocarcinoma cases (41%), 48 squamous cell carcinoma cases (36%), four large cell carcinoma cases (3%), and 26 unclassifiable adenocarcinoma cases (20%). Our patient cohort was enriched with male (66%), smokers (64%), early stages (50%), non-recurrent (61%), and primary (62%) patients. In accordance with the previous reports [19], the success rate of establishing PDX mouse models was higher in squamous cell carcinoma than in adenocarcinoma. Histopathological examination of patient tumor and PDX tumor tissues concluded that tumors were consistent between patient and PDX mouse in 97 cases (73.5%). The discrepancy in histology maybe presumably ascribed to the lymphomagenesis that had been reported to occur frequently in NSG or NOG mice transplanted with Epstein-Barr virus infected tumor tissues [20,21]. This concordance and discrepancy were further investigated by comparing the mutation and copy number profiles between patient and PDX mouse tumors.

We performed WES on tumor, normal, and PDX samples of 132 lung cancer patients, with a mean coverage of 30×. In order to compare the somatic mutation profiles of patient and PDX mouse tumors, it is essential to test the reliability of mutation calls in the PDX mouse tumors because human stromal cells are replaced with the mouse stromal cells during engraftment. Thus, we checked the portion of mouse-originated reads from the WES data of PDX tumors using the Xenome software to separate the human and mouse reads specifically. A summary of the Xenome alignment results is provided in Supplementary Table 2. The median portion of human reads was 95.1% and that of mouse reads was below 5% except a few cases (Fig. 1A). This implied that the PDX mouse tumors contained a sufficient amount of human cells. Thus we used the well-known MuTect2 algorithm after BWA-MEM mapping to call somatic mutations in PDX mouse tumors.

The mutation rates were significantly higher in PDX mouse tumors than in patient tumors (p = 7.36e-06). The median values of exonic mutations in PDX and patient tumors were 136 and 102, respectively (Fig. 1B). Since we removed the mouse-originated reads before calling somatic mutations, the difference can be attributed to clonal selection and evolutionary processes in establishing PDX mouse tumors. For example, human stromal cells would be replaced with the mouse stromal cells during engraftment of the patient's tumor tissue into immunodeficient mice [22]. As a result, tumor cells are enriched in the PDX mouse tumors, which leads to more somatic mutations in variant calling. To test this hypothesis, we examined the variant allele frequencies (VAFs) of major cancer genes (TP53, KRAS, and PIK3CA) identified in both patient and PDX tumors (Fig. 1C). All three genes showed that the VAFs of PDX mouse tumors are larger than those of patient tumors, which supports our hypothesis of clonal enrichment in PDX mouse tumors. However, the extent of clonal selection pressure varied for different genes. Interestingly, the VAFs of TP53 gene were close to 100% in PDX mouse tumors. The biological meaning of this observation warrants further studies.

Next, we examined how well the mutations identified in patients were reproduced in PDXs. In all 132 samples, it was found that 63% of the exonic mutations in patients were also identified in PDXs on average (Fig. 2A). The portion of common mutations, however, varied tremendously from 0% to 98%. Low rates of common mutations were mostly observed in cases where patient and PDX tumors showed different pathology in histological analysis. Considering 92 cases with consistent histopathological result, 78% of the exonic mutations in patients were also identified in PDXs on average.

We have also examined the overlap of functional mutations, including non-synonymous SNVs, stop-gain mutations, stop-loss mutations, frameshift insertions, and frameshift deletions, between patient and PDX mouse tumors using 517 cancer-related genes curated from OncoKB and other literatures (Fig. 2B) [23-25]. The most commonly mutated gene was TP53. For cases with consistent histology, major portion of driver gene mutations were observed commonly in both patient and PDX tumors. Concordant driver genes included many mutations targeted by cancer drugs, including PIK3CA, EGFR, BRAF, and ALK. Thus, our bank of PDX mouse models can be a useful resource for testing drug efficacy in a preclinical setup.

Then we investigated how the histological analysis results are associated with the rate of common mutations. Out of 119 cases with the histological result available, we had 97 consistent and 22 inconsistent cases (Fig. 2B). The common mutation rates were over 20% in most consistent cases, whereas they were below 3% in most inconsistent cases. Thus, we were able to find 10%–20% of common mutation rate as a general threshold for determining whether the PDX mutations recapitulate the patient’s somatic mutations faithfully (Fig. 2A).

However, there existed several exceptional cases from this guideline. We found only one case where the mutation profiles were concordant, but the histological examination result was different tumors in patient and PDX (Fig. 2B). In contrast, we found nine cases in the opposite direction where the histological examination concluded concordant tumors, but the mutation profiles were vastly different between patient and PDX (Fig. 2B). Reasons for this discrepancy are not clear, but it is difficult to imagine that no common mutation was found if the patient and PDX tumors were truly of the same histology. Histological examination is not a perfect, but error-prone procedure. In conclusion, it is not necessary to compare the mutation profiles if the histological examination result is inconsistent, but when the histological examination gave a good result, comparing mutation profiles can be useful in determining if the PDX tumor is truly identical to the patient tumor. Thus, the mutation profile plays complementary roles to histological examination.

The main advantage of PDX tumor model is that it is possible to amplify the amount of tumor cells by engrafting tumor tissue into other immunodeficient mice. Molecular characteristics are usually expected to be preserved in the passage process, but detailed examination at the genome scale is quite rare.

We examined the mutation and copy number profiles for six cases where PDX tumor samples were available over several passages/generations (Fig. 3). We had four histologically consistent cases and two inconsistent cases. Both the mutation and copy number profiles were well reproduced throughout many generations in four good cases. Interestingly, the copy numbers showed much larger changes than somatic mutations especially between patient tumor and the first generation of xenograft tumor (Fig. 3B). It seems that PDX tumors harbor a number of new copy number losses that were maintained over many passages, which again implied that the clonal selection occurred in establishing the PDX mouse tumors.