2.1 Clinical Cohort

The study samples were collected from the patients with NSCLC who received antitumor therapy at Beijing Chest Hospital, Capital Medical University from January 2015 to December 2021. For the primary LUAD and transformed SCLC cohorts, the inclusion criteria were as follows: (1) initial histological or cytological diagnosis of LUAD; (2) EGFR mutation was confirmed by Amps qPCR or NGS testing of the biopsy tissue or blood; (3) have received prior EGFR-TKIs targeted therapy; (4) after drug resistance, a second biopsy showing SCLC confirmed by pathological test; (5) complete clinical information such as medical history, imaging examination, TNM staging and therapeutic process and efficacy evaluation; and (6) paraffin-embedded sections with biopsy tissue or embedded pleural effusion containing more than 90% tumor, with sufficient tissue quantity for library construction and genome sequencing. A total of 5 primary LUAD tissues before SCLC transformation and 12 transformed SCLC tissues after resistance to EGFR-TKIs treatment were obtained for subsequent whole exome sequencing analysis. For the de novo SCLC cohort, the inclusion criteria were as follows: (1) initial histological or cytological diagnosis of SCLC; (2) no prior treatment had been performed; (3) complete clinical information such as medical history, imaging examination and stage; and (4) paraffin-embedded sections with surgical tissue containing more than 90% tumor and sufficient tissue volume for library construction and genome sequencing. We randomly selected 18 patients with de novo SCLC for subsequent whole exome sequencing analysis. The clinical features of all patients are detailed in Table S1. This retrospective study was approved and informed written consent was waived by the Ethics Committee at Beijing Chest Hospital, Capital Medical University. The research was conducted in compliance with the Declaration of Helsinki.

2.2 Sample Collection and DNA Extraction

All tumor samples were obtained from paraffin-embedded sections of surgical, biopsy or embedded pleural effusion specimens. Hematoxylin and eosin (H&E) staining, along with histological microscopy, were conducted on these sections. Expert pathologists identified the tissue type and assessed the tumor purity. Genomic DNA was extracted from formalin-fixed paraffin-embedded (FFPE) tumor samples using the QIAamp DNA FFPE Tissue Kit (Qiagen GmbH, Hilden, Germany), following the manufacturer's instructions. The DNA concentration was determined using the Qubit dsDNA HS (High Sensitivity) assay kit with a Qubit fluorometer (Invitrogen; Thermo Fisher Scientific Inc., Waltham, MA, USA). To evaluate DNA integrity, 200 ng of the extracted DNA was loaded onto a 1% agarose gel alongside a DNA marker (l-Hind III) (Takara Biotechnology Co. Ltd., Dalian, China). Samples were deemed intact if they exceeded the second-largest bands (9416 bp) of the l-Hind III marker and were then used for further analysis. All genomic DNA was extracted from the tissue samples, and quality control was performed before sequencing. Two cases of transformed SCLC tissue did not pass quality control and were excluded from sequencing, as noted in Table S2.

2.3 Whole Exome Sequencing

For each sample, 1.0 μg of genomic DNA was used as the starting material for library preparation. Sequencing libraries were constructed using the Agilent SureSelect Human All Exon V6 kit (Agilent Technologies, San Diego, CA) according to the manufacturer's guidelines, and index codes were assigned to each sample. Briefly, DNA fragmentation was performed using a hydrodynamic shearing system (Covaris, Massachusetts, USA) to produce fragments ranging from 180 to 280 bp. The resulting overhangs were converted to blunt ends through exonuclease and polymerase activities, after which the enzymes were removed. Following the adenylation of the 3′ ends of the fragments, adapter oligonucleotides were ligated. Fragments with adapters on both ends were then selectively amplified via PCR. After amplification, the library was hybridized in a liquid phase with biotin-labeled probes, and streptavidin-coated magnetic beads were utilized to capture the exons of the target genes. The captured libraries underwent additional PCR enrichment to add index tags for hybridization preparation. Final products were purified using the AMPure XP system (Beckman Coulter, Beverly, USA) and quantified using the Agilent high sensitivity DNA assay on the Agilent Bioanalyzer 2100 system. The qualified exome capture libraries were then sequenced on the Illumina NovaSeq 6000 platform, following standard protocols for 150 bp paired-end multiplexed sequencing.

2.4 Processing of WES Sequencing Data

After filtering out low-quality sequencing reads and adapter sequences using FASTP, clean reads were aligned to the human reference genome (UCSC hg19) with the Burrows-Wheeler Aligner (bwa-0.7.17). The aligned genomes were sorted using Sambamba (v0.6.7), and duplicate reads were identified with Picard tools (v2.18.9). Somatic single nucleotide variants (SNVs) and insertions/deletions (INDELs) were identified using VarScan2 and MuTect2 in tandem. Specifically, VarScan2 somatic (v2.3) was employed for variant calling with standard parameters, except for adjustments: minimum coverage was set to 10 for normal samples and 8 for tumor samples, the minimum variant frequency was set to 0.01, and tumor purity was established at 0.5. For the MuTect2 analysis, we utilized the version included in the GATK bundle (4.0.5.1) with default settings. Functional annotation of the variants was performed using ANNOVAR tools.

2.5 SNV and INDEL Filtering

To minimize false positive variant calls, additional filtering methods were applied to the mutation detection results from both MuTect2 and VarScan2. A variant was considered valid if it was identified by both tools, with a somatic p value ≤ 0.1 for SNVs and ≤ 0.05 for INDELs, and a variant allele frequency (VAF) > 2%. Alternatively, a variant detected solely by VarScan2 needed a VAF exceeding 5%. For matched tumor sequencing data, the VAF should be less than 1%, and the number of reads for alternative alleles must be fewer than 5 for SNVs and 2 for INDELs. Additionally, variants found in regions of simple repeats and segmental duplications were excluded. The population frequency of the SNV was required to be below 1% in any of the following databases: 1000 Genomes, EXAC, or ESP6500. For tumors without paired blood samples, a more stringent criteria was used to remove any possible germline mutations detected on SCLC or NSCLC driver genes. First, a panel of 17 normals was used to build a PON to filter somatic variant calls from artifacts and germline leakage. Then, variants with population allelic frequency greater than 0.001 in 1000 genomes or gnomAD databases were removed. Moreover, only pathogenic coding mutations, which were evaluated using SIFT or Polyphen as damaged, and truncated mutations were reserved for further analysis.

2.6 CNV Analysis

We employed ASCAT to assess somatic copy number alterations (SCNA) in paired tumor-normal sequencing data. The tumor purity and ploidy estimates were further refined through manual validation of the ABSOLUTE results. Allele counts at specific positions were derived from the 1000 Genomes Project using AlleleCounter, with a minimum coverage of 20 for the normal samples applied during filtering. LogR (the depth ratio of tumor to normal) values underwent GC wave correction via ASCAT, retaining only heterozygous B-allele frequency (BAF) values for subsequent analysis. Allele-specific segmentation was conducted to produce segmented logR and BAF data using ascat.aspcf. Manual checks were performed to determine the optimal model for ploidy and cellularity based on ABSOLUTE findings. For tumors sequenced without matched normal samples, CNVkit was utilized to identify copy number variations (CNV) with default parameters, excluding a pooled reference created from multiple FFPE normal samples. Allele-specific copy number was inferred from GATK variant calling results with threshold parameters set at −1.1, −0.4, 0.3, and 0.7. Significant somatic CNVs were identified using GISTIC2.0 based on outputs from ASCAT or CNVkit, with significant copy number alteration (CNA) regions defined as having amplifications or deletions with a residual q value of < 0.05.

2.7 HRD Score Calculation

Homologous recombination deficiency (HRD) scores are calculated using the scarHRD R package. The HRD score, derived from allele-specific copy number data, combines the scores for loss of heterozygosity (LOH), telomeric allelic imbalance (TAI), and large-scale transitions (LST). The HRD-LOH score reflects the count of 15 Mb regions exhibiting LOH that do not encompass entire chromosomes. The HRD-TAI score indicates allelic imbalances that reach the telomeric ends of chromosomes. Lastly, HRD-LST is characterized by chromosomal breaks occurring between adjacent regions of at least 10 Mb, with a maximum distance of 3 Mb separating them.

2.8 Analysis of Mutational Signatures

The mutational signatures were assessed using the R package deconstructSigs (v1.8.0). A multiple linear regression model was employed to evaluate the activity levels of predefined signatures. The contribution of each mutational signature for individual patients was compared against the COSMIC SBS signature V2 (https://cancer.sanger.ac.uk/cosmic/signatures_v2.tt).

2.9 Statistical Analysis

To compare continuous values across different groups, the Fisher exact test was utilized. Figures were generated using either the ggplot2 or ComplexHeatmap packages. All statistical analyses were conducted in the R environment (v4.2.1), with statistical significance set at a two-sided p value of < 0.05.

3.1 Clinical Characteristics of Patients

According to the process of SCLC transformation, we collected 5 adenocarcinoma tissue samples from EGFR-mutated LUAD patients before SCLC transformation, 12 transformed SCLC tissue samples after EGFR-TKI treatment, and 18 de novo SCLC tissue samples for whole exome sequencing analysis. All tested DNA samples were sourced from FFPE tissue sections from patient's surgical, biopsy tissues or embedded pleural effusion. The clinical characteristics are presented in Table S1. It can be observed that transformed SCLCs are predominantly found in females (2/12, 83.3%) and non-smokers (2/12, 83.3%), whereas de novo SCLCs are predominantly found in males and smokers. This is consistent with the epidemiological characteristics of LUADs and SCLCs. Among the transformed SCLC cases, four cases were at the limited stage and eight cases were at the extensive stage. After an average treatment duration of 17.8 months, SCLC components were detected during patient biopsies when resistance to EGFR-TKIs (Table S2). The pathological diagnosis were made by professional pathologists. As shown in Figure S1, transformed SCLC tissue shares the same morphology with de novo SCLCs, the tumor cells are small, round oval or spindle-shaped with scant cytoplasm and ill-defined cytoplasmic borders. Immunohistochemical staining for the typical SCLC marker CD56, Syn, Napsin A were positive. Among them, six cases were identified as pure SCLC, whereas the other six cases were identified as mixed type, containing both adenocarcinoma and SCLC components.

Among the 12 patients who underwent SCLC transformation, the EGFR mutation type was predominantly exon 19 deletion (11/12, 91.7%), with one cases exhibiting simultaneous T790M mutations and only one case of L858R mutation at their adenocarcinoma stage. Unfortunately, for the transformed SCLCs tissues, we could not obtain the paired primary LUAD tissues due to the limited amount or substandard quality.

3.2 Mutational Landscape of Transformed SCLCs With Primary LUAD and De Novo SCLC

Genomic DNA was extracted from all tissue samples, and quality control was performed prior to sequencing. Two cases of transformed SCLC tissue failed quality control and were not sequenced, as shown in Table S2. The average sequencing coverage is 99.9%, with an average sequencing depth of 158.89× and an average clean base yield of 30G.

Focusing on the somatic mutational landscape of primary LUADs, transformed SCLCs and de novo SCLCs, we found the mutational pattern of transformed SCLCs was more closely aligned with that of primary LUADs than with de novo SCLCs, especially for EGFR mutations (Figure 1A). In both primary LUADs and transformed SCLCs, all patients harbored at least one mutation in EGFR, whereas in de novo SCLCs, only 3 out of 18 (17%) patients harbored EGFR mutation. In the evolutionary process, transformed SCLCs reserved EGFR mutation originating from their own primary LUADs, which is a key distinguishing feature compared to de novo SCLCs. Additionally, the prevalence of TP53 mutations showing a trend of increase among these three cohorts (Figure 1A), with 60%, 70% and 89% in primary LUADs, transformed SCLCs and de novo SCLCs, respectively. For another tumor suppressor gene, RB1, the prevalence didn't display obvious differences (Figure 1A), with 40%, 30%, and 50% of primary LUADs, transformed SCLCs and de novo SCLCs. Intriguingly, mutations in the genes COL22A1 and ALMS1 occurred only in transformed SCLCs and de novo SCLCs (Figure 1A), indicating their common roles in the originating of these two types of SCLCs.

To further confirm this observation, we also compared mutational landscape between our transformed SCLC cohort and three other published de novo SCLC cohorts (Figure S2). The similar phenomenon was observed, in which the prevalence of TP53 mutations was comparable across all cohorts, whereas the prevalence of EGFR mutations was significantly higher in transformed SCLCs.

Next, we compared the components of somatic signatures between transformed SCLCs and de novo SCLCs. The age-related signature (SBS1) appeared in almost all samples, regardless of whether they were transformed SCLCs or de novo SCLCs. On the other hand, the somatic signature associated with smoking (SBS4) was exclusively detected in de novo SCLCs (Figure 1B).

3.3 Distinct Pattern of Omics Features Between De Novo and Transformed SCLCs

Based on a systematic comparison of omics features between de novo and transformed SCLCs, we identified the distinct pattern (Figure 2). First, the tumor mutational burden (TMB) was significantly higher in de novo SCLCs compared to that in transformed ones (Wilcoxon rank sum test, p value = 0.01). Second, in contrast, multiple omics features related to genomic instability were significantly higher in transformed SCLCs, including HRD (Wilcoxon rank sum test, p value = 0.025), uniparental disomy (UPD, Wilcoxon rank sum test, p value = 0.003), LOH (Wilcoxon rank sum test, p value = 0.008) and TAI (Wilcoxon rank sum test, p value = 0.02). The HRD score is the combination of loss of heterozygosity (LOH), telomeric-allelic imbalance (TAI) and large-scale state transitions (LST).

3.4 Extensive UPD Events Happened in Transformed SCLCs

As shown in the analysis of omics features, the UPD value was significantly higher in transformed SCLCs compared to de novo ones (Figure 2). Since UPD acts as a “second hit” in Knudson's model to achieve biallelic inactivation of tumor suppressor genes [11], we comprehensively compared CNV landscape between transformed SCLCs and de novo SCLCs to study this phenomenon in detail. Firstly, we analyzed the CNV events on chromosome arms. When focusing on chromosome arms with deleted CNV, we found that transformed SCLCs harbored significantly more UPD events compared to either primary LUAD or de novo SCLCs (Figure 3A). The extensive UPD events are likely an important driving factor for the oncogenesis of transformed SCLCs. Conversely, the distribution pattern of amplified CNV was similar among these three cohorts (Figure 3A). Meanwhile, the occurrence frequency of CNV gain events was lower than that of CNV loss events.

Since UPD has the potential to lead to homozygosity of existing aberrations such as mutations, deletions, methylations, histone-modifications, or imprinted genes, it could contribute to the development and/or progression of cancer by inactivating tumor suppressor genes or doubling the copy number of oncogenic alleles [12]. Therefore, we integrated the mutational status with copy number loss and UPD profiles for each gene and selected the most frequently mutated ones to compare across the study cohort (Figure 3B). According to the hypothesis, the mutation was considered the first event (Event 1) and copy number loss and UPD were considered the second event (Event 2). Our results indicated that TP53 was simultaneously mutated and affected by UPD in both transformed SCLC and de novo SCLC. For transformed SCLCs, 50.0% of the samples showed a mutation at TP53 also presented copy number loss or UPD affecting 17p13.1, while for de novo SCLCs, the rate was 77.8%. As for mutation at RB1, both transformed SCLCs and de novo SCLCs displayed a higher frequency of copy number loss or UPD than mutation. While copy number loss and UPD occurred frequently in both transformed SCLCs and de novo SCLCs, they were rare in primary LUADs. Moreover, we found that many high-frequency mutated genes (such as CLTCL1, LRIG3, PABPC1, PTCH1) shared by transformed SCLCs and de novo SCLCs but excluded by primary LUADs, with their main type of mutation being either UPD or genomic losses as a “second hit” event. The high frequency UPD in transformed SCLC further supports the theory that transformed SCLC originates from lineage transformation. Moreover, the similarity of gene profiles between transformed SCLC and de novo SCLC might contain key factors in lineage-specific drivers of lung neuroendocrine cell carcinoma.

Segmental UPD is commonly identified in tumors with impaired chromosome repair mechanisms. To be more detailed, we visualized the distribution of various kinds of CNV happened in genes belonging to six DNA damage repair (DDR) pathways (Figure S3A), including base excision repair (BER), nucleotide excision repair (NER), Fanconi anemia (FA), homology-dependent recombination repair (HRR), non-homologous end joining (NHEJ) and mismatch repair (MMR). As expected, UPD extensively happened in genes involved in DDR pathways, implying defective function of DNA repair in transformed SCLCs. In addition, considering the important role of the Notch signaling pathway in the determination of intra-tumoral heterogeneity of SCLCs, we also visualized the distribution of different categories of CNV happened in genes of the pathway. UPD of genes involved in Notch signaling pathway was also predominantly found in transformed SCLCs, implying the relationship between loss of heterozygosity of these genes and unique features of transformed SCLCs (Figure S3B). Lastly, we focused on genes of the stem cell relevant pathway, as some antibody drug conjugates (ADCs) have been developed to target these genes for the treatment of SCLCs. Consistent with two pathways mentioned above, UPD of genes involved in this pathway was also predominantly found in transformed SCLCs (Figure S3C), which might provide more opportunities for drug development.

3.5 Potential Genomic Biomarkers Differentiating Transformed SCLCs From De Novo SCLCs

Unlike the high frequency of UPD events observed in both transformed SCLC and de novo SCLC, based on the G-score distribution plot produced by GISTIC2, we found the pattern of peaks indicating deeply amplified and deleted genomic regions was quite different between transformed SCLCs and de novo SCLCs, without any overlap (Figure S4). First, at the level of mutated genes, we found that mutations in EGFR, PTCH2, CNGB3, SPTBN5, CROCC, and MYO15A were prevalent in transformed SCLCs, whereas mutations in PABPC3 and MUC19 were dominant in de novo ones (Figure 4A,B). Second, at the level of somatic copy number alterations, it was evident that genomic loss, especially for UPD events, in arm regions 18q, 10p, 3q, 9q, 12p, 19p, 11q, and 13q occurred at a striking higher frequency in transformed SCLCs compared to de novo ones, whereas genomic gain in the arm region 3q was predominantly observed in de novo SCLCs (Figure 4A,B). The clinical utility of these potential biomarkers should be confirmed in prospective clinical trials.