Detecting Small Cell Transformation in Patients with Advanced EGFR Mutant Lung Adenocarcinoma through Epigenomic cfDNA Profiling

Advanced EGFR-mutant (EGFRm) lung adenocarcinoma (LUAD) is an archetype of precision oncology, initially treated with selective EGFR tyrosine kinase inhibitors (TKI; ref. 1). While there are currently five EGFR TKIs approved in the United States, osimertinib, a highly selective third-generation EGFR TKI, is the preferred first-line choice based on improved progression-free and overall survival compared to earlier generation TKIs (2). Despite initial robust responses to EGFR TKIs, acquired resistance inevitably occurs. Subsequent treatment decisions are ideally made following evaluation for potentially targetable resistance mechanisms on a post-progression tumor biopsy, which uncovers potential therapeutic approaches in about one-third of patients (3–6).

Histologic transformation from LUAD to small cell lung cancer (SCLC) is one well-characterized resistance mechanism that occurs in up to 15% of patients with EGFRm LUAD following progression on an EGFR TKI, and appears to be increasing in prevalence with the use of more selective EGFR TKIs, such as osimertinib (4–7). EGFRm LUADs that undergo SCLC transformation are aggressive cancers that portend poor prognosis and require a change in therapeutic regimen (8). Transformed (t)SCLC clinically mimics de novo SCLC—a distinct diagnosis which, unlike EGFRm LUAD, is typically associated with smoking history. Both de novo and tSCLC are treated with platinum-etoposide chemotherapy (1).

Current guidelines recommend that patients with EGFRm LUAD progressing on targeted therapy undergo tumor biopsy to evaluate for actionable mechanisms of TKI resistance, including acquired genomic alterations as well as histologic transformation (1). However, tumor biopsies pose risks to patients and are often not clinically feasible. Consequently, less than half of patients with metastatic EGFRm LUAD undergo tumor tissue biopsy at the time of TKI resistance (9). Further, due to intrapatient tumor heterogeneity, targetable mechanisms of resistance (including histologic transformation) may be missed by sampling a single metastatic focus, resulting in lost opportunities to deliver guideline-recommended histology-directed therapy.

For these reasons, molecular profiling at the time of EGFR TKI resistance is increasingly performed via liquid biopsies, which can detect somatically acquired genomic alterations in circulating tumor cell-free DNA (cfDNA). Compared to tumor biopsies, liquid biopsies are minimally invasive, can easily be repeated at multiple timepoints, and may better capture intrapatient tumor heterogeneity (10). However, limitations in current commercially available liquid biopsies preclude detection of certain clinically actionable resistance phenotypes that lack defining genomic alterations (6). In particular, the diagnosis of tSCLC cannot be made by currently available liquid biopsy assays.

The development of noninvasive diagnostic approaches to detect SCLC transformation has the potential to usher in a new paradigm of diagnostic and therapeutic precision for patients with advanced lung cancer (8, 11–14). In particular, histologic transformation is accompanied by widespread epigenomic reprogramming, which presents an opportunity to detect tSCLC through epigenomic analysis of cfDNA (11). We, and others, have developed tools to profile tumor epigenomic features from patient plasma, including DNA methylation (15–18), chromatin accessibility (19, 20), and histone modifications (21, 22). Herein, we build upon this work, generating and analyzing 351 tissue and plasma epigenomic libraries to demonstrate the clinical utility of epigenomic cfDNA profiling to noninvasively detect small cell transformation in patients with EGFRm LUAD (Fig. 1).

Lung cancer patient-derived xenografts (PDX) were derived from patients with LUAD, de novo SCLC, and tSCLC as previously described (23–26). Tissue samples were obtained from donors who provided explicit written consent per the Declaration of Helsinki under an approved Institutional Review Board (IRB) protocol at Massachusetts General Hospital (MGH). All mouse studies were conducted through Institutional Animal Care and Use Committee–approved animal protocols in accordance with institutional guidelines (MGH Subcommittee on Research Animal Care, OLAW Assurance A3596-01). tSCLC PDXs were reviewed by a staff thoracic pathologist to confirm SCLC histology. Plasma samples were collected from patients with advanced lung cancer diagnosed and treated at the Dana-Farber Cancer Institute (DFCI) or MGH between November 2016 and March 2023. All patients provided written informed consent. The use of samples was approved by DFCI (01-045 and 09-171) and MGH (13-416) IRB. Studies were conducted in accordance with recognized ethical guidelines.

Peripheral blood was collected in EDTA Vacutainer tubes (BD) or Streck Cell-Free DNA tubes and processed within 3 hours of collection. Plasma was separated by centrifugation at 1,600 g for 10 minutes, transferred to microcentrifuge tubes, and centrifuged at 3,000 g at room temperature for 10 minutes. The supernatant was aliquoted and stored at −80°C until the time of DNA extraction. cfDNA was isolated from 1 mL of plasma, using the QIAGEN Circulating Nucleic Acids Kit (Qiagen), eluted in AE buffer, and stored at −80°C. Low-pass whole genome sequencing (LPWGS) was performed on all cfDNA samples. The ichorCNA R package was used to infer copy-number profiles and cfDNA tumor content from read abundance across bins spanning the genome using default parameters (27). Per the published limit of detection of ichorCNA, estimated cfDNA tumor content cut-off of greater than or less than 0.03 were used to characterize samples as having detectable or undetectable circulating tumor DNA, respectively.

Frozen tissue was pulverized using the Covaris CryoPrep system and fixed with 2 mmol/L disuccinimidyl glutarate for 10 minutes followed by 1% formaldehyde buffer for 10 minutes and quenched with glycine. Chromatin was sheared to 300 to 500 bp using the Covaris E220 ultrasonicator and then incubated overnight with the following antibodies coupled with 40 μL protein A and protein G beads (Invitrogen) at 4°C overnight: H3K27ac (Abcam #ab4729, Lot GR3442890-1), H3K4me3 (Thermo Fisher Scientific #PA5-27029, Lot XI3696063), and H3K27me3 (Cell Signaling #9733S, Lot 19). Five percent of the sample was not exposed to antibody and was used as a control. Beads were washed three times each with Low-Salt Wash Buffer (0.1% SDS, 1% Triton X-100, 2 mmol/L EDTA, 20 mmol/L Tris-HCl pH 7.5, 150 mmol/L NaCl), High-Salt Wash Buffer (0.1% SDS, 1% Triton X-100, 2 mmol/L EDTA, 20 mmol/L Tris-HCl pH 7.5, 500 mmol/L NaCl), and LiCl Wash Buffer (10 mmol/L Tris pH 7.5, 250 mmol/L LiCl, 1% NP-40, 1% Na-Doc, 1 mmol/L EDTA) and rinsed with TE buffer (pH 8.0) once. Samples were then de-cross-linked, treated with RNase and proteinase K, and DNA was extracted (Qiagen). DNA sequencing libraries were prepared from the purified immunoprecipitated and non-immunoprecipitated DNA using the ThruPLEX DNA-seq Kit (TakaraBio). Libraries were sequenced on an Illumina HiSeq 4000 to generate 150 bp paired-end reads (Novogene Corporation).

Chromatin immunoprecipitation sequencing (ChIP-seq) reads were aligned to the human genome build hg19 using the Burrows-Wheeler Aligner version 0.7.17 (RRID:SCR_010910; ref. 28). Non-uniquely mapped and redundant reads were discarded. MACS v2.1.1.20140616 (RRID:SCR_013291) was used for ChIP-seq peak calling with a q-value threshold of 0.01 (29). IGV v2.8.2 was used to visualize normalized ChIP-seq read counts at specific genomic loci (30). ChIP-seq heatmaps were generated with deepTools v3.3.1 (RRID:SCR_016366) and show normalized read counts at the peak center ±2 kb unless otherwise noted (31). Overlap of ChIP-seq peaks was assessed using BEDTools v2.26.0. Peaks were considered overlapping if they shared one or more base pairs.

Frozen tissue was resuspended and dounce homogenized in 1,000 μL of homogenization buffer. Nuclei were filtered using a 70-μm Flowmi strainer, isolated using iodixanol density-gradient centrifugation method, and washed with RSB buffer (10 mmol/L Tris-HCl pH 7.4, 10 mmol/L NaCl, and 3 mmol/L MgCl₂ in water). Fifty thousand nuclei were resuspended in 50 μL of transposition mix [2.5 μL transposase (100 nmol/L), 16.5 μL PBS, 0.5 μL 1% digitonin, 0.5 μL 10% Tween-20, and 5 μL water; ref. 32]. Transposition reactions were incubated at 37°C for 30 minutes in a thermomixer shaking at 1,000 rpm. Reactions were cleaned with Qiagen columns. Libraries were amplified using the Omni-ATAC protocol and sequenced on an Illumina platform (Novogene Corporation) using 150-base paired-end reads (33).

Sample clustering, principal component analysis, and identification of lineage-enriched peaks were performed using Cobra v2.0 (RRID:SCR_005677), a ChIP-seq analysis pipeline implemented with Snakemake (34, 35). ChIP-seq data from LUAD, de novo SCLC, and tSCLC PDXs were compared to identify H3K27ac, H3K4me3, and H3K27me3 peaks with significant enrichment in the three tumor subtypes. A union set of peaks for each histone modification was created using BEDTools (RRID:SCR_006646). narrowPeak calls from MACS were used for H3K27ac and H3K4me3 while broadPeak calls were used for H3K27me3. The number of unique aligned reads overlapping each peak in each sample was calculated from BAM files using BEDtools. Quantile normalization was applied to this matrix of normalized read counts for clustering and PCA analysis. Unsupervised hierarchical clustering was performed based on Spearman correlation between samples. Principal component analysis was performed using the prcomp R function. Raw read counts for each peak were normalized to the total number of mapped reads for each sample. Then using DEseq2 v1.14.1 (RRID:SCR_015687), histology-enriched peaks were identified at the indicated FDR-adjusted P value (p_adj) < 0.001 and log₂ fold-change >2 (36).

RNA was extracted from frozen tumor samples using the Qiagen RNeasy Mini Kit (Cat No./ID: 74104). RNA sequencing (RNA-seq) libraries were constructed from 1 μg RNA using the Illumina TruSeq Stranded mRNA LT Sample Prep Kit. Barcoded libraries were pooled and sequenced on the Illumina HiSeq 2,500 generating 50 bp paired-end reads. FASTQ files were processed using the VIPER workflow (37). Read alignment to human genome build hg19 was performed with STAR (RRID:SCR_004463; ref. 38). Cufflinks (RRID:SCR_014597) was used to assemble transcript-level expression data from filtered alignments (39). Differential gene expression analysis was conducted using DESeq2 (36).

Methylated DNA immunoprecipitation sequencing (MeDIP-seq) was performed on tissue and plasma following published methods (15–18). Library preparation was performed on 10 ng of DNA using the KAPA HyperPrep Kit (KAPA Biosystems). We then performed end-repair, A-tailing, and ligation of NEBNext adaptors (NEBNext Multiplex Oligos for Illumina kit, New England BioLabs). Libraries were digested using the USER enzyme (New England BioLabs). λ DNA, consisting of unmethylated and in vitro methylated DNA, was added to prepared libraries to achieve a total amount of 100 ng DNA. Methylated and unmethylated Arabidopsis thaliana DNA (Diagenode) was added for quality control. MeDIP was performed using the MagMeDIP Kit (Diagenode) following the manufacturer’s protocol. Samples were purified using the iPure Kit v2 (Diagenode). Success of the immunoprecipitation was confirmed using qPCR to detect recovery of the spiked-in Arabidopsis thaliana methylated and unmethylated DNA. KAPA HiFi Hotstart ReadyMix (KAPA Biosystems) and NEBNext Multiplex Oligos for Illumina (New England Biolabs) were added to a final concentration of 0.3 μmol/L and libraries were amplified. Samples were pooled and sequenced (Novogene Corporation) on Illumina HiSeq 4,000 to generate 150 bp paired-end reads.

Quality and quantity of raw MeDIP-seq reads were examined using FastQC version 0.11.5 (http://www.bioinformatics.babraham.ac.uk/projects/fastqc) and MultiQC version 1.7 (40). Raw reads were quality and adapter trimmed using Trim Galore! version 0.6.0 (http://www.bioinformatics.babraham.ac.uk/projects/trim_galore/) using default settings in paired-end mode. The trimmed reads then were aligned to hg19 using Bowtie2 version 2.3.5.1 in paired-end mode and all other settings default (41). The SAMtools version 1.10 software suite was used to convert SAM alignment files to BAM format, sort and index reads, and remove duplicates (42). The R package RSamtools version 2.2.1 was used to calculate the number of unique mapped reads. Saturation analyses to evaluate reproducibility of each library were carried out using the R Bioconductor package MEDIPS version 1.38.0 (43).

One microgram of antibody was coupled with 10 μL protein A (Invitrogen, cat #10002D) and 10 μL protein G (Invitrogen, cat #10004D) for at least 6 hours at 4°C with rotation in 0.5% BSA (Jackson Immunology, cat #001-000-161) in PBS (Gibco, cat #14190250), followed by blocking with 1% BSA in PBS for 1 hour at 4°C with rotation. The following antibodies were used: H3K27ac (Abcam #ab4729) and H3K4me3 (Thermo Fisher Scientific #PA5-27029). A total of 800 μL of thawed plasma was centrifuged at 3,000 g for 15 minutes at 4°C. The supernatant was precleared with the magnetic beads with 20 μL protein A and 20 μL protein G for 2 hours at 4°C. Then, the precleared and conditioned plasma was subjected to antibody-coupled magnetic beads overnight with rotation at 4°C. The reclaimed magnetic beads were washed with 1 mL of each washing buffer twice. Three washing buffers were used in following order: low salt washing buffer (0.1% SDS, 1% Triton X-100, 2 mmol/L EDTA, 150 mmol/L NaCl, 20 mmol/L Tris-HCl pH 7.5), high salt buffer (0.1% SDS, 1% Triton X-100, 2 mmol/L EDTA, 500 mmol/L NaCl, 20 mmol/L Tris-HCl pH 7.5), and LiCl washing buffer (250 mmol/L LiCl, 1% NP-40, 1% Na Deoxycholate, 1 mmol/L EDTA, 10 mmol/L Tris-HCl pH 7.5). Subsequently, the beads were rinsed with TE buffer (Thermo Fisher Scientific, cat #BP2473500) and resuspended and incubated in 100 μL of DNA extraction buffer containing 0.1 mol/L NaHCO3, 1% SDS, and 0.6 mg/mL Proteinase K (Qiagen, cat #19131) and 0.4 mg/mL RNaseA (Thermo Fisher Scientific, cat #12091021) for 10 minutes at 37°C, for 1 hour at 50°C and for 90 minutes at 65°C. DNA was purified through phenol extraction (Invitrogen, cat #15593031) and ethanol precipitation was performed with 3 mol/L NaOAc (Ambion, cat #AM9740) and glycogen (Ambion, cat #AM9510). Cell-free ChIP-seq (cfChIP-seq) libraries were prepared with ThruPLEX DNA-Seq Kit (Takara Bio, cat #R400675) following the manufacturer’s instructions. After library amplification, the DNA was purified by AMPure XP (Beckman coulter, cat# A63880). The size distribution of the purified libraries was examined using Agilent 2100 Bioanalyzer with a high sensitivity DNA Chip (Agilent, cat #5067-4626). The library was submitted for the 150 base-pair paired-end sequencing on an Illumina NovaSeq6000 system (Novogene Corporation).

For cfDNA histone data, we computed reads per kilobase per million mapped reads for each cfChIP-seq library using bed files of the SCLC-enriched and LUAD-enriched peaks in the PDXs (FDR-adjusted P < 0.001 and log₂ fold-change >2). We excluded the top and bottom 0.5% of sites based on the average signal across all cfChIP-seq samples. For cfDNA methylation data, we used the MeDIPs R package to calculate CpG-normalized relative methylation scores across 300 bp windows across the genome for each cell-free methylated DNA immunoprecipitation and sequencing (cfMeDIP-seq) library (43, 44). We then summed relative methylation scores in cfDNA at SCLC-enriched and LUAD-enriched differentially methylated regions from the PDXs (FDR-adjusted P < 1.0 × 10⁻⁶ and log₂ fold-change >3) for each cfDNA sample and normalized this value to the sum of rms values across all 300 bp windows (18). For the cfDNA chromatin accessibility analysis, we applied Griffin (v0.1.0), a computational tool that profiles nucleosome protection and accessibility from cfDNA, to all LPWGS libraries (19). Fragments aligning to the region within ±5,000 bp from the site of interest [defined as differential assay for transposase-accessible chromatin sequencing (ATAC-seq) peaks between SCLC and LUAD PDXs] were extracted. Duplicate and low-quality alignments (mapping quality <20) were filtered out. Fragments within the nucleosome size range (140–250 bp) were preserved. We then extracted the “mean window coverage” at the SCLC-enriched and the LUAD-enriched regions of open chromatin in the PDXs (FDR-adjusted P < 0.001 and log₂ fold-change >2). For each data type, the SCLC risk score was calculated as the ratio of signal at the SCLC-enriched sites over the LUAD-enriched sites. This value was normalized to the median score across all samples for that data type and the log₂ value was calculated, which was reported as the cfDNA SCLC risk score. To generate the integrated epigenomic SCLC risk score, we first calculated Z-scores for each sample within each individual data type (DNA methylation, H3K27ac, and chromatin accessibility). The integrated epigenomic cfDNA SCLC risk score was calculated as the sum of the Z-scores for each of the individual cfDNA SCLC risk scores for each sample.

χ² test was used to calculate the P value for the correlation of epigenomic features with differentially expressed genes. All statistical tests were two-sided except where otherwise indicated.

The PDX epigenomic data have been deposited in Gene Expression Omnibus (GSE269746). The cfDNA data generated from patient samples that support the findings of this study are available upon request from the corresponding authors (J.E. Berchuck, M.L. Freedman) to comply with the DFCI ethics regulations to protect patient privacy. All requests for raw and analyzed data will be promptly reviewed by the Belfer Office for Dana-Farber Innovations to verify if the request is subject to any intellectual property or confidentiality obligations. Any data and materials that can be shared will be released via a Data Transfer Agreement.

We first sought to comprehensively characterize the epigenomic landscape of tSCLC relative to LUAD and de novo SCLC. We performed this analysis on PDXs developed from biopsies obtained from patients with LUAD, de novo SCLC, and tSCLC. We chose this model system for several reasons. First, compared to primary tumor biopsies whose cellular content often contains a high proportion of stromal cells, PDXs facilitate analysis of pure tumor cell populations to more cleanly study tumor-intrinsic molecular features. Second, PDXs can produce large tumor volumes to facilitate molecular analyses that would be infeasible to perform on primary tumor biopsies, which often produce insufficient material for the epigenomic profiling studies performed herein. Importantly, we previously demonstrated that the PDXs in this study recapitulate their original tumor molecular profiles (24). Our analysis focused on 26 lung cancer PDXs, comprising LUAD (n = 13), tSCLC (n = 4), and de novo SCLC (n = 9) tumors (Supplementary Table S1). The SCLC tumors were predominantly ASCL1 (SCLC-A) and/or NEUROD1 (SCLC-N) subtype, with one sample each of the POU2F3 (SCLC-P) and YAP1 [SCLC-Y, also referred to as triple negative or inflamed (SCLC-I)] subtype (Supplementary Fig. S1; refs. 45–47). Consistent with prior reports, all tSCLC PDXs exhibited loss of TP53 and/or RB1 (Supplementary Table S2; refs. 11–14). For each PDX, we performed ChIP-seq to profile the histone modifications H3K27ac (a mark of active gene promoters and enhancers), H3K4me3 (a mark of active gene promoters), and H3K27me3 (a mark of repressed regulatory elements), MeDIP-seq to profile DNA methylation, and ATAC-seq to profile chromatin accessibility, resulting in 126 epigenome-wide libraries (Fig. 1; Supplementary Fig. S2; refs. 23, 24). We also performed bulk RNA-seq on PDXs for which this data had not already been generated (23, 24).

Across every epigenomic feature assessed, unsupervised analyses demonstrated that tSCLC tumors clustered with de novo SCLC tumors and were distinct from LUAD tumors (Fig. 2A; Supplementary Fig. S3). tSCLC tumors exhibited a gain of signal associated with active gene transcription—e.g., open chromatin, promoter H3K27ac and H3K4me3, gene body methylation, and loss of gene body H3K27me3—at neural lineage-defining genes, such as ASCL1, NEUROD1, and DLL3 (Fig. 2B). Likewise, more of the H3K27ac ChIP-seq and ATAC-seq peaks in tSCLC PDXs were also peaks in de novo SCLC than in LUAD PDXs (Supplementary Fig. S4). Collectively, these data affirm a prior report of shifts in DNA methylation and gene expression (11) and demonstrate that histologic transformation from LUAD to SCLC is characterized by widespread epigenomic reprogramming, converging on a profile that resembles de novo SCLC.

We sought to leverage these divergent epigenomic profiles to develop a diagnostic test to detect SCLC transformation in patients with EGFRm LUAD through epigenomic analysis of cfDNA. Building upon our prior work that tissue-informed analysis improves the detection of tumor-specific epigenomic features in cfDNA (18), we first identified epigenomic features enriched in tSCLC compared to EGFRm LUAD. Given the similarity in epigenomic profiles of EGFRm and non-EGFRm LUAD, and tSCLC and de novo SCLC, we included all 26 PDXs in this analysis (Supplementary Fig. S3). Comparative analysis of the LUAD and SCLC tumors resulted in a set of 24,424 H3K27ac, 2,272 H3K4me3, 3,298 DNA methylation, and 16,352 accessible chromatin sites with significantly greater signal in one or the other histologic subtypes. To identify features that would be clinically informative for a liquid biopsy, we removed sites with peaks in white blood cells (WBC), the primary source of background contamination. This step removed 50% of all sites across features, resulting in 10,907 H3K27ac, 833 H3K4me3, 1,210 DNA methylation, and 9,995 accessible chromatin sites enriched in SCLC or LUAD without signal in WBCs (Fig. 3A; Supplementary Fig. S5). Importantly, these epigenomic features strongly correlated with transcriptional activity. There was a strong association of upregulated genes in one histologic subtype having a nearby H3K27ac, H3K4me3, or open chromatin peak in that subtype (Fig. 3B; Supplementary Fig. S6).

We next performed multianalyte epigenomic cfDNA profiling on 48 plasma samples collected from 32 patients with metastatic lung cancer. This cohort comprised 20 patients with EGFRm LUAD who never developed tSCLC (EGFRm LUAD) and 12 patients with EGFRm LUAD following diagnosis of biopsy-proven tSCLC (tSCLC), 6 of whom also had plasma collected prior to being diagnosed with tSCLC (Supplementary Table S3). Consistent with prior reports, all of the tumors from patients with tSCLC included in this cohort exhibited loss of TP53 and/or RB1 (Supplementary Table S2; refs. 11–14). On all plasma samples, we generated epigenome-wide data for four analytes—H3K27ac, H3K4me3, DNA methylation, and chromatin accessibility—from 1 mL of plasma (Fig. 1). In brief, we utilized a novel assay (cfChIP-seq) to identify nucleosomal cfDNA fragments bound to H3K27ac and H3K4me3, respectively (22). We then performed cfMeDIP-seq to profile cfDNA methylation (15–18). Finally, we performed LPWGS and utilized Griffin to infer regions of open chromatin based on cfDNA coverage patterns (19, 20). All subsequent analyses include samples with detectable circulating tumor DNA, defined as estimated tumor fraction >3% based on the lower limit of detection of ichorCNA applied to LPWGS data (27). More than 95% of plasma samples collected from patients with EGFRm LUAD progressing on osimertinib have cfDNA tumor fraction >3% using this method, supporting the clinical relevance of this criteria (48). Notably, there was no difference in cfDNA tumor fraction between samples from patients with tSCLC and EGFRm LUAD (P = 0.10; Supplementary Table S3). Filtering based on this cut-off and quality control metrics resulted in a final set of 105 epigenomic cfDNA libraries, which were included in subsequent analyses (Supplementary Fig. S7).

To evaluate the ability to noninvasively detect tSCLC in patients with EGFRm LUAD, we developed individual SCLC risk scores for each epigenomic analyte based on cfDNA signals at the SCLC- and LUAD-enriched sites in the PDXs for that feature (Fig. 1; Supplementary Fig. S5). The following SCLC risk scores were calculated as the normalized ratio of signals at the SCLC- versus LUAD-enriched sites (see “Materials and Methods”).

We first assessed the ability to accurately detect tSCLC through cfDNA nucleosome profiling. We observed significantly higher H3K27ac SCLC risk scores in cfDNA samples from patients with tSCLC than those with EGFRm LUAD (P = 0.0042; Fig. 4A). H3K27ac SCLC risk scores discriminated between plasma samples from patients with tSCLC versus EGFRm LUAD with an area under the receiver operating characteristic curve (AUROC) of 0.87 (P = 0.0056). H3K4me3 SCLC risk scores trended in the same direction, but did not achieve statistical significance (AUROC = 0.70; P = 0.073; Fig. 4B). We hypothesized that the superior performance of H3K27ac relative to H3K4me3 was due in large part to the presence of more than 13 times the number of differential H3K27ac sites (n = 10,907) between SCLC and LUAD than H3K4me3 sites (n = 833), since both features mark active promoters, but only H3K27ac marks active enhancers as well. This is illustrated by the distribution of these two histone modifications near INSM1, a neural lineage–defining gene, wherein an H3K4me3 peak marks the gene promoter and H3K27ac peaks mark the promoter as well as the two most proximal enhancers in tSCLC (GH20J020396 and GH20J020399; Supplementary Fig. S8). In cfDNA, H3K27ac not only marked neural-lineage genes but also potential SCLC therapeutic targets such as DLL3 (Supplementary Fig. S9). These data demonstrate the ability to noninvasively detect tSCLC and potentially therapeutic target expression, through cfDNA nucleosome profiling and highlight the value of profiling gene enhancers in addition to promoters.

To further validate the ability of the PDX-derived H3K27ac SCLC risk score to discriminate between histology subtypes, we performed H3K27ac ChIP-seq on 20 LUAD cell lines and 13 SCLC cell lines (Supplementary Table S4). Notably, all four molecular subtypes of SCLC were represented, including SCLC-A (n = 5), SCLC-N (n = 3), SCLC-P (n = 3), and SCLC-Y (n = 2; ref. 49). H3K27ac SCLC risk scores were significantly higher in the SCLC than the LUAD cell lines (P = 9.5 × 10⁻⁷) and achieved an AUROC of 0.95 (P = 1.4 × 10⁻⁵) for accurate discrimination of the two histologic subtypes (Supplementary Fig. S10A and S10B). Notably, the SCLC-A (P = 3.8 × 10⁻⁵), SCLC-N (P = 0.0011), and the SCLC-P (0.0023) cell lines all individually exhibited higher SCLC risk scores than the LUAD cell lines. In contrast, the SCLC risk scores for the SCLC-Y cell lines were not different from the LUAD cell lines (P = 0.36) and were significantly lower than the scores for the other three SCLC subtypes (P = 0.026; Supplementary Fig. S10A). When excluding the SCLC-Y cell lines, the SCLC risk score achieved an AUROC of 0.995 (P = 6.8 × 10⁻⁶) for accurately discriminating SCLC-A, SCLC-N, and SCLC-P cell lines from LUAD cell lines (Supplementary Fig. S10C).

cfDNA methylation and chromatin accessibility analysis also discriminated between patients with tSCLC and those with EGFRm LUAD. We again observed significantly higher SCLC risk scores in cfDNA samples from patients with tSCLC than those with EGFRm LUAD for both DNA methylation (P = 0.00090; Fig. 4C) and chromatin accessibility (P = 0.0023; Fig. 4D) data. Classifiers based on signal in cfDNA at differential methylation and chromatin accessibility sites between SCLC and LUAD tumors discriminated between patients with tSCLC and EGFRm LUAD with AUROCs of 0.85 (P = 0.0015) and 0.82 (P = 0.0033), respectively. These data add to a growing body of literature demonstrating that cfDNA methylation and chromatin accessibility analysis can detect clinically actionable tumor biology (15–17, 19, 20, 50, 51).

We hypothesized that integrated cfDNA analysis of multiple epigenomic features would improve our ability to discriminate between patients with EGFRm LUAD and tSCLC. To characterize the extent to which multiple epigenomic features would provide additive versus redundant information, we assessed the overlap of differential sites between SCLC and LUAD PDXs for H3K27ac, methylation, and open chromatin. Of the 20,079 differential sites across the three features without background signal in WBCs, the vast majority (90%; n = 18,078) were unique to one epigenomic data type (Fig. 5A). The number of combined sites was 1.8, 2.0, and 16.6 times greater than the number of H3K27ac, accessible chromatin, and DNA methylation sites, respectively, and spanned more than 60 Mb—approximately two times the number of base pairs in the protein-coding genome. Given the largely non-overlapping information from these three epigenomic features, we tested a multi-analyte classifier integrating H3K27ac, DNA methylation, and chromatin accessibility data. This integrated epigenomic SCLC risk score discriminated between cfDNA samples from patients with tSCLC and EGFRm LUAD with greater accuracy than any of the individual analytes (AUROC = 0.94; P = 0.0095; Fig. 5B). The optimal diagnostic cut-off (−0.029) demonstrated 89% sensitivity and 91% specificity to identify cfDNA samples from patients with tSCLC versus EGFRm LUAD (likelihood ratio of 9.8).

Circulating tumor DNA levels are a critical factor in the ability of liquid biopsies to accurately detect and characterize cancers. As such, we assessed the relationship of the integrated epigenomic cfDNA SCLC risk score with cfDNA tumor content. We observed that SCLC risk scores strongly correlated with cfDNA tumor fraction in patients with tSCLC (R² = 0.79; P = 4.5E−5) but not in patients with EGFRm LUAD (R² < 0.01; P = 0.99; Fig. 5C). This finding suggests that the SCLC risk score reflects tumor biology (i.e., the amount of signal in cfDNA at the site of epigenomic features enriched in a histologic subtype) rather than solely reflecting levels of circulating tumor DNA.

A unique feature of our cohort is plasma samples collected prior to and at the time of small cell transformation in 2 patients with EGFRm LUAD, providing the opportunity to correlate longitudinal assessment of the integrated epigenomic cfDNA SCLC risk scores with the emergence of tSCLC. The first patient (Fig. 6A) was a 73-year-old woman diagnosed with metastatic LUAD with EGFR exon 19 deletion involving the brain. She was started on first-line erlotinib and remained in remission 36 months later when the first plasma timepoint was collected showing an SCLC risk score of −0.4. Four months later, she experienced radiographic progression with cfDNA analysis showing an EGFR T790M mutation. A second plasma sample at this timepoint showed an SCLC risk score of −1.0. Treatment was switched from erlotinib to osimertinib. A third plasma timepoint collected 10 months later showed a marked increase in the SCLC risk score to 5.2. Scans at that time revealed new liver metastases and a liver biopsy showed tSCLC. The second patient (Fig. 6B), a 62-year-old woman, was diagnosed with metastatic LUAD with EGFR exon 19 deletion involving the liver, bone, and brain for which she was started on first-line osimertinib. The first two plasma samples collected 3 months apart at the time of minor radiographic progression on osimertinib showed SCLC risk scores of −2.6 and −1.1, respectively. A third plasma sample 3 months later showed an increase in the SCLC risk score to 0.6. Scans at that time revealed new liver metastases and a liver biopsy showed tSCLC. Notably, the rise in the SCLC risk score preceded clinical diagnosis of tSCLC by 92 days in this patient. These patient vignettes illustrate that longitudinal epigenomic cfDNA analysis reflects the emergence of tSCLC in patients with EGFRm LUAD.

Histologic transformation to SCLC is an aggressive, clinically actionable resistance phenotype that emerges in a subset of patients with EGFRm LUAD. Limitations of current clinical tools result in delays in diagnosis and underdiagnosis of tSCLC, and consequently, missed opportunities to deliver optimal guideline-recommended histology-directed systemic therapy. Herein, in the largest study to date of cfDNA samples from patients with tSCLC, we demonstrate for the first time the ability to non-invasively detect small cell transformation in patients with EGFRm LUAD through epigenomic profiling of 1 mL of plasma. With the limitations of tissue biopsy resulting in fewer than half of patients with metastatic EGFRm LUAD undergoing histologic tumor assessment at the time of EGFR TKI resistance, these data highlight the potential to advance diagnostic and therapeutic precision for patients with advanced lung cancer by augmenting the current diagnostic paradigm with epigenomic cfDNA analysis.

Liquid biopsies are now widely utilized in clinical oncology to detect cancer recurrence and inform therapeutic decisions. However, most commercially-available cfDNA assays only detect tumor genomic alterations. The lack of genomic alterations exclusive to tSCLC limits the utility of these genomic-based cfDNA approaches to detect small cell transformation in patients with EGFRm LUAD. To address this limitation, we and others have developed tools to analyze several tumor epigenomic features from patient plasma, including DNA methylation (15–18), chromatin accessibility (19, 20), and histone modifications (21, 22). Several recent studies highlight the potential of epigenomic cfDNA profiling to provide dynamic insights into tumor biology in patients with advanced lung cancer. Haq and colleagues analyzed cfDNA methylation profiles from patients with advanced de novo SCLC, identifying two methylation-defined subsets that associate with distinct tumor biology and clinical prognosis (52). Heeke and colleagues (50) and Chemi and colleagues (51) demonstrated that SCLC subtypes (based upon predominant transcription factor expression, i.e., ASCL1, NEUROD1, POU2F3) harbor distinct methylation profiles that can be detected through cfDNA methylation analysis. Our study is distinct in that we demonstrate for the first time that epigenomic cfDNA profiling can be used to detect small cell transformation in patients with EGFRm LUAD progressing on EGFR TKIs. Diagnosing tSCLC by cfDNA profiling would be immediately clinically actionable, as guidelines recommend that tSCLC be treated with a de novo SCLC regimen of platinum-etoposide chemotherapy which would not otherwise be used in patients with LUAD (1, 53).

An accurate and easily implementable diagnostic test to detect tSCLC in the clinic, would not only facilitate timely delivery of standard of care therapy but could also potentially identify patients for tSCLC-directed therapeutic clinical trials. The first cohort of clinical trials designed specifically for patients with tSCLC are currently enrolling, testing addition of programmed death-ligand 1 inhibitors in combination with either platinum-etoposide chemotherapy or poly-ADP ribose polymerase 1 inhibitors (NCT04538378; NCT05957510; NCT03944772). Notably, we present data supporting the ability of epigenomic cfDNA analysis to noninvasively detect expression of potential therapeutic targets for tSCLC, such as DLL3, the target of tarlatlamab, a bispecific antibody that recently demonstrated a 40% response rate in patients with previously treated de novo SCLC (53). Finally, the potential feasibility of detecting an emerging signal of tSCLC histology with our assay lends itself to the possibility of early adaptive treatment strategies to prevent or delay outright SCLC transformation, similar to the concept behind an ongoing trial (NCT03567642) treating patients with EGFRm LUAD at higher risk for SCLC transformation (concurrent TP53/RB1 mutations) with four cycles of platinum/etoposide chemotherapy in addition to osimertinib.

A major strength of our study is the breadth and depth of epigenomic data generated from tSCLC tumors. To our knowledge, only one study has previously investigated the epigenomic landscape of tSCLC (11). This multiomic analysis of tSCLC tumors by Quintanal-Villalonga and colleagues found that small cell transformation is primarily driven by transcriptional reprogramming, with integrated methylation analysis providing insights into epigenomic changes that underlie histologic transformation. Our work builds upon this observation, generating epigenome-wide data on DNA methylation, chromatin accessibility, and three histone modifications involved in gene regulation on a series of 26 LUAD, de novo SCLC, and tSCLC PDXs. These data affirmed that the observed transcriptional reprogramming that accompanies histologic transformation is characterized by widespread epigenomic reprogramming, converging on an epigenomic profile that resembles de novo SCLC. While the focus of this manuscript was to develop a tumor-informed cfDNA-based epigenomic classifier to detect tSCLC, we believe that this publicly available epigenomic-transcriptomic dataset will be a valuable resource to further our understanding of the biology that underlies histologic transformation in lung cancer and clinically relevant differences between de novo and transformed SCLC.

The platform developed and employed for cfDNA analysis in this study generated new insights into the feasibility and clinical utility of profiling multiple epigenomic features in cfDNA. Because epigenomic cfDNA profiling approaches have been developed independently, it is not known whether integrating multiple epigenomic analytes is feasible in real-world patient samples or whether this results in better ability to detect clinically relevant tumor biology. We demonstrate for the first time the ability to generate high-quality epigenome-wide cfDNA data on DNA methylation, histone modifications, and chromatin accessibility from 1 mL of patient plasma. Notably, some of the plasma samples included in this study were collected as long as 8 years prior to analysis. Combined with the minimal sample requirement, these technologies potentially unlock the ability to obtain multianalyte epigenome-wide cfDNA data from real-world plasma samples. Additionally, to our knowledge, this is the first study to suggest that combining multiple cfDNA epigenomic features may improve the ability to non-invasively detect a clinically actionable resistance phenotype. The observation that histology-specific H3K27ac, DNA methylation, and open chromatin sites were largely nonoverlapping led us to develop a multi-analyte classifier integrating these three epigenomic analytes that resulted in better diagnostic accuracy (AUROC of 0.94) than any of the individual features (AUROC of 0.82–0.87). While further studies are needed to understand the value of multianalyte versus single-analyte epigenomic analysis, we demonstrate for the first time the feasibility of profiling multiple epigenomic features in cfDNA from 1 mL of real-world plasma samples with data suggesting that integrated multianalyte cfDNA profiling may improve the ability to noninvasively detect clinically relevant tumor biology.

We acknowledge important limitations of this study. First is the modest number of patient samples in the cfDNA cohort. While small in absolute terms, this represents the largest study to date of plasma samples from patients with pathologically confirmed tSCLC and the first to demonstrate the ability to noninvasively detect small cell transformation in patients with EGFRm LUAD through cfDNA analysis. Given the rarity of these specimens, plasma samples in this study were collected across 8 years at two institutions under different conditions, e.g., EDTA tubes for some samples and Streck tubes for others, and different plasma extraction methods. The results of the cfDNA analysis in this real-world cohort—despite the variability in the several pre-analytical conditions of the samples—demonstrate the robustness of the epigenomic assays deployed in this study. A second limitation is the lack of an independent validation cohort. Given the rarity of plasma samples from patients with biopsy-proven tSCLC, this was not feasible. We would like to highlight, however, that the epigenomic classifiers were developed solely from an independent cohort of tumors (i.e., comparative analysis of unrelated SCLC and LUAD PDXs) and applied to the cfDNA samples, thus the results are less subject to diagnostic biases, such as overfitting. Nevertheless, the performance of the described epigenomic classifiers, as well as the optimal cut-offs that maximize diagnostic accuracy for identifying patients with EGFRm LUAD who have undergone small cell transformation, need to be validated in independent cohorts. Additionally, further validation in cohorts with representation of all molecular subtypes of SCLC will be important (45–47). The majority of the SCLC PDXs from which the classifier was derived, and the tumors from patients with tSCLC in the cfDNA cohort, were of the SCLC-A or SCLC-N subtypes. Encouragingly, the PDX-derived H3K27ac SCLC risk score demonstrated excellent diagnostic accuracy for not only SCLC-A and SCLC-N, but also the SCLC-P in our cell line validation experiment. However, the inability of this classifier to distinguish SCLC-Y from LUAD is an important limitation. For now, tumor tissue biopsy remains the gold standard to diagnose histologic transformation in patients with advanced LUAD with liquid biopsy representing a complementary approach that may better assess intratumoral spatial heterogeneity, can provide diagnostic information when tissue is unavailable, and allows for serial sampling over a patient’s disease course. Evaluation of the performance of the SCLC risk score in cfDNA cohorts from patients with annotated tSCLC tumor molecular subtypes is needed. Finally, the retrospective nature of our cohort resulted in variability in timing of collection of plasma samples relative to biopsy-proven tSCLC diagnosis, so we were unable to methodically determine how far epigenomic cfDNA-based identification of SCLC histology precedes clinical tissue diagnosis. Intriguingly, we observed in one patient that a rise in the cfDNA SCLC risk score preceded clinical diagnosis of tSCLC by more than 3 months. To more formally evaluate this and other clinically relevant questions around the clinical utility of epigenomic cfDNA-based diagnostics, we strongly encourage incorporation of plasma collection into prospective clinical trials.

In summary, we observed widespread epigenomic reprogramming in tSCLC tumors relative to EGFRm LUAD tumors and leveraged these divergent molecular profiles to demonstrate for the first time the ability to non-invasively detect tSCLC in patients with EGFRm LUAD progressing on an EGFR TKI through epigenomic cfDNA analysis. With clinical validation, this epigenomic cfDNA-based approach to detect small cell transformation could usher in a new paradigm of diagnostic and therapeutic precision for patients with advanced lung cancer.

Y.P. Hung reports honoraria from Elsevier and American Society of Clinical Pathology on textbook writing and continuing medical education–related activity, both of which are unrelated to this study. N.R. Mahadevan reports stock ownership in AstraZeneca and Roche. D.A. Barbie reports personal fees from Qiagen/N of One and other support from Xsphera Biosciences outside the submitted work. Z. Piotrowska reports grants from NIH during the conduct of the study. Z. Piotrowska also reports personal fees from Eli Lilly, Boehringer Ingelheim, Bayer, Sanofi, C4 Therapeutics, and Taiho Pharmaceuticals; grants, personal fees, and other support from Janssen and AstraZeneca; grants and personal fees from Takeda, Cullinan Oncology, Daiichi Sankyo, and Blueprint Medicines; grants from Novartis, Spectrum Pharmaceuticals, AbbVie, GlaxoSmithKline/Tesaro, and Phanes Therapeutics; and grants and other support from Genentech/Roche outside the submitted work. T.K. Choueiri reports personal fees and other support from Precede Bio during the conduct of the study, as well as grants and other support from Precede Bio outside the submitted work; in addition, T.K. Choueiri has a patent for Precede Bio with royalties paid. T.K. Choueiri also reports institutional and/or personal, paid and/or unpaid support for research, advisory boards, consultancy, and/or honoraria past 5 years, ongoing or not, from Alkermes, Arcus Bio, AstraZeneca, Aravive, Aveo, Bayer, Bristol Myers Squibb, Calithera, Circle Pharma, Deciphera Pharmaceuticals, Eisai, EMD Serono, Exelixis, GlaxoSmithKline, Gilead, HiberCell, IQVA, Infinity, Ipsen, Janssen, Kanaph, Lilly, Merck, Nikang, Neomorph, Nuscan/Precede Bio, Novartis, Oncohost, Pfizer, Roche, Sanofi/Aventis, Scholar Rock, Surface Oncology, Takeda, Tempest, Up-To-Date, CME events (Peerview, OncLive, MJH, CCO and others), outside the submitted work; institutional patents filed on molecular alterations and immunotherapy response/toxicity, and ctDNA; equity from Tempest, Pionyr, Osel, Precede Bio, CureResponse, InnDura Therapeutics, and Primium; committees for NCCN, GU Steering Committee, ASCO (BOD 6-2024-), ESMO, ACCRU, and KidneyCan; medical writing and editorial assistance support may have been funded by communications companies in part; mentored several non-US citizens on research projects with potential funding (in part) from non-US sources/foreign components; and the institution (Dana-Farber Cancer Institute) may have received additional independent funding of drug companies and/or royalties potentially involved in research around the subject matter. T.K. Choueiri is supported in part by the Dana-Farber/Harvard Cancer Center Kidney SPORE (2P50CA101942-16) and Program 5P30CA006516-56, the Kohlberg Chair at Harvard Medical School and the Trust Family, Michael Brigham, Pan Mass Challenge, Hinda and Arthur Marcus Fund, and Loker Pinard Funds for Kidney Cancer Research at DFCI. S.C. Baca reports personal fees and other support from Precede Biosciences outside the submitted work. A.N. Hata reports grants and personal fees from Amgen, Nuvalent, and Pfizer; grants from BridgeBio, Bristol-Myers Squibb, C4 Therapeutics, Eli Lilly, Novartis, and Scorpion Therapeutics; and personal fees from Engine Biosciences, Oncovalent, TigaTx, and Tolremo outside the submitted work. M.L. Freedman reports personal fees and other support from Precede Biosciences outside the submitted work; in addition, M.L. Freedman has a patent for 'Methods, kits and systems for determining the status of lung cancer and methods for treating lung cancer based on same' pending. J.E. Berchuck reports non-financial support and other support from Precede Biosciences during the conduct of the study. J.E. Berchuck also reports grants, personal fees, and non-financial support from Guardant Health; personal fees and other support from Genome Medical; and other support from Oncotect, TracerDx, and Musculo outside the submitted work. In addition, J.E. Berchuck has an institutional patent on methods to detect neuroendocrine prostate cancer through tissue-informed cell-free DNA methylation analysis issued, licensed, and with royalties paid from Precede Biosciences and an institutional patent on methods to detect small cell lung cancer through epigenomic cfDNA analysis pending. No disclosures were reported by the other authors.

T. El Zarif: Data curation, formal analysis, investigation, writing–review and editing. C.B. Meador: Data curation, formal analysis, investigation, writing–original draft, project administration. X. Qiu: Software, formal analysis, investigation, methodology, writing–review and editing. J.-H. Seo: Formal analysis, investigation, methodology, writing–review and editing. M.P. Davidsohn: Investigation. H. Savignano: Investigation. G. Lakshminarayanan: Investigation. H.M. McClure: Investigation. J. Canniff: Investigation. B. Fortunato: Formal analysis. R. Li: Formal analysis. M.K. Banwait: Data curation. K. Semaan: Formal analysis. M. Eid: Investigation. H. Long: Supervision, methodology. Y.P. Hung: Data curation, writing–review and editing. N.R. Mahadevan: Writing–review and editing. D.A. Barbie: Conceptualization, resources, writing–review and editing. M.G. Oser: Conceptualization, resources, writing–review and editing. Z. Piotrowska: Resources, writing–review and editing. T.K. Choueiri: Resources, writing–review and editing. S.C. Baca: Software, methodology, writing–review and editing. A.N. Hata: Conceptualization, resources, supervision, funding acquisition, writing–review and editing. M.L. Freedman: Conceptualization, resources, formal analysis, supervision, methodology, writing–review and editing. J.E. Berchuck: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, investigation, methodology, writing–original draft, project administration, writing–review and editing.

J.E. Berchuck is supported by the Dana-Farber/Harvard Cancer Center Lung Cancer SPORE Research Development Award, the Department of Defense (W81XWH-20-1-0118, HT9425-23-1-0048), and the Dave Page Cancer Research Fund. C.B. Meador is supported by an Institutional Research Grant from the American Cancer Society (2022A019104). N.R. Mahadevan is supported by NCI/NIH (K08CA270077). J.E. Berchuck, C.B. Meador, D.A. Barbie, M.G. Oser, Z. Piotrowska, and A.N. Hata are supported by the LUNGSTRONG Foundation and Pan-Mass Challenge Team 3G. Z. Piotrowska and A.N. Hata are supported by the NIH (R01CA137008). S.C. Baca is supported by the Department of Defense (W81XWH-21-1-0358), the Damon Runyon Cancer Research Foundation, the Fund for Innovation in Cancer Informatics, and the Kure It Cancer Research Foundation. M.L. Freedman is supported by the National Institute of Health (R01CA262577, R01CA251555), the Claudia Adams Barr Program for Innovative Cancer Research, the Dana-Farber Cancer Institute Presidential Initiatives Fund, the H.L. Snyder Medical Research Foundation, the Cutler Family Fund for Prevention and Early Detection, the Donahue Family Fund, the Department of Defense (W81XWH-21-1-0339, W81XWH-22-1-0951), and the Movember PCF Challenge Award.