Liquid versus tissue biopsy for detecting acquired resistance and tumor heterogeneity in gastrointestinal cancers

Abstract

During cancer therapy, tumor heterogeneity can drive the evolution of multiple tumor subclones harboring unique resistance mechanisms in an individual patient^1–3. Prior case reports and small case series have suggested that liquid biopsy (specifically, cell-free DNA (cfDNA)) may better capture the heterogeneity of acquired resistance^4–8. However, the effectiveness of cfDNA versus standard single-lesion tumor biopsies has not been directly compared in larger scale prospective cohorts of patients following progression on targeted therapy. Here, in a prospective cohort of 42 patients with molecularly-defined gastrointestinal cancers and acquired resistance to targeted therapy, direct comparison of post-progression cfDNA versus tumor biopsy revealed that cfDNA more frequently identified clinically-relevant resistance alterations and multiple resistance mechanisms, detecting resistance alterations not found in the matched tumor biopsy in 78% of cases. Whole-exome sequencing of serial cfDNA, tumor biopsies, and rapid autopsy specimens elucidated substantial geographic and evolutionary differences across lesions. Our data suggest that acquired resistance is frequently characterized by profound tumor heterogeneity, and that the emergence of multiple resistance alterations in an individual patient may represent the “rule” rather than the “exception.” These findings have profound therapeutic implications and highlight the potential advantages of cfDNA over tissue biopsy in the setting of acquired resistance.

The inevitable emergence of acquired resistance is a major limitation of current targeted therapies⁹. Several key studies highlight the potential role that tumor heterogeneity plays in the emergence of resistance^{3,4,7,10–14}. In acquired resistance, the evolutionary pressure of therapy can drive outgrowth of distinct tumor subclones harboring independent resistance mechanisms in individual patients, within different metastatic lesions or within the same lesion^1,5,7,10,13. Genomic analysis of standard single-lesion tumor biopsies upon disease progression has been the mainstay of identifying mechanisms of acquired resistance, but recent studies suggest tumor biopsies may vastly under-represent the heterogeneity of resistance in a single patient^5,7,10,15,16. In particular, analyzing a core biopsy from one region of a single metastatic lesion may fail to detect clinically-relevant resistance mechanisms, leading to mixed responses or failure of subsequent therapy^3,7,17.

Liquid biopsy, specifically, cell-free DNA (cfDNA), may offer advantages for assessing tumor heterogeneity^4,18,19. Tumor-derived cfDNA, also termed circulating tumor DNA (ctDNA), is shed from tumor cells throughout the body. Therefore, cfDNA analysis can potentially identify multiple concurrent heterogeneous resistance mechanisms in individual patients that single-lesion tumor biopsies may miss^8,20–22. While case reports and small case series have suggested advantages of cfDNA^4,5,7,23, cfDNA and tumor biopsy have not been directly compared in larger scale prospective patient cohorts following progression on targeted therapy.

Therefore, we evaluated a prospective cohort of molecularly-defined gastrointestinal (GI) cancer patients who developed acquired resistance to targeted therapy. Through a systematic, disease center–wide liquid biopsy program, 42 patients who achieved stable disease or partial response to targeted therapy underwent liquid biopsy. When possible, a matched tumor biopsy at the time of eventual disease progression was obtained (23 patients). Patients encompassed seven molecular subtypes across three tumor types (Supplementary Table 1), offering a general assessment of cfDNA upon acquired resistance. To identify candidate acquired resistance mechanisms, cfDNA isolated from post-progression plasma was analyzed by next-generation sequencing (NGS) to identify emergent alterations not detected in pre-treatment tumor or cfDNA. Targeted NGS was performed in all cases, with parallel whole-exome sequencing (WES) when tumor fraction in cfDNA was sufficient (>5%). Only previously reported and functionally-validated resistance alterations were counted as resistance mechanisms^{5,7,8,12,13,15,21–35} (Supplementary Table 2).

Post-progression cfDNA identified at least one previously validated resistance alteration in 32/42 (76%) patients. Notably, 17/32 (53%) of these patients (40% of all) exhibited >1 detectable resistance alteration (range 2–9, median 3 per patient), suggesting frequent and profound tumor heterogeneity associated with acquired resistance (Fig. 1, Supplementary Table 3). In total, 78 clinically-relevant resistance alterations were found in cfDNA across multiple molecularly-defined tumor types receiving various targeted therapies.

Figure 1: Identification of acquired resistance mechanisms in liquid versus tumor biopsy.

A comparison of specific resistance alterations identified in plasma cfDNA (N=42) versus tumor biopsy (N=23) for each patient. Patients are grouped according to tumor type (N=3) (CRC = colorectal, GE = gastroesophageal, biliary) and molecular subtype (N=5) (FGFR2 = FGFR2 fusion, MET = MET amplification, HER2 = HER2 amplification, BRAF^V600E, RAS wild type. Red represents alterations identified in plasma, but not in the tissue biopsies; green represents alterations identified in tissue biopsies, but not in plasma; and purple represents alterations identified in both plasma and tissue biopsies. Pale purple represents alterations identified in plasma that were not detected in the post-progression tissue biopsy but were eventually detected in subsequent tissue biopsies from the same patient. The alterations detected in cfDNA versus tumor biopsy are quantified in a histogram across the top of the panel and are summarized graphically on the right, depicting specifically the percentage of patients with one, more than one, or no experimentally validated resistance alterations identified by cfDNA (top) or tumor biopsy (bottom).

To compare the effectiveness of cfDNA versus standard tumor biopsy, we analyzed matched post-progression tumor biopsies from 23 patients by WES and/or targeted NGS and compared to pre-treatment tumor tissue. Tumor biopsy identified resistance alterations less frequently than cfDNA in only 11/23 (48%) patients (Fig. 1, Extended Data Fig. 1), whereas cfDNA identified at least one resistance alteration in 20/23 (87%) patients, and in 76% of all patients. Moreover, multiple resistance alterations were identified in the post-progression biopsy in only 2/23 cases (9%). This contrasts with cfDNA, where 40% of patients harbored multiple resistance alterations, with up to nine alterations in a single patient. Overall, post-progression cfDNA identified additional clinically-relevant resistance alterations not detected in the post-progression tumor biopsies in 18/23 (78%) cases.

Conversely, only 1/23 (4%) post-progression tumor biopsy identified a resistance alteration that was not detected in matched post-progression cfDNA by NGS. In TPS125—a RAS wild type colorectal cancer (CRC) patient treated with the anti-EGFR antibody cetuximab—an EGFR^K489E mutation, affecting its antibody-binding domain, was identified in a post-progression biopsy at 11.9% allele frequency (cancer cell fraction: 78%). EGFR^K489E was not detected by clinical NGS in post-progression cfDNA, but was detected upon re-analysis using high-sensitivity ddPCR at 0.53% allele frequency (near the lower detection limit of clinical NGS assays for cfDNA, Supplementary Table 4). Thus, some low-level resistance alterations in rare tumor subclones may exist below the detection level in cfDNA with current technologies. Still, remarkably, every resistance alteration identified in a post-progression tumor biopsy was also detectable in matched cfDNA, suggesting that, while possible, missing resistance alterations by cfDNA analysis may be infrequent.

Interestingly, five cases with post-progression tumor biopsies had subsequent tumor specimens available for analysis, obtained during standard clinical care or through an institutional rapid autopsy program. In all cases, additional resistance mechanisms detected in post-progression cfDNA were later found in distinct metastatic lesions (Extended Data Fig. 2).

TPS037, a 53-year-old male with metastatic BRAF^V600E CRC, was treated with a combination of EGFR, BRAF, and PI3Kα inhibitors (cetuximab, encorafenib, and alpelisib; ) for 16 months before tumor progression¹⁵. Biopsy of a progressing liver lesion revealed the original BRAF^V600E mutation and an emergent KRAS^G12S mutation (Fig. 2a). Post-progression cfDNA analysis by clinical NGS (Guardant360®) showed not only the original BRAF^V600E and the emergent KRAS^G12S mutations, but also NRAS^Q61R and low-level EGFR amplification—all known resistance mechanisms in BRAF^V600E CRC^15,24,32. Subsequently, the patient underwent: (1) resection of a brain metastasis; (2) a second protocol-related liver biopsy; and (3) biopsy of a symptomatic subcutaneous soft-tissue lesion.

Figure 2: Comparison of multiple tumor biopsies versus liquid biopsy in a BRAF-mutant colorectal cancer patient.

(a) cfDNA (N=1) and tumor tissue biopsy specimens (N=4) included in the analysis for patient TPS037. For each specimen, the alterations and associated allelic fractions as determined by either ddPCR, targeted NGS, or WES are shown (Supplementary Tables 3–5). All ddPCR analyses were performed to a minimum coverage depth of 300X. Layered pie charts represent likely clonal composition of each specimen with the color of each subclone matching the color of the respective gene and branch in the phylogenetic tree. (b) Phylogenetic tree representing clonal architecture present in the specimens (using PhylogicNDT⁴⁴). The number of somatic alterations assigned to each cluster and detected events in known cancer genes appear on the branches. (c) Representative clonal and subclonal coding somatic alterations detected in each plasma or tissue specimen. Size of each square represents the estimated cancer cell fraction of each alteration with an empty box indicating no detection. Events in known cancer genes are highlighted in blue.

To better understand the subclonal architecture of resistance, post-progression tumor specimens and cfDNA were analyzed by WES. Consistent with prior reports, the original, likely clonal, BRAF^V600E mutation persisted in all specimens (Fig. 2a). KRAS^G12S was detected by WES in both liver biopsies, but not in the brain or soft-tissue metastases. Interestingly, the brain lesion harbored an EGFR amplification not detected in any other tumor specimen but detected at low level in cfDNA by clinical NGS. High-sensitivity ddPCR analysis of each tumor specimen also identified KRAS^G12S (allele frequency 0.17%) in the subcutaneous lesion and NRAS^Q61R (allele frequency 3.1%) in one liver biopsy (Supplementary Table 4). Thus, all three resistance alterations identified in cfDNA were identified in at least one of the post-progression tumor samples, though no single biopsy harbored all alterations. The clonal architecture and phylogenetic relationship of the four post-progression tumor biopsies and cfDNA by WES (Figs. 2b,c; Supplementary Table 5) showed shared clonal (truncal), private clonal, and subclonal alterations present across various samples. Consistent with the detectability in cfDNA of all resistance alterations present in the individual tumor biopsies, each clonal and the majority of subclonal mutation clusters found in different tumor biopsies were represented (at clonal or subclonal frequencies) in cfDNA, again highlighting the ability of cfDNA to capture heterogeneous molecular alterations present in distinct tumor lesions in an individual patient.

Similarly, TPS177, a 58-year-old woman with metastatic gastric adenocarcinoma harboring a fusion between CD44 (exons 1–8) and FGFR2 (exons 3–18), received the FGFR inhibitor Debio1347 (), achieving stable disease for four months before progression. Serial cfDNA analysis (Fig. 3a, Supplementary Table 4) revealed an initial decrease in the abundance of a clonal founder mutation RHOA^Y42C—representative of overall tumor burden—which rebounded rapidly upon progression and was accompanied by the emergence of four FGFR2 mutations: (1) FGFR2^V564L, affecting the gatekeeper residue in the inhibitor binding site, (2) FGFR2^L617V and (3) FGFR2^E565A—all known mechanisms of resistance to FGFR inhibition, and (4) FGFR2^S780L, which resides outside the kinase domain and may not represent a true resistance alteration⁷. As high-level FGFR2-fusion amplification was present (>50–150 copies), the existence of a potential passenger mutation would not be surprising. The patient underwent biopsy of a progressing paraaortic lymph node, but unfortunately expired due to disease complications. Rapid autopsy yielded 17 tumor lesions from various anatomic sites.

Figure 3: Serial liquid biopsy and autopsy in an FGFR2-fusion positive gastric cancer patient.

(a) The allele fraction of specific alterations in cfDNA isolated from serial plasma specimens during therapy with FGFR inhibitor were assessed by ddPCR to a minimum coverage depth of 300X. A truncal RHOA^Y42C mutation is shown in green. Emergent candidate resistance alterations are also shown. (b) Phylogenetic tree representing clonal architecture across all specimens. The number of somatic alterations assigned to each cluster and detected events in known cancer genes appear on the branches. FGFR2-fusion and amplification events have likely occurred in the cyan subclone 2 (based on expression and copy-number patterns). FGFR2^L617V mutations are found at high levels in subclones 4 and 9. (c) Clonal and subclonal alterations detected in each plasma (N=2), tumor biopsy (N=1), or autopsy specimens (N=17). Size of each square represents the estimated cancer cell fraction of each alteration with an empty box indicating no detection. FGFR2 expression (TPM), copy number (CN), and the presence of supporting CD44-FGFR2 fusion reads as determined by WES and RNAseq are shown, along with the number of reads for the two major FGFR2 resistance mutations (blue = mutant, green = wild type). Boxes with diagonal lines indicate no RNAseq data available. FGFR2 fusions were confirmed in the cfDNA samples based on off-target WES reads. (d) Diagram of the locations of autopsy specimens included in the analysis with likely clonal migration patterns across lesions. Layered pie charts represent the estimated clonal composition of each specimen with the color of each subclone matching the respective gene color in the branch of phylogenetic tree.

We analyzed WES data from pre-treatment and post-progression cfDNA, post-progression biopsy, and 17 autopsy specimens (Figs. 3b–d; Supplementary Table 5), and studied the clones and their phylogenetic structure with PhylogicNDT³⁶. Interestingly, FGFR2^V564L (detected in post-progression cfDNA) was found predominantly in liver metastases, whereas FGFR2^L617V (also in progression cfDNA) had an anti-correlated pattern, predominantly detected in other metastatic sites, suggesting these alterations confer resistance in different cell populations. FGFR2^E565A, detected in cfDNA, was identified at low level in a single liver lesion (D). Interestingly, autopsy specimens from the stomach (M/N/O), the primary tumor site, harbored only low levels of either resistance alteration. WES and RNAseq of stomach lesions also did not detect clear evidence of the CD44–FGFR2 fusion and showed markedly reduced FGFR2 expression levels and local copy number relative to other lesions (Fig. 3c). Phylogenetic analysis suggested that subclones present in these lesions likely represent early ancestors (Fig. 3b, upper right branch) that may have existed prior to development of the CD44–FGFR2 fusion–containing clone that later seeded the majority of metastases. Thus, in these stomach lesions, resistance to FGFR inhibition may have occurred primarily due to outgrowth of an early fusion-negative clone. FGFR2 fusion as a late event is consistent with the biology of gastric cancer, where RTK amplifications (e.g., ERBB2 or MET) can occur as late events, leading to outgrowth of non-amplified clones during targeted therapy^30,37–39. However, the presence of FGFR2 resistance alterations at low levels suggests that some fusion-positive subclones may have remained in these lesions, but at levels too low for detection by WES or RNAseq. Overall, all three FGFR2 mutations and the majority of private alterations present in individual tumors were detected in post-progression cfDNA (Figs. 3c,d).

Our data illustrate that acquired resistance to targeted therapy in GI cancers is highly heterogeneous, often with multiple resistance alterations per patient, and that liquid biopsy can effectively detect multiple resistance alterations residing concurrently in distinct tumor subclones and different metastatic lesions. Remarkably, a resistance alteration detected in a tumor biopsy but not in the matched cfDNA by clinical NGS occurred in only 1/23 cases. However, this alteration indeed existed in cfDNA and was detected using a higher sensitivity method. Thus, while cfDNA analysis may miss rare tumor subclones, our study suggests effective representation of alterations from multiple lesions in cfDNA. However, cfDNA detection effectiveness may differ by tumor type, metastatic site, or in tumors with low rates of ctDNA shedding.

This study has several potential limitations. Although it represents one of the largest direct comparisons of post-progression cfDNA and matched tumor biopsy to date, overall patient numbers remain limited. Thus, additional studies may further define the degree of heterogeneity upon acquired resistance and its relevance beyond gastrointestinal cancers. Additionally, in some cases, multiple biopsies and rapid autopsy confirmed that the multiple resistance alterations detected in cfDNA were derived from specific tumor lesions, but multiple tumor specimens were not available for most patients. However, in a comparator cohort of patients with similar clinical characteristics (Supplementary Table 6) with cfDNA analysis after progression on standard cytotoxic chemotherapy alone, without targeted therapy, we did not observe emergence of any of the clinically-relevant resistance alterations listed in Supplementary Table 2, suggesting that resistance alterations detected in cfDNA in our study represent tumor-derived alterations emerging under selective pressure from targeted therapy. Finally, cfDNA analysis is unlikely to detect non-genetic resistance mechanisms; in 24% of cases, no resistance mechanisms were detected by cfDNA or tumor biopsy. Although analysis of alterations in these patients revealed shared patterns of copy-number alterations with tumors having a known mechanism, a clear genomic driver of resistance could not be identified (Extended Data Fig. 3, Supplementary Table 7). Thus, tumor biopsy will still be key in assessing acquired resistance, particularly for novel or non-genetic resistance mechanisms.

In conclusion, direct comparison of cfDNA versus tumor biopsy illustrates how single-lesion tumor biopsies in the setting of acquired resistance frequently fail to identify the presence of multiple clinically-relevant resistance mechanisms, with cfDNA identifying additional concurrent resistance mechanisms in 78% of cases. Thus, our data suggest that single-lesion tumor biopsy alone is inadequate for characterizing acquired resistance, and that cfDNA regularly identifies heterogeneous clinically-relevant resistance alterations critical to the selection of subsequent therapy. This finding has important clinical implications: a treatment strategy aiming to overcome resistance will need to address the fact that a patient likely harbors multiple subclones with different resistance mechanisms. A recent study demonstrated the effectiveness of cfDNA analysis for selecting initial molecularly-directed therapy⁴⁰. Similarly, our results support that clinical incorporation of post-progression liquid biopsy may have a valuable role in efforts to assess and overcome acquired resistance.

ONLINE METHODS

Patients and specimen collection

All biopsies, tumor specimens, and peripheral blood draws for plasma isolation were collected in accordance with Institutional Review Board–approved protocols, to which patients provided written informed consent, and all studies were conducted in accordance with the Declaration of Helsinki. Patients provided informed written consent for blood and tissue collection and for genomic analysis of these specimens. Through a systematic liquid biopsy program in the Center for Gastrointestinal Cancers at the Massachusetts General Hospital Cancer Center, plasma collection for post-progression cfDNA analysis was attempted for all patients receiving targeted therapy for molecularly-defined GI cancer subtypes who achieved radiologic response or stable disease between September 2014 and March 2018. In total, 42 patients were studied. Peripheral blood was collected in two 10 mL Streck tubes upon disease progression for cfDNA analysis. When possible, post-progression tumor biopsies (solid and blood) were obtained in parallel at the time of progression, which were obtained in 23/42 patients. Pre-treatment tumor tissue was available for analysis in all 42 cases, and pre-treatment cfDNA in 19 cases. Rapid autopsies were performed within the first 3 hours postmortem. Targeted exome sequencing on clinical tissue specimens using a Clinical Laboratory Improvement Amendment (CLIA)–certified clinical NGS assay was performed in the Department of Molecular Pathology at the Massachusetts General Hospital. A subset of patients for whom tumor and cfDNA specimens harbored sufficient tumor DNA content were analyzed by WES, as described below. Imaging studies, including CT and MRI scans, were obtained as part of routine clinical care. A comparator cohort of consecutive GI cancer patients who did not receive targeted therapy, who underwent clinical NGS analysis of cfDNA (Guardant360®) as part of routine clinical care after disease progression on standard cytotoxic chemotherapy (including patients who received anti-angiogenic therapy, such as bevacizumab and two patients with no prior therapy), and who had comprehensive baseline clinical tumor NGS testing and provided informed written consent as above were also evaluated.

cfDNA isolation and analysis

Whole blood was collected by routine phlebotomy in two 10 mL Streck tubes. Plasma was separated within 1–4 days of collection through two different centrifugation steps (the first at room temperature or 10 minutes at 1600 x g and the second at 3000 x g for the same time and temperature). Plasma was stored at −80°C until cfDNA extraction. cfDNA was extracted from plasma using the QIAamp Circulating Nucleic Acid Kit (QIAGEN) according to the manufacturer’s instructions.

Clinical NGS analysis of cfDNA was conducted using the Guardant360® assay. Peripheral blood was collected in two 10 mL Streck tubes and shipped for centralized processing and testing (Guardant Health, Redwood City, CA) in accordance with standard clinical testing procedures. Sequencing libraries were prepared from up to 30 ng cfDNA with custom in-line barcode molecular tagging, and complete sequencing at 15,000 × read depth of the critical exons in a targeted panel of 70+ genes was performed at a Clinical Laboratory Improvement Amendments (CLIA)–certified, College of American Pathologists–accredited laboratory (Guardant Health)⁴¹. Targeted NGS of cfDNA for 226 cancer-related genes using the IRCC-Target panel was performed as previously reported^7,8. To identify candidate mechanisms of acquired resistance, in each case cfDNA sequencing data from plasma obtained post-progression was compared to sequencing data from pre-treatment tumor or cfDNA as available to identify emergent alterations not detected in the pre-treatment specimens.

For droplet digital PCR (ddPCR) experiments, DNA template (8 to 10 μL) was added to 10 μL of ddPCR Supermix for Probes (Bio-Rad) and 2 μL of the custom primer/probe mixture. This reaction mix was added to a DG8 cartridge together with 60 μL of Droplet Generation Oil for Probes (Bio-Rad) and used for droplet generation. Droplets were then transferred to a 96-well plate (Eppendorf) and then thermal cycled with the following conditions: 5 minutes at 95°C, 40 cycles of 94°C for 30 seconds, 55°C for 1 minute followed by 98°C for 10 minutes (Ramp Rate 2°C/sec). Droplets were analyzed with the QX200 Droplet Reader (Bio-Rad) for fluorescent measurement of FAM and HEX probes. Gating was performed based on positive and negative controls, and mutant populations were identified. The ddPCR data were analyzed with QuantaSoft analysis software (Bio-Rad) to obtain Fractional Abundance of the mutant DNA alleles in the wild-type/normal background. The quantification of the target molecule was presented as the number of total copies (mutant plus wild-type) per sample in each reaction. Allelic fraction is calculated as follows: AF % = (N_mut/(N_mut+N_wt))*100), where N_mut is the number of mutant alleles and N_wt is the number of wild-type alleles per reaction. A minimum coverage depth of 300X was achieved for each analysis. ddPCR analysis of normal control plasma DNA (from cell lines) and no DNA template controls were always included. Probe and primer sequences are available upon request.

Whole exome sequencing data

For whole-exome sequencing (WES), the AllPrep DNA/RNA Mini Kit (Qiagen, Hilden, Germany) was used for dual extraction of both genomic DNA and RNA. DNA was quantified in triplicate using a standardized PicoGreen® dsDNA Quantitation Reagent (Invitrogen, Carlsbad, CA) assay. The quality control identification check was performed using fingerprint genotyping of 95 common SNPs by Fluidigm Genotyping (Fluidigm, San Francisco, CA). Library construction from double stranded DNA was performed using the KAPA Library Prep kit, with palindromic forked adapters from Integrated DNA Technologies. Libraries were pooled prior to hybridization. Hybridization and capture were performed using the relevant components of Illumina’s Rapid Capture Enrichment Kit, with a 37Mb target. All library construction, hybridization and capture steps were automated on the Agilent Bravo liquid handling system. After post-capture enrichment, library pools were denatured using 0.1N NaOH on the Hamilton Starlet. Cluster amplification of DNA libraries was performed according to the manufacturer’s protocol (Illumina) using HiSeq 4000 exclusion amplification chemistry and HiSeq 4000 flowcells. Flowcells were sequenced utilizing Sequencing-by-Synthesis chemistry for HiSeq 4000 flowcells. The flowcells were then analyzed using RTA v.2.7.3 or later. Each pool of whole exome libraries was sequenced on paired 76 cycle runs with two 8 cycle index reads across the number of lanes needed to meet coverage for all libraries in the pool.

Phylogenetic analysis of multiple samples from the same patient

WES data were analyzed on the FireCloud cloud-based analysis platform (https://portal.firecloud.org/). Somatic mutations for each tumor/normal pair were detected using the Cancer Genome Analysis WES Characterization Pipeline available on FireCloud. The Characterization Pipeline includes multiple steps, including: MuTect⁴² for detection of somatic single nucleotide variants (SSNVs), Strelka⁴³ for detecting small insertions and deletions (INDELs), deTiN⁴⁴ estimates potential tumor-in-normal (TiN) contamination, ContEst⁴⁵ for detecting cross-patient contamination, AllelicCapSeg⁴⁶ for assessing allele-specific copy-number alterations, and ABSOLUTE⁴⁶ for estimating tumor purity, ploidy, absolute allelic copy-number and cancer cell fractions (CCFs). The data for this study consisted of multiple samples collected from each patient (pre- and post-treatment) as well as autopsy samples. Phylogenetic analysis, subclonal reconstruction and tree building was done with the PhylogicNDT package³⁶.

Tumor clustering based on patterns of commonly acquired alterations

To better characterize patients with an unknown mechanism of resistance, we wanted to identify groups of patients that share similar patterns of alterations that were acquired between the pre-treatment and post-progression samples. We define, for each patient with pre- and post-treatment samples, a set of features that reflect the changes in their somatic alterations between the pre-treatment and post-progression samples, which we call δ (Extended Data Fig. 3a). We focused the analysis on either a list of 609 known cancer genes (Cancer Gene Census, downloaded Oct 2017) or a set of 5,144 genesets (MSigDB class c2.cp and c5.bp, v6 with at most 200 genes). We considered both point mutations (SSNVs or INDELs) and somatic copy-number alterations. When analyzing point mutations, we set δ_gene=1 to reflect a significant change in CCF, i.e., if the change in CCF of the mutation in the gene was at least 2 standard deviations above the mean CCF change across all mutations and the post-progression sample had a CCF>0.2. Similarly, when analyzing copy-number, we set δ_gene=1 to reflect a significant copy-number change, i.e., if the change was greater than 2 standard deviations above or below the mean change in copy-number levels, and the gene was either amplified (copy-number level>4) or deleted (copy-number level <1) in the post-progression sample, respectively. To study genesets, we set δ_geneset to reflect the fraction of genes in the geneset that have a significant change in either copy-number or mutation CCF, accounting for the size of the geneset and the total number of significant genes in the patient.

Next, in order to find groups of patients that have similar changes in their somatic alterations, we applied the BiMax biclustering algorithm⁴⁷ to each of the four δ matrices: ${\mathrm{\delta }}_{\mathrm{c}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}\mathrm{r}\mathrm{g}\mathrm{e}\mathrm{n}\mathrm{e}}^{\mathrm{c}\mathrm{o}\mathrm{p}\mathrm{y}-\mathrm{n}\mathrm{u}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{r}},{\mathrm{\delta }}_{\mathrm{g}\mathrm{e}\mathrm{n}\mathrm{e}\mathrm{s}\mathrm{e}\mathrm{t}}^{\mathrm{m}\mathrm{u}\mathrm{t}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}},$ ${\mathrm{\delta }}_{\mathrm{g}\mathrm{e}\mathrm{n}\mathrm{e}\mathrm{s}\mathrm{e}\mathrm{t}}^{\mathrm{c}\mathrm{o}\mathrm{p}\mathrm{y}-\mathrm{n}\mathrm{u}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{r}}$, and ${\mathrm{\delta }}_{\mathrm{g}\mathrm{e}\mathrm{n}\mathrm{e}\mathrm{s}\mathrm{e}\mathrm{t}}^{\mathrm{c}\mathrm{o}\mathrm{p}\mathrm{y}-\mathrm{n}\mathrm{u}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{n}\mathrm{d}\mathrm{m}\mathrm{u}\mathrm{t}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}$. This analysis yielded between 3 and 11 biclusters (Extended Data Figs. 3b–e). To evaluate the significance of the resulting biclusters, we used a two-sided t-test to compare the mean size of biclusters observed in the matrix against results from n=10000 permuted matrices, where the values of the matrix were shuffled and biclustered.

Data availability

Sequencing data from WES and RNA-seq will be available in dbGaP in accession number phs001853.v1.p1.

Extended Data

Extended Data Figure 1: Phylogenetic trees for paired tumor biopsies.

Phylogenetic trees for patients with WES for pre- and post-treatment samples. The number of somatic alterations assigned to each cluster and detected events in known cancer genes appear on the branches. A 2-D plot showing the cancer cell fraction distribution of subclones in the pre- and post-treatment samples. Events in known cancer are shown next to their subclonal cluster.

Extended Data Figure 2: Patients with multiple post-progression tumor biopsies.

Three of five patients with multiple post-progression tumor biopsies are shown. The other two patients (TPS037 and TPS177) are shown in Figures 2 and 3, respectively. For each patient, the location of each tumor biopsy and the resistance alterations (blue) and truncal alterations (orange) detected in cfDNA and in each tumor specimen are shown. TPS001 is a RAS wild type colorectal cancer patient who developed resistance to an anti-EGFR antibody⁸. Biopsy of one liver lesion revealed an activating MEK1 (MAP2K1) mutation, and biopsy of a second liver lesion identified a KRAS mutation. Both resistance alterations were detected in post-progression cfDNA. TPS007 is an FGFR2-fusion positive cholangiocarcinoma patient who developed resistance to an FGFR inhibitor⁵. Five FGFR2 resistance mutations were identified in post-progression cfDNA, and three of these alterations were identified in distinct liver lesions harvested through a rapid autopsy program, with one lesion harboring two FGFR2 alterations. TPS011 is a RAS WT colorectal cancer patient who developed resistance to an anti-EGFR antibody. A recurrent colon tumor harbored a KRAS mutation, and an EGFR extracellular domain mutation--known to interfere with antibody binding--was identified in an ovarian metastasis, whereas both alterations were detected in cfDNA. Importantly, in all patients, individual resistance mechanisms emerging in distinct metastatic lesions were detectable in cfDNA.

Extended Data Figure 3: Biclusters of patients based on similar changes (δ) in somatic alteration.

Biclustering of four δ matrices reflecting changes in cancer cell fraction of mutations or copy-number in known cancer genes or genesets yielded significant biclusters (all empirical p-values<0.0001; Online Methods). The biclusters from all four δ matrices included at least one bicluster with patient TPS130, a patient with an unknown mechanism of resistance. Patient TPS130 consistently biclustered together with TPS021 and TPS037, patients with known mechanisms of resistance, across all matrices, highlighting the possibility that additional genomic alterations contribute to resistance beyond the identified resistance alterations. (a) The change in somatic alterations, δ, is calculated based on WES data of the samples closest to the start and end of therapy. We biclustered four δ matrices: ${\delta }_{cancergene}^{copy-number},{\delta }_{geneset}^{mutation},$ ${\delta }_{geneset}^{copy-number}$, and ${\delta }_{geneset}^{copy-numberandmutation}$ and assessed their significance by comparing the size of biclustersagainst n=10000 permuted matrices with a two-sided t-test (Online Methods). (b-e) Illustration of the biclustering results (using BiMax) of the four δ matrices (biclusters listed in Supplementary Table 6). Outlined in red are biclusters containing TPS130 observed in all four δ matrices.

Supplementary Material

1

Supplementary Table 1: Patient and clinical data

Supplementary Table 2: Validated resistance alterations used in study

Supplementary Table 3: Post-progression cfDNA data

Supplementary Table 4: ddPCR data for TPS037, TPS125, and TPS177

Supplementary Table 5: Coding gene alterations and their CCF and cluster IDs for all mutations for WES. Total Samples N=56. Pre-treatment tissue (N=10). Post-treatment tissue (N=32). Pre-treatment blood (N=5). (Post-Treatment blood) N=9.

Supplementary Table 6: Demographic and clinical data of comparison cohort of patients receiving chemotherapy alone

Supplementary Table 7: Biclustering results from four δ matrices with patients and features found in each bicluster.