Supplementary Materials for

Tumor-derived circulating endothelial cell clusters in colorectal cancer

Igor Cima, Say Li Kong, Debarka Sengupta, Iain B. Tan, Wai Min Phyo, Daniel Lee,

Min Hu, Ciprian Iliescu, Irina Alexander, Wei Lin Goh, Mehran Rahmani, Nur-Afidah

Mohamed Suhaimi, Jess H. Vo, Joyce A. Tai, Joanna H. Tan, Clarinda Chua, Rachel Ten,

Wan Jun Lim, Min Hoe Chew, Charlotte A.E. Hauser, Rob M. van Dam, Wei-Yen Lim,

Shyam Prabhakar, Bing Lim, Poh Koon Koh, Paul Robson, Jackie Y. Ying,

Axel M. Hillmer, Min-Han Tan*

*Corresponding author. Email: mhtan@ibn.a-star.edu.sg

Published 29 June 2016, Sci. Transl. Med. 8, 345ra89 (2016)

DOI: 10.1126/scitranslmed.aad7369

The PDF file includes:

Materials and Methods

Fig. S1. Device setup, retrieval efficiency, and sample output purity.

Fig. S2. scrmPCR methodology and proof of principle in cell lines.

Fig. S3. CD45− clusters express EMT markers, do not share common mutations

with the primary tumor, and have normal chromosomal structures.

Fig. S4. Cytomorphology of CD45− clusters used for RNA-seq.

Fig. S5. Cell-type inference workflow from RNA-seq data.

Fig. S6. Validation of the cell-type inference algorithm.

Fig. S7. Comparison between endothelial clusters (this study) and CTC clusters in

Aceto et al. (9).

Fig. S8. Circulating endothelial clusters are tumor-derived.

Fig. S9. tEC clusters in a mouse model of CRC.

Fig. S10. tEC clusters do not correlate with selected CRC tumor and patient

characteristics or comorbidities.

Legends for tables S1 to S14

References (38–44)

Other Supplementary Material for this manuscript includes the following:

(available at www.sciencetranslationalmedicine.org/cgi/content/full/8/345/345ra89/DC1)

Table S1 (Microsoft Excel format). scrmPCR primers used in this study.

www.sciencetranslationalmedicine.org/cgi/content/full/8/345/345ra89/DC1

Table S2 (Microsoft Excel format). Whole genome application (WGA) false-

positives are detected only at VAF < 10%.

Table S3 (Microsoft Excel format). Circulating CD45− clusters do not mirror

matching primary tumor mutations.

Table S4 (Microsoft Excel format). Sporadic mutations in circulating CD45−

clusters are not detected in matching primary tumor tissues.

Table S5 (Microsoft Excel format). Germline variants in tumor tissues and

circulating CD45− cell clusters (coverage and zygosity).

Table S6 (Microsoft Excel format). RNA-seq data: number of uniquely mapped

reads to hg19 exons.

Table S7 (Microsoft Excel format). RNA-seq data: processed data (FPKM) of all

samples described in the study.

Table S8 (Microsoft Excel format). Definition of the cell types used for the

inference algorithm.

Table S9 (Microsoft Excel format). Top 80 genes ordered from highest to lowest

specificity index (S) for each cell type selected for the inference algorithm.

Table S10 (Microsoft Excel format). Positive published RNA-seq control samples

used to validate the inference algorithm.

Table S11 (Microsoft Excel format). Characteristics of tEC clusters isolated in

this study.

Table S12 (Microsoft Excel format). Circulating endothelial cell cluster count

before and after surgery.

Table S13 (Microsoft Excel format). Baseline patients’ and healthy donors’

characteristics.

Table S14 (Microsoft Excel format). CD45− cluster count for each baseline

sample type and number of single clusters analyzed for each technique in the

corresponding samples.

Materials and Methods

Clinical samples processing and data collection

All samples were collected in EDTA Vacutainer tubes (Becton-Dickinson), and processed within

6 hr at IBN. Two CRC cases were excluded from the analysis because of a technical issue with

the microfiltration. Wherever available, matched tumor and metastatic samples, along with

normal mucosa sampled at least 5 cm away from the tumor, were immediately frozen after

resection, and stored at −80C until use. Clinicopathologic data for participating subjects are

described in table S13, and were collected retrospectively after completion of circulating clusters

counts. Furthermore, table S14 includes circulating clusters counts for each baseline sample and

includes number of single clusters analyzed for each technique and patient throughout the study.

Clinical data was collected without prior knowledge of clusters counts. Similarly, clinical data

for CRC patients were not known when determining clusters counts, except for the diagnosis and

preoperative status of all FSH samples.

Device assembly

The retrieval device was assembled as follows. Firstly, the rubber plug of a 3-ml syringe plunger

was removed, and a 5.5 mm-diameter hole was created using a punch cutter. The perforated

rubber plug was placed in the 3-ml syringe. Next, an O-ring was placed in the slot of the sleeve

insert, followed by the microsieve and another O-ring. Finally, the sleeve insert, together with

the microsieve and O-rings were placed in the 3-ml syringe above the perforated rubber plug.

This arrangement enabled the microfiltration of cells by size from the whole blood, and the

subsequent retrieval of captured cells from the upper surface of the microsieve.

Microfiltration optimization

To optimize blood microfiltration with microsieve, cells labeled with 5 µM CellTracker (Life

Technologies) were added to donor blood at 10–50 cells per ml of whole blood. Blood was

microfiltered at various flow rates by means of a peristaltic pump (Ismatec). After 6 washes with

PBS, 0.5% BSA and 2 mM EDTA, cells were resuspended in culture medium. Cell nuclei were

then stained using Hoechst 33342 (Life Technologies), and cells were retrieved to determine

retrieval efficiency and fold-depletion of contaminating WBCs. In some experiments, we also

counted CellTracker-positive cells remaining on the microsieve. Percent retrieval efficiency was

calculated as follows:

% Retrieval Efficiency = (Retrieved cells) × 100 / (Spiked Cells)

Fold depletion from 1 ml whole blood was calculated as follows:

Fold Depletion = (WBCs in Whole Blood) / (WBCs in Microfiltrate)

WBC count in the microfiltrate was defined as the number of any Hoechst 33342

positive, CellTracker-negative event in the case of experimental enrichment, or as any CD45-

positive event in the case of clinical samples. All clinical samples were immediately processed in

downstream applications using the optimized parameters described in Figure 1C and D. To

estimate the ideal target WBC depletion, we performed single-cell micromanipulation on serial

dilution of PBMCs containing 50 CellTracker-positive HCT 116 cells. Micromanipulation of

pure HCT 116 cells without contaminating white blood cells was easily performed after a 5,000-

fold dilution of PBMCs. The ideal target retrieval efficiency was based on a literature search of

label-free CTC isolation devices (17). Microfiltration of clinical samples was performed using 2

ml of whole blood for each microsieve device and the optimized microfiltration conditions.

Single cell and single cluster micromanipulation

Cells and clusters were identified using bright field or phase contrast imaging, nuclear staining

and specific fluorescence signals. Target single cells or clusters were then manually

micropipetted in a 10-µl droplet of wash buffer using a mouth pipette operated through a 25-ml

syringe. Subsequently, single cells or clusters were deposited in 0.2-ml PCR tubes containing an

appropriate buffer: 5 µl of 2× Reaction buffer (CellsDirect One-Step qRT-PCR Kit, Life

Technologies) for scrmPCR; 2 µl of PBS for whole genome amplification; or 2 µl of SuperBlock

buffer (Thermo Scientific) for low-input RNA-seq. Cells were immediately stored at −80 C

until use. In some cases, the complete microfiltrate was spun down without micromanipulation

and stored at −80 C until further use.

Cell lines and culture

HCT 116, COLO 201, SW480, SW620, DLD-1, RKO, and CT26 colorectal cancer cell lines and

BJ-5ta immortalized human foreskin fibroblasts were purchased from ATCC and cultured in

DMEM (Life Technologies) supplemented with 10% FBS. HUVECs were purchased from

ATCC, used at passage 1 to 3, and cultured in EGM-2 medium (Lonza). HDMECs and HPrECs

were purchased from ScienCell, cultured in Endothelial Cell Medium (ScienCell) on fibronectin-

coated dishes, and used at passage 1 for RNA-seq analysis and at passage 5 for the FOLH1

analysis. iPSCs were donated by Y. Chouduri (Institute of Bioengineering and Nanotechnology).

All cells were maintained in a humidified incubator at 37 C in the presence of 5% CO2.

Reagents for on-sieve immunofluorescence

Following fluorescence-labeled antibodies were used: anti-CD45 1:200 (clone 2D1;

eBioscience), anti-EpCAM 1:20 (9C4, BioLegend), anti-CD31 1:20 (WM59, BioLegend), anti-

CD144 1:10 (55-7H1, BD Biosciences), anti-CD41 1:20 (HIP8, BioLegend), anti-CD42B 1:20

(HIP1, BioLegend), anti-CD133 1:20 (AC133, BD Biosciences). Live iPSCs in suspension were

used as positive controls for the CD133 staining. Fibrin was stained using anti-Fibrin 1:200

(Clone UC45, Pierce antibodies) and secondary goat anti-mouse IgG/IgM conjugated to Alexa

488 1:1000 (Pierce antibodies). Fibrin-positive controls included fibrin gels obtained by mixing

20 µl of fibrinogen droplets (20 mg/ml in PBS, pH = 7.2, Sigma-F3879) deposited on a well of a

96-well plate, mixed with 2 µl of thrombin (50 U/ml, Sigma-T4648) and incubated for 10 min at

room temperature. For intracellular antigens, the Inside Stain Kit (Miltenyi Biotec) comprising a

fixative and a permeabilization solution was used together with human FcR Blocking Reagent

with the following antibodies: anti-VWF 1:200 (rabbit polyclonal A 0082, DAKO, conjugated

in-house to Alexa 488 or Alexa 555 using Life Technologies APEX Antibody Labelling Kit),

anti-Vimentin (V9, Santa Cruz Biotechnology), and anti-pan Cytokeratin (C11, Cell Signaling

Technology). Nuclei were stained using Hoechst 33342 (Life Technologies). In some

experiments, Calcein AM (Life Technologies) was used to identify live cells.

Nucleic acid extraction

Total RNA from tissues was isolated using the RNeasy mini kit (Qiagen). DNA from tissues was

isolated using the DNeasy mini kit (Qiagen). Complete microfiltrates or laser-capture

microdissected (LCM) samples were subjected to RNA extraction using the RNAqueous-Micro

Total RNA Isolation Kit (Ambion) following the manufacturer’s instructions. Single cells and

single clusters did not undergo nucleic acid extraction, but were lysed and used directly in

downstream assays as described for the scrmPCR method, DNA sequencing and aCGH and

RNA-seq.

scrmPCR

This method can simultaneously detect and quantify RNA transcripts and sequence DNA

hotspots in the same cell. Briefly, single-cell RNA transcripts were reverse transcribed at 50 C

for 30 min using SuperScript III Reverse Transcriptase (Invitrogen) and a mix of 500 nM target

reverse primers. A preamplification round was then performed using Platinum Taq DNA

polymerase (Invitrogen) by adding a matching mix of forward primers to the transcript-specific

reverse primers and primer pairs for targeted genomic regions. Preamplification cycling was

conducted by alternating annealing and denaturation steps without extension as follows: 6 cycles

at 60 C, 4 min, 95 C, 1 min; 6 cycles at 55 C, 4 min, 95 C, 1 min; 6 cycles at 50 C, 4 min,

95 C, 1 min. In some cases, we used 18 cycles at 60 C, 4 min, 95 C, 1 min. Primer clean-up

was performed using the Axyprep PCR Clean-up Kit (Axygen) or using Exonuclease I

enzymatic digestion (Thermo Fisher). Samples were diluted 1/20 and stored at −20 C until

further use.

To quantify RNA transcripts, quantitative PCR was performed on a ViiA7 Instrument

(Applied Biosystems) using 2 µl of the preamplification mixture, semi-nested primer pairs

appropriate for the target transcript (table S1) and the SensiFAST SYBR Lo-ROX Kit (Bioline)

according to the manufacturer’s protocol. Relative gene expression was normalized using ACTB

as the reference gene. Single cells or single clusters that did not show ACTB expression

quantifiable in < 25 cycle threshold (Ct) values were discarded from the analysis. To analyze

selected DNA mutational hotspots, PCR was performed using 2 µl of the preamplification

mixture, nested PCR primer pairs (table S1) and a master mix containing a proof-reading

polymerase (KOD Hot Start Master Mix, EMD Millipore) in line with the manufacturer’s

instructions. For KRAS exon 2 sequencing of tumor and normal tissues shown in Fig. 2C, PCR

amplification was performed using a primer pair different from the ones indicated in table S1):

TTTGTATTAAAAGGTACTGGTGGAG as forward primer, and

CCTTTATCTGTATCAAAGAATGGTC as reverse primer. PCR products were separated on

agarose gel; specific bands were excised and sequenced using the Sanger method.

Droplet digital PCR (ddPCR)

ddPCR was performed using the BioRad QX200 system and manufacturer instructions. Briefly,

primers and probes were designed to target variants detected in three clusters from three different

patients (fig. S3, B and C). For each sample, the PCR reaction was performed in duplicate using

450 nM forward and reverse primers, 250 nM probes targeting the mutant and the wild type

alleles, 1x of ddPCR Supermix and 5ng of HindIII-digested DNA template using the following

thermal cycling protocol: 95°C for 10 min; 40 cycles at 94°C, 30 s and 63°C, 1 min; 98°C for 10

min. Results were analyzed on the BioRad QX200 Droplet Reader using the QuantaSoft

software. Following primers and probes were used: AKT c.G27A: Forward primer

TACTACCCCCATCTCTCCCT, reverse primer CTGAGAGGAGCGCGTGAG, probe 1

(mutant) FAM-ACCTCGTTTGTGCAGCCAACCTTCCT-BLACK HOLE, probe 2 (wildtype)

HEX-ACCTCGTTTGTGCAGCCAACCCTCCT-BLACK HOLE; KRAS c.G196A: Forward

primer ATACACAAAGAAAGCCCTCC, reverse primer CAGACTGTGTTTCTCCCTTC,

probe 1 (mutant) FAM-CCCTCATTGTACTGTACTCCTCTTGACCTG-BLACK HOLE, probe

2 (wild type) HEX-CCCTCATTGCACTGTACTCCTCTTGACCTG-BLACK HOLE; NRAS

c.G97A: Forward primer TGGGTAAAGATGATCCGACA, reverse primer GACTGAGTACA

AACTGGTGG, probe 1 (mutant) FAM-CTGGGCCTCACCTCTATGGTGGGATTATAT-

BLACK HOLE, probe 2 (wild type) HEX-CTGGGCCTCACCTCTATGGTGGGATCATAT-

BLACK HOLE.

EPC assay

Colony-forming EPC assays were performed as previously described (22). Live endothelial

clusters were counted in 2-ml microfiltrates after CD144 and Calcein AM fluorescence staining.

Unstained microfiltrates from 2 ml of blood from a second device were then placed in culture on

a 96-well plate coated with fibronectin (1 µg/cm2) (Sigma-Aldrich) in the presence of EGM-2

cell culture medium (Lonza). The presence of clusters was confirmed by bright field microscopy

before incubation. HUVECs were used as positive controls after spiking 10,000 HUVECs into 2

ml of donor blood and isolating them by microfiltration using two microsieve devices. In one

device, retrieved HUVECs were quantified by CD144 and Calcein AM staining. HUVECs

retrieved from the other device were seeded at 5, 10, 20, 40, 80, or 160 cells in octuplicate wells.

After 2 days, the medium was changed and cell cultures were maintained for a total of 30 days,

with half of the medium changed every other day. The presence and viability of colonies were

monitored every week under bright field microscopy. After 30 days, cells were stained using

CD144 antibodies, Calcein AM and Hoechst 33342, and quantified under an IX81 (Olympus)

inverted fluorescence microscope.

D-dimer ELISA

D-dimer ELISA was performed using a RayBio human D-dimer ELISA kit (Raybiotech) with

manufacturer instructions. Plasma samples were diluted 1:2,000,000 in the provided buffer and

incubated overnight at 4°C.

Isolation of endothelial cells from fresh tissues

Tissues were minced before digestion with collagenase, dispase and DNAseI for 60 min at 37C.

After Ficoll-Paque density centrifugation, a two-step magnetic selection was performed using

MACS reagents and materials (Miltenyi Biotec) in line with the manufacturer’s instructions.

First, CD45-expressing cells were depleted by retention on LD columns after labeling the cells

with anti-CD45 magnetic beads and human FcR Blocking Reagent. Next, the CD45-depleted

fraction was collected and re-labeled by adding anti-CD31 magnetic beads and human FcR

Blocking Reagent. A fraction enriched for CD31+CD45– cells was harvested on MS columns

before elution and storage at −80 C until further use.

Laser capture microdissection of endothelial cells from archival frozen tissues

Eight-µm tissue sections were cut from fresh frozen samples of tumor and matched normal

tissues, and placed on PEN-LCM membrane slides (Thermo Fisher). Slides were washed twice

in buffer (PBS/EDTA/1%BSA/RNAse inhibitor). After a 10-min peroxidase blocking step using

Peroxidase Blocking Reagent (DAKO) and three washes in buffer, slides were stained for 20 min

with anti-CD31 antibodies at 1 µg/ml in buffer (WM59, BioLegend). After three washing steps,

bound antibodies were visualized using the REAL EnVision Detection System (Dako Cat. 5007)

in line with the manufacturer’s instructions. Laser capture microdissection was immediately

performed on the dry slides using the Arcturus LCM system (Thermo Fisher) following

manufacturer’s instructions. Endothelia were captured on adhesive CapSure caps (Thermo

Fisher) and stored at −80°C until further use. RNA was isolated using RNAqueous-Micro Total

RNA Isolation Kit (Ambion) and stored at −80C until further use. Low-input RNA-seq libraries

and sequencing were performed as described using 10 pg of RNA from each sample.

Cell type inference method

First, we generated a map of specific genes for each cell type of interest. The primary cell atlas

data set (GSE49910) (38) was used for this purpose. From this data set we extracted 215

experiments of interest, corresponding to n = 36 different cell types or ‘lineages’ (table S8)

without reprocessing of the data. Technical replicates were averaged. For each gene, g, in each

lineage, l, a ‘specificity index’, 𝑆, was calculated based on Shannon entropy and the Q statistics

introduced by (39),

𝑆(𝑙|𝑔) = − ∑ 𝑝(𝑙|𝑔)

𝑁

𝑙=1

⋅ log2(𝑝(𝑙|𝑔)) − log2(𝑝(𝑙|𝑔))

where 𝑝(𝑙|𝑔) is the relative expression of gene g in lineage l. Gene specificity was confirmed by

visualizing gene expression data with high specificity indices using the primary cell line data

base from BioGPS (40) (this data base is also dervied from GSE49910) (fig. S6A). For each cell

type, the top 80 genes with the highest specificity index (‘specific genes’) were selected and

were reported in table S9, which represents the map of specific genes for each lineage of interest.

Next, for each RNA-seq query sample, we set a predefined threshold to define expressed genes.

In our dataset, we used FPKM > 0 as expressed genes. For the positive controls in figs. S6 and

S7, we used the cut-offs reported in the corresponding studies. Next, in the query RNA-seq list

of expressed genes, we counted the occurrences of the top 80 specific genes for each cell type

present in our map of specific genes (table S9). To determine if the number of enriched genes

was different from enrichment by chance, we generated 1,000 lists of 80 randomly selected

genes from the Affymetrix HG-U133_Plus_2 gene symbol list (‘random genes’) and counted the

average number of genes present by chance in each experimental RNA-seq profile for each

lineage. Lastly, for each cell type in each experimental sample, a Fisher’s exact test was applied

to determine whether the number of enriched specific genes was equal to the number of

randomly enriched genes. The odds ratios for each test were mean-centered, scaled and

visualized in a heat map comprising all tested cell types. The final results were used to generate

hypotheses of cell types based on the distribution of the normalized odds ratios. The algorithm

was validated using lists of expressed genes from published RNA-seq data from various cell

types and tissues (fig. S6C).

Principal component analysis on RNA-seq data

Principal component analysis (PCA) was performed in the R environment (version 3.1.0) (32) on

RNA-seq data set of single clusters and tissue samples (table S7). The prcomp function was used

with Log2(FPKM+1) data and with default parameters after excluding variables (genes) with 0

standard deviation and quantile normalization of the samples.

Differential gene expression (DGE) and linear discriminant analysis (LDA) of RNA-seq

data

DGE analysis and LDA of RNA-seq data from laser-dissected nECs and tECs was performed by

‘NOISeq’ (41) and the lda function of the MASS package respectively, in the R environment

(version 3.1.0). For DGE, input data were log2(FPKM+1) normalized before analysis. To obtain

differentially expressed genes, we applied the noiseqbio function with default parameters, except

for using r = 100 permutations and setting the smoothing parameter adj to 0.5. The list of genes

presented in fig. S8D was selected using NOISeq probability >0.8 (corresponding to a q-value

<0.2) and no filter on the log2 fold change (log2FC) data. For LDA, the data corresponding to the

list of differentially expressed genes in the LCM samples were extracted from the log2

normalized RNA-seq dataset. Variables with >80% missing values [log2(FPKM+1) = 0] in the

clusters data were excluded from the analysis (CDH2 and ACAT). A linear discriminant model

using the remaining 6 variables, ST6GALNAC2, WDR36, ARMC10, PSMB2, STOM, and

NUCKS1 was computed on the nECs and tECs data using the lda function of the ‘MASS’

package, after verifying the absence of multicollinearity (Pearson’s r < 0.8 for each variable pair)

and the assumptions of normality and homogeneity of the covariance matrices by the Box’s M

test (P = 0.55). The model accuracy was tested using leave-one-out cross-validation (LOOCV)

on the same data set and was reported in Fig. 4A. Finally, the model was applied to classify each

single cluster. To calculate the expected probability of cluster classification by chance, we

generated a series of 1000 random 6-gene signatures with Box’s M test P values > 0.1 and

identical missingness probabilities, and tested each signature using the workflow described for

the 6-gene signature. The exact binomial test was used to compare the observed probability (Po)

of classified clusters as tECs with the mean expected probability (Pe) of the randomly generated

signatures classified as tECs.

Microvessel analysis and lumen count

Briefly, after scanning of the tissue sections using an Ariol SL-50 scanning platform Leica

Biosystems, the tumor regions for all sections were marked and quantified. Within these regions,

all areas with the greatest number of distinctly highlighted micro-vessels (hotspots) were marked

and quantified. Next, nine images covering all hotspot areas were analyzed at 20× magnification

for microvessel density (MVD), lumen counts and lumen size. MVD was determined using an

automated ImageJ macro as follows. Ruifrok’s color deconvolution was applied to extract the

CD31-positive signal. After image binarization, a common threshold level was chosen to

represent correct vascular morphology and minimize background levels across all slides.

Morphological post-processing steps were applied to join nearby objects, split objects and close

gaps within objects similarly as described (36). Objects larger than a minimal size (default: 105

pixels, corresponding to 50 µm2) were counted as microvessels. Smaller size objects were

considered as noise and excluded from the analysis. The ImageJ macro was validated by

comparing it with manually counted images (Pearson’s r = 0.951, n = 30 images). Lumen count

and perimeters were determined manually for each image using ImageJ measurement and count

features. Mosaic and damaged vessels were counted in the tumor area and in available normal

tissues present on the same slide by screening each section using a 20× magnification. False-

positive stains (e.g. hematopoietic cells positive for CD31) were excluded from the analysis.

These cells were identified in each tissue by morphology and lower staining intensity compared

to microvessels. Mean values of 9 images per section were used for statistical assessments.

Patient’s IDs and clusters counts were blinded during the analysis to avoid subjective bias during

acquisition of the data.

Tail vein injection of tumor-derived endothelial cells in experimental mice. CT26 cells

(500,000 in 100 µl of sterile PBS) were subcutaneously injected into the left flank of 7-week-old

Balb/c mice. At 14 days after injection, mice were euthanized and the tumors (about 2 cm-

diameter) were isolated for downstream processing. Tumor samples were embedded in optimal

cutting temperature compound (OCT, Sakura Finetek). The remaining tumor tissues were

minced using scalpel blades and digested with a solution of collagenase and dispase (Liberase

DH, Roche) for 2 hr at 37 C in cell culture medium with occasional stirring. After two washing

steps, the tissues were incubated with DNase I (Thermo Fisher) for 15 min, and filtered through a

40-µm cell strainer to obtain single-cell suspensions. Cells were stained for 45 min at 4 C using

1 µg/ml of allophycocyanin (APC)-labeled anti-CD309 antibodies (Avas12a1, eBioscience),

diluted in PBS/EDTA/1%BSA or matching Rat IgG2a K Isotype control (eBioscience), in the

presence of anti-mouse FC Receptor Block (eBioscience). After two washing steps, cells were

stained with Propidium Iodide (PI) Readyprobes (Thermo Fisher). Cell suspensions were then

sorted using a MoFlo XDP Cell Sorter to obtain a total of 500,000 CD309+PI- single cells. Cells

were subsequently divided into two equal fractions. One fraction was labeled using Celltracker

Orange (Invitrogen), while the other fraction was labeled using Celltracker Green (Invitrogen).

Cell suspensions were mixed and 50,000 cells were injected into the tail vein of nine 7-week-old

Balb/c mice. Mice were euthanized at the indicated time points, and the presence of single

fluorescence events (single cells) or dual fluorescence events (clusters) were analyzed in blood

samples by a FACS Calibur instrument (BD Biosciences), after lysis of the erythrocytes fraction

using BD Pharm Lyse (BD Biosciences) and following manufacturer’s instructions. The acquired

data were analyzed using FlowJo version 10.0.7 (Tree Star).

Tumor-derived endothelial cell clusters in experimental mice. CT26 cells (500,000 in 100 µl

of sterile PBS) were subcutaneously injected into the left flank of two 6-week-old

L2G85.BALB/c mice (The Jackson Laboratory). This mouse strain carries the GFP transgene

under the ubiquitous CAG promoter. After 14 days, mice were euthanized and tumors were

subcutaneously transplanted into five wild-type 6-week-old Balb/c mice. Fourteen days later, the

mice were euthanized, and the blood samples were immediately collected in EDTA Vacutainer

tubes (Becton-Dickinson) by cardiac puncture. Tumor tissues were collected and dissected in

three segments. Each segment was tested for the presence of the GFP transgene by genotyping

its genomic DNA with the recommended protocols and the following primers: oIMR0872

AAGTTCATCTGCACCACCG as GFP forward primer, oIMR1416

TCCTTGAAGAAGATGGTGCG as GFP reverse primer, oIMR7338

CTAGGCCACAGAATTGAAAGATCT as internal positive control forward primer, and

oIMR7339 GTAGGTGGAAATTCTAGCATCATCC as internal positive control reverse primer.

One mouse was excluded from subsequent analysis because of the absence of GFP signal in any

one of the three segments. Blood samples were processed within 2 h of collection. Erythrocytes

were lysed using BD Pharm Lyse (BD Biosciences) and following manufacturer’s instructions.

Cells were washed in buffer (PBS/EDTA/1%BSA), and stained using 1 µg/ml of anti-CD309-

APC (Avas12a1, eBioscience) or matching rat IgG2a K Isotype control (eBioscience) and

Hoechst 33342, in the presence of anti-mouse FC Receptor Block (eBioscience). Stained cells

were analyzed using a LSR II instrument (BD Biosciences) with independent laser sources for

each fluorescent signal. Acquired data were analyzed using FlowJo version 10.0.7 (Tree Star).

Blood from three control Balb/c mice was used as negative control.

Supplementary Figures:

Fig. S1. Device setup, retrieval efficiency, and sample output purity. (A) Device setup.

Microfiltration devices enclosing a silicon microsieve (inset scale bar, 10 µm), connected to a

peristaltic pump (B) Capture efficiency of SW620 cells from whole blood, indicating percentage

of captured cells on the microsieve that could be retrieved for downstream assays, or that were

lost because the cells adhered to the microsieve. Four independent experiments are shown. (C)

Size distribution of SW620, HCT 116, SW480, and COLO201 cells (n = 50 cells each). Arrows

indicate median size of WBCs and CTCs isolated from colorectal, prostate and breast cancer

patients respectively, reported in (14). (D) Contaminating nucleated cells after microfiltration of

clinical samples using the optimized protocol shown in Fig. 1, B and C (n = 13 blood samples).

(E) Retrieval efficiency from additional cell lines using optimized protocol shown in Fig. 1, B

and C. Data are means ± SEM (n = 3 experiments for each cell line).

Fig. S2. scrmPCR methodology and proof of principle in cell lines. (A) scrmPCR workflow

for single cells or single clusters. (B) Concurrent DNA amplification (gel electrophoresis) and

RNA quantitation (heat map) for the indicated genes and representative single or bulk cells. (C)

Sequence chromatograms from DNA hotspots of known mutant alleles in RKO and DLD-1

single cells. (D) Melting curve plots for the indicated genes from qPCR reactions. sc, single cell;

NTC, no-template control; Rn, normalized relative to reporter signal.

TP53 (exon5)

TP53 (exon6)

TP53 (exon7)

TP53 (exon8)

BRAF (exon15)

DL

D1

B

DL

D1

SC

1

DL

D1

SC

2

DL

D1

SC

3

NT

CCDH1

SERPINE1

EPCAM

K

CDX2

DL

D1

SC

3

RK

RK

RK

RK

KRAS (exon2)

NT

C

DLD-1 RKO

Bu

lk

sc#

1

sc#

2

sc#

3

Bu

lk

sc#

1

sc#

3

sc#

2

CDH1

SERPINE1

EPCAM

KRT19

CDX2

SE

RP

INE

1

20

10

0 10

20

Relative gene expression

(normalized to ACTB)

-20 0 -10 10 20

bulk

sc#

1

RKO

BRAFmut DLD-1

KRASmut TP53mut

sc#

1

bu

lk

#"%!"#!"%$"#

KRT19 CDH1 ACTB

CDX2 EPCAM

Temperature (°C)

-Rn

’ -R

n’

AG AGA GGT GAG TTTCT

SERPINE1

A

Single-cell/cluster

Gene

expression

DNA

mutations

QPCR Proof-reading PCR

& Sanger seq.

On

e-t

ube

re

actio

n Lysis

Targeted reverse transcription

Multiplex preamplification

of both DNA and cDNA targets

B C

D

Fig. S3. CD45– clusters express EMT markers, do not share common mutations with the

primary tumor, and have normal chromosomal structures. (A) Representative four-color

immunofluorescence staining for CD45, vimentin (VIM), pan-cytokeratin (CK), and DAPI,

indicating heterogeneous mesenchymal and epithelial marker expression by 2 clusters from two

WG

Acon

trol

No

rma

l tissue

DN

A

WG

Acon

tro

l

Tum

or

tissu

e D

NA

Chromosomes (1-22)

pa

nC

KV

IMC

D45

Sin

gle

Clu

ste

r

Sin

gle

Clu

ste

r

Cl.5

Cl.1

Cl.2

Cl.3

Cl.4

Cl.5

Sin

gle

Clu

ste

r Cl.6

Cl.8

Cl.9

D Patient 7

Patient 13

Patient 15

* **

Tu

mo

r

tissue

Tum

or

tissue

*** *

Tu

mo

r

tissu

e

* * * * *** ** * *

Norm

al

tissu

e

No

rmal

tissue

Norm

al

tissu

e

Cluster#1 Cluster#2 B

02000400060008000

10000

02000400060008000

10000

Ch

an

nel 1

am

plit

ud

e

Channel 2 amplitude

0

1000

2000

3000

4000

5000

6000

P10-Cl.5

Tumor tissue

Normal tissue

Blank

Target Sample Replicate #1 Replicate #2

P08 tumor 0 0

P08 normal 0 0

P08-Cl.12 4.1 3.3

P25 normal 0 0

P86 normal 0 0

P87 normal 0 0

Blank 0 0

P10 tumor 0 0

P10 normal 0 0

P10-Cl.5 14.5 15.9

P25 normal 0 0

P86 normal 0 0

P87 normal 0 0

Blank 0 0

P13 tumor 0 0

P13 normal 0 0

P13-Cl.5 43.2 40.9

P25 normal 0.08 0

P86 normal 0 0

P87 normal 0 0.13

Blank 0 0

Fractional abundance (%)

KRAS

c.G196A

AKT

c.G27A

c.G97A

NRAS

02000400060008000

10000

02000400060008000

10000

C

Chromosomes (1-22)

Chromosomes (1-22)

Chromosomes (1-22)

E

A

different patients. (B) Representative 2-D plot of droplet digital PCR confirming the absence of

matching mutations between clusters and tumor tissues. Green dots indicate droplets containing

PCR products from wild-type alleles; red and blue dots indicate droplets containing PCR

products with mutant allele amplification; gray dots are no amplification. (C) Fractional

abundance of each replicate and tissue sample or clusters assayed by ddPCR on 5 ng DNA

template. Red values indicate positive signals; values <0.2% were below the assay maximum

sensitivity. (D) Control experiment to assess the impact of whole genome amplification (WGA)

for aCGH experiments. (E) Array comparative genomic hybridization (aCGH) of single clusters

and matched tissues for three patients, in addition to patient 10 in Fig. 2E.

Fig. S4. Cytomorphology of CD45– clusters used for RNA-seq. Phase contrast and Hoechst

33342 images of clusters used for RNA-seq for each patient. No images are available for patient

P1 because the cells were directly deposited in RNA-seq buffer without imaging. Scale bars, 10

µm.

Cl.5

Cl.4

Cl.2

Cl.6 Cl.11

Cl.2

Cl.10

Cl.14

Cl.15

Cl.16

Cl.7

Cl.2

Pa

tie

nt

18

Pa

tie

nt

10

Pa

tien

t 1

6P

atien

t 2

1

Pa

tie

nt 8

Pa

tie

nt

20

Pa

tien

t 1

9

Cl.10

Cl.11

Fig. S5. Cell-type inference workflow from RNA-seq data.

Fig. S6. Validation of the cell-type inference algorithm. (A) Selected genes with the highest

specificity index for representative cell types were verified for specificity using BioGPS. (B)

Gene expression levels for markers commonly used in CTC research to define epithelial cells.

Note KRT18 expression in the endothelial lineages, and EPCAM expression in embryonic stem

L9

4_

Oo

cyte

L9

5_

Sp

erm

ato

cyte

L1

_E

SC

L2

_E

SC

L3

_E

SC

L4

_E

SC

L5

_E

SC

L1

4_

iPS

CL

21

_M

eso

de

rma

l_S

tem

.Ce

llL

7_

Ne

uro

ep

ithe

lial

L1

0_

Ne

ura

l_P

recu

rso

rL

11

_N

eu

ral_

Pre

cu

rso

rL

8_

Astro

cyte

L9

_A

stro

cyte

L5

5_

MS

C_

Fe

tal.B

loo

dL

56

_M

SC

_F

eta

l.Bo

ne.M

arro

wL

57

_M

SC

_F

eta

l.Liv

er

L2

6_

Fib

robla

st_

Fo

reskin

L2

7_

Fib

robla

st_

Bre

ast.T

issu

eL

43

_S

mo

oth

.mu

scle

.ce

ll_B

ron

ch

ial

L3

6_

Ch

on

dro

cyte

L3

8_

Oste

obla

st

L3

9_

Ad

ipo

cyte

L3

0_

En

do

the

lial_

Hu

mb

ilica

l.Ve

inL

31

_E

nd

oth

elia

l_Lym

ph

.Ve

sse

lL

33

_E

nd

oth

elia

l_H

um

bilic

al.V

ein

L3

4_

En

do

the

lial_

Lym

ph

atic

L2

2_

Ep

ithe

lial_

Bla

dd

er

L2

3_

Ep

ithe

lial_

Bla

dd

er

L2

4_

Ep

ithe

lial_

Bro

nch

ial

L2

5_

Ep

ithe

lial_

Bro

nch

ial

L4

0_

He

pa

tocyte

L4

4_

HS

C_

Bo

ne.M

arro

wL

45

_C

MP

_B

on

e.M

arro

wL

46

_G

MP

_B

on

e.M

arro

wL

47

_M

EP

_B

on

e.M

arro

wL

50

_C

FU

_B

one.M

arro

wL

54

_E

ryth

robla

st_

Pro

L4

8_

Mye

locyte

_B

on

e.M

arro

wL

49

_M

ye

locyte

_B

on

e.M

arro

wL

52

_P

late

let

L5

8_

Mo

no

cyte

L5

9_

Mo

no

cyte

_C

D1

4L

61

_M

on

ocyte

_C

D1

6L

62

_M

on

ocyte

L6

3_

Ma

cro

ph

age

_M

on

ocyte

.de

rive

dL

64

_M

acro

ph

age

_M

on

ocyte

.de

rive

dL

65

_M

acro

ph

age

_M

on

ocyte

.de

rive

dL

66

_M

acro

ph

age

_M

on

ocyte

.de

rive

dL

67

_M

acro

ph

age

_A

lve

ola

rL

68

_D

en

dritic

.Ce

ll_M

on

ocyte

.de

rive

dL

69

_D

en

dritic

.Ce

ll_M

on

ocyte

.de

rive

dL

71

_D

en

dritic

.Ce

ll_P

lasm

acyto

idL

72

_D

en

dritic

.Ce

ll_P

lasm

acyto

idL

87

_N

eu

trop

hil

L8

8_

Ne

utro

ph

ilL

73

_T.C

ell_

CD

4.N

aiv

eL

74

_T.C

ell_

CD

4.C

en

tr..Me

mo

ryL

75

_T.C

ell_

CD

4.C

en

tr..Me

mo

ryL

76

_T.C

ell_

CD

4.E

ff..Me

mo

ryL

77

_T.C

ell_

CD

8.C

en

tr..Me

mo

ryL

78

_T.C

ell_

CD

8L

79

_T.C

ell_

CD

4L

80

_T.C

ell_

CD

8.E

ff..Me

mo

ryL

81

_T.C

ell_

CD

8.N

aiv

eL

82

_T.C

ell_

CD

8L

51

_B

.Ce

ll_B

one.M

arro

wL

89

_B

.Ce

llL

90

_B

.Ce

ll_N

aiv

eL

91

_B

.Ce

llL

92

_B

.Ce

ll_M

em

ory

Oocyte (GSM1160112)Spermatozoa (GSM1383918)HSCs (GSM1306651)Mesodermal stem cellsNeural stem cells (GSM953382)Astrocytes (GSM1901321)Mesenchymal stem cells (GSM921002)Fibroblasts (GSM1668455)Smooth muscle cells (GSM1275863)Osteblasts (GSM1333401)

Adipocytes (GSM1468495)Endothelial cells (GSM1009635)Epithelial cells (GSM1857416)

GMP (GSM1341669)HSC (GSM1341670)MEP (GSM1341671)Promyelocytic cell HL-60 (GSM1663005)PlateletsMacrophages (GSM1586130)pDC(GSM2090431)Neutrophils(GSM996197)T lymphocytes (GSM1631665)B lymphocytes (GSM1260318)

Em

bry

on

ic s

tem

cell

Me

sen

chym

al

ste

m c

ell

iPS

CM

eso

de

rmal

ste

m c

ell

Ne

uro

na

lp

recu

rsor

Astr

ocyte

Cho

ndro

cyte

Oste

obla

st

Ad

ipocyte

Fib

robla

st

Sm

oo

th m

uscle

cell

End

oth

elia

l

Ep

ith

elia

l

He

pa

tocyte

Pro

-Ery

thro

bla

st

Pla

tele

t

Neu

tro

phil

T lym

ph

ocyte

B lym

pho

cyte

ME

P

Mo

no

cyte

Macro

pha

ge

mD

C

Oocyte

Spe

rma

tocyte H

SC

CM

PG

MP

Mye

lob

last

CF

U

pD

C

Normalized odds ratio

Min Max

Cell typesC

D

MPO

(Hematopoietic)

CCL18

(Monocyte-derived)

CD3D

(T cell)

MS4A1

(B cell)

PLAC1L

(Oocyte)

Prim

ary

cells

Pri

ma

ry c

ells

Gene expression (RMA)

Cell types

Embryonic stem cell

iPSC

Epithelial (mesoderm-derived)

Endothelial

Hepatocyte (endoderm-derived epithelial)

Epithelial (ectoderm-derived)

Hematopoietic pluripotent

Macrophages

Dendritic cells

T cells

B cells

Oocyte

LEFTY

(ESC)

KRT13

(Epithelial)

CDH5

(Endothelial)

GC

(Hepatocyte)

KRTDAP

(Keratinocyte)

Prim

ary

cells

KRT18 EPCAM

Gene expression (RMA)

Gene

(Cell type)

Gene

(Cell type)

Gene

A B

Chondrocytes (GSM1695254)

Hepatocytes (GSM54066)

Positiv

e c

on

trols

cells and hematopoietic cells. Data are represented as horizontal barplots derived from the output

of the primary cell atlas dataset of BioGPS (C) Heat map comparing the number of genes

enriched for each published positive control sample (rows) and cell type (columns) over random

enrichment. Each colored box represents a normalized odds ratio of the respective Fisher’s exact

test ranging from 0 (black) to 1 (red). mDC, myeloid dendritic cell; pDC, plasmacytoid dendritic

cell. (D) IDs for each cell types reported in table S8. Order of cell types corresponds to the

heatmap columns shown in (C).

Fig. S7. Comparison between endothelial clusters (this study) and CTC clusters in Aceto et

al. (9). Selected breast cancer cell lines with epithelial and mesenchymal lineage profiles and

primary endothelial cells were used as positive controls for epithelial, mesenchymal stem cells

and endothelial lineages. Lineage inference of CTC clusters reported in (9) shows the presence

of epithelial-derived cell clusters. Platelet and red blood cell signals were removed from this

analysis to highlight cell types belonging to the above-mentioned lineages of interest.

Fig. S8. Circulating endothelial clusters are tumor-derived. (A) Principal component analysis

reveals an association of endothelial clusters with tumor tissues. Left panel, principal component

1 (PC1) vs PC2. PC1 correlates with the standard deviation of inputed variables for single

clusters (r = -0.455) and not does distinguish tumor and normal tissues (batch effect). Right

panel, PC2 vs PC3 indicating association of endothelial clusters with the tumor tissues but not

with the normal tissues. Colored areas represent k-means clustering. (B) Cell type inference of

whole normal and tumor tissues and matched laser-captured endothelial cells, showing

enrichment of the target cell population. (C) Unsupervised hierarchical clustering of endothelial

cell populations show significant association of circulating endothelial clusters with intestinal

endothelial cells. nECs, colon-derived normal endothelial cells; tECs, CRC-derived tumor

endothelial cells. (D) Genes differentially expressed between nECs and tECs. PNOI probability of

differential expression as computed by NOISeq. Log2FC, log2(fold change). (E) Heatmap

showing the expression of published tumor endothelial markers (23, 25, 42) in single circulating

endothelial clusters, medians of normal tissue, tumor tissue and matched normal and tumor laser-

capture dissected endothelial cells. Rows of log2(FPKM+1) values were normalized by scaling

between 0 and 1. Values are shown along with color-coding of the scale.

Fig. S9. tEC clusters in a mouse model of CRC. (A) Experimental approach. (B) GFP

genotyping in CT26 xenografts at passage 0 and passage 1. Mouse 2 was excluded from

subsequent analysis. (C) Gating strategy for the analysis of tumor-derived (GFP+) endothelial

cell clusters. Biex, Biexponential scale. (D) Immunofluorescence staining of tumor vasculature

using CD309+ antibodies at passage 0 and passage 1. Scale bar, 50 µm. (E) Boxplot represent

quantification of CD309+GFP+ single and clustered cells in the blood of P1 xenografts and

controls. Data are individual animals.

ACT26 s.c.

500,000

14 d 1) Blood sampling

2) Flow cytometry

14 d

Transplantation s.c.

5x Balb/c2x L2G85 Balb/c (GFP +)

B

FSC

SS

C

Hoechst 33342 - Width

Ho

ech

st 3

33

42

- A

rea

GFP - Alexa 488 (Biex)C

D3

09

- A

PC

(B

iex)

GFP - Alexa 488 (Biex)

CD

30

9 -

AP

C (

Bie

x)

Single Cells Clusters

Single

Cells

Clusters

GFP+

CD309+GFP+

CD309+

C

E Single cells

Passage 1Passage 0

#1 #2

Passage 1

xenograft tumor

wt

173 bp (GFP)324 bp (control)

a

b

c

L2G85

Passage 0

#2

#4

#5

D

Passa

ge 1

Pa

ssa

ge

0

* *

a b c

#1

* *

#3

* * *

*

**

G

FP

+ /

CD

30

9 +

Sin

gle

cells

(%

of sin

gle

ce

lls)

Con

trol (

n = 3

)

Xenog

raft

(n =

4)

0.0

0.5

1.0

1.5

2.0

0.0

0.5

1.0

1.5

2.0

Con

trol (

n = 3

)

Xenog

raft

(n =

4)

G

FP

+ /

CD

30

9 +

Clu

ste

rs (

% c

luste

rs)

Clusters

0.0

0.5

1.0

1.5

2.0

0.0

0.5

1.0

1.5

2.0 P=0.14

CD309 / Hoechst

*

Fig. S10. tEC clusters do not correlate with selected CRC tumor and patient characteristics

or comorbidities. (A) Association of circulating endothelial cell cluster counts with selected

tumor characteristics (tumor location or size, immune cells, blood markers). A/T/D, ascending,

transverse and descending colon; R/S, rectum and sigmoid colon. (B and C) Association of

circulating endothelial cell cluster counts with general population characteristics (age, BMI, sex)

(B) and with comorbitidies (hypertension, diabetes, heart disease) (C). Data refer to patients, for

which all available clinical data were used in the analysis. Correlations were tested using the

Kendall’s Tau method. Unpaired samples were tested using a two-tailed Wilcoxon-Mann-

Whitney U test.

SUPPLEMENTARY TABLES

Tables S1-S14 are provided in one Supplementary Excel spreadsheet, with captions below.

Table S1. scrmPCR primers used in this study.

Table S2: Whole genome application (WGA) false-positives are detected only at VAF <

10%. WGA and unamplified DNA derived from the same sample (Patient 15) were sequenced to

detect false positive mutations arising from the amplification procedure.

Table S3. Circulating CD45– clusters do not mirror matching primary tumor mutations. Targeted high-throughput sequencing of single CD45– clusters. Numbers in colored boxes

indicate variant allele frequency (VAF) (coverage). Cl, cluster; ND, not detected; NA, not

available.

Table S4. Sporadic mutations in circulating CD45– clusters are not detected in matching

primary tumor tissues. Blue rows indicate variants tested by digital PCR and reported in fig S3.

Numbers in colored boxes indicate variant allele frequency (VAF) and coverage. ND, not

detected.

Table S5. Germline variants in tumor tissues and circulating CD45– cell clusters (coverage

and zygosity). NA, not available; ND, not detected.

Table S6. RNA-seq data: number of uniquely mapped reads to hg19 exons.

Table S7. RNA-seq data: processed data (FPKM) of all samples described in the study.

Table S8. Definition of the cell types used for the inference algorithm.

Table S9. Top 80 genes ordered from highest to lowest specificity index (S) for each cell

type selected for the inference algorithm.

Table S10. Positive published RNA-seq control samples used to validate the inference

algorithm.

Table S11. Characteristics of tEC clusters isolated in this study. Staining intensity: –,

negative; +, low; ++, average; +++, strong.

Table S12. Circulating endothelial cell cluster count before and after surgery. Data are from

Fig. 4B. Pre: Clusters count in blood 0-24 hrs before surgery. Post: Clusters count in blood 24-72

hrs after surgery.

Table S13. Baseline patients’ and healthy donors’ characteristics.

Table S14. CD45– cluster count for each baseline sample type and number of single clusters

analyzed for each technique in the corresponding samples. Sources: FORTIS, Fortis Surgical

Hospital, Singapore; NCC, National Cancer Center, Singapore; IBN, Institute of Bioengineering

and Nanotechnology, Singapore; SPHS, Singapore Population Health Studies. CRC, colorectal

cancer.