adjustedP < 0.05) ( 12 ). Finally, the relative
expression levels of predicted key TFs in
each of the four groups were confirmed by a
quantitative polymerase chain reaction (qPCR)
(fig. S3C).
Classification of CRPC patients using
transcriptomic signatures of the four subtypes
Next, we examined RNA-seq datasets from
366 CRPC patients to assign each patient to
the four subtypes ( 21 ). We derived the sig-
nature genes for each of the four subtypes
as the ones with higher expression in one
group relative to others in organoids and cell
lines, and filtered out genes with low expres-
sion or low variance in CRPC patient samples
Tanget al., Science 376 , eabe1505 (2022) 27 May 2022 5of13
Top 25 highest ranked TFs
Outdegree
low
high
Expr
low
high
Chromatin
accessibility
low
high
Outdegree
Expr
TF rank
ChromatinaccessibilityOutdegree
Expr
TF rank
Chromatinaccessibility
CRPC-AR
CRPC-NE
CRPC-WNT
CRPC-SCL
0
1000
2000
0
1000
2000
Step1. Construct peak-gene links
peak 1 peak 2gene X peak 3
TSS
gene Y
TSS
TF A TF A
Step2. Predict TF binding to peaks
TF B
peak 1 peak 2 peak 3
TF A
Step3. Predict target genes of TF
peak 1 peak 2gene X peak 3
TSS
gene Y
TSS
TF A
Step4. Infer gene regulatory network
TFs
non-TF genes
0
10000
20000
30000
40000
0 102030
genes mapped per peak
Frequency
0
1000
2000
3000
0204060
peaks mapped per gene
Frequency
genes per peak:
mean: 1.4
peaks mapped to
only one gene: 75.2%
peaks per gene:
mean: 8.46
median: 3
D
A
BC
AR
FOXA1
PGR
FOXO3FOXB2FOXK1GATA2FOXO4HOXA13HOXB13FOXJ2HOXC12FO
XN3
NFYBFOXP3FOXL2NFIC
HOXC13FOXC2FOXS1
HSF4
ARID5AHNF1BHOXC10HNF1ATCF7L
2
KL
F^4
SOX4SNAI1TCF7KLF2ZIC1LMO2LEF1TCF4
ID4ZIC
4
TCF7
L1
TWIST1
TCF3SP5KL
F^7 SP9
SOX13
KLF5SP3
RUNX3TGI
F1
HINFP
SP4
NEUROD1
ASCL1TCF12MYOG
MSC
NHLH1ATO H 8NKX24NKX22OLIG2
RFX2EBF1
NHLH2
NEUROG3BHLHE22
RFX3
NEUROG2NEU
ROD2HMX3RFX5HMX2MEO
X2
E2F6
ZBTB18
TAL2
FOS
L1
BAT F
FOSL2JU
NB
TP63JDP2
NFE2L3CEBPBTFAP2C
FOSMAFFNFE2
TEAD3FOSBTFAP2BCEB
PA
NFICATF3
RUN
X2FLI1
ETS2
CEBPDNR
2E1
NFE
2L2
NFKB1
Fig. 3. Identification of the key transcription factors (TFs) of each subtype.(A) Schematic illustrating the construction of sample-specific regulatory networks
using ATAC-seq and RNA-seq data. (B) Distribution of the number of genes linked per peak. (C) Distribution of the number of peaks linked per gene. (D) Rank
order of the top 25 TFs for each of the four subtypes. For each TF, the relative contributions of three metrics to TF rank are shown (Expr, expression). In CRPC-SCL,
FOSL1, BATF, FOSL1, JUNB, JDP2, FOS, MAFF, FOSB, and ATF3 belong to the AP-1 family.
RESEARCH | RESEARCH ARTICLE