Nature - USA (2020-01-16)

422 | Nature | Vol 577 | 16 January 2020

Article

upon treatment with G12Ci) (Fig. 1c, Extended Data Fig. 3a, b). Treat-
ment with G12Ci sequestered most cells in a state with low output.
Some drug-treated cells had high output (arrow in Fig. 1c), indicating
diverging responses across the population. The change in output score
revealed different fates along the trajectories (Fig. 1d, e, Extended Data
Fig. 3c–e): cells in path 1 had inhibited output, and cells in path 2 had
an initial inhibition followed by reactivation.
The trajectories correlated with changes in cell-cycle-specific expres-
sion signatures (Extended Data Fig. 4a). By classifying cells along the
cell cycle, we found that treatment with G12Ci induced a quiescent
state (G0) that was transcriptionally distinct from G1 (Fig. 1f, Extended
Data Fig. 4b–d). This induction mirrored the inhibition of KRAS(G12C)
output along the trajectories (Fig. 1g). Treatment with G12Ci also led to
higher levels of protein expression for p21 and p27 (Fig. 1h, Extended
Data Fig. 4e), two markers of quiescence^18. Consistent with this, cell-
cycle analysis of double-thymidine-synchronized cells showed that the
G12Ci treatment arrests cells in a G0 or G1 state (Extended Data Fig. 4f ).
We used a quiescence biosensor^19 based on a cyclin-dependent-kinase
binding-defective p27 mutant (p27K−) to monitor the subpopulations
and to validate the results of the scRNA-seq analysis (Fig. 1i, Extended
Data Fig. 5a). As predicted, the G12Ci treatment led to a bimodal cell
distribution, comprising a subpopulation of quiescent cells with high
p27K− (about 80% of the total population) and a subpopulation of rap-
idly adapting cells with low p27K− (about 20%). This differed from the
effect of inhibitors that target MEK or ERK, two kinases that are down-
stream of KRAS(G12C) (Extended Data Fig. 5b, c). Although exposed
to the G12Ci treatment for the same duration, cells with low p27K− had
higher active KRAS than cells with high p27K− (Fig. 1j, Extended Data
Fig. 5d) and were able to progress past the G1 restriction point (Fig. 1i,
inset). Rechallenge with the G12Ci attenuated the adapting response
to some degree, but it could not eliminate this population (Extended
Data Fig. 5a, e). Furthermore, the inhibition of levels of KRAS–GTP by
the G12Ci during rechallenge was less than its initial effect (Extended
Data Fig. 5f ).

To identify the adaptive signals that are responsible for the divergent

response to the G12Ci treatment, we used two orthogonal approaches

(Methods). A differential expression analysis, which contrasted single

cells from the inhibited and adapting trajectories, revealed many tran-

scripts with trajectory-specific expression (Extended Data Fig. 6a). A

genome-wide knockout screen identified single-guide RNA (sgRNA)

targets that enhanced the effect of treatment with G12Ci (Extended

Data Fig. 6b). After integrating the results from both approaches, we

focused on genes with subpopulation-specific expression that were also

functionally related to proliferation on treatment (Fig. 2a, b, Extended

Data Fig. 6c–e). Of these, heparin-binding epidermal growth factor

(HBEGF), aurora kinase A (AURKA) and KRAS were studied in more detail.

The expression of HBEGF mRNA—which encodes a ligand of the epi-

dermal growth factor receptor (EGFR)^20 —was downregulated shortly

after treatment with G12Ci, but rebounded at 48–72 h (Extended Data

Fig. 7a). The single-cell analysis ‘localized’ this rebound to a cluster

of quiescent cells (Fig. 2a, b, Extended Data Fig. 6d). This change

was associated with an approximately twofold increase in secreted

HBEGF during the G12Ci treatment (Extended Data Fig. 7a). Consistent

with a potential role in mediating adaptation to the G12Ci treatment,

sgRNAs targeting EGFR signalling intermediates were depleted on

treatment with G12Ci (Fig. 2c). Likewise, siRNA-mediated knockdown

of HBEGF enhanced the antiproliferative effect of the drug (Extended

Data Fig. 7b).

The secretion of HBEGF could affect a broader population of cells in

an autocrine and/or paracrine fashion by activating EGFR, which—simi-

lar to other receptor tyrosine kinases—drives nucleotide exchange to

activate RAS^20. Indeed, adapting cells had higher levels of EGFR pathway

activation than quiescent cells (Fig. 2d). Stimulation with recombinant

EGF induced KRAS activation in sorted quiescent cells and enhanced

signalling in an EGFR- and SHP2-dependent manner (Extended Data

Fig. 7c–e). EGF also enhanced the escape from quiescence (Fig. 2e)

when the ligand was added during the adaptive phase of the G12Ci

treatment (24–48 h), but not when it was added at the beginning. This

abc

d

h

fg

ij

Hours

Cells (%)

100

75

50

25

0

042472

n = 2,565n = 3,259n = 1,006n = 3,347

G12C-suppressed

G12C-induced

Untreated Treated Cells

0.8

0.6

0.4

0.2

0.20.40.60.8 0.20.40.60.8

16

12

8

4

p27

p21

GAPDH

24624324872

G12Ci (h)

0

H358

KRAS(G12C), p53DEL, LKB1WT

H2122

KRAS(G12C), p53MT, LKB1DEL

SW1573

KRAS(G12C), p53WT, LKB1WT

DC2

DC1 n = 10,177 cells

Clusters:

Path 1

Path 3 Path 2

768

9

2

3541

10

G12C output

Cluster

1.0

0.5

0

76910354182

n = 1,052n = 1,270n = 594n = 771n = 1,488n = 1,467n = 829n = 938n = 1,058n = 710

Cells

e^04 Hours^2472

100

75

50

25

Cells (%) 0

G1/S

S

G2/M

M

M/G1

G0

100

75

50

25

Cells (%) 0

G1/S

S

G2/M

M

M/G1

(^76910354182) G0
Cluster
p27K–, log(AU)
240
160
80
0
123456
Baseline
G12Ci, 72 h
100
72 h, L72 h, H
0 h
G12Ci
S
(^500) G0/G1
G2/M
Cells (%)
G12C output
0.6
0.4
0.2
Pseudotime
02 5 50 75100
Suppressed
Induced
0.6
0.4
0.2
—Path 1—Path 2—Path 3
0472
p27K– sort
p21
GAPDH
KRAS-GTP
KRAS
72
L
72 G12Ci, h
NA H
Fig. 1 | Divergent single-cell fates after conformation-specific KRAS(G12C)
inhibition. a, Diffusion component (DC) analysis of single cells from models of
KR AS(G12C) tumours, treated with a G12Ci for 0, 4, 24 and 72 h. The arrows
indicate inhibitory trajectories derived by the Slingshot algorithm. b, Cluster
composition across treatment time. c, The distribution of KRAS(G12C)-
dependent transcriptional output score across single cells. d, The trend in
G12C output as a function of pseudotime was established by fitting a spline to
single-cell data. The 95% confidence interval is shown. n = 4,759, n = 8,653 and
n = 4,050 in path 1, path 2 and path 3, respectively; n, number of cells. e, G12C
output score across clusters. Median, upper and lower quartiles, and outliers
are shown. f, g, Cell-cycle-phase distribution over time (f) or across clusters (g).
h, Extracts from drug-treated KR AS(G12C)-mutant cells (H358) were analysed
to determine the expression of the indicated proteins. i, H358 cells expressing
the quiescence biosensor (mVenus–p27K−) were analysed by f luorescence-
activated cell sorting (FACS). Inset, cell-cycle distribution of the indicated
populations. j, Biosensor-expressing cells were treated, sorted and analysed to
determine the levels of active and total KR AS. A representative of three
independent experiments is shown in h–j. H, high p27K− expression; L, low
p27K− expression; NA, not applicable.