604 | Nature | Vol 585 | 24 September 2020
Article
possess an ether linkage at the glycerol sn-1 position (Fig. 2a,
Extended Data Fig. 3a). Ether phospholipids comprise two sub-
types: 1-O-alkyl-glycerophospholipids (R^1 CH 2 CH 2 OCH 2 R^2 ); and
1-O-alkenyl-glycerophospholipids (R^1 CH=CHOCH 2 R^2 ), known as plasm-
alogens^12. Most plasmalogens possess an ester-linked polyunsaturated
fatty acyl (PUFA) chain at the sn-2 position^12 (Fig. 2a). Consistent with the
lipid-synthesis function of peroxisomes, lipidomic profiling revealed
selective loss of ether glycerolipids in both PEX3- and PEX10-depleted
cells; this reduction is most prominent for polyunsaturated ether phos-
pholipids (PUFA-ePLs), which are presumably largely plasmalogens
(Fig. 2b, Extended Data Fig. 3b, Supplementary Data 3, 4).
Because AGPS and FAR1—two peroxisome-associated enzymes
that catalyse the biosynthesis of ether lipids—also emerged as top
hits in both CRISPR screens, we proposed that ether lipid biosyn-
thesis might be integral to the pro-ferroptotic functions of per-
oxisomes. Indeed, CRISPR–Cas9-mediated depletion of either the
AGPS or the FAR1 gene, performed either in bulk populations or in
derived single-cell clones, recapitulated the ferroptosis-resistance
phenotype and the concomitant changes in ether phospholipids
that were observed upon peroxisome depletion (Fig. 2c–g, Extended
Data Figs. 3c–m, 4a–e, Supplementary Data 3, 4). The specificity
of the CRISPR–Cas9 knockout procedure was verified by reversal
of these phenotypes through the introduction of sgRNA-resistant
Agps or Far1 cDNAs into these cells (Extended Data Fig. 5a–c). The
pro-ferroptotic roles of the ether lipid biosynthesis pathway were
further confirmed by overexpression of ectopic cDNAs in wild-type
cells, short hairpin RNA (shRNA)-mediated gene knockdown, and the
use of a small-molecule inhibitor of AGPS^13 (Extended Data Fig. 5d–h).
By contrast, knockout of two peroxisomal enzymes that are unre-
lated to ether lipid metabolism—superoxide dismutase 1 (encoded
by SOD1) and catalase (encoded by C AT)—did not alter the sensitivity
of the cells to ferroptosis (Extended Data Fig. 5i–k), consistent with
the pro-ferroptotic role of peroxisomes being specifically mediated
by ether lipid biosynthesis.
Notably, depletion of either peroxisomes or ether lipids also reduced
the sensitivity of GPX4-dependent HuH-7 hepatocellular carcinoma and
SNU-685 endometrial carcinoma cells to ferroptosis^14 (Extended Data
Fig. 6a–e). Moreover, analysis of previous lipidomics data revealed that
ccRCC tumours—a cancer type known to be susceptible to ferropto-
sis^3 —exhibit higher levels of polyunsaturated ePLs and increased AGPS
expression relative to normal kidney tissues^15 (Extended Data Fig. 6f ).
These analyses suggest that the pro-ferroptotic role of the ether lipid
biosynthesis pathway might be applicable to other carcinoma cell
types and to human cancers.
It is noteworthy that the loss of ether phospholipids in
peroxisome-depleted cells occurred without significant changes in
the levels of either ACSL4 or lysophosphatidylcholine acyltransferase 3
(LPCAT3)—rate-limiting enzymes in the biosynthesis of most cellular pol-
yunsaturated lipids^8 ,^16 (Extended Data Fig. 7a, b). Co-deletion of ACSL4
together with peroxisome and ether lipid biosynthesis genes led to fur-
ther protection against ferroptosis—similar to the effect conferred by
the ferroptosis inhibitor ferrostatin-1 (Fer-1) (Extended Data Fig. 7c, d).
These observations implicate non-peroxisome-derived PUFA lipids
in ferroptosis, and point towards a potential combinatorial strategy
that involves the inhibition of both ACSL4 and PUFA-ePL biosynthetic
enzymes to block ferroptosis.
Ether lipid biosynthesis is achieved by functional collaboration of
peroxisomes with the endoplasmic reticulum (ER)^11. Within peroxi-
somes, the enzymes FAR1, GNPAT and AGPS act to synthesize the ether
lipid precursor 1-O-alkyl-glycerol-3-phosphate (AGP)^17. AGP is then
dispatched to the ER, where it is acylated at the sn-2 position of glyc-
erol, head groups are added to the sn-3 position and—in the case of
plasmalogens—a double bond is created by dehydrogenation to form
an alkenyl-ether linkage^11 ,^17 –^19. Although PUFA addition at the sn-2 posi-
tion is critical for the synthesis of ferroptosis-relevant ether lipids, it
is not known which ER enzyme(s) mediate this process. We identified
the ER-resident enzyme 1-acylglycerol-3-phosphate O-acyltransferase
3 (encoded by AGPAT3) as a hit in both CRISPR screens (Fig. 1b, Extended
Data Fig. 1a, b, f, g). AGPAT3 was also a pro-ferroptotic hit in a CRISPR
screen in human haploid KBM7 cells^16. Notably, AGPAT3 has been
reported to selectively incorporate arachidonic acid or docosahexae-
noic acid into lysophosphatidic acids, leading to the synthesis of diacyl
PUFA-phosphatidic acids^20. However, whether AGPAT3 also contributes
to AGP acylation and PUFA-ePL biosynthesis is unknown.
For this reason we performed lipidomic analysis, which revealed that
depletion of AGPAT3 selectively reduced the levels of polyunsaturated
species among both ether-linked and diacyl phospholipids (Fig. 2h,
Extended Data Fig. 7e, Supplementary Data 3, 5). Consistent with the
above, genetic depletion of AGPAT3 suppressed sensitivity to ferrop-
tosis (Fig. 2i, Extended Data Fig. 7f–j), confirming the pro-ferroptotic
role of AGPAT3. Moreover, expression of a CRISPR-resistant, wild-type
Agpat3 cDNA restored ferroptosis sensitivity in AGPAT3-depleted
cells, whereas the catalytically dead Agpat3E176A mutant^20 failed to do
so (Extended Data Fig. 7k, l). These results suggest that AGPAT3 func-
tions downstream of the peroxisomal pathway to synthesize PUFA-ePLs
in the ER (Extended Data Fig. 8a).
To confirm that PUFA-ePLs are necessary and sufficient to induce sen-
sitivity to ferroptosis, we formulated various diacyl and ether phospho-
lipids as liposomal nanoparticles for delivery to cells (Extended Data
Fig. 8b). Among the ethanolamine-containing phospholipids, application
of either C18(plasm)-C20:4PE, C18(plasm)-C22:6PE or C18:0-C20:4PE—
but not C18(plasm)-C18:1PE—led to ferroptosis sensitization in OVCAR-
8 cells (where (plasm) indicates the alkenyl-glycerophospholipid;
Extended Data Fig. 8c). Among the choline-containing phospholip-
ids, C18(plasm)-C20:4PC—but not C18(plasm)-C18:1PC—exhibited
significant ferroptosis sensitization activity, whereas C18:0-C20:4PC,(ccRCC)786-O
DMSO(Ref.^3 )OVCAR-8(HGSOC)DMSORSL3 ML210
Day 4
Day 6
Day 8Day 4
Day 8
Day 120.5 μM
1 μM
2 μM
Identify pro-ferroptosis genes
by sequencing0246Average log 2 (RSL3/DMSO)Average –log(P 10
)OVCAR-8AGPSPEX3
FAR1 PEX12
PEX19 PEX2PEX10PEX14PEX13PEX16
AGPAT6 ACSL4
AGPAT3
LPCAT30246810Average log 2 (ML210/DMSO)(8 days)786-OFAR1AGPSPEX5GNPATPEX19PEX12
PEX16
PEX2PEX7PEX13PEX1
PEX10PEX6c(^0) –1
0.4
0.8
1.2
log 10 ([ML210], μM) log 10 ([ML210], μM) log 10 ([ML210], μM)
Relative viability
sgNC
PEX10 PEX10 sg1sg2
01 –1 01
sgNC
PEX3sg1
PEX3sg2
sgNC
PEX12 PEX12 sg1sg2
–2 –1 0
0
0.4
0.8
1.2
Relative viability
–2 –1 0 –2 –1 0
log 10 ([RSL3], μM) log 10 ([RSL3], μM) log 10 ([RSL3], μM)
ab
(12 days)
(^0) –1 01
0.5
1.0
1.5
0 1 2 3 4 0 2 4 6 8
ACSL4
AGPAT3
0
0.4
0.8
1.2
0
0.4
0.8
1.2
0
0.4
0.8
1.2
Fig. 1 | Genome-wide CRISPR screens identify peroxisome components as
contributors to ferroptosis susceptibility. a, Schematic summarizing the
CRISPR screens in OVCAR-8 and 786-O cells to identify ferroptosis regulators.
HGSOC, high-grade serous ovarian carcinoma; ML210 and RSL3 are covalent
small-molecule inhibitors of GPX4 and inducers of ferroptosis. b, Volcano plots
showing the top genes in 12-day RSL3-treated OVCAR-8 (left) and 8-day
ML210-treated 786-O (right) cells. For presentation purposes, only genes that
are enriched ≥1.5-fold (log 2 fold change ≥ 0.585) in the RSL3- or ML210-treated
condition are plotted. Blue highlights lipid synthesis genes; red highlights
peroxisome genes. See Methods for data analysis methods. c, Viability curves
of OVCAR-8 cells expressing a non-targeting negative control sgRNA (sgNC) or
sgRNAs targeting PEX10, PEX3 or PEX12 after treatment with the indicated
concentrations of ML210 or RSL3 for 72 h. n = 3 biologically independent
samples. Data are mean ± s.d., and show representative results of experiments
performed in triplicate.