SCIENCE science.org 10 DECEMBER 2021 • VOL 374 ISSUE 6573 1319cytokinesis) accumulate and intercellular
bridges (ICBs), the final membranous con-
nection between two daughter cells, persist
in cells from PIK3C2A-null patients. These
phenotypes suggest a role for PI3K-C2a
in late stages of cell division. Indeed, the
authors found that PI3K-C2a localizes to
ICBs. Direct binding to both PI(4,5)P 2 in the
plasma membrane and g-tubulin restricts
PI3K-C2a to the ICB central bulge, a struc-
ture called the Flemming body. Furthermore,
catalytic activity and proper localization of
PI3K-C2a are needed for abscission. Gulluni
et al. determined that PI3K-C2a at the ICB
uses PI(4)P as a substrate to produce a local-
ized pool of PI(3,4)P 2 , a lipid not previously
known to act in cytokinesis.
The final stage of cytokinesis involves
recruitment of the endosomal sorting com-
plexes required for transport III (ESCRT-III)
family of charged multivesicular body pro-
tein 4 (CHMP4) proteins to the Flemming
body ( 9 ). Studies in tissue culture cells have
revealed that CHMP4 is recruited to the
Flemming body through centrosomal pro-
tein of 55 kDa (CEP55) and ALG-2–interact-
ing protein X (ALIX). CHMP4B is then re-
leased from the Flemming body owing to its
polymerization into helical filaments that
constrict the membrane of the ICB, leading
to membrane fission and physical separa-
tion of the daughter cells ( 10 ). Recruitment
and subsequent release of CHMP4B from
the Flemming body must be tightly regu-
lated to ensure that abscission occurs at
the correct time and not, for example,
while lagging chromosomes persist in the
ICB, which would create genomic instabil-
ity ( 11 ). Notably, heterozygous mutations
in CHMP4B are linked with early-onset
cataracts. However, mice that lack CEP55
develop normally except within the brain.
CEP55 is thus not essential for cell division
in all cells but is specifically required in
neurons, raising the possibility of parallel,
cell type–specific pathways to abscission.
Gulluni et al. find that lens cells express
low amounts of ALIX but can still recruit
CHMP4B to the ICB, suggesting that these
cells use a variation of the ESCRT-III abscis-
sion pathway. In human tissue culture cells,
the authors discovered that the requirement
for ALIX can be bypassed by the ESCRT-II
protein vacuolar protein-sorting–associated
protein 36 (VPS36), which binds PI(3,4)P 2 in
the Flemming body. Depletion of PI3K-C2a
results in a corresponding loss of PI(3,4)P 2
from this structure, along with VPS36 and
CHMP4B. These observations suggest that a
specific lipid microdomain is needed to by-
pass ALIX function during cytokinesis.
The redundancy of the ALIX and PI(3,4)P 2
pathways raises the question of why certain
cell types might be particularly susceptible
to abscission defects. Answers may lie in
either the frequency of divisions or the tis-
sue context in which the cell divisions occur.
Adherent cells grown in culture and de-
pleted of abscission factors, including fibro-
blasts lacking CEP55, display only low levels
of cytokinesis failure. By contrast, mice that
lack CEP55 and ALIX exhibit more severe
cell division defects in neurons ( 12 , 13 ). One
possibility could relate to forces within the
tissue. Adherent cells in culture can migrate
away from each other upon mitotic exit and
consequently undergo cytofission simply by
pulling themselves apart ( 14 ). Within a tis-
sue such as the brain or the lens, geometric
constraints to motility may render cells more
susceptible to abscission defects.
Differences in cell type– and tissue-specific
cell division pathways could also stem from
the need to respond to different regulatory
pathways. Abscission is thought to be sub-
ject to a checkpoint mechanism, delaying the
final step of daughter cell separation until all
the chromosomes have left the intercellular
bridge. Recent findings suggest that activa-
tion of abscission is also susceptible to the
state of nuclear envelope reassembly and
membrane tension across the ICB ( 9 ). Could
it be that each of these events feeds into a dif-
ferent arm of the abscission pathway? PI3K-
C2a has been found to generate PI(3)P in
response to shear stress and increased mem-
brane tension ( 15 ). By using independent
pathways or those with overlapping function,
abscission may have evolved to respond to a
variety of events to ensure its correct timing
and the maintenance of genome stability. jREFERENCES AND NOTES- T. Fujiwara et al., Nature 437 , 1043 (2005).
- D. Tiosano et al., PLOS Genet. 15 , e1008088 (2019).
- F. Gulluni et al., Science 374 , eabk0410 (2021).
- G. R. V. Hammond et al., Science 337 , 727 (2012).
- F. Gulluni, M. Martini, E. Hirsch, J. Cell. Biochem. 118 ,
3561 (2017). - E. M. Storck, C. Özbalci, U. S. Eggert, Annu. Rev. Biochem.
87 , 839 (2018). - S. Frémont, A. Echard, Curr. Biol. 28 , R458 (2018).
- A. Panopoulos et al., Mol. Biol. Cell 25 , 3105 (2014).
- M. Vietri, M. Radulovic, H. Stenmark, Nat. Rev. Mol. Cell
Biol. 21 , 25 (2020). - N. Elia, R. Sougrat, T. A. Spurlin, J. H. Hurley, J. Lippincott-
Schwartz, Proc. Natl. Acad. Sci. U.S.A. 108 , 4846 (2011). - N. T. Umbreit et al., Science 368 , eaba0712 (2020).
- J. N. Little et al., J. Neurosci. 41 , 3344 (2021).
- M. H. Laporte et al., Sci. Rep. 7 , 44767 (2017).
- C. L. Dix et al., Dev. Cell 45 , 132 (2018).
- A. Boukhalfa et al., Nat. Commun. 11 , 294 (2020).
ACK NOWLEDGMENTS
The authors are grateful to J. Ashkenas for help with preparing
the manuscript. The authors are funded by Natural Science
and Engineering Research Council discovery grants (RGPIN-
2016-06775 to J.A.B. and RGPIN-2019-05782 to A.W.).10.1126/science.abm7949BIOCHEMISTRYInteractomes
in the era of
deep learning
D eep learning provides an
atomic snapshot of the yeast
protein interactome
By Joana Pereira1,2 and Torsten Schwede1,2C
haracterizing macromolecular in-
teractions allows better understand-
ing of the inner workings of a cell.
However, all methods available today
have limitations: Some tell us whether
two macro molecules interact, others
provide atomic detail about the interaction
partners or, at best, the structures of iso-
lated assemblies without cellular context.
On page 1340 of this issue, Humphreys et
al. ( 1 ) describe a new computational ap-
proach, founded on the ongoing deep learn-
ing revolution in structure bioinformatics
( 2 , 3 ), to predict the composition and model
the three-dimensional (3D) structure of pro-
tein-protein interactions at the same time.
They apply their approach to a eukaryotic
system—baker’s yeast (Saccharomyces cere-
visiae)—and predict and accurately model
more than 1500 protein-protein interactions,
106 of which were not seen before, paving the
way to high-throughput, high-accuracy mod-
eling of entire cells.
Determining the 3D structures of macro-
molecules and their interactions provides
important information about macromo-
lecular mechanisms, which can be used, for
example, in drug development or exploited
in biotechnology. Experimental structural
biology methods such as macromolecular
crystallography (MX) and high-resolution
cryo–electron microscopy (cryo-EM) provide
atomic-level detail of macromolecular struc-
tures and their assemblies ( 4 ). Such experi-
ments are laborious and require purification
of the macromolecules from their cellular
context. Although techniques such as yeast
two-hybrid (Y2H) and cross-linking mass
spectrometry (XL-MS) allow for large-scale
detection of interaction partners, methods
such as site-directed mutagenesis or Förster
energy resonance transfer (FRET) experi-
ments characterize individual interactions
and interfaces. This information can be used
to guide the modeling of assemblies by, for
example, macromolecular docking in integra-Cataracts are characterized by opacity forming
in the lens, which can be caused by disrupted
cytokinesis that changes the structure of the lens.