Microscopic Dynamics of Liquid-Liquid Phase Separation and Domain Coarsening
in a Protein Solution Revealed by X-Ray Photon Correl ation Spectroscopy
Anita Girelli ,
1
Hendrik Rahmann,
2
Nafisa Begam ,
1
Anastasia Ragulskaya,
1
Mario Reiser,
2,3
Sivasurender Chandran ,
1,4
Fabian Westermeier ,
5
Michael Sprung,
5
Fajun Zhang ,
1,*
Christian Gutt ,
2,†
and Frank Schreiber
1,‡
1
Institut für Angewandte Physik, Universität Tübingen, Auf der Morgenstelle 10, 72076 Tübingen, Germany
2
Department Physik, Universität Siegen, Walter-Flex-Strasse 3, 57072 Siegen, Germany
3
European X-Ray Free-Electron Laser XFEL, Holzkoppel 4,22869 Schenefeld, Germany
4
Department of Physics, Indian Institute of Technology Kanpur, Uttar Pradesh 208016, India
5
Deutsches Elektronen-Synchrotron DESY, Notkestraße 85, 22607 Hamburg, Germany
(Received 19 October 2020; accepted 23 February 2021; published 2 April 2021)
While the interplay between liquid-liquid phase separation (LLPS) and glass formation in biological
systems is highly relevant for their structure formation and thus function, the exact underlying mechanisms
are not well known. The kinetic arrest originates from the slowdown at the molecular level, but how this
propagates to the dynamics of microscopic phase domains is not clear. Since with diffusion, viscoelasticity,
and hydrodynamics, distinctly different mechanisms are at play, the dynamics needs to be monitored on the
relevant time and length scales and compared to theories of phase separation. Using x-ray photon
correlation spectroscopy, we determine the LLPS dynamics of a model protein solution upon low
temperature quenches and find distinctly different dynamical regimes. We observe that the early stage
LLPS is driven by the curvature of the free energy and speeds up upon increasing quench depth. In contrast,
the late stage dynamics slows down with increasing quench depth, fingerprinting a nearby glass transition.
The dynamics observed shows a ballistic type of motion, implying that visco elasticity plays an important
role during LLPS. We explore possible explanations based on the Cahn-Hilliard theory with nontrivial
mobility parameters and find that these can only partially explain our findings.
DOI: 10.1103/PhysRevLett.126.138004
Recent work suggests that structure formation in biology
can take place, inter alia, through liquid-liquid phase
separation (LLPS) [1–3]. Phase separation in crowded
environments thus represents a mechanism for intracellular
organization via the formation of biomolecular condensates
[4]. The biological functions of these condensates—includ-
ing steering biochemical reactions rates, sensing, or signal-
ing—are being intensely investigated [3]. LLPS is also
associated with a variety of diseases caused by a loss and/or
change of function of the condensates [5,6].
The state of the condensates depends on the dynamic
processes during their formation, often involving non-
equilibrium processes over a hierarchy of length and time
scales [2,7]. A case in point is the slowdown of the
dynamics on molecular length scales caused by concen-
tration and its influence on the dynamical and structural
properties of the condensate on mesoscopic length scales.
Ultimately, such a microscopic slowdown can lead to the
arrest of LLPS on larger length scales accompanied by the
formation of bicontinuous gel network structures [8–10].
The kinetics of arrested phase separations in protein
solutions have been studied successfully in the past,
demonstrating that the ensemble-averaged structure factor
ceases to develop further in q position and intensity upon
low temperature quenches [11–15]. However, the dynamics
of protein solutions en route to an arrested LLPS is largely
unknown, mainly because of the requirement to monitor an
exceptionally broad range of time and length scales
simultaneously. This, in turn, prevented the experimental
validation of models of the dynamics of critical phenomena
during LLPS, such as the Cahn-Hilliard equation and
related models, especially in the vicinity of glass-gel
transitions displaying large dynamical asymmetries
between the species involved [2].
LLPS is a general phenomenon, which is relevant not
only for protein systems but also in many other fields of
science [16]. LLPS domains have been studied with
different microscopy techniques [17,18]. Their macro-
scopic properties such as turbidity and viscosity [19–21]
were monitored, as well as their molecular properties
[20,22–27]. Other scattering techniques were used to
access the kinetics of the phase separation [14,15,19,21].
X-ray photon correlation spectroscopy (XPCS) employ-
ing coherent x-rays can resolve the collective dynamics on
the required length scales, ranging simultaneously from
nanometers to microns and timescales from microseconds
to hours [28–35]. Here, we demonstrate that a combination
of scanning techniques, large beams, and long sample
detector distances [36–38] allows us to reduce the required
x-ray doses to values below the critical dose of many
PHYSICAL REVIEW LETTERS 126, 138004 (2021)
0031-9007=21=126(13)=138004(7) 138004-1 © 2021 American Physical Society
protein systems [39]. More details on the experimental
parameters are provided in the Supplemental Material (SM)
[40]. With this approach, we are able to follow the
dynamics during an LLPS of γ globulin (Ig) in a concen-
trated aqueous polyethylene glycol (PEG) solution. The
experimental results are compared to simulations based on
the Cahn-Hilliard equation, taking the gel transition into
account (in the spirit of model C according to Ref. [2]; see
the SM) [2,45,46]. Our work paves the way for future
XPCS experiments exploring the full spatiotemporal win-
dow of LLPS, allowing one also to benchmark and guide
computer simulations of complex protein dynamics in
crowded environments.
XPCS experiments were conducted at the P10 Coherence
Applications Beamline at PETRA III, Deutsches Electronen-
Synchrotron, employing an x-ray beam of photon energy
8.54 keV, a size of 100 × 100 μm
2
, and a maximum photon
density of 10
7
photons=s=μm
2
. The key for performing low
dose XPCS experiments is to make use of large beams with a
sufficient degree of coherence. Time series of coherent
diffraction patterns were collected with an EIGER 4-mega-
pixel detector covering a q range from 0.003 to 0.05 nm
−1
.
Samples of Ig with PEG and NaCl have been prepared as
described in [15] and in the SM and quenched from 37 °C to
different quench temperatures T
q
below the binodal line.
Figure 1(a) displays the temporal evolution of the
scattering intensity as a function of scattering vector q
for a quench temperature of T
q
¼ 10 °C [for other quench
temperatures, see Figs. S1(a) and S1(b) in the SM],
capturing the LLPS process during the first 60 s. The
x-ray intensity increases rapidly during the early time of the
phase separation, and the position of the spinodal peak q
max
shifts to smaller values, indicating an increase in length
scales of the concentration fluctuations. Appropriately
normalizing the intensity and wave vectors by the respec-
tive peak intensities Iðq
max
Þ and positions q
max
, we obtain a
master plot as expected for spinodal decomposition [47]
[see inset, Fig. 1(a) and Figs. S1(c) and S1(d)]. Iðq
max
Þ
shows a rapid increase during the early stage, with a rate
determined by the quench depth, while at around 10 s the
growth starts to slow down considerably [Fig. 1(b)]. This
slowdown is more pronounced when quenched to lower
temperatures, while at higher temperatures, IðqÞ continues to
grow even beyond 40 s, albeit at a slower rate. Insights into
the dynamics during the LLPS are obtained by analyzing a
time series of the coherent x-ray speckle patterns. For this, we
calculate two-time correlation functions (TTCs) for specific
scattering vectors q and different quench depths via
Cðt
1
;t
2
;qÞ¼
h½Iðt
1
Þ −
¯
Iðt
1
Þ½Iðt
2
Þ −
¯
Iðt
2
Þi
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
½
I
2
ðt
1
Þ −
¯
I
2
ðt
1
Þ½I
2
ðt
2
Þ −
¯
I
2
ðt
2
Þ
q
; ð1Þ
with h·i being an average over detector pixels corresponding
to a specific range q δq (calculated for q from 5 to 11 μm
−1
and δq ¼ 0.3 μm
−1
) and
¯
Iðt
1
Þ¼hIðt
1
Þi. From this quantity,
the time dependent g
2
ðt; t
w
;qÞ¼Cðt þ t
w
;t
w
;qÞ have
been extracted by horizontal cuts along t
1
, starting at the
diagonal of the respective TTC [48].
Figure 2 shows in the upper panel the TTCs and in the
lower panels the correlation function g
2
ðt; t
w
;qÞ for quench
temperatures of T
q
¼ 15, 8, and 4 °C (from left to right),
respectively. We identify three stages of the dynamics
during LLPS: the first stage appears directly after the
quench in the first 20 s, and its very fast dynamics is visible
only by a thin line in the TTC and by the final part of the
decay in the correlation function. After this early stage, the
TTC shows a pronounced slowing down of the dynamics
accompanied by a rising background level visible by the
appearance of a squarelike feature. This corresponds to a
second relaxation mode with a much slower relaxation time
appearing in the g
2
functions. The contribution of this
relaxation channel to the overall decay is increasing rapidly.
After ca. 40 s, the fast process in the LLPS is coming to an
end, as evidenced by the disappearance of the first decay.
(a)
0.005 0.01 0.015 0.02
q (nm
-1
)
0
100
200
300
400
I(q) (a.u.)
0
10
20
30
40
50
t
w
(s)
123
q/q
max
0
0.5
1
I(q)/I(q
max
)
(b)
FIG. 1. (a) Intensity as a function of scattering vector q for a
quench temperature T
q
¼ 10 °C; in the inset, the rescaled
intensity. The different colors correspond to different waiting
times t
w
as indicated in the color bar. The time t
w
¼ 0 is the time
at which the quench temperature T
q
was reached. (b) The
intensity at the peak position as a function of time for different
temperatures. The error bars are within the symbol size if they are
not visible.
FIG. 2. In the upper panels, the two-time correlation function
for T
q
¼ 15, 8, and 4 °C (left to right) at q ¼ 0 .005 nm
−1
are
displayed, and the lower panels show the corresponding g
2
functions at different waiti ng times t
w
.
PHYSICAL REVIEW LETTERS 126, 138004 (2021)
138004-2
In the third stage, later also called the late stage, the TTC
evidences a second slowing down process, with the
relaxation time depending on the final quench temperature
with slower dynamics visible for lower temperatures (from
T
q
¼ 15 °C to T
q
¼ 4 °C, on the top row). The g
2
ðt; t
w
;qÞ
functions have been fitted by a sum of Kohlrausch-
Williams-Watts (KWW) functions [49], resulting in the
following equation:
g
2
ðt; t
w
;qÞ¼A
1
exp
−
2
t
τ
1
γ
1
þ A
2
exp
−
2
t
τ
2
γ
2
;
ð2Þ
with decorrelation times τ
1
and τ
2
, relaxation amplitudes A
1
and A
2
, and KWW exponents γ
1
and γ
2
. We note that for
some g
2
functions a third exponential decay with small
amplitude and slow decay is needed to describe additional
tails in g
2
at very long timescales. However, due to the low
statistics of the third decay, we evaluate only the two
leading decays here.
Figure 3(a) displays the relaxation times as a function of
waiting time t
w
. We identify an exponential increase of the
relaxation time in the early stage of the LLPS and faster
dynamics with increasing quench depths [see Fig. 3(c) and
inset of Fig. 3(a)]. At this early phase of the spinodal
decomposition, the dynamics is driven by the curvature of
the free energy ∂
2
F=∂c
2
, which increases in magnitude
with deeper quenches (lower temperatures) and thus speeds
up the dynamics. This picture reverses when the slow
relaxation sets in [τ
2
in Fig. 3(c)]. Now the relaxation times
for the lower temperatures are considerably larger than the
high temperature quenches.
The transition from dynamics dominated by the fast
decay to dynamics dominated by the slow decay is
quantified by the nonergodicity parameter [50], here
defined as f ¼ A
2
=ðA
1
þ A
2
Þ. A rather sharp increase of
the nonergodicity parameter around t
w
¼ 30 s is observed
[Fig. 3(b)], with a later rise time and slower increase for
lower temperatures, suggesting that the transition to coars-
ening dynamics is already slowed down by the lower
mobility at low temperatures.
These three stages in the dynamics can be linked to the
evolution of the LLPS. The first corresponds to the early
stage of a spinodal decomposition, displaying an enhance-
ment of the concentration fluctuations and the formation of
an interface between two different phases. The third one is
the late stage in which the domains are growing. For our
system, this seems to be controlled by the mass transport
mechanism and by the mobility of the involved compo-
nents. In other cases in condensed matter physics, different
domain growth mechanisms have been identified [51]. The
second stage is a transition between the early and late stage.
Important for the assessment of the form of the dynamics
and relevant for computing quantities such as diffusion and
transport coefficients during the LLPS are the relaxation
rate ΓðqÞ¼1=τðqÞ and the KWW exponents, including
their q dependence [32]. We find in the early phase of the
LLPS a linear relationship Γ ∝ q [Fig. S2(a)] and γ
1
around
1.4 to 1.7 (Fig. S2) with no pronounced q dependence. In
contrast, the dynamics during the late coarsening stage also
displays a linear (Γ ∝ q) behavior but now with a pro-
nounced q dependence of the KWW exponents, which are
decreasing from values of γ
2
¼ 2 at small q values to
γ
2
¼ 1 at large values of q [Fig. S2(d)]. Our results clearly
show that the coarsening dynamics of the protein droplets
at this length scale are not governed by Brownian dynamics
but instead by a ballistic and partially cooperative motion of
the protein droplets driven by the spinodal decomposition.
Similar super diffusive ballistic types of dynamics have
been observed frequently in soft matter systems, e.g., in the
late stage of colloidal gelation processes [52–54], during
spinodal decomposition of colloidal systems [27], and in
the framework of MD simulations of metallic glasses [55]
with strongly interconnected clusters moving together.
Phenomenological models of single microcollapse events
(a)
10
1
10
2
t
w
(s)
10
0
10
1
10
2
(s)
5 10152025
2
4
6
8
(b)
0 1020304050
t
w
(s)
0
0.2
0.4
0.6
0.8
1
Non-ergodicity parameter
T
q
=15
o
C
T
q
=12
o
C
T
q
=10
o
C
T
q
=8
o
C
T
q
=6
o
C
T
q
=4
o
C
(c)
5
10
15
1
(s)
4 6 8 10121416
T
q
(
o
C)
100
200
300
2
(s)
t
w
=22s
t
w
=96s
FIG. 3. (a) Decorrelation time for q ¼ 0.005 nm
−1
as a function of waiting time t
w
. The open symbols correspond to the slow mode
(τ
2
) and the filled symbols to the fast one (τ
1
). Inset: linear representation. (b) Nonergodicity parameter f ¼ A
2
=ðA
1
þ A
2
Þ as a function
of t
w
. The legend also applies to (a). (c) Decorrelation time of the fast and slow modes as a function of quench temperature T
q
,
respectively.
PHYSICAL REVIEW LETTERS 126, 138004 (2021)
138004-3
have been put forward as possible explanations [56]
underlying such unusual dynamics and subsequently
extended to series of such intermittent events to account
for both q dependence of the relaxation rate and the KWW
exponents [57,58]. The connection of these phenomeno-
logical models to the dynamics during an LLPS is not
obvious, especially with regard to typical field theories
such as the Cahn-Hilliard equation (CHE) used to model
spinodal decomposition. A shared feature between colloi-
dal gels and the LLPS, investigated here, is the presence of
elastic deformations. These can arise in an LLPS from the
viscoelastic properties, which are caused by a dynamical
coupling of diffusion and stresses, and could explain
the similarities between these observations [57]. In fact,
the dynamical asymmetry of the two phases present in the
system can lead to viscoelastic phase separation [59].
To obtain more insight into the underlying dynamical
mechanisms, we compare the experimental data to numeri-
cal simulations of the temporal evolution of the CHE
[60,61] (here in 2D). The dynamics of the ordinary CHE
shows faster dynamics for quenches to lower temperatures.
As this is not what we observe in the experimental data, we
introduce a dependence of the mobility μ on protein density
ρðr; tÞ, representing the slowdown of the dynamics when the
density of the dense phase is increasing and its mobility
freezes out [45] in the spirit of model C according to Ref. [2].
The parameter c
g
sets the concentration at which the mobility
decreases (for details, see the SM). The simulations then
yield a time dependent real space configuration ρðr; tÞ of the
protein density [Fig. 4(c)], which is converted into an x-ray
speckle pattern by means of its Fourier components jρðq; tÞj
2
and analyzed with the same time correlation methods as
applied to the experimental data [45,46].
Figure 4(a) displays the temporal evolution of the
spatially averaged mobility ¯μðt
w
Þ in the dense phase and
Fig. S10 the spatially averaged density
¯
ρðt
w
Þ of the dense
and dilute phases for different values of c
g
. A point in the
2D image was considered part of the dense phase if its
concentration was higher than the initial concentration. The
mobility of the dense phase drops quickly during the LLPS
upon lowering the c
g
value. The densities in the dense and
dilute phases do not reach their equilibrium values any-
more, which is considered an arrest of the LLPS [45,46].
Typical real space configurations are shown in Fig. 4(b)
with the LLPS visible via the formation of domains of the
diluted phase (blue) in a host matrix of the dense phase
(red). The corresponding density profiles (see Fig. S9)
show the typical hallmarks of the spinodal decomposition,
with density fluctuations developing quickly and reaching
rather smooth profiles at the end of the LLPS for a mobility
that is independent of protein density and equal to 1. In
contrast, for c
g
¼ 0.6, we observe that the density fluctua-
tions become immobile when reaching the threshold value,
leading to smaller domains.
TTC and g
2
functions were computed following the
procedure of experimental data. Figure 4(c) displays the
TTC for the simulations and the corresponding g
2
func-
tions. We identify a fast dynamic process during the early
time of the LLPS, which represents the dynamics of
spinodal interface formation between the dense and diluted
phases. The dynamics quickly slows down with a sudden
appearance of a second much slower relaxation process
when the dense phase has formed. The relaxation times
during the first stage are determined by the quench depth
(smaller for deeper quenches) because the curvature of the
free energy is here the main driving force for the velocity of
interface formation. The corresponding KWW values γ
1
and γ
2
[Figs. S11(c) and S11(d)] are between 1.5 and 2.5,
which is in good agreement with our experimental values.
Comparing the real space data and the parameters
describing the dynamics, it is possible to see that the rise
(a)
10
2
10
3
t
w
0
0.2
0.4
0.6
0.8
1
(b)
(r,t
w
=10)
(r,t
w
=70)
(r,t
w
=90)
(r,t
w
=130)
0
0.5
1
(c)
FIG. 4. The simulated ¯μ of the dense phase (a) as a function of waiting time. In brown, the case with constant mobility and the quench
depth for this case was set to T
q
¼ 0.2 T
c
, with T
c
being the critical temperature [62,63]. In red, c
g
¼ 0.8 and T
q
¼ 0.12 T
c
, and in
yellow, c
g
¼ 0.6 and T
q
¼ 0.08 T
c
. (b) Examples of real space configurations for T
q
¼ 0.14 T
c
and mobility at different waiting times
with the densities indicated by the color bar. (c) Upper panel: two-time correlation function of the simulated data for q ¼ 1. From left to
right, the parameters T
q
and c
g
of the simulations are T
q
¼ 0.2 T
c
and c
g
¼ ∞, which correspond to a constant mobility set to 1,
T
q
¼ 0.12 T
c
and c
g
¼ 0.8, and T
q
¼ 0.08 T
c
and c
g
¼ 0.6. (c) Lower panel: corresponding g
2
functions at different waiting times t
w
.
PHYSICAL REVIEW LETTERS 126, 138004 (2021)
138004-4
of the nonergodicity parameter starts during the onset of
domain and coarsening dynamics. In this short transition
time (see Fig. 3), domain coarsening is progressing in
parallel with the final interface formation. After this
transition, the domain coarsening is taking place. The
dynamics seems to be composed of three processes: the
coalescence of different domains, domain growth, and
the spatial movement that is guided by the surface tension
and precedes the coalescence of two domains. Its decorre-
lation time is larger for deeper quenches because of the
lower transition concentration of the mobility [Fig. 4(c) and
Fig. S11], representing a situation in which the temper-
ature-dependent glass line of the dense host phase intersects
the spinodal phase region. For the lower value of c
g
, the
diluted phase is essentially trapped inside the frozen host
matrix [red in Fig. 4(b)].
Based on the simulation r esults, we can conclude that
thekineticsofthesystemsiscapturedbymodelCas
suggested before [2]. The d ynamics, however, is only
partially reproduced. While the slowdown of the late s tage
dynamics with increasing quench depth is correctly
described, as well as the presence of two relaxation
modes, model C does not reproduce the Γ ∝ q behavior
observed in the experiment or the q dependence of the
KWW values of the slow dynamics (Fig. S11). We
speculate that this is due to the neglect of the viscoelastic
properties of the system in the CHE, which would lead to
elastic deformation and motion inducing a linear
dispersion relation.
In conclusion, we demonstrated the technique of low
radiation dose XPCS and used it to study protein dynamics
during an LLPS. The method delivers simultaneously
information on the collective dynamics via XPCS and
the structural evolution via the ensemble averaged scatter-
ing IðqÞ. The two-time correlation maps provide a high
level of detail of the dynamics during a spinodal decom-
position of Ig in solution with PEG. We identify distinctly
different dynamical regimes of the LLPS with different
temperature behaviors. The early stage dynamics reflects
concentration fluctuation and interface formation and is
faster for lower temperatures, reflecting stronger quench
depths. In contrast, the later stage of coarsening is slower
for lower temperatures, which is caused by the reduced
mobility of the slowed down proteins comprising the host
matrix. With simulations, we were able to identify a
concentration and time dependence of the molecular-scale
mobility that connects the dynamics of the condensate to
molecular-scale quantities.
The authors acknowledge discussions with M. Oettel,
financial support from the DFG and the BMBF
(05K20VTA), and the allocation of beamtime by DESY.
N. B. acknowledges the Alexander von Humboldt-Stiftung
for a postdoctoral fellowship. A. R. acknowledges the
Studienstiftung des Deutschen Volkes for a Ph.D.
fellowship. C. G. acknowledges BMBF (Grants
No. 05K19PS1 and No. 05K20PSA) for financial support.
*
fajun.zhang@uni-tuebingen.de
†
christian.gutt@uni-siegen.de
‡
frank.schreiber@uni-tuebingen.de; https://publons.com/
researcher/2502617/frank-schreiber/
[1] C. P. Brangwynne, C. R. Eckmann, D. S. Courson, A.
Rybarska, C. Hoege, J. Gharakhani, F. Jülicher, and
A. A. Hyman, Germline p granules are liquid droplets that
localize by controlled dissolution/condensation, Science
324, 1729 (2009).
[2] J. Berry, C. P. Brangwynne, and M. Haataja, Physical
principles of intracellular organization via active and passive
phase transitions, Rep. Prog. Phys. 81, 046601 (2018).
[3] Y. Shin and C. P. Brangwynne, Liquid phase condensation
in cell physiology and disease, Science 357, eaaf4382
(2017).
[4] A. A. Hyman and C. P. Brangwynne, Beyond stereospeci-
ficity: Liquids and mesoscale organization of cytoplasm,
Dev. Cell 21, 14 (2011).
[5] L. Malinovska, S. Kroschwald, and S. Alberti, Protein
disorder, prion propensities, and self-organizing macromo-
lecular collectives, Biochim. Biophys. Acta Proteins Pro-
teomics 1834, 918 (2013).
[6] S. C. Weber and C. P. Brangwy nne, Getting RNA and
protein in phase, Cell 149, 1188 (2012).
[7] E. Zaccarelli, Colloidal gels: Equilibrium and non-equilib-
rium routes, J. Phys. Condens. Matter 19, 323101 (2007).
[8] S. Manley, H. M. Wyss, K. Miyazaki, J. C. Conrad, V.
Trappe, L. J. Kauf man, D. R. Reichman, and D. A. Weitz,
Glasslike Arrest in Spinodal Decomposition as a Route to
Colloidal Gelation, Phys. Rev. Lett. 95, 238302 (2005).
[9] J. C. Conrad, H. M. Wyss, V. Trappe, S. Manley, K.
Miyazaki, L. J. Kaufman, A. B. Schofield, D. R. Reichman,
and D. A. Weitz, Arrested fluid-fluid phase separation in
depletion systems: Implications of the characteristic length
on gel formation and rheology, J. Rheol. 54, 421 (2010).
[10] P. J. Lu, E. Zaccarelli, F. Ciulla, A. B. Schofield, F.
Sciortino, and D. A. Weitz, Gelation of particles with
short-range attraction, Nature (London) 453, 499 (2008).
[11] T. Gibaud and P. Schurtenberger, A closer look at arrested
spinodal decomposition in protein solutions, J. Phys. Con-
dens. Matter 21, 322201 (2009).
[12] T. Gibaud, F. Cardinaux, J. Bergenholtz, A. Stradner, and P.
Schurtenberger, Phase separation and dynamical arrest for
particles interacting with mixed potentials—The case of
globular proteins revisited, Soft Matter 7, 857 (2011).
[13] F. Cardinaux, T. Gibaud, A. Stradner, and P. Schurtenberger,
Interplay between Spinodal Decomposition and Glass For-
mation in Proteins Exhibiting Short-Range Attractions,
Phys. Rev. Lett. 99, 118301 (2007).
[14] S. Da Vela, M. K. Braun, A. Dörr, A. Greco, J. Möller, Z.
Fu, F. Zhang, and F. Schreiber, Kinetics of liquid-liquid
phase separation in protein solutions exhibiting LCST phase
behavior studied by time-resolved USAXS and VSANS,
Soft Matter 12, 9334 (2016).
PHYSICAL REVIEW LETTERS 126, 138004 (2021)
138004-5