research papers
Automated matching of two-time X-ray photon correlation maps from phase-separating proteins with Cahn–Hilliard-type simulations using auto-encoder networks
aDepartment Physik, Universität Siegen, Walter-Flex-Strasse 3, 57072 Siegen, Germany, bInstitut für Angewandte Physik, Universität Tübingen, Auf der Morgenstelle 10, 72076 Tübingen, Germany, cDepartment of Physics, AlbaNova University Center, Stockholm University, SE-106 91 Stockholm, Sweden, and dDeutsches Elektronen-Synchrotron DESY, Notkestraße 85, 22607 Hamburg, Germany
*Correspondence e-mail: sonja.timmermann@uni-siegen.de
Machine learning methods are used for an automated classification of experimental two-time X-ray photon correlation maps from an arrested liquid–liquid
of a protein solution. The correlation maps are matched with correlation maps generated with Cahn–Hilliard-type simulations of liquid–liquid phase separations according to two simulation parameters and in the last step interpreted in the framework of the simulation. The matching routine employs an auto-encoder network and a differential evolution based algorithm. The method presented here is a first step towards handling large amounts of dynamic data measured at high-brilliance synchrotron and X-ray free-electron laser sources, facilitating fast comparison with phase field models of phase separation.Keywords: protein dynamics; X-ray photon correlation spectroscopy; XPCS; Cahn–Hilliard; auto-encoders; machine learning.
1. Introduction
X-ray photon correlation spectroscopy (XPCS) is an experimental method capable of accessing the dynamics of protein systems over length scales that range from the atomic to micrometre scale, and timescales from microseconds to hours (X. Lu et al., 2008; Zhang et al., 2011, 2017; Madsen et al., 2016; Möller et al., 2016; Zinn et al., 2018; Perakis & Gutt, 2020; Ruta et al., 2012; Begam et al., 2021). In XPCS, time series of coherent X-ray speckle patterns are measured, giving access to dynamical properties via time-resolved correlation maps – the two-time correlation function (TTC) (Brown et al., 1997; Bikondoa, 2017). With fast megapixel 2D X-ray detectors, XPCS experiments can produce large quantities of data in a short time interval with up to thousands of TTCs per hour of beamtime. Given the typical duration of a synchrotron experiment of a few days, the resulting quantities of TTCs are difficult to handle. Therefore, there is a generic need for methods facilitating fast and reliable analysis including classification of the data. This is especially important with regards to steering, selecting and controlling the experimental parameters during the beamtime. However, in the aftermath of the experiment, quick classification methods are important for benchmarking models which, for example, simulate a gel transition or the solidification process of a protein solution upon liquid–liquid (LLPS).
Phase separation is a ubiquitous process in nature, with applications and consequences for a wide range of scientific disciplines such as solid-state physics, material sciences and biology. ; Schurtenberger et al., 1989; Broide et al., 1991; Berland et al., 1992; Muschol & Rosenberger, 1997) constitutes a possible pathway for organizing membraneless structures in living cells (Brangwynne et al., 2009; Shin & Brangwynne, 2017; Berry et al., 2018). Detailed investigations were carried out with respect to the biological functions of these protein condensates, encompassing biochemical reaction rates, buffering protein concentrations, and sensing or signaling (Shin & Brangwynne, 2017). A variety of diseases which result from a loss or change of function of these condensates are accompanied by phase transitions (Malinovska et al., 2013; Weber & Brangwynne, 2012).
in biological systems attracted considerable attention with the discovery that LLPS in protein solutions (Ishimoto & Tanaka, 1977Phase separation is initiated by quenching a system into the metastable region of the phase diagram, launching a process of out-of-equilibrium self-organization (Dong & Granick, 2021). The state of the condensates depends on the dynamic and kinetic processes during their formation, with dynamical asymmetries between the two phases on a hierarchy of length and timescales and invoking viscoelastic properties of the resulting network structures (Berry et al., 2018; Zaccarelli, 2007; Tanaka, 2000). In the biological context, an understanding of the component timescales is important as they are expected to meet the intrinsic timescales of biochemical processes in the condensates.
Upon et al., 2005; Conrad et al., 2010; P. J. Lu et al., 2008). Employing time-resolved scattering experiments, the kinetics during arrested phase separations have been observed frequently as a slowdown of the growth of the static S(Q) in both Q position and peak height for low-temperature quenches (Gibaud & Schurtenberger, 2009; Gibaud et al., 2011; Cardinaux et al., 2007; Da Vela et al., 2016, 2017, 2020; Bucciarelli et al., 2015). In contrast, the dynamics of protein solutions evolving into an arrested are still poorly understood as studies require the concurrent monitoring of an extraordinarily broad range of length and timescales. Likewise, experimental validation of models of the dynamics of critical phenomena during LLPS like the Cahn–Hilliard equation (CHE) (Cahn, 1965) and other correlated models is still needed, particularly with respect to glass–gel transitions displaying considerable dynamical asymmetries between the dilute and concentrated phase (Berry et al., 2018; Girelli et al., 2021; Ragulskaya et al., 2021).
the dynamics slow down on molecular length scales due to local concentration changes. This microscopic deacceleration can eventually lead to the arrest of on larger length scales complemented by the formation of bicontinuous gel network structures – as observed in colloidal and protein systems (ManleyMachine learning (ML) has already proven to be useful in the analysis of various structural X-ray data from small-angleX-ray scattering (Archibald et al., 2020; Chen & Pollack, 2020; Franke et al., 2018; Wang et al., 2017) and wide-angle X-ray scattering (Chen & Pollack, 2020; Wang et al., 2017), diffraction (Berntson et al., 2003; Oviedo et al., 2019; Vecsei et al., 2019), and reflectometry (Greco et al., 2019) experiments. The algorithms provide real-time feedback (Wang et al., 2017; Ke et al., 2018; Greco et al., 2019) during the experiment or identify and remove bad images to reduce the stored data volume (Wang et al., 2017; Ke et al., 2018). In many cases the ML algorithms have been trained with synthetic structural data generated from existing databases on proteins (Franke et al., 2018; Archibald et al., 2020), RNA (Chen & Pollack, 2020) or crystal structures (Oviedo et al., 2019; Vecsei et al., 2019), which enables them to classify the results obtained in experiments without an expert spending time on labeling large quantities of data for training (Berntson et al., 2003). Recently, ML algorithms have also been trained to reduce the noise of experimental TTCs (Konstantinova et al., 2021). The CHE has also been used in different ML tasks (Wight & Zhao, 2021; Pokuri et al., 2019; Zhang & Garikipati, 2020; Farimani et al., 2018).
The focus of X-ray scattering applications of ML has been mainly on static properties up to now. We note, however, that many processes in nature, such as the LLPS of protein solutions, display dynamic phenomena which require the development of schemes and methods capable of classifying the corresponding dynamic X-ray signatures of the processes. ML-based classification of process data evolving in time is computationally very expensive due to the large quantity of training data required. Employing a minimum 2D model which allows a network with finite computer resources to be trained, we report on a first step towards this goal and present a neural network method for analysis and classification of dynamical X-ray information during LLPS.
The CHE in a simplified 2D version still captures the essence of the physics involved (Sciortino et al., 1993; Sappelt & Jäckle, 1997; Lamorgese & Mauri, 2009), while allowing us to generate a large number of simulations on the timescale of hours, and couple it with a concentration-induced gelation as proposed by Sciortino et al. (1993) and Sappelt & Jäckle (1997). We aim for an automated assignment of the measured TTCs in the work of Girelli et al. (2021) with TTCs from simulations that differ in the parameters ε, i.e. proportional to reduced quench depth, and , i.e. the at which the mobility drops by a factor of 1/2. We train an auto-encoder neural network with TTC simulation data and employ a differential evolution based algorithm for matching encoded experimental TTCs with simulations. Our work constitutes a first step to making use of the static, kinetic and dynamical information contained in XPCS data for ML-based analysis of processes in phase separation.
2. Analysis and results
2.1. Simulation of the Cahn–Hilliard equation
We model the dynamic processes during the et al., 1993; Gunton et al., 1983),
with the help of the CHE (SciortinoHere is the t and position which is related to the local protein concentration Φ by (Sciortino et al., 1993). The symmetry breaking is introduced by the parameter ε which can be identified as being proportional to the reduced quench depth , where is the upper of the system. If ε fulfills the condition (e.g. Gunton et al., 1983), the system is in an unstable state below the spinodal curve and separates into regions with higher and lower local protein concentrations [Figs. 1(a) and 1(b)].
at a given timeThe M in equation (1) (Sappelt & Jäckle, 1997) which reduces the mobility of the highly concentrated phase via
by is coupled to a gelation process by introducing a concentration-dependent mobility parameterIn this model the mobility decreases from values M = 1 to when the Ψ exceeds the value of [Fig. 1(c)]. The parameter α is fixed to providing a steep decrease of the mobility as required for an arrested (Sappelt & Jäckle, 1997).
We perform the simulation on a 2D grid of size 256×256. The grid uses periodic boundary conditions and is initialized with a mean of such that we observe droplets with a low concentration instead of an interconnected network, which would be the case for . The fluctuations necessary for LLPS are added as random noise with a maximum amplitude of . Ten thousand time steps are calculated for each simulation with a fixed time step of . The temporal evolution of the for different parameter configurations is depicted in Figs. 1(a) and 1(b). Apart from the initial noise configuration, the simulations differ only in the values of the parameters ε and . Both were systematically varied within the intervals
These intervals were chosen with , thus avoiding an arrest of the ε as for small values of ε the concentration of the dense phase is too low [gray line in Fig 1(a)] to introduce visible changes in the simulations for different values of .
before the formation of droplets. We restrict our analysis to large values ofA single simulation run yields a time series of the protein density's real-space configuration [Fig. 2(a)] which is converted to an X-ray speckle pattern by means of Fourier transform, yielding a 256×256 image. The temporal evolution of the square of the azimuthal integration of these Fourier images is shown in Fig. 2(b) and is equivalent to the We introduce RQ as the distance from the center of the Fourier image in units of pixels. In Fig. 2(b) the occurrence of a peak indicates the phase separation.
The dynamics on a certain length scale in Fourier space can be traced by correlating pixelwise the corresponding intensities with a TTC,
with being the average over pixels within a distance to the center.
The azimuthally integrated intensity profile [Fig. 2(b)] shows the growth dynamics of the droplets by a shift of the peak position towards smaller values of RQ. These dynamics can be observed in the TTC by choosing a value of RQ such that it is located at the falling slopes of the peaks in Fig. 2(b), which is . For our training data, we chose RQ to be an integer in this range, while increases linearly from 0.7 at RQ = 26 to 3 at RQ = 45, as done in experimental analyses as well to compensate for lower photon counts at higher Q values.
2.2. Experimental data
The XPCS experiments (Fig. 3) were conducted at the Coherence Applications beamline P10 at PETRA III, DESY, employing an X-ray beam of photon energy 8.54 keV, size 100 × 100 µm and maximum photon density 107 photons s−1 µm−2 [for further experimental details, see Girelli et al. (2021)]. We measured time series of coherent diffraction patterns that were collected with an EIGER 4 Mpixel detector covering a Q range from 3 to 50 µm−1. The samples consisted of G (IgG) with polyethylene glycol (PEG) and NaCl, and were quenched from 37°C to six different quench temperatures below the binodal line. Further details about the sample preparation can be found in the work of Da Vela et al. (2017).
In the experiment, 4000 frames were recorded for every quench temperature. The sample was illuminated for 0.02 s followed by 0.1 s where the shutter was closed such that the measurement covers a time range of 4000 × 0.12 s = 480 s. The experimental TTCs were calculated similarly to the TTCs from the simulation using equation (4). Regarding the temporal shift of the integrated intensity profiles as a function of Q, the falling edge can be restricted to [7.5, 10] µm−1. The contrast of the experimental TTCs is lower than that of the simulated ones where we assume completely coherent scattering and detection. This issue is addressed by normalizing the experimental TTCs to the speckle visibility which is determined by the values of the pixels adjacent to the diagonal of the experimental TTC, as the values directly on the diagonal are distorted by shot noise (Inoue et al., 2012).
2.3. Data preparation and auto-encoding
For the training data, 12 000 pairs of the parameters ε and were sampled from a uniform distribution within the intervals. To increase the simulation robustness against the random initial conditions, TTCs from five simulations with different initial noise were averaged. Thus, 5 × 12 000 = 60 000 simulations were performed in total, which took 7 h on two NVIDIA 2080 Ti graphic cards. TTCs were calculated for 20 RQ ranges, so that the training data consisted of 20 × 12 000 = 240 000 TTCs.
All simulated TTCs display a broad, slow part in the beginning (Fig. 4) which is absent in the experimental data. This part is associated with the still homogeneous phase at early times before the onset of density differences [Fig. 2(a), t = 2000]. Approaching smaller values of RQ this slow part becomes even more pronounced.
These broad parts are eliminated by cropping the TTCs. The maximum of the intensity in the associated Q rings is well defined for both simulated and experimental TTCs and is chosen to be the time where the TTCs are cropped. By doing this, the early part of the is removed from the TTC, such that only the coarsening stages of simulated and experimental TTCs are compared. The cutting procedure is illustrated in Fig. 4 for three simulated TTCs. For different models, this preprocessing via cropping needs to be adapted to remove, for example, well known deficiencies from simulated TTCs or make sure that the timescales of simulation and experiment are comparable. We derive the latter from the facts that the intensity curves show similar behavior within the coarsening stage for both experimental and simulated data and the TTCs show the same overall features in the whole measurement time (Girelli et al., 2021). The cropped TTCs are finally scaled down to a 64×64 resolution.
Instead of matching raw TTCs between the experiment and the simulations, we compare their encoded representations via an auto-encoder (Rumelhart et al., 1986). The auto-encoder network (Starostin, 2022) was trained with the simulated TTCs, encoding them in 32-dimensional vectors . For the encoder, the first three layers of the pretrained ResNet-18 model (He et al., 2016) were used as a feature extractor, followed by adaptive average pooling with an output size of 3×3 and a fully connected network. The decoder architecture follows the work of Hou et al. (2017).
The performance of the auto-encoder network on simulated and experimental data is shown in Fig. 5, from which one can see that the network is capable of reducing the experimental TTCs to the main dynamic components. Thus, the encoded representations allow comparison of TTCs in a more reliable and faster way. If a series of TTCs from the same simulation but at different values of RQ is created, the entries of the vector change continuously (Fig. 6) which enabled us to interpolate TTCs for arbitrary values of RQ in between the selected values.
2.4. Results
Our goal is to develop tools for a fast classification of experimental TTCs in the context of different models. We approach this task via matching encoded TTCs from measurements at different quench temperatures with encoded TTCs from simulations of our simplified model. During this matching process, we find the quench temperature dependence of the two simulation parameters and the values of the three additional calibration parameters , where and are the times at which the experimental TTCs are cropped such that they match the timescale of the simulation. denotes the factor for converting the Q range in units of pixels, which were used in the simulation, to units of inverse micrometre as they were used in the experiment via
A differential evolution (Price, 1996) based algorithm was employed for this matching process. For each of the six experimental measurements, TTCs were calculated at five different Q values,
such that we obtain a set of 30 experimental TTCs. The differential evolution algorithm optimizes the mean dot product D between the normalized encoded representations of 30 experimental TTCs and 30 simulated TTCs from the training data set with respect to , , , and :
The choice of boundaries for the calibration parameters was based on the assumption that the duration of the experiment is longer than that of the simulation, and that the chosen Q range [7.02; 8.2] µm−1 lies within the used RQ range for the simulations. Knowing that ε is proportional to , we apply additional boundaries to consider only those solutions where the average value of ε is decreasing as a function of experimental quench temperature.
During the fitting procedure, the experimental TTCs were cropped and encoded for each generated combination of the parameters. A caching technique was used to accelerate these calculations. The encoded vectors of the simulated TTCs were taken from the training data and interpolated on the basis of the generated calibration coefficient . After 1500 iterations the result of the fit converged to .
Fig. 7(a) shows the averaged outcomes for the simulation parameters and ε. The rise in for shallower quenches indicates that the gelation is occurring at higher concentrations at these temperatures. The () pairs were integrated into a phase diagram [Fig. 7(b)] which displays the spinodal and binodal lines as determined from the Landau free energy [equation (1)]. The predictions of the neural network enable us to estimate the corresponding gel line [line in the lower-right part of Fig. 7(b) as a guide to the eye]. This line bends towards the spinodal line, resulting in lower concentrations of the dense phase for deeper quenches. Such a behavior of the gel line in the coexistence region has also been observed for the LLPS of BSA-YCl3 (Da Vela et al., 2020), lysozyme (Cardinaux et al., 2007) and γ-globulin (Da Vela et al., 2017).
Averaging over the 30 outcomes for the calibration parameters yields
A further validation of these results with respect to ε and for the experimental TTCs is difficult as neither parameter is directly accessible in the experiment.
3. Discussion and conclusion
In conclusion, we successfully trained an auto-encoder neural network with TTC data originating from Cahn–Hilliard simulations. This auto-encoder was used to extract the main dynamic components out of TTCs originating from XPCS experiments of LLPS in a model protein solution of IgG. We used a differential evolution based algorithm for matching encoded TTCs from the experiment and the simulation and determine thereby the parameters necessary for connecting length and timescales of simulation to experiment. The simulation parameters of the classified experimental TTCs were used to construct a phase diagram with the experimental data indicating the position of a gel line.
The training of neural networks based on dynamical simulation data is challenging. The Cahn–Hilliard model employed here, including the gelation, is rather simplistic and uses two labeling parameters only. However, we emphasize the possibility of extending this methodology to more complicated models capable of making strong predictions and eventually also benchmarking of theoretical models based on very large XPCS data sets, possibly by including the kinetic information in the model as well. This is a first step towards handling large amounts of XPCS data on short timescales and will be of general interest not only for protein dynamics but for material science aspects of spinodal decomposition.
Acknowledgements
ST and VS contributed equally to this work. Open access funding enabled and organized by Projekt DEAL.
Funding information
ST and CG acknowledge BMBF for financial support under grant Nos. 05K19PS1 and 05K20PSA. FZ and FS acknowledge BMBF for financial support under grant No. 05K20VTA. AR thanks the Studienstiftung des deutschen Volkes for a personal fellowship. NB thanks the Alexander von Humbold Foundation for a postdoctoral research fellowship.
References
Archibald, R. K., Doucet, M., Johnston, T., Young, S. R., Yang, E. & Heller, W. T. (2020). J. Appl. Cryst. 53, 326–334. Web of Science CrossRef CAS IUCr Journals Google Scholar
Begam, N., Ragulskaya, A., Girelli, A., Rahmann, H., Chandran, S., Westermeier, F., Reiser, M., Sprung, M., Zhang, F., Gutt, C. & Schreiber, F. (2021). Phys. Rev. Lett. 126, 098001. Google Scholar
Berland, C. R., Thurston, G. M., Kondo, M., Broide, M. L., Pande, J., Ogun, O. & Benedek, G. B. (1992). Proc. Natl Acad. Sci. USA, 89, 1214–1218. CrossRef PubMed CAS Web of Science Google Scholar
Berntson, A., Stojanoff, V. & Takai, H. (2003). J. Synchrotron Rad. 10, 445–449. Web of Science CrossRef CAS IUCr Journals Google Scholar
Berry, J., Brangwynne, C. P. & Haataja, M. (2018). Rep. Prog. Phys. 81, 046601. Web of Science CrossRef PubMed Google Scholar
Bikondoa, O. (2017). J. Appl. Cryst. 50, 357–368. Web of Science CrossRef CAS IUCr Journals Google Scholar
Brangwynne, C. P., Eckmann, C. R., Courson, D. S., Rybarska, A., Hoege, C., Gharakhani, J., Jülicher, F. & Hyman, A. A. (2009). Science, 324, 1729–1732. Web of Science CrossRef PubMed CAS Google Scholar
Broide, M. L., Berland, C. R., Pande, J., Ogun, O. O. & Benedek, G. B. (1991). Proc. Natl Acad. Sci. USA, 88, 5660–5664. CrossRef CAS PubMed Web of Science Google Scholar
Brown, G., Rikvold, P. A., Sutton, M. & Grant, M. (1997). Phys. Rev. E, 56, 6601–6612. Web of Science CrossRef CAS Google Scholar
Bucciarelli, S., Casal-Dujat, L., De Michele, C., Sciortino, F., Dhont, J., Bergenholtz, J., Farago, B., Schurtenberger, P. & Stradner, A. (2015). J. Phys. Chem. Lett. 6, 4470–4474. Web of Science CrossRef CAS PubMed Google Scholar
Cahn, J. W. (1965). J. Chem. Phys. 42, 93–99. CrossRef CAS Web of Science Google Scholar
Cardinaux, F., Gibaud, T., Stradner, A. & Schurtenberger, P. (2007). Phys. Rev. Lett. 99, 118301. Web of Science CrossRef PubMed Google Scholar
Chen, Y.-L. & Pollack, L. (2020). IUCrJ, 7, 870–880. Web of Science CrossRef CAS PubMed IUCr Journals Google Scholar
Conrad, J. C., Wyss, H. M., Trappe, V., Manley, S., Miyazaki, K., Kaufman, L. J., Schofield, A. B., Reichman, D. R. & Weitz, D. A. (2010). J. Rheol. 54, 421–438. Web of Science CrossRef CAS Google Scholar
Da Vela, S., Begam, N., Dyachok, D., Schäufele, R. S., Matsarskaia, O., Braun, M. K., Girelli, A., Ragulskaya, A., Mariani, A., Zhang, F. & Schreiber, F. (2020). J. Phys. Chem. Lett. 11, 7273–7278. Web of Science CrossRef CAS PubMed Google Scholar
Da Vela, S., Braun, M. K., Dörr, A., Greco, A., Möller, J., Fu, Z., Zhang, F. & Schreiber, F. (2016). Soft Matter, 12, 9334–9341. Web of Science CrossRef CAS PubMed Google Scholar
Da Vela, S., Exner, C., Schäufele, R. S., Möller, J., Fu, Z., Zhang, F. & Schreiber, F. (2017). Soft Matter, 13, 8756–8765. Web of Science CrossRef CAS PubMed Google Scholar
Dong, R.-Y. & Granick, S. (2021). Nat. Commun. 12, 911. Web of Science CrossRef PubMed Google Scholar
Farimani, A., Gomes, J., Sharma, R., Lee, F. L. & Pande, V. (2018). arXiv:1803.08993. Google Scholar
Franke, D., Jeffries, C. M. & Svergun, D. I. (2018). Biophys. J. 114, 2485–2492. Web of Science CrossRef CAS PubMed Google Scholar
Gibaud, T., Cardinaux, F., Bergenholtz, J., Stradner, A. & Schurtenberger, P. (2011). Soft Matter, 7, 857–860. Web of Science CrossRef CAS Google Scholar
Gibaud, T. & Schurtenberger, P. (2009). J. Phys. Condens. Matter, 21, 322201. Web of Science CrossRef PubMed Google Scholar
Girelli, A., Rahmann, H., Begam, N., Ragulskaya, A., Reiser, M., Chandran, S., Westermeier, F., Sprung, M., Zhang, F., Gutt, C. & Schreiber, F. (2021). Phys. Rev. Lett. 126, 138004. Web of Science CrossRef PubMed Google Scholar
Greco, A., Starostin, V., Karapanagiotis, C., Hinderhofer, A., Gerlach, A., Pithan, L., Liehr, S., Schreiber, F. & Kowarik, S. (2019). J. Appl. Cryst. 52, 1342–1347. Web of Science CrossRef CAS IUCr Journals Google Scholar
Gunton, J., San Miguel, M. & Sahni, P. (1983). Phase Transitions and Critical Phenomena, edited by C. Domb & J. Lebowitz, Vol. 8, 2nd ed., pp. 269–466. London: Academic Press. Google Scholar
He, K., Zhang, X., Ren, S. & Sun, J. (2016). Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 770–778. IEEE. Google Scholar
Hou, X., Shen, L., Sun, K. & Qiu, G. (2017). 2017 IEEE Winter Conference on Applications of Computer Vision (WACV), pp. 1133–1141. IEEE. Google Scholar
Inoue, I., Shinohara, Y., Watanabe, A. & Amemiya, Y. (2012). Opt. Express, 20, 26878–26887. Web of Science CrossRef CAS PubMed Google Scholar
Ishimoto, C. & Tanaka, T. (1977). Phys. Rev. Lett. 39, 474–477. CrossRef CAS Web of Science Google Scholar
Ke, T.-W., Brewster, A. S., Yu, S. X., Ushizima, D., Yang, C. & Sauter, N. K. (2018). J. Synchrotron Rad. 25, 655–670. Web of Science CrossRef CAS IUCr Journals Google Scholar
Konstantinova, T., Wiegart, L., Rakitin, M., DeGennaro, A. M. & Barbour, A. M. (2021). Sci. Rep. 11, 14756. Web of Science CrossRef PubMed Google Scholar
Lamorgese, A. G. & Mauri, R. (2009). Phys. Fluids, 21, 044107. Web of Science CrossRef Google Scholar
Lu, P. J., Zaccarelli, E., Ciulla, F., Schofield, A. B., Sciortino, F. & Weitz, D. A. (2008). Nature, 453, 499–503. Web of Science CrossRef PubMed CAS Google Scholar
Lu, X., Mochrie, S. G. J., Narayanan, S., Sandy, A. R. & Sprung, M. (2008). Phys. Rev. Lett. 100, 045701. Web of Science CrossRef PubMed Google Scholar
Madsen, A., Fluerasu, A. & Ruta, B. (2016). Synchrotron Light Sources and Free-electron Lasers, edited by E. J. Jaeschke, S. Khan, J. R. Schneider & J. B. Hastings, pp. 1617–1641. Cham: Springer International Publishing. Google Scholar
Malinovska, L., Kroschwald, S. & Alberti, S. (2013). Biochim. Biophys. Acta, 1834, 918–931. Web of Science CrossRef CAS PubMed Google Scholar
Manley, S., Wyss, H. M., Miyazaki, K., Conrad, J. C., Trappe, V., Kaufman, L. J., Reichman, D. R. & Weitz, D. A. (2005). Phys. Rev. Lett. 95, 238302. Web of Science CrossRef PubMed Google Scholar
Möller, J., Chushkin, Y., Prevost, S. & Narayanan, T. (2016). J. Synchrotron Rad. 23, 929–936. Web of Science CrossRef IUCr Journals Google Scholar
Muschol, M. & Rosenberger, F. (1997). J. Chem. Phys. 107, 1953–1962. CrossRef CAS Web of Science Google Scholar
Oviedo, F., Ren, Z., Sun, S., Settens, C., Liu, Z., Hartono, N. T. P., Ramasamy, S., DeCost, B. L., Tian, S. I. P., Romano, G., Gilad Kusne, A. & Buonassisi, T. (2019). npj Comput. Mater. 5, 60. Web of Science CrossRef Google Scholar
Perakis, F. & Gutt, C. (2020). Phys. Chem. Chem. Phys. 22, 19443–19453. Web of Science CrossRef CAS PubMed Google Scholar
Pokuri, B. S. S., Ghosal, S., Kokate, A., Sarkar, S. & Ganapathysubramanian, B. (2019). npj Comput. Mater. 5, 95. Web of Science CrossRef Google Scholar
Price, K. (1996). Proceedings of North American Fuzzy Information Processing, pp. 524–527. IEEE. Google Scholar
Ragulskaya, A., Begam, N., Girelli, A., Rahmann, H., Reiser, M., Westermeier, F., Sprung, M., Zhang, F., Gutt, C. & Schreiber, F. (2021). J. Phys. Chem. Lett. 12, 7085–7090. Web of Science CrossRef CAS PubMed Google Scholar
Rumelhart, D. E., Hinton, G. E. & Williams, R. J. (1986). Learning Internal Representations by Error Propagation, pp. 318–362. Cambridge: MIT Press. Google Scholar
Ruta, B., Chushkin, Y., Monaco, G., Cipelletti, L., Pineda, E., Bruna, P., Giordano, V. M. & Gonzalez-Silveira, M. (2012). Phys. Rev. Lett. 109, 165701. Web of Science CrossRef PubMed Google Scholar
Sappelt, D. & Jäckle, J. (1997). Physica A, 240, 453–479. CrossRef CAS Web of Science Google Scholar
Schurtenberger, P., Chamberlin, R. A., Thurston, G. M., Thomson, J. A. & Benedek, G. B. (1989). Phys. Rev. Lett. 63, 2064–2067. CrossRef PubMed CAS Web of Science Google Scholar
Sciortino, F., Bansil, R., Stanley, H. E. & Alstrøm, P. (1993). Phys. Rev. E, 47, 4615–4618. CrossRef CAS Web of Science Google Scholar
Shin, Y. & Brangwynne, C. P. (2017). Science, 357, eaaf4382. Web of Science CrossRef PubMed Google Scholar
Starostin, V. (2022). XPCS Autoencoder, https://github.com/schreiber-lab/xpcs_autoencoder. Google Scholar
Tanaka, H. (2000). J. Phys. Condens. Matter, 12, R207–R264. Web of Science CrossRef CAS Google Scholar
Vecsei, P. M., Choo, K., Chang, J. & Neupert, T. (2019). Phys. Rev. B, 99, 245120. Web of Science CrossRef Google Scholar
Wang, B., Yager, K., Yu, D. & Hoai, M. (2017). 2017 IEEE Winter Conference on Applications of Computer Vision (WACV), pp. 697–704. IEEE. Google Scholar
Weber, S. C. & Brangwynne, C. P. (2012). Cell, 149, 1188–1191. Web of Science CrossRef CAS PubMed Google Scholar
Wight, C. & Zhao, J. (2021). Commun. Comput. Phys. 29, 930–954. Google Scholar
Zaccarelli, E. (2007). J. Phys. Condens. Matter, 19, 323101. Web of Science CrossRef Google Scholar
Zhang, F., Allen, A. J., Levine, L. E., Ilavsky, J., Long, G. G. & Sandy, A. R. (2011). J. Appl. Cryst. 44, 200–212. Web of Science CrossRef CAS IUCr Journals Google Scholar
Zhang, Q., Bahadur, D., Dufresne, E. M., Grybos, P., Kmon, P., Leheny, R. L., Maj, P., Narayanan, S., Szczygiel, R., Ramakrishnan, S. & Sandy, A. (2017). Phys. Rev. Lett. 119, 178006. Web of Science CrossRef PubMed Google Scholar
Zhang, X. & Garikipati, K. (2020). Comput. Methods Appl. Mech. Eng. 372, 113362. Web of Science CrossRef Google Scholar
Zinn, T., Homs, A., Sharpnack, L., Tinti, G., Fröjdh, E., Douissard, P.-A., Kocsis, M., Möller, J., Chushkin, Y. & Narayanan, T. (2018). J. Synchrotron Rad. 25, 1753–1759. Web of Science CrossRef CAS IUCr Journals Google Scholar
This is an open-access article distributed under the terms of the Creative Commons Attribution (CC-BY) Licence, which permits unrestricted use, distribution, and reproduction in any medium, provided the original authors and source are cited.