Biomolecular condensate phase diagrams with a combinatorial microdroplet platform

The assembly of biomolecules into condensates is a fundamental process underlying the organisation of the intracellular space and the regulation of many cellular functions. Mapping and characterising phase behaviour of biomolecules is essential to understand the mechanisms of condensate assembly, and to develop therapeutic strategies targeting biomolecular condensate systems. A central concept for characterising phase-separating systems is the phase diagram. Phase diagrams are typically built from numerous individual measurements sampling different parts of the parameter space. However, even when performed in microwell plate format, this process is slow, low throughput and requires significant sample consumption. To address this challenge, we present here a combinatorial droplet microfluidic platform, termed PhaseScan, for rapid and high-resolution acquisition of multidimensional biomolecular phase diagrams. Using this platform, we characterise the phase behaviour of a wide range of systems under a variety of conditions and demonstrate that this approach allows the quantitative characterisation of the effect of small molecules on biomolecular phase transitions.


Comparison between PhaseScan and manual measurements Supplementary Figure 2. Comparison between PhaseScan and manual measurements.
The crowding-induced homotypic phase separation of EGFP-tagged FUS G156E in the presence of PEG 6000 was analysed by independent PhaseScan (dots and colour map) and manual pipetting experiments (crosses). All points were observed to match the phase diagram as determined by PhaseScan.

Supplementary Note 2
To verify that the phase diagram generated by the PhaseScan platform is in accordance with the phase diagram obtained from bulk experiments, we performed manual pipetting experiments with FUS G156E and compared the phase behaviour with the data obtained by PhaseScan (Supplementary Figure 2). We performed manual bulk measurements at twelve different points in the phase diagram, stock solutions of buffer, 6k PEG (20%) and FUS (18.23 μM as verified by Nanodrop measurements at 488 nm) were prepared and mixed to give 10 μL of each of the sample compositions as shown in Supplementary Figure 2. The sample was pipetted onto a microscope slide equipped with imaging chambers, which was sealed with a coverslip, before imaging EGFP fluorescence with a 10× objective. The images were manually classified as containing phase-separated or homogeneous protein and compared to the corresponding region in the phase diagram produced by the PhaseScan platform. All manually determined datapoints match well with the boundary as determined by PhaseScan platform.

Supplementary Note 4
To investigate the reproducibility and time-dependency of phase diagram generation, three phase scan experiments were conducted with the same reagent inputs, before imaging data was acquired 5, 30 and 60 min after droplet generation (Supplementary Figure 4). Only negligible differences are observed in the integrated phase separation probability between the repeats (average difference 1.51 ± 0.33 % (mean ± standard deviation, N = 3)), which were randomly sampled to contain the same number of datapoints in each dataset (N = 2754). This demonstrated reproducibility of droplet production, trapping, data acquisition and analysis.
Notably, although a small number of incorrectly classified or barcoded outliers exist in all three datasets, these datapoints have negligible effect on the reported position of the phase boundary since they exist as only a small (< 2%) proportion of the overall dataset.
The negligible differences between the phase diagrams shows that there is no timedependence in the observed phase behaviour, confirming that the phase diagrams are collected under equilibrium conditions. In addition, by randomly sub-sampling each dataset further (i.e. bootstrapping) to include N = 1500 datapoints, less than the smallest dataset presented in this manuscript ( Figure   2(f), N = 1599), we found that the probability difference between sampling repeats remained < 2.75 ± 0.74 % (mean ± standard deviation, N = 5). To investigate whether droplet size influences the observed phase behaviour, an experiment was performed in which a phase diagram for EGFP-FUS G156E and PEG was recorded using two different populations of droplet size (Supplementary Figure 5(a)). Droplets with mean radius of 43 μm (small droplets) and 65 μm (large droplets) were produced by operating the droplet generator with oil flow rates of 300 μL/h and 60 μL/h, respectively (Supplementary Figure   4(b)). Notably, this range of droplet size (3.5-fold difference in volume and 2.3-fold difference in surface area) is much larger than that produced during a conventional PhaseScan experiment, where droplet radii typically vary by <10%. Droplet imaging and analysis was executed as described in the Methods, with the data segregated between the two droplet sizes to afford phase diagrams for each of these droplet populations (Supplementary Figure 5(c, d)). By overlaying the computed phase boundaries for the small and large droplets, it is apparent that no significant difference in phase behaviour between the two populations is observable (Supplementary Figure 5(e)). Phase separation in the binodal region of phase-space occurs through a nucleation/growth mechanism, and phase separation could therefore be assumed to display volume and/or surface dependence as observed previously. 1 However, in the experimental procedure outlined here, droplets are assayed several minutes after generation and mixing. Therefore, we propose that each droplet microenvironment has sufficient time to reach chemical equilibrium, with the PhaseScan measurement invariant to droplet size over the range investigated here. Figure 6. PhaseScan characterisation of phase separation of (PR)25 and polyU RNA. To demonstrate the applicability of the PhaseScan assay to simple peptide systems as well as full-length proteins, we examined the phase behaviour of a dipeptide repeat system derived from the hexanucleotide repeat expansion in the chromosome 9 open reading frame 72 (C9orf72) gene, implicated in ALS. 2 The peptide consisted of 25 repeats of proline-arginine dipeptide (PR)25. This type of peptide is well known to phase separate when mixed with negatively charged polymers, including single-stranded RNA. We assayed the formation of (PR)25 with poly uridine (PolyU100) RNA and observed RNA-dependent phase separation above 1.5 ng/μL RNA for ~20 μM (PR)25. Source data are provided as a Source Data file.

Supplementary Note 6
A calibration procedure was employed to convert the intensity per unit volume to the concentration of the corresponding barcode fluorophore (Supplementary Figure 9) in order to allow each droplet to be accurately located in chemical space in the resultant phase diagram.
The same fluorophore-containing solutions used in each experiment were injected into microchannels of known dimensions (cross sectional area typically 150 × 30 μm) and were imaged under the same conditions as the droplets (Supplementary Figure 9(a, b)). From these images, the relationship between intensity and fluorophore concentration was determined on an experiment-by-experiment basis (to account for variation in pipetting, imaging, light source intensity etc.), as well as demonstrating the linear relationship between fluorophore concentration and fluorescence intensity (Supplementary Figure 9(c, d)).
To demonstrate the efficacy of this approach, a control experiment was conducted where droplets containing known concentrations of Alexa647 fluorophore (as determined by UV-vis spectroscopy/NanoDrop) were produced, trapped, imaged and subjected to the analysis and calibration procedure (Supplementary Figure 9(e, f, g)). The nominal fluorophore concentration present in each droplet population was observed to be in very good agreement with that determined by the analysis procedure (Supplementary Figure 9(h)). Example droplet classification procedure for phase separated and homogeneous droplets, respectively. From left to right: (i) detected droplet, (ii) padding, (iii) convolution, (iv) cropping of the padded part, and (v) thresholding and identification of pixel clusters. For any droplet containing a cluster with >1 pixels, the droplet is considered as phase separated, otherwise mixed.

Supplementary Note 7
The collected images were analysed using a custom-written Python script (Supplementary Figure 10). The raw images were processed to remove camera dark noise (Background) and flattened to correct non-uniform epifluorescence illumination by division with a calibration image taken of a homogeneous fluorescent solution (Illumination) according to Processed Image =

Image -Background
Illumination -Background . The raw image was enhanced by taking the logarithm of the image, applying rolling ball background subtraction, taking the logarithm again, and thresholding the darkest 10% and brightest 10% pixels. Droplets were fitted as circles by finding the bright peaks in the image (Supplementary Figure 10(a)). Any incomplete circles that are partly out of the image boundary as well as those smaller or larger than the radius thresholds were removed. Non-circular droplets or erroneous detections were removed by comparing a corresponding perfect sphere of same size and the image of the droplet, where the brightness multiplied by a proportional constant is used as the z-axis. The total intensity was calculated and normalised to afford intensity per unit volume (calculated using fitted diameter).
To distinguish droplets with condensates formed via phase separation and those which are wellmixed (Supplementary Figure 10(b, c)), the droplet images were convoluted with an edge detection kernel after padding. The result image was made binary with a given threshold ratio of quartiles and medians. If there were at least two connected bright pixels, the droplet was classified as phase separated. Otherwise, the bright pixels were determined as noise and the droplet was labelled as well-mixed. The phase boundary was estimated using a support vector machine (SVM) algorithm with a radial basis function (RBF) kernel. The parameters of SVM were selected according to grid search scores. The parameter space was sampled as a 2D/3D mesh grid and predicted by the SVM model, which was then used to generate iso-boundary or iso-surface as the phase diagram boundary.