Main

Compared with mechanical sectioning, gallium- and plasma-based focused ion beam (FIB) milling yields virtually artifact-free views of frozen-hydrated biological material. Provided that the initial sample is vitreous, lamellas that contain cellular components of interest can routinely be produced from plunge-frozen specimens for cryo-electron tomography (cryo-ET) and high-resolution subtomogram averaging (STA). While there is some concern regarding sample damage by FIB milling1,2, in situ STA can reveal near-atomic-level details of biological processes3,4. To expand the concept of lamella milling to larger samples than single cells, high-pressure freezing (HPF) is required to ensure complete vitrification5. HPF planchettes, however, are not compatible with standard on-grid milling procedures.

Cryo-lift-out addresses this need with micromanipulators that operate at submicrometer precision to handle small portions of the high-pressure-frozen material6. Adapted from the material sciences, they grant access to biological material that may be too large for conventional lamella milling or the hybrid waffle method7. Due to the more widespread availability and improved hardware, methods development on these systems has gained a lot of momentum over the past years8,9. However, several aspects, including ease of operation, throughput and subcellular targeting, still need to be improved so cryo-lift-out can become a routine method for structural biology.

Current best practices of needle-based lift-out employ a sacrificial adaptor made from copper (‘adaptor chunk’) to attach the sample to the micromanipulator10,11,12. Although this improves success rates substantially, connecting the initial coarse lamella to the support grid requires specialized grids (Fig. 1a and Supplementary Fig. 1a,b). While frequently used at room temperature, these pin grids are problematic under cryo-conditions, mainly due to three factors. First, lamellas are attached only on one side and, therefore, break easily (Supplementary Fig. 1c). Second, due to the lack of support, the beam-induced motion in tomograms acquired far from the attachment site is higher than in those closer to the pin (Fig. 1g). While motion correction can compensate for this sample movement to an extent, increasing mechanical stability and minimizing drift already during data acquisition is preferable. Custom-made slot grids, which support lamellas on both sides, have been explored to address this problem8,13. However, their preparation is time-consuming, and their use is cumbersome for needle-based lift-out systems. Lastly, ice crystals frequently contaminate pin grid-mounted lamellas. This may be due to their exposed location and the higher electrostatic field strength on the pins (Supplementary Fig. 1d,e). Based on these observations, ideal support grids for cryo-lift-out lamellas should be available off-the-shelf, compatible with different chunk geometries, and guarantee sample integrity for extended workflows.

Fig. 1: Concept and key features of SOLIST.
figure 1

a, Schematic representation of the classical lift-out procedure using pin grids. b, By dividing the sample chunk into multiple sections, SOLIST effectively multiplies the number of lamellas obtained. c,d, FIB view images of the first (c) and the last (d) slice of a four-lamella SOLIST session being attached to the grid (see also Supplementary Video 1). e, Success rates of each step of the SOLIST procedure: section attachment (Rough), fine milling (Thin) and transfer to the transmission electron microscope (TEM; n = 46 lamellas). f, Box plot of lamella thickness of the classical lift-out (PIN) and the yeast test data set (SOLIST). Mean ± s.d.: PIN, 184 ± 9.3 nm; SOLIST, 187 ± 31 nm (PIN, n = 6; SOLIST, n = 32). g, Comparison of beam-induced motion in SOLIST and classical lift-out lamellas attached to pin grids (n = 9 tomograms per group; each 41 tilts; P of two-sided t-test indicated in plot). h, The resolution potential of SOLIST exemplified by a S. cerevisiae ribosome average at 7.2 Å obtained from 65 tomograms (gold-standard Fourier shell correlation (FSC) at 0.143). For the box plots: the whiskers indicate maximum and minimum values not considered outliers, the top of the box represents the 75th percentile and the bottom of the box the 25th percentile; the red line indicates the median, and the red plus markers (if plotted) indicate individual outliers.

Source data

Full size image

Recently, grids made entirely from gold have been shown to improve high-resolution single-particle datasets14. The combination of mechanical stability, matched expansion coefficient of grid and film, and electrical conductivity, resulting in reduced beam-induced motion, makes them ideal sample supports. Many of the same qualities would be desirable for cryo-lift-out applications. But could lamellas even be fixed on the thin, holey gold film? Moreover, would they be stable?

Results

The SOLIST approach

To explore this idea, we used a needle micromanipulator to place individual lift-out chunks of ca. 25 µm × 10 µm × 3 µm on squares of all-gold grids. To prevent them from moving, we connected the pieces to the film using microsutures, which act similarly to the microwelding spots used for attaching chunks to pin grids (Supplementary Fig. 2a–e). After clearing small areas of the film in front of the biological material to prevent uncontrolled ripping (Supplementary Fig. 2b), we polished the leading edges of the coarse lamellas and coated them with an organometallic platinum layer using the gas injection system (GIS) to further improve attachment and reduce curtaining (Supplementary Fig. 2c). In the final step, we milled the samples analogously to cells on grids, yielding lamellas transparent to transmission electron microscopy (TEM) (Supplementary Fig. 2f–i) and confirming that commercially available gold grids are stable enough to support lift-out lamellas (Fig. 1e).

Since preparing the initial lift-out site is very time-consuming, producing only a single lamella from this material is wasteful (Table 1). In fact, following the classical procedure6, much of the chunk is discarded and milled away while it could be used to prepare additional lamellas. We instead envisioned a method to enhance throughput by making use of this excess material. To this end, we divided the initial sample chunk into several 2–4 µm sections and placed them individually on grid squares (Fig. 1b). Each time, a 1–2 µm cut was made to release the coarse lamella. The procedure was then repeated until the whole sample block had been used up (Fig. 1c,d). Following this sequential attachment, a 25-µm-tall piece can thus be split into four to six individual lamellas, each consuming about 4–5 µm of the lift-out chunk (Supplementary Video 1). This serialized on-grid lift-in sectioning for tomography (SOLIST) considerably enhances sample use, as several lamellas are obtained from one initial lift-out. SOLIST is therefore more time efficient since it effectively multiplies the number of lamellas obtained in each microscope session, while increasing sample stability simultaneously (Table 2 and Supplementary Video 2).

Table 1 Typical milling time per step for the classical lift-out (nonserialized)
Full size table
Table 2 Typical milling time per step for SOLIST
Full size table

In parallel to our work, a similar concept to SOLIST was published, which also relies on dividing and attaching chunks sequentially on bare grids without support film. However, it requires specialized slot grids that are oriented with respect to the FIB and that the size and geometry of the lift-out chunk matches the grid dimensions15.

Quality controls

Lamellas prepared with SOLIST are highly stable during storage and can be subjected to several loading–unloading cycles (Fig. 1e). Their thickness is comparable to regular on-grid lamellas (Fig. 1f), and compared with pin grids, contamination with ice crystals is reduced and similar to normal on-grid lamellas (Supplementary Fig. 3). Furthermore, beam-induced motion is significantly lower for SOLIST than for the one-sided pin attachment of classical lift-out, for which the drift varies depending on the distance from the attachment point (Fig. 1g). Our approach, therefore, improves data quality even before frame motion correction.

In general, SOLIST can be applied to any sample amenable to HPF and does not distinguish between planchettes and waffles as sources for the chunk preparation (Supplementary Fig. 4). While some of the lift-out material is required for attachment and support (~4–5 µm on either side), sufficiently large lamellas (15–25 µm × 15 µm) can regularly be produced (Supplementary Fig. 3a). High-throughput approaches for cryo-ET data acquisition may be used to optimize microscope time16,17. Owing to their stability and consistent geometry, automating the lamella milling process18,19 is straightforward. Lift-out sites can be prepared without targeting for high-content samples (for example, cell suspensions). Provided that the structure of interest is abundant enough, such as ribosomes in yeast, we found that subnanometer subtomogram averages can be obtained already from a small number of tomograms (Fig. 1h and Extended Data Fig. 1). However, to identify scarcer targets and work in more complex regions of cells, additional steps would be required.

3D-targeted FIB milling of SOLIST slices

Considering the stability of the coarse SOLIST lamellas, we reasoned that targeting could be achieved by 3D correlation (3DC)20, which so far had not been demonstrated in HPF samples. In fact, our method reduces several problems commonly associated with classical 3DC FIB milling: (1) the attachment on clean grids is less prone to ice contamination, which can interfere with detecting the fiducial beads; (2) the well-defined geometry of the lamellas, and especially their initial thinning to 3–4 µm, reduces fluorescence imaging artifacts such as out-of-plane fluorescence and image distortions due to the refractive index mismatch between the sections and their surrounding; (3) since HPF samples can be prepared with a higher density than cells grown individually on grids, the chance of finding an area of interest in the coarse lamella should be higher, thereby improving throughput and success rate.

Hence, we devised a 3DC procedure to use these potential benefits and tested it on in vitro reconstituted biomolecular condensates (BMC) of chromatin with the fusion oncoprotein BRD4–NUT21. Nucleosome condensates have been investigated by cryo-ET before using plunge freezing22. However, the observed chromatin organization in the periphery of the droplets, thin enough for direct cryo-ET, may have been altered since the samples were placed directly on grids and blotted. Keeping the sample in suspension and instead using HPF for cryopreservation should leave their organization unchanged. It could, therefore, provide new insights into the remodeling action of chromatin-binding proteins in near-native state, and for the entire BMC droplet.

While integrated fluorescence light microscopes are emerging as a promising new option for correlative FIB milling23, their reduced axial resolution and limited availability prompted us to develop a three-step 3DC technique using a dedicated external cryo-fluorescence microscope. In its first step, we prepared coarse SOLIST lamellas of the BMC droplets as described above. To allow later registration of fluorescence and FIB or scanning electron microscope (SEM) images, the chunks were placed on gold support grids coated with fiducial beads to serve as landmarks for the mapping procedure (Fig. 2a). Due to the geometry of the FIB shuttle and the angle between the electron and the ion beam, sections need to be placed toward the back of the grid to be accessible by all imaging modes (SEM, FIB, fluorescence light microscopy and TEM) during SOLIST (Fig. 2b). While this reduces the number of potential attachment sites, about 48 positions remain available on a regular 200-mesh cryo-electron microscopy (cryo-EM) grid (Fig. 2c). After the initial attachment with microsutures and GIS coating, we transferred the support grids to a cryo-confocal light microscope to acquire z-stacks on the sections. Finally, we moved the samples back to the FIB/SEM microscope, where we performed 3D-correlative milling analogously to cells on grids (Fig. 2d)20,24. Compared with cellular lamellas, the correlation is more straightforward as beads remain visible and are not covered by excess buffer. Furthermore, the process is supported by the defined geometry of the coarse SOLIST sections, which enables automated FIB milling and ensures that the fluorescent region of interest is included in the final lamella (Fig. 2e,f).

Fig. 2: SOLIST enables 3D-correlative targeting in HPF samples.
figure 2

a, Coarse SOLIST lamellas are screened for the target of interest at the cryo-fluorescence microscope, and 3DC is performed using the fiducial beads. b,c, The shallow milling angle (b) makes only an array of ca. 4 × 12 squares on the grid (c) accessible for attachment and correlation. d, 3D CLEM on a coarse SOLIST lamella using the registration of the fiducial beads to overlay fluorescence light microscopy (FLM) and FIB views. The coarse sample chunk (S and dashed outline) contains the GFP signal as seen from the projection (the crosshair shows the center of the original image). e, TEM montage of the targeted SOLIST lamella produced in d, overlaid with the fluorescence data. f, Higher-magnification TEM montage partially overlaid with the fluorescence data showing that the target is contained within the lamella. Parts of the image were created with BioRender.com.

Full size image

Imaging human brain organoids using high resolution cryo-ET

Having established that SOLIST can produce high-resolution subtomogram averages and that specific sample areas can be targeted by correlative milling, we next sought to explore the scope of possible targets. While there are certain benefits to applying HPF even to single cells, in particular, full vitrification, plunge freezing or, alternatively, the hybrid waffle method should be used whenever possible to improve throughput25. However, for samples too large for either approach, SOLIST could be used to prepare cryo-lamellas and make them accessible to high-resolution cryo-ET more efficiently. In particular, this applies to organoids and tissues, for which—despite a recent report26—adapting the waffle method is difficult.

We, therefore, sought to first apply SOLIST to human forebrain organoids. These miniature organs are emerging as important medical model systems for neurodevelopmental and neuropsychiatric diseases27,28,29,30. However, brain organoids are large (~3 mm in diameter) and must be sectioned on a vibratome to fit into HPF planchettes. To prevent excessive damage to the slice surface during HPF, they cannot be cut to the exact dimensions of the cavity. Hence, we cut them to ~75 µm, leading them to be covered by a substantial amount of freezing buffer, which needed to be removed to access the biological material (Supplementary Fig. 5). To prepare them for SOLIST, the frozen organoid slices were first localized in x and y using reflected light, and fluorescence overviews were acquired directly on the HPF planchette (Supplementary Fig. 5b,c). However, no stacks are acquired due to the poor resolution in z, which is aggravated by reflection artifacts of the gold-coated planchettes. Lift-out sites are therefore prepared on the basis of x/y coordinates only. The serialized sectioning of SOLIST compensates for the lack of targeting in z by turning depth into a horizontal series of lamellas, which one can quickly screen by TEM for the desired structures with the deeper sections appearing earlier (Fig. 3a,b). To further reduce the time spent on site preparation, planchettes may be trimmed (‘planed’) on a cryo-ultramicrotome before or after the correlation to remove excess buffer on top of the sample (Supplementary Fig. 5a). Without these improvements, previous work on brain organoids had been limited to their thin periphery26. SOLIST instead now provides the first high-resolution snapshots of their interiors, showing among other structures ribosomes, mitochondria and microtubules (Fig. 3c–f and Supplementary Video 3).

Fig. 3: SOLIST enables investigations in developing human forebrain organoids.
figure 3

a, A chunk of high-pressure-frozen organoid slice is translated into a series of SOLIST slices. The sequential sectioning compensates for the lack of accuracy in z-targeting. b, Within the series of lamellas, organoid material is found in the deeper layers and, therefore, in the earlier attached sections. c–f, Representative tomograms (c,e) and corresponding segmentations (d,f) from adjacent SOLIST sections show cellular structures in brain organoid: membranes (gray), vesicles (yellow and blue), mitochondria (black), ribosome (cyan) and microtubules (turquoise); see also Supplementary Video 3. Parts of the image were created with BioRender.com.

Full size image

Sampling native tissue at molecular resolution

We next explored if SOLIST could also make native tissues amenable to high-resolution cryo-electron microscopy. To this end, we selected three targets in mice for this task: brain, liver and heart, each with their unique cellular compositions. Using variations of traditional cryo-FIB milling and lift-out, human brain and mouse liver have recently been studied by cryo-ET31,32. However, sample preservation and resolution were limited. While alternative approaches such as treating tissue with glycerol for plunge freezing may be viable in some contexts33, they do not offer a general solution and so far lack the throughput for high-resolution structural studies directly in native specimen. Both bottlenecks could be addressed by coupling HPF and SOLIST. Therefore, we subjected the mouse liver tissue and parts of the corpus callosum, a bundle of nerve fibers connecting the left and right hemispheres in the brain, to our cryo-lift-out protocol.

To reduce processing time on the FIB/SEM microscope, we removed excess material around the brain tissue by cryo-trimming, gaining direct access to the nerve material during cryo-lift-out (Supplementary Fig. 5d,e). After successful SOLIST, cryo-ET revealed a crowded network of vesicles and microtubules, some of which could be followed over several micrometers. However, none of these structures was abundant enough to furnish a high-resolution average (Fig. 4a, Extended Data Fig. 2 and Supplementary Video 4).

Fig. 4: SOLIST visualizes native mouse tissue organization.
figure 4

a, Tomograms of mouse brain reveal membranes (gray), axonal microtubules (turquoise) and vesicles (yellow and blue). b, The sampled liver tissue contained highly complex membranes associated with abundant ribosomes (cyan). c, From only 15,000 particles, a subnanometer ribosome subtomogram average was reconstructed, revealing several well-resolved helices at 8.3 Å (gold-standard Fourier shell correlation (FSC) at 0.143).

Full size image

Next, we turned to the liver tissue on which we performed SOLIST without prior trimming. Since it is involved in the body’s energy metabolism, the liver is rich in glycogen in granules, which our cryo-tomograms revealed in abundance and engaged with various complex membrane structures (Extended Data Fig. 3a,b). Moreover, most areas contained a high number of ribosomes, many of which were associated with membranes (Fig. 4b, Extended Data Fig. 3b,d and Supplementary Video 5). Presented with this opportunity, we sought to examine how well these structures could be resolved in the SOLIST lamellas compared with the 18 Å structure published previously32. We, therefore, collected 66 tomograms on six lamellas, yielding ~15,000 ribosome particles, which refined to 8.3 Å (Fig. 4c and Extended Data Fig. 4). This shows that, with SOLIST, adequate resolutions can be reached not only in test samples such as yeast (Fig. 1h and Extended Data Fig. 1) but also in native tissue and within ca. 24 h of microscope time using conventional dose-symmetric tilt schemes (based on three tomograms per hour). Further improvements are possible using bigger datasets from parallelized data acquisition, for example, using PACEtomo or on-axis multishot tomography16,34.

To explore additional structures besides ribosomes, we turned to our last tissue: mouse heart muscle from the left ventricle. The sample was processed as described before, and lamellas were prepared with SOLIST without prior planchette cryo-trimming. While muscle architecture is highly complex, cryo-ET has recently resolved some parts at resolutions ranging from 15 to 20 Å (refs. 35,36). This includes the thin filaments made up of F-actin and tropomyosin and the thick filaments consisting of a complex array of components, hosting the myosin motors responsible for muscle activity. Screening our lamellas, TEM overviews quickly revealed fibers traversing large areas (Extended Data Fig. 5 and Supplementary Video 6). Some parts contained cross-section views of muscle cells, with the thick and thin filaments clearly visible (Fig. 5a).

Fig. 5: SOLIST resolves native mouse cardiac filaments at molecular resolution.
figure 5

a, A cross-section tomogram of a muscle cell from the left heart ventricle. Thick (large circles) and thin filaments (small circles) are clearly visible. b, Subtomogram averages of the thin filaments at 18 Å from only three cross-section tomograms reveal the actin (red) and tropomyosin components (orange). c, Their highly ordered arrangement becomes apparent after averaging and pasting back the thin fibers. d, In addition to the bare thin filament, a myosin-bound actin (salmon) class is found. It represents an average of all potential motor binding sites rather than an actual state (composite maps shown).

Full size image

Thus, we selected three tomograms recorded on those sites and subjected them to a pipeline focused on the thin filaments (Extended Data Fig. 6). After 3D classification, two main fiber types became visible. The first appears as an actin-like class and features tropomyosin, resolved at 19 Å, which appears as a second filament wrapped around the central actin density (Fig. 5b). Overall, our structure is highly similar to subtomogram averages obtained from FIB-milled cardiomyocytes grown on grids. However, it is from considerably less particles36. Pasting back these fibers into the tomographic volumes reveals their dense and highly ordered packing within this region of the muscle (Fig. 5c). The second class, on the other hand, resembles an in vitro reconstituted fiber with myosin heads bound to the central actin filament37 (Fig. 5d). Even without the use of helical symmetry, each site appears occupied. It, therefore, represents an average of all potential motor binding sites rather than an actual state (Extended Data Fig. 6). Further classification and per-fiber averages would be required to resolve individual myosin-bound positions, a task that appears feasible considering that this level of detail was already obtained from just three tomograms on SOLIST lamellas.

Discussion

In summary, SOLIST enables the efficient preparation of several cryo-FIB lamellas from high-pressure-frozen samples with optional fluorescence-guided targeting. Using our new approach, we provide a first glimpse into a developing human brain organoid and reveal the cellular organization in three different mouse tissues, including native heart muscle fibers at high resolution. While SOLIST for tomography is not yet as efficient as traditional serial sectioning of plastic-embedded samples at room temperature, it nonetheless represents a significant improvement over previous cryo-lift-out implementations that suffered from low throughput, ice contamination and poor sample stability. With SOLIST lifting these restrictions, we anticipate new high-resolution cryo-ET datasets soon to be generated for even more complex samples such as native patient specimens. Access to this information will no doubt help advance our structural understanding of healthy and pathological molecular processes in humans, and it will be the next step toward a ‘biopsy at the nanoscale’.

Methods

Sample preparation

Yeast

Yeast extract peptone dextrose (YPD) media plates and yeast Saccharomyces cerevisiae samples were provided by the Biomass Production Unit of Human Technopole. In brief, single colonies were picked from the YPD plate and grown overnight into 10 ml YPD media at 30 °C while shaking at 200 rpm in Multitron incubator (Infors HT).

Organoids

Forebrain organoids were generated from H9 embryonic stem cells as previously described38. Embryoid bodies were generated from embryonic stem cell colonies incubated in the neural induction medium consisting of 2 μM dorsomorphine (Stem Cell Technologies) and 2 μM A83-01 (Stem Cell Technologies). From day 7 to day 14, embryoid bodies were embedded in Matrigel (Corning) and patterned toward a forebrain fate in the medium consisting of Dulbecco’s modified Eagle medium (DMEM):F12 (Gibco), 1× N2 Supplement (Life Technologies), 1× penicillin–streptomycin (Euroclone), 1× non-essential amino acids (Gibco), 1× GlutaMax (Gibco), 1 μM CHIR99021 (Tocris) and 1 μM SB-43154 (Stem Cell Technologies). On day 14, embedded organoids were mechanically removed from Matrigel and incubated in the differentiation medium composed of DMEM:F12, 1× N2 and B27 Supplements (Life Technologies), 1× penicillin–streptomycin, 1× 2-mercaptoenthanol (Gibco), 1× non-essential amino acids and 2.5 μg ml−1 human insulin (Sigma-Aldrich) until day 35. From day 35 to day 70, the differentiation medium was supplemented with 1% Matrigel. From day 70 onward, organoids were incubated in the maturation medium composed of Neurobasal medium (Life Technologies), 1× B27 Supplement, 1× penicillin–streptomycin, 1× 2-mercaptoenthanol, 0.2 mM ascorbic acid (Sigma-Aldrich), 20 ng ml−1 brain-derived neurotrophic factor (Peprotech), 20 ng ml−1 glial cell line-derived neurotrophic factor (Peprotech) and 1 μM dibutyryl-cAMP (Stem Cell Technologies). Brain organoids were extracted from Matrigel as described by Qian et al.38. They were then embedded in low-melting-point agarose (Invitrogen) and sectioned into 75-μm-thick slices using a VT 1200S vibratome (Leica Microsystems).

In vitro chromatin sample

Condensation of chromatin droplets was performed in vitro as described previously21. In short, reconstituted chromatin was mixed with eGFP-labeled BRD4 containing only two bromodomains (BD1–BD2) in a phase separation buffer (25 mM Tris, pH 7.5, 150 mM KOAc, 1 mM Mg(OAc)2, 5% glycerol, 5 mM dithiothreitol, 0.1 mM EDTA and 0.1 mg ml−1 bovine serum albumin). Just before freezing, 10% Ficoll 400 (Sigma-Aldrich) was added to the solution as a cryo-protectant. To confirm droplet formation, the sample was observed on a Zeiss LSM980 Airyscan2 confocal microscope with a 63×/1.4 numerical aperture (NA) oil objective and using an excitation wavelength of 488 nm.

Mouse tissues

Tissue was obtained from Charles River Italia. After brief CO2 exposure, two 4-month-old male and one female C57BL/6N mice (housed at Charles River Italia) were decapitated, and the heads were placed immediately on ice. The entire heart was removed, placed in phosphate-buffered saline (PBS), and stored on ice. The liver was removed and placed in 4 °C PBS on ice.

All animals were housed, handled and procedures performed in accordance with European legislation (2010/63/UE) at Charles River Italia. The protocol was approved by the Italian authorities under authorization number 5AD83.N.I74.

Liver mouse tissue

Upon reception of the liver at Human Technopole, one lobe was transferred to cold PBS and then manually dissected with a precold razor blade and microscissors using a Leica S9E microdissection microscope equipped with light epi-illumination.

Brain and heart mouse tissues

Tissues from the mouse heart was processed by the Tissue Processing Unit (Human Technopole). In brief, the tissue was mounted on the VT 1200S vibratome with the support of a block of low-melting-point agarose (Invitrogen) in ice-cold Tyrode’s solution and sectioned into 80-μm-thick slices. The slices were then manually microdissected into smaller pieces to fit into a 100-μm-deep well of 3 mm type A planchettes. In the case of the brain, tissues from the corpus callosum region were embedded in low-melting-point agarose (Invitrogen) and sectioned into 80-μm-thick slices using a VT 1200S vibratome.

HPF

Planchette preparation

Before use, 3 mm type A and type B planchettes were cleaned by sonication for 5 min, 30 s on, 30 s off, 40% amplitude on a Branson SFX 550 (Thermo Fisher Scientific). Type B lids were polished with 1 μm lapping paper (Leica Microsystems) and coated with 0.1% soy lecithin (Sigma-Aldrich) dissolved in chloroform (Thermo Fisher Scientific).

Yeast

Overnight cultures were mixed with 25% Ficoll 400 (Sigma-Alrich) dissolved in YPD medium to reach a final concentration of 10% Ficoll, followed by centrifugation at 3,723g for 3 min at room temperature. Pellets were transferred to a 100-μm-deep well of a type A planchette that was subsequently covered with the coated (flat) side of type B planchette. The sandwich was then subjected to HPF on Leica EM ICE (Leica Microsystems).

Organoids

Just before freezing, slices were stained with Hoechst (Invitrogen) for later targeting. They were then loaded in 100-μm-deep wells of type A planchettes and frozen on a Leica EM ICE with 20% dextran (Sigma-Aldrich) and 5% sucrose (Sigma-Aldrich) supplied in the freezing medium.

In vitro droplets

The assembled condensation reaction supplemented with Ficoll 400 was loaded onto a 100-μm-deep well of type A planchette and immediately frozen using Leica EM ICE.

Mouse tissues

Just before freezing, brain slices were incubated with SiR-tubulin (Spirochrome) in Tyrode’s solution at a final concentration of 1,000 nM for sample targeting. All the slices from mouse tissues were then incubated in a freezing solution of 20% dextran (Sigma-Aldrich) and 5% sucrose (Sigma-Aldrich) for 5 min before loading into a clean 100-μm-deep well of 3 mm type A planchette to HPF on a Leica EM ICE.

Planchette cryo-trimming

Copper HPF planchettes were cryo-trimmed using a EM UC7/FC7 ultramicrotome (Leica Microsystems) operated at −170 °C. The tapered trimming method was used to remove approximately 30 µm of the surface layer from the 3 mm carrier. This procedure adequately eliminated tissue damage incurred during vibratome sectioning, resulting in a uniformly smooth surface appropriate for subsequent lift-outs. Additionally, a 1.5 mm × 1.5 mm square centering around the frozen tissue was built by progressively removing material starting at the periphery of the carrier using a Trim45 diamond trimming knife (DiATOME). The trimming process was conducted in three distinct stages, implementing a 25 µm taper: initial removal at 100 µm, followed by 75 µm and concluding with 50 µm. The carrier–knife interaction width was limited to 250 μm, together with the feed of 100 nm, and the cutting speed of 100 mm s−1 was applied to produce final surfaces free of cracks and reduced chattering. The final block face was finely trimmed at 50 nm feed and 30 mm s−1 cutting speed. In some cases, the surplus buffer surrounding the specimen was additionally trimmed away to unveil an edge closely encompassing the sample area of interest, as determined through epi-illumination assessment (Supplementary Fig. 5a,d,e).

Planchette fluorescence microscope prescreening

High-pressure-frozen brain organoids were imaged at the cryo-fluorescence light microscope Thunder (Leica Microsystems) with a 50×/0.9 NA objective and equipped with a cryo-stage. A customized adaptor was used to fit the HPF planchette into the Thunder loading cartridge. Planchette overviews in reflected light, GFP (green) and Hoechst (blue) channels were acquired to define regions enriched with target structures (Supplementary Fig. 5b,c).

Cryo-lift-out and SOLIST procedure

Cryo-lift-out was performed at the identified areas on an Aquilos 2 cryo dual-beam FIB/SEM microscope equipped with the EasyLift system (Thermo Fisher Scientific) and using the 45° shuttle12. In general, the ‘milling position’ refers to the shuttle orientation for regular on-grid FIB milling and is defined as relative rotation 0°. The ‘perpendicular position’ can be obtained by a stage tilt of 7° followed by a relative stage rotation of 180°. Milling patterns such as cleaning cross-section patterns and regular cross-section patterns are abbreviated as CCS and RCS, respectively. For the 35° shuttle, angles need to be adjusted accordingly.

In this work, the milling was performed at a voltage of 30 kV while the SEM imaging was performed at a voltage of 3–5 kV and a current of 25–50 pA. To confirm the appropriate thickness of the final lamellas, SEM imaging was done at 3 kV voltage and 50 pA current.

Adaptor chunk preparation

An adaptor gold chunk of ca. 5–10 µm × 5 µm × 6 µm was prepared from a gold EM grid (UltrAuFoil, R2/2, 200 mesh, Quantifoil) and attached to the Tungsten needle of the EasyLift system12. Gold was chosen over copper because: (1) thus the chunk could directly be prepared from the loaded all-gold support grid; (2) we observed that the milling of gold was easier and faster than copper; and (3) sputter rates appeared higher for gold and, therefore, it was easier to microweld the biological material to the adaptor chunk. With the stage in the perpendicular position, 4 µm of material on the grid bar was removed from each side of the chunk side using RCS patterns at 5 nA current but maintaining a 2 µm piece on the right side as an attachment (Supplementary Fig. 6a). The surface of the chunk was briefly cleaned from the milling position at an angle of 12° stage tilt with 0.3 nA current using CCS patterns. An undercut was milled 5 µm from the surface with an RCS pattern of 1 µm in y using 0.3–0.5 nA current. The same current was used to clean 2 µm of material on each side of the chunk, leaving the top 1 µm on the right side to connect the gold chunk to the bulk (Supplementary Fig. 6b). The stage was rotated back to the perpendicular position, and an RCS pattern was used at 0.3 nA current to clean up material redeposition in the back (Supplementary Fig. 6c). At the milling position with the stage tilt of 12°, the needle was inserted above the EM grid and lowered to touch the surface of the chunk. The needle was attached by milling microstitches on the gold chunk (see ‘Lift-out’ section for more details). Eventually, the connecting bridge was milled away with an RCS pattern at 0.5 nA current (Supplementary Fig. 6d).

Sample site preparation

For lift-out from the milling direction

With the stage at the perpendicular position, the sample chunk was prepared by removing a 40-μm-long area at the front and a 4–5-μm-long area at the back of the target site using 1–5 nA currents with RCS patterns. The same current and patterns were then used to clean up 5-μm-wide trenches on the left and right side of the target site but ensuring that a 3 µm piece was preserved on the left side as the connecting bridge (Supplementary Fig. 7a). The surface of the sample chunk was cleaned from the milling position at an angle of 12° stage tilt using an RCS pattern with 0.3 nA current (Supplementary Fig. 7b). The same current and patterns were then used to clean the leading edge from the perpendicular position (Supplementary Fig. 7c). An undercut was made 20–25 µm below the chunk surface from the milling direction at the stage tilt of 35° using 1 nA current with an RCS pattern. At the same time, the right and left side of the sample chunk were cleaned, preserving 3 µm of sample on the top of the left side for connection (Supplementary Fig. 7d).

For lift-out from the perpendicular direction

When the stage was at the perpendicular position, a sample chunk of 25 μm (x) in width and 80 μm in length (y) was prepared. Eighty micrometers of the frozen material at the front of the lift-out site was first removed using 5 nA current with an RCS pattern. With the same current and pattern, trenches of 5–10 μm were then cut at the sides and at the back of the sample site, but ensuring that 3 µm piece was preserved at the bottom right for connection (Supplementary Fig. 8a). In case the surface was not flat, it was cleaned from the milling position at a stage tilt of 12° using an RCS pattern and 0.5–1 nA currents (Supplementary Fig. 8b). With the stage at the perpendicular position, the same current and pattern were used to clean the front edge to improve the attachment of the needle (Supplementary Fig. 8c). An undercut was milled ~10 µm below the chunk surface from the milling direction at the stage tilt of 18–25° with an RCS pattern at 1 nA current (Supplementary Fig. 8d).

Lift-out

Lift-out from the milling direction

At the milling position, the needle with the adaptor chunk was inserted at a 12° stage tilt and lowered until it touched the sample surface without exerting too much force. Four to six microstitches of 0.8 µm × 1 µm × 0.5 μm were cut on the gold chunk using CCS patterns, 0.3 nA current and milling direction toward the top. Milling was performed twice or three times, and the gold redeposition between the sample and the adaptor chunk was monitored during the process. When the material started to redeposit, the patterns were shifted upward slightly from the sample and milled again.

To release the sample chunk, the remaining bridge was milled with an RCS pattern at 0.5 nA current. As soon as the sample chunk was free from the bulk, it was lifted out by raising the needle ~50 µm above the planchette surface before retracting the needle.

For a detailed illustration see Supplementary Fig. 9a.

Lift-out from the perpendicular direction

At the perpendicular position, the needle with the adaptor chunk was inserted with the stage tilt of 15°–18° and lowered until it touched the sample front edge and was attached to the sample chunk as described above. The rear bridge connecting the sample to the bulk was milled with 0.5 nA current using an RCS pattern, and the chunk was lifted out. For a detailed illustration, see Supplementary Fig. 9b.

SOLIST procedure

To optimize lamella milling, suitable attachment sites were identified preferably four or five squares above the grid center. At the milling position and stage tilt of 12° or 13°, the needle with the attached sample chunk was inserted on top of the support grid (UltrAuFoil, R2/2, 200 mesh, Quantifoil). The needle was lowered until the sample chunk touched the grid foil, and slightly backed up again to avoid stress on the gold film (as visualized from the bending of the film on FIB view).

For the correlative workflow, sections of 3–4 µm were cut but minimal sections of 2 µm can also be safely attached to the grids. An RCS pattern of at least 2 µm in y direction was placed on the sample chunk, and the section was released by cutting a 1.5–2-µm-high section with a RCS pattern using 0.3–0.5 nA current. This procedure was repeated until the lift-out chunk was exhausted.

With all the sample sections attached, the stage was rotated to the perpendicular position and several microstitches of 1.0 µm × 0.5 µm × 0.25 µm were placed on the left and right side of the sample and done with CCS patterns at 30 pA current, and the milling direction toward the center of the lamella. Parts of the film in front of the section were also removed to prevent ripping. The leading edge of each section was cleaned from the same stage position with an RCS pattern at 0.3 nA current. The stage was then rotated to the milling position, and the surface of each section could be cleaned at the stage tilt of 12° or 13° using a CCS pattern at 0.3 nA current.

The grid was coated with the Pt deposition GIS for 20 s from the milling direction and 20 s from the perpendicular, and again 20 s from the milling direction before unloading (for storage or correlative workflow) or fine milling.

Fine milling

With RCS patterns, lamellas were symmetrically thinned down to 800 nm at 0.3 nA current and to 300 nm at 0.1 nA current. Afterward, rectangular patterns in the parallel milling mode were used at 30–50 pA current to reach 200 nm final pattern offset. To ensure homogeneous thickness along the lamella, the stage was tilted 1° up with respect to the initial milling angle, and a rectangular pattern at 30–50 pA was used to progressively mill from the rear toward the front of the lamella (‘overtilt milling’). Final lamellas were sputter coated for 2 s at 30 mA current and 0.1 mbar pressure before unloading for data acquisition.

Support grid preparation for the correlative workflow

The support grid (UltrAuFoil, R2/2, 200 mesh, Quantifoil) was glow-discharged at 30 mA for 30 s (GloQube Plus Glow discharge system, Quorum Technologies). Autofluorescent beads (Dynabeads MyOne, Life Technologies-Thermo Fischer Scientific) were washed following the manufacturer’s instructions and diluted in deionized water to obtain a proper distribution on the grid. From this suspension, 3 µl of beads were applied to the grids, and after buffer evaporation, the grid was clipped in a FIB-type Autogrid and loaded into the Aquilos 2 shuttle at position 2.

Cryo-fluorescence light microscopy

Fluorescent image stacks of the coarse lamellas were acquired on a cryo-confocal microscope (Stellaris 5, Leica Microsystems) equipped with 50×/0.9 NA objective and a cryo-stage. The pixel size was 216 nm (x, y), and the z-step was 330 nm. Images were acquired using an excitation wavelength of 630 nm and 489 nm for reflected light and eGFP-labeled protein, respectively. The bead signal was acquired separately in the same channel of eGFP using a higher intensity level.

Correlative fine milling

Grids were transferred to the Aquilos 2. A FIB image of the sample section acquired at the chosen milling position was registered to the fluorescence data with the 3D Correlative Toolbox software as described previously20,24. In brief, ‘Gaussian fit’ was used to identify the x, y and z positions of five to ten beads in the eGFP channel acquired at high intensity. The same beads were identified in the FIB view to calculate a transformation matrix. Then, the target structures were identified in the eGFP channel that had been recorded at lower intensity. Their z position was fit by the software (‘Gaussian Fit z’) and predicted on the FIB view. Lamellas of ca. 200 nm (final pattern distance) were made at the predicted sites as described above (see ‘Fine milling’ section).

Classical cryo-lift-out procedure

To compare the classical lift-out method with SOLIST, 29 lamellas from waffle-frozen samples were prepared with a semi-classical approach using half-moon grids.

Sample preparation

Apoferritin at a concentration of 84.4 µM was provided by the Biophysics Unit of the Structural Biology Center (Human Technopole). The protein solution was mixed with Ficoll 400 (Sigma-Aldrich) at 40% to obtain a final concentration of 25.6 µM of apoferritin in 20% Ficoll.

Waffle freezing

Before use, 6 mm type B planchettes were cleaned and coated as described above (see ‘Planchette preparation’ section). The support cryo-EM grids (UltrAuFoil, R2/2, 200 mesh, Quantifoil) were prepared as described previously (see ‘Support grid preparation for the correlative workflow’ section). Waffles were prepared following the method described before7.

Attachment site preparation

A copper half-moon grid (Pelco)| with four posts was clipped in a FIB-type Autogrid with the pins toward the ring notch and loaded in position 2 of the Aquilos 2 shuttle. To obtain a straight surface and ensure a safe attachment of the lifted-out lamellas, the right and left sides of each pin were cleaned. With the stage in the perpendicular position, a 20-µm-long CCS pattern was placed on the edges of the pins and 15 nA current was used for milling. The surface was then polished with 3 nA current.

Sample site preparation and lift-out

The sample site was prepared as described above for performing lift-out from the perpendicular direction (see ‘Sample site preparation’ of the ‘Cryo-lift-out and SOLIST procedure’ section). Due to the presence of the grid bars, the sample site was 15 µm in x and 50 µm in y, and 30 µm of material was cleaned in the front. Lift-out from the perpendicular direction was performed as above (see ‘Lift-out’ of the ‘Cryo-lift-out and SOLIST procedure’ section).

Section attachment

With the stage at the milling position, the EasyLift needle with the sample chunk was inserted above the half-moon grid. The needle was adjusted in the x–y plane and lowered in z until the bottom part of the sample chunk was at the same height as the pins. It was then moved in the x direction until it contacted the first pin. The first section of 3–4 µm in y was then attached to the pin with 2-µm-high microstitches created by milling the sample with 30 pA current. An RCS pattern of 1 µm in the y direction was placed on the sample chunk, and 1 nA current was used to release the section.

To increase the throughput of the classical procedure, a serialized approach similar to SOLIST was used: after the first section was attached, the stage position was adjusted, and the attachment was repeated on the right and left side of each pin until the lift-out chunk was exhausted. This method could routinely produce eight sections on the same pin grid.

With all the sample sections attached, the stage was rotated to the perpendicular position, and several microstitches of 2.5 µm × 0.25 µm × 0.25 µm were placed on the sample close to the pin and milled with CCS patterns at 50 pA current. The milling direction was set toward the center of the section.

To minimize lamella vibration and to standardize the comparison with SOLIST, the free side of the lamella, which is not attached to the pin, was milled with a CCS pattern at 0.5 nA current to obtain sections of exactly 10 µm in width (excluding the area used for the microstitches). The same current was then used to clean the leading edge. The stage was rotated to the milling position, and the surface of each section was cleaned from 12–15° stage tilt using a CCS pattern at 0.3 nA current. The grid was GIS-coated for 20 s from the milling direction, 20 s from the perpendicular position and again 20 s from the milling direction before fine milling.

Fine milling

Lamellas were symmetrically thinned down and sputter coated as described above (see ‘Fine milling’ of the ‘Cryo-lift-out and SOLIST procedure’ section). A time comparison of the two methods is presented in Tables 1 and 2.

Cryo-ET

Data acquisition

The support grids with lamellas were loaded on a Titan Krios G4i operated at 300 kV equipped with Falcon4i direct electron detector and SelectrisX energy filter (ThermoFisher Scientific). For the correlative workflow, 3D Correlative Toolbox was used to register fluorescence data to SEM view of the final lamella or to TEM montage to guide correlative TEM acquisition20,24. Cryo-tomograms were recorded at 1.5 Å per pixel (81,000× magnification) for the in vitro LLPS droplets and at 1.953 Å per pixel (64,000× magnification) for samples of classical cryo-lift-out, yeast, brain organoids and mouse tissues. The defocus range was set to −1 μm to −4 μm. A dose-symmetric tilting scheme with an angle increment of 3° was used from −60° to +60° with respect to the lamella pretilt. Tilt series were acquired using the SerialEM tilt series controller or PACEtomo and 3 e− Å−2 per tilt (total dose for 41 images: 123 e− Å−2)16,39.

Data preprocessing

Tilt series were processed in a modified version of TOMOgram MANager (TOMOMAN) (https://github.com/williamnwan/TOMOMAN.git). Frame alignment of the electron event representation (EER) data was performed in Relion5 with EER groups adjusted to 0.25–0.3 e− Å−2 (ref. 40). The final tomograms were denoised using cryoCARE41 and deconvolution-filtered with PyTOM (https://github.com/dtegunov/tom_deconv.git) to enhance the contrast of biologically relevant structures. Aligned tomograms were exported to WARP using TOMOMAN, where the contrast transfer function (CTF) was estimated during preprocessing (tilt CTF; tilt series CTF) and used later for the ad hoc reconstruction of subtomograms42.

STA of yeast ribosomes

Template matching was performed in STOPGAP on a small set of bin8 tomograms43. For the full yeast dataset, crYOLO was trained on the template matching results of ten representative tilt series and used to pick the entire dataset of 65 tomograms44. After particle positions were extracted, an initial classification was performed in Relion (4.0.1) to remove false positives and yielding a clean list of ~17,000 particles45. After this cleaning step, successive rounds of global 3D refinement followed by gradually unbinning to a pixel size of 2.5 Å in WARP resulted in an initial S. cerevisiae ribosome average42. After refinement in M and one more round of 3D alignment in Relion, the final consensus map was obtained at 7.2 Å (refs. 45,46).

STA of mouse liver ribosomes

Template matching was performed with the GPU-accelerated routine implemented in PyTOM at 7° angular sampling on bin8 tomograms47. After particle positions were extracted, an initial classification was performed in Relion (4.0.1) to remove false positives and yielding a clean list of ~15,000 particles45. After this cleaning step, successive rounds of global 3D refinement followed by gradually unbinning to a pixel size of 3.906 Å (bin1) resulted in an initial Mus musculus ribosome average. After refinement in M and one more round of 3D alignment in Relion, the final consensus map was obtained at 8.3 Å (refs. 45,46).

STA of mouse heart filaments

For the three selected tomograms, thin filaments were tracked through the volume using TrackMate in FIJI48,49,50. Traces with a minimum length of 35 pixels were exported and converted into IMOD model files using a custom Matlab (Mathworks) script51. Using the spline tracing implemented in STOPGAP, each object was sampled with a point every 28 Å along the filament43. Tilt and psi angles were set according to the filament orientation while randomizing the rot angle. An initial 3D refinement was performed in Relion with T = 1 (--tau2_fudge 1) and with a prior on the rot and tilt angles (--sigma_tilt 3 --sigma_psi 3) to prevent filament orientations to diverge too much from the initial pick45. Three-dimensional classification at 10 Å per pixel (T = 1) yielded averages from which the filament direction could be deduced on the basis of the orientation of the ‘myosin like’ densities. All particles were flipped to a common direction and 3D refined at 3.906 Å per pixel (bin2). After distance cleaning (cutoff 28 Å), the remaining particles were 3D classified (T = 1) into eight classes. Combining similar classes produced the ‘actin-like’ average mostly devoid of extra signals (14,000 particles) and the ‘myo-like’ average enriched in the the densities of the motor domains (10,000 particles). Each class was then 3D refined using the helical parameters obtained from relion_helical_toolbox (--search) to furnish the final subtomogram averages45.

Tomogram drifting analysis

To compare the stability of the lifted-out lamellas produced with the classical procedure and with SOLIST, 18 tomograms were acquired on 9 lamellas produced with the former approach. For each lamella, one tomogram was acquired close to the pin and one close to the hanging side. The heavy ice contamination prevented us from finding more sites for data acquisition. After processing the tilt series in WARP, a custom script was used to extract and measure the drift accumulated for each tomogram and normalized by the number of EER frames42. The same script was also used to measure the drift of nine tomograms acquired on SOLIST lamellas, randomly selected from a pool of 88 tomograms. The data were plotted, and the statistical analysis was performed with the two-sample t-test in Matlab.

Lamella thickness measurement

A total of 32 lamellas from 10 different SOLIST sessions were used to measure the thickness of lamellas. At least two tomograms per lamella were examined. Tomograms were opened in IMOD and then flipped (to view the y–z plane); the size in z was measured with the measuring tool51. For each tomogram, measurements were performed at three different positions that were averaged per tomogram and then per lamella. A custom Matlab script was used to plot the results.

Usable lamella area measurements

For each of the 18 classical lift-out lamellas and 18 randomly selected SOLIST lamellas, the total area and regions of contamination were segmented in FIJI49,50. Using the measurement tool, the respective areas (pixel2) were determined and the relative usable area was calculated as (total area − contaminated area)/total area.

Segmentation and animation

Initial membrane segmentations were generated using MemBrain52. To match with the training model, denoised bin8 tomograms were scaled to the pixel size of 10 Å and rescaled to the original pixel size after. Further polishing was performed in Chimera with the Volume eraser tool to remove broken membrane segments53. Densities were then segmented in AMIRA (Thermo Fisher Scientific), and the combined label field was separated into individual densities using a custom Matlab script. Labels were then polished with a low-pass filter of 20 Å (RELION5 image handler) and a Gaussian filter of 1 pixel in Chimera before being eventually imported to ChimeraX for visualization and animation40,53,54.

Figure composition

Figures were created in Adobe Illustrator, and parts of Fig. 3 and Supplementary Figs. 4 and 5 were created with BioRender.com.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.