Hitting the Lamella: Depth-Aware Light Targeting for Cryo-Tomography of Stressed Gut Bacteria in Frozen Tissue

Cryo-electron tomography images the molecular machinery inside cells in their frozen, near-native state. It works by tilting a sample in an electron microscope and reconstructing a 3D volume. Electrons cannot pass through more than a few hundred nanometres of ice, so a slab of frozen tissue must first be carved down with an ion beam into a sheet about 200 nm thick, called a lamella. To place that sheet over a target such as one bacterium sitting on a gut wall, researchers find the target with a fluorescence microscope at cryogenic temperature and then mill around it. On flat, thin samples this optical step is already marginal. Liu et al. measured fluorescent beads on electron-microscopy grids and found that the standard optical model of how a point of light blurs (the point spread function, or PSF) misplaces emitters by up to 90 nm laterally and more than 300 nm in depth. That depth error is larger than the lamella itself. Inside tissue tens of micrometres thick, light also scatters and bends at refractive-index boundaries, so the error should grow with depth in a way that varies from sample to sample. Nobody has measured how large it becomes.

We propose to measure that depth-dependent error directly in frozen tissue and to correct it. The correction is a PSF model conditioned on how deep the target is and on how strongly the local tissue attenuates light, which we estimate from the same optical stack. We then show the correction changes what fraction of milled lamellae actually contain the intended target. The test application comes from environmental stress biology. Wang et al. showed that acute salinity shock in the sea cucumber Stichopus monotuberculatus destroys the intestinal brush border and reshapes a gut community dominated by Proteobacteria, with 48 h as the critical time point. That work used fixed 2D slides and bulk sequencing, so it cannot say what happens to the bacteria themselves at the damaged surface. Answering that requires a tomogram of a specific bacterium in a specific place, which is exactly the targeting problem.

Core Claim

In high-pressure-frozen tissue, cryo-fluorescence localization error is dominated by a component that grows with depth and with local attenuation. A model fitted on flat grids, even a spatially matched one, cannot remove this component. A PSF conditioned on depth and on locally estimated attenuation can bring axial error below half the lamella thickness at the depths that high-pressure freezing can vitrify. This should raise the targeted-lamella hit rate from a level that makes in-tissue bacterial tomography impractical to one that makes it routine.

The link between localization error and hit rate is simple and is the reason axial error is the quantity to attack. Suppose the residual axial error $\delta z$ is roughly Gaussian with standard deviation $\sigma_z$, and the lamella has thickness $t$ centred on the estimated target position. Then the probability that a point-like target survives milling is

$$ P_{\text{hit}} = \operatorname{erf}\!\left(\frac{t/2}{\sqrt{2}\,\sigma_z}\right). $$

With $t = 200$ nm, $\sigma_z \approx 300$ nm gives a hit rate near one in four, and $\sigma_z \approx 30$ nm gives nearly certain hits. The 300 nm figure is the magnitude of the uncorrected error and the 30 nm figure is the corrected one, both from Liu et al. on grids. A systematic bias, as opposed to random spread, is worse still, because it moves the lamella off the target in every attempt. A bacterium is about a micrometre across, so it tolerates more error than a bead does. Its subcellular features, such as the envelope facing the host surface, do not.

Research Program

The work has three parts. The first is a measurement study, the second builds the correction, and the third applies it to the biology.

The measurement study embeds 200 nm fluorescent beads with an electron-dense core into sea cucumber foregut tissue before high-pressure freezing. These are the same bead size Liu et al. used, which keeps the two studies comparable. We image the frozen block with the cryo-fluorescence microscope, record each bead's apparent 3D position, and then mill progressively while imaging the exposed face with the scanning electron beam. The scanning images give each bead's true position to a few nanometres, so we obtain ground-truth localization error as a function of depth below the tissue surface (0–150 µm) and of tissue type (epithelium, lumen contents, connective layer). This dataset does not currently exist. On its own it answers the question Liu et al. left open about thick samples.

The correction model has three ingredients. It starts from Liu et al.'s Zernike-mode description of aberrations, which is a standard decomposition of a distorted wavefront into named shapes such as astigmatism and coma. It adds a depth-dependent defocus and spherical term, which models refractive-index mismatch between the objective's medium and vitreous tissue. It adds a third term that broadens the PSF as a function of a local attenuation coefficient. That coefficient is estimated from the decay of reflected or autofluorescent signal with depth along each column of the optical stack. The estimation follows the per-line attenuation fitting used in intravascular optical coherence tomography (Automatic Classification of A-Lines…), which recovers the local attenuation coefficient from single depth profiles.

The physical justification is the Beer–Lambert picture. Unscattered light decays roughly as $e^{-\mu_t z}$, where $\mu_t$ is the total attenuation coefficient. Index mismatches between membranes and cytoplasm drive scattering in tissue (Chen et al.), so the scattered fraction that biases localization should grow with $\mu_t z$. The model is fitted per sample, using the embedded beads in that sample as calibration, and is checked on held-out beads.

The biological application compares tissue held at 30‰ salinity with tissue at 40‰ for 48 h, matching Wang et al.'s design. Bacteria are labelled before freezing with a fluorescent cell-wall precursor. We target bacteria in direct contact with the epithelial surface, mill lamellae, and collect tomograms. The readouts are envelope geometry (periplasm width, outer-membrane continuity), the density of ribosomes inside the cell, and whether contact sites sit on intact microvilli or on denuded epithelium.

Ribosome counting uses subtomogram classification. This is the step where small 3D boxes cut from a tomogram are sorted into molecule types. The method is adapted from Cryo-shift (Cryo-shift), which classifies real tomograms with a model trained on simulated ones.

The same frozen optical volumes also give a fixation-free measurement of microvillus length and mucosal thickness. We compare these with paraffin-section histology of matched animals. This tests whether the "absent brush border" Wang et al. report at 40‰ is partly a chemical-fixation artifact. Fast freezing preserves cell states that slow chemical fixation alters (Time-deterministic cryo-optical microscopy), so the question is worth asking, and it costs nothing extra.

Within this program the four directions play different roles. Cryo-tomography is the endpoint. Optical penetration depth is the constraint being modelled. Environmental stress supplies the biological question. Histopathology supplies the tissue-scale readout and a check on fixation artifacts.

Baselines

The naive reference point is the theoretical PSF model shipped with commercial cryo-fluorescence systems, used with no depth correction. For this model on grids, Liu et al. report localization biases of up to 90 nm laterally and over 300 nm axially. We expect it to be substantially worse in tissue. We will report it as measured, because the file contains no reported result found on Lacuna for cryo-fluorescence localization error in thick vitrified tissue.

The strongest existing competitor is Liu et al.'s spatially matched experimental PSF, calibrated on beads on flat grids with the same microscope and objective. On grids it reaches about 6 nm lateral and 30 nm axial error at typical photon counts, and sub-nanometre lateral and sub-5 nm axial error at high photon counts. It is the strongest competitor because it already captures the instrument's field-dependent astigmatism and coma, which Liu et al. identify as the dominant aberrations. If the tissue adds little beyond the instrument, this model is all one needs.

The comparison will be kept fair in four ways. Both models will be fitted with the same number of calibration beads and the same photon budget. The grid-calibrated model will be allowed the same per-session recalibration. Both will be evaluated on identical held-out tissue beads. Beating it means a statistically clear reduction in axial RMSE (root mean square error) at depths beyond roughly 20 µm, and a hit rate that improves by an amount consistent with the erf relation above.

A second competitor needs no optical model at all. In stop-on-sight milling, one mills slowly while imaging the exposed face with the scanning electron beam and stops when the target appears. It gets the same instrument time per lamella as our method.

Two ablations isolate the key ingredient. The first removes the attenuation term and keeps only depth-dependent defocus and spherical correction. The second keeps the attenuation term but uses a single sample-average attenuation coefficient instead of a local one. If the full model is no better than the first ablation, the contribution reduces to known index-mismatch optics and the attenuation idea should be dropped.

Why not just…?

Why not just use Liu et al.'s spatially matched PSF in tissue? This is the most serious threat, and the proposal's first experiment exists to settle it. Liu et al. already found that cryogenic imaging widens the PSF by 1.4–1.7× and that aberrations vary across the field even on flat grids. Tissue adds index-mismatch and scattering that change with depth, and a grid calibration cannot see either. If the depth-resolved bead data show the grid-calibrated model keeping axial error under about 50 nm throughout the vitrifiable depth, we will report that as the result. The project then becomes the hit-rate and biology study using Liu et al.'s method, with no new model.

Why not just mill while watching the face with the electron beam and stop when the bacterium appears? For micrometre-sized targets that differ clearly from their surroundings, this may work, and we treat it as a baseline rather than a straw man. Its weakness here is contrast. In frozen, unstained tissue, a bacterium pressed against microvilli and mucus may not stand out from host organelles in scanning-electron images. Repeated face imaging also costs time and can damage the sample. The comparison will show whether this approach identifies targets reliably. If it does, optical correction matters only for features smaller than the bacterium.

Why not just use resin-embedded FIB-SEM after in vivo or room-temperature imaging, as in correlative two-photon and FIB-SEM? Here FIB-SEM means milling and imaging a plastic-embedded block layer by layer with an ion beam and a scanning electron beam. For the question of where bacteria sit relative to damaged microvilli, it would plausibly work. We therefore do not claim cryo-tomography is needed for that spatial question. The claim is restricted to readouts that heavy-metal staining and resin replacement destroy: native envelope spacing and the identification of individual ribosomes inside the bacteria.

Why not just extend Wang et al. with more sequencing and H&E? Bulk sequencing of homogenized gut tells you who is present, not where they are or what state their cells are in. H&E is limited to roughly 200 nm resolution and depends on fixation. The fixation-free morphometry comparison in this proposal directly tests one of Wang et al.'s central histological observations.

Why not clear the tissue or add hardware adaptive optics? Chemical clearing takes weeks and distorts tissue (SOCM), and it is incompatible with vitrification. Deformable-mirror correction works in living tissue (hierarchical adaptive optics), and so does accumulation of singly scattered light (closed-loop single-scattering accumulation). Neither has been put on a cryo stage, which typically uses a long-working-distance air objective. A computational correction can be applied to existing instruments today.

Proposed Experiments

The first and decisive experiment is the depth-resolved bead study. We will measure lateral and axial localization error against depth and local attenuation, using four models: the theoretical PSF, the grid-calibrated spatially matched PSF, the two ablations, and the full model. Error will be scored against scanning-electron ground truth on held-out beads, across at least several independent frozen tissue blocks so that sample-to-sample variation is visible.

The second experiment converts localization accuracy into practical outcome. Matched sets of lamellae will be targeted at labelled bacteria using each targeting method, including stop-on-sight milling. The primary metric is the fraction containing the target in the overview image taken in the transmission electron microscope. The secondary metric is the fraction where the bacterium–host contact plane lies within the lamella. We will check whether observed hit rates follow the erf prediction from measured $\sigma_z$. Agreement would mean bead studies alone can predict targeting success in future tissues.

The third experiment is the biology. We will compare bacterial envelope geometry and ribosome density between 30‰ and 40‰ tissue at 48 h. Ribosome classification uses a Cryo-shift-style model adapted from simulated to real tomograms. As a reference for what such classifiers achieve, Cryo-shift reports 0.84 average accuracy on experimental rat-neuron tomograms. SHREC 2021 (SHREC 2021) reports a best localization F1 of 0.868 for DeepFinder on simulated data, and smaller particles score worse. Because ours are tissue tomograms of bacteria, we will validate the classifier on a manually annotated subset rather than assume these numbers transfer.

The fourth experiment measures microvillus length and mucosal thickness in the frozen optical volumes and compares them with paraffin histology from matched animals. A large discrepancy at 40‰ would show that part of the reported brush-border loss arises during fixation, not from the stress itself.

The project succeeds if the full model clearly beats the grid-calibrated PSF at depth, and if that advantage appears as a measurable rise in hit rate. It fails informatively if grid calibration already suffices. In that case the field learns that tissue targeting is limited by contrast or milling rather than by optics, and the biological study proceeds using existing tools.