Why Science? | Correlative Light and Electron Microscopy, Multimodal Registration and Ultrastructure Evidence

Section 1 of 36

1. Start with one place, two contrasts

CLEM is valuable when molecular identity and ultrastructure must be connected at the same location. Fluorescence can indicate a labelled target or event, while electron microscopy reveals fine structure. The correlation itself is the experiment.

Contents · Next section

Section 2 of 36

2. Define what counts as the same target

The claim may concern the same cell, organelle, membrane domain or molecular complex. Each scale needs a different relocation and registration accuracy. Write the acceptable target-registration error before collecting the first image.

Contents · Previous section · Next section

Section 3 of 36

3. Distinguish relocation from registration

Relocation finds the approximate region again using grid squares, landmarks or coordinates. Registration mathematically aligns images. Successful relocation does not prove pixel-level correspondence, and a low residual among fiducials does not guarantee low error at the target.

Contents · Previous section · Next section

Section 4 of 36

4. Model transformations explicitly

Rigid transforms permit translation and rotation; affine transforms add scale and shear; non-rigid transforms model local deformation. Use the least flexible model that fits validated landmarks. Extra flexibility can manufacture convincing overlap.

Contents · Previous section · Next section

Section 5 of 36

5. Respect different resolution and contrast

A diffraction-limited fluorescent spot can cover many electron-microscopy structures. Electron contrast may show membranes but not label identity. Correlation narrows interpretation; it does not give both modalities identical resolution.

Contents · Previous section · Next section

Section 6 of 36

6. Treat preparation as a joint optimisation

Fixation, cryogenic handling, dehydration, embedding, sectioning and staining affect ultrastructure and fluorescence differently. The best compromise depends on the question. Report what was preserved, what was sacrificed and how it was tested.

Contents · Previous section · Next section

Section 7 of 36

7. Define the correlation claim

Predeclare the target, modalities, registration model, landmarks, error threshold and independent specimens. Decide whether the endpoint is proximity, contact, enclosure or identity. Those words require different spatial evidence.

Contents · Previous section · Next section

Section 8 of 36

8. Choose a workflow deliberately

CLEM may be live-to-fixed, cryogenic, resin-embedded, serial-section or on-section. Each workflow changes time, deformation and accessible labels. Name the specific route instead of treating CLEM as one uniform method.

Contents · Previous section · Next section

Section 9 of 36

9. Preserve fluorescence through processing

Fluorophores can fade, shift spectra or be quenched by fixation and embedding. Measure retention in matched controls and image before and after critical steps where possible. A missing fluorescent signal after processing is not proof of biological absence.

Contents · Previous section · Next section

Section 10 of 36

10. Protect ultrastructure

Fixation strength, osmium, dehydration and resin influence membranes and contrast. Evaluate known structures and include preparation controls. Avoid optimising only the fluorescent channel while accepting unexplained ultrastructural damage.

Contents · Previous section · Next section

Section 11 of 36

11. Add fiducials that survive both modalities

Beads, patterned substrates, grid features or endogenous landmarks can link coordinate systems. Fiducials should surround the target, span the field and remain identifiable without altering biology. Near-collinear landmarks give unstable transforms.

Contents · Previous section · Next section

Section 12 of 36

12. Record serial-section order

For array or serial-section CLEM, lost, folded or reversed sections change the three-dimensional narrative. Keep section identifiers, cutting order, thickness and orientation. The 2025 protocol emphasises systematic marking and registration across sections.

Contents · Previous section · Next section

Section 13 of 36

13. Use coordinate maps and stage metadata

Save overview images, stage positions, magnification, orientation and grid coordinates at every scale. Screenshots and handwritten arrows are fragile. A relocation chain should lead from specimen overview to the final electron field.

Contents · Previous section · Next section

Section 14 of 36

14. Plan timing in live-to-fixed work

The fluorescent event and fixed ultrastructure occur at different times. Record the delay from event to fixation and the kinetics of fixation. Registration cannot remove biological motion during that interval.

Contents · Previous section · Next section

Section 15 of 36

15. Practise with an invented CLEM table

The fictional registrations below show why a small residual is not the whole answer.

A transform can fit landmarks beautifully and still miss the biological target.

Contents · Previous section · Next section

Section 16 of 36

16. Register from coarse to fine

Align overview maps first, then intermediate and high-magnification images. Each level limits the search space and exposes mismatches. Save every transform rather than applying an undocumented final warp.

Contents · Previous section · Next section

Section 17 of 36

17. Measure landmark uncertainty

Fiducial centres are not exact, especially when point-spread functions and electron contrast differ. Record annotation variation and size. Weighting landmarks by confidence can be more honest than treating every click as perfect.

Contents · Previous section · Next section

Section 18 of 36

18. Validate with withheld landmarks

Fit the transform on one set of fiducials and calculate error on separate landmarks near the target. This tests prediction rather than only fit. Report target-registration error or a comparable spatial uncertainty.

Contents · Previous section · Next section

Section 19 of 36

19. Inspect local deformation

Sectioning, drying and beam exposure can stretch or compress regions. Overlay displacement vectors and compare local distances. A global affine fit can hide a warped membrane beside the target.

Contents · Previous section · Next section

Section 20 of 36

20. Interpret overlap probabilistically

When the fluorescent spot is wider than several organelles, multiple candidates may fall inside it. Use the combined localisation and registration uncertainty to rank possibilities. State ambiguity rather than selecting the most attractive structure.

Contents · Previous section · Next section

Section 21 of 36

21. Preserve replicate hierarchy

Many correlated fields may come from one specimen and many structures from one cell. The cell, preparation or organism may be the independent unit. Report successful and failed correlations, not only showcase matches.

Contents · Previous section · Next section

Section 22 of 36

22. Challenge fluorescence linkage

Antibodies, tags and expressed reporters can displace or alter the target. Validate function and specificity with negative controls. Correlation to the label position is not necessarily correlation to the protein’s active site.

Contents · Previous section · Next section

Section 23 of 36

23. Challenge fixation delay

A vesicle or organelle can move between live fluorescence and immobilisation. Use rapid fixation, time controls or cryogenic workflows when timing matters. Registration error and biological displacement are separate uncertainties.

Contents · Previous section · Next section

Section 24 of 36

24. Challenge landmark selection

Choosing landmarks after seeing the desired overlap invites bias. Predefine types and exclusions, blind annotators where possible and show all candidates. A transform should not be tuned to make the hypothesis look true.

Contents · Previous section · Next section

Section 25 of 36

25. Challenge non-rigid warping

A flexible transform can force unrelated features together. Limit degrees of freedom, regularise, inspect deformation fields and validate on withheld points. Preserve the unwarped images for audit.

Contents · Previous section · Next section

Section 26 of 36

26. Challenge serial-section continuity

A structure appearing in adjacent sections may branch, end or be misordered. Track landmarks through the sequence and report missing sections. Three-dimensional continuity is an inference that needs its own checks.

Contents · Previous section · Next section

Section 27 of 36

27. Compare independent labels and modalities

Immunogold, genetic perturbation, tomography or biochemical assays can test identity and mechanism. CLEM supplies spatial correlation, not causation. Follow-up should attack the most plausible alternative interpretation.

Contents · Previous section · Next section

Section 28 of 36

28. Learn with transparent overlays

Students can align two printed maps using landmarks, reserve one point for validation and compare rigid with distorted overlays. They observe that a good fit at landmarks may still fail elsewhere, without laboratory hazards.

Contents · Previous section · Next section

Section 29 of 36

29. Connect Primary Science to matching evidence

A simple map-matching exercise shows how shared landmarks identify the same place at different scales. Learners explain which features are dependable and why one matching point is insufficient.

Contents · Previous section · Next section

Section 30 of 36

30. Build PSLE Science process skills

Students identify landmark coordinates as observations, the transform as a method and colocalisation as an inference. They suggest label-negative controls, repeated annotations and withheld landmarks to make the test fair.

Contents · Previous section · Next section

Section 31 of 36

31. Extend into Secondary and O-Level Science

Physics contributes optics and electrons; Biology contributes cells and labels; Mathematics contributes coordinates and transformations; Computing contributes registration. Science enrichment can connect them without claiming CLEM is a prescribed syllabus topic.

Contents · Previous section · Next section

Section 32 of 36

32. Use the topic for school choices

Verify current school programmes through official sources and value foundational practical and data reasoning. Never invent electron-microscope access, partnerships, admission requirements or career outcomes. A low-cost registration exercise can still teach authentic science.

Contents · Previous section · Next section

Section 33 of 36

33. See the career ecosystem

CLEM connects cell biology, pathology, microscopy, image analysis, software, specimen preparation and research facilities. Work is collaborative and requires careful records, safety and current training appropriate to each role.

Contents · Previous section · Next section

Section 34 of 36

34. Did You Know? The overlay is a hypothesis

A coloured fluorescence spot placed on a grey electron micrograph is not automatically proof of identity. The transform, fiducials and measured error determine whether that overlay is a defensible spatial hypothesis.

Contents · Previous section · Next section

Section 35 of 36

35. Did You Know? Better fit can mean worse science

A more flexible warp usually lowers landmark residuals, but it can also distort the specimen and overfit. Withheld landmarks and deformation maps show whether the apparent improvement predicts real locations.

Contents · Previous section · Next section

Section 36 of 36

36. Keep landmark-to-identity reasoning visible

Define the same-target scale, select a compatible workflow, measure fluorescence and ultrastructure preservation, record relocation, distribute fiducials, fit the simplest justified transform, validate near the target, quantify timing and deformation, preserve unwarped data, report all correlation attempts and seek orthogonal tests before claiming molecular identity in ultrastructure.

A defensible CLEM project begins with a correlation budget. Specify the target, field size, expected motion or deformation, light-microscopy localisation, electron-microscopy pixel size and maximum acceptable target-registration error. The coarsest term often sets the limit. There is little value in nanometre electron pixels if relocation uncertainty is several micrometres.

Workflow choice should follow the biological question. Live-to-fixed CLEM preserves temporal information but introduces fixation delay; resin workflows support serial sections but may reduce fluorescence; cryogenic workflows can preserve hydrated structure but demand specialised transfer and alignment. Write the scientific compromise, not merely the protocol name.

Fluorescence validation should measure intensity, localisation and specificity before and after critical preparation steps. Use label-negative, single-label and known-positive controls. If signal disappears preferentially from one compartment, the surviving fluorescence is a biased sample. Retention percentage alone may miss spatially selective loss.

Ultrastructure validation should use recognised membranes, organelles or standards and blinded scoring. Stronger fixation or staining can improve contrast while changing antigenicity and dimensions. Measure shrinkage or expansion in landmarks. A perfect overlay between two deformed modalities can still describe altered biology.

Fiducial design matters geometrically. Landmarks should surround the region of interest and span both axes; three-dimensional work needs depth information as well. Points clustered in one corner make the target an extrapolation. Synthetic patterns can improve localisation, but biological compatibility and survival through preparation must be tested.

Landmark annotation should be repeated by independent observers or an algorithm with validated uncertainty. A bead centre can appear different in wide fluorescence and sharp electron contrast. Preserve the original clicks and confidence, not only the final transform. Weighted fitting can reflect landmark quality when justified.

Registration should proceed from coarse overview to fine detail, with transformations saved at every level. The 2025 serial-section protocol’s emphasis on marking, relocation and computational alignment makes this chain explicit. Skipping intermediate scales increases the chance of a convincing match to the wrong region.

Transform complexity needs evidence. Start with rigid or similarity models when physical preparation should preserve shape, then test affine or non-rigid terms only when residual patterns justify them. Plot the deformation field. A flexible spline that produces extreme local stretch near the target should not be trusted because its landmark residual is small.

Validation points must be withheld from fitting and placed near the biological target. Cross-validation can rotate which points are held out. Report target-registration error or an interval based on landmark uncertainty and model fit. Mean fiducial residual is a training statistic, not a prediction guarantee.

Serial sections add order and thickness uncertainty. Record every section, missing or folded sections, compression and knife marks. Use landmarks that persist through several sections and test continuity. If a target lies near a section surface, label loss or cutting damage can dominate the interpretation.

Live-to-fixed experiments should log the last fluorescence frame, fixation start, immobilisation estimate and electron-microscopy time. Fast organelles can move farther than the registration error before fixation. The spatial uncertainty should combine biological motion with image alignment instead of reporting only one.

Colocalisation should be stated at the scale supported. ‘Within the same cell’, ‘within 300 nanometres of this membrane’ and ‘on this molecular complex’ are different claims. Use the joint uncertainty and the number of candidate structures inside it. When multiple candidates remain, report them rather than choosing one.

Correlation yield is a result. Track how many labelled events were identified, relocated, survived processing, registered within threshold and produced interpretable ultrastructure. Failure may depend on condition or morphology. A gallery of successful targets without denominators exaggerates robustness.

Biological inference needs independent specimens. Many correlated organelles in one cell do not replace replication across preparations. Balance conditions across sessions and microscope operators. Show all specimen-level outcomes and model nested structure. CLEM is especially vulnerable to long, selective workflows that quietly shrink sample size.

Pre-registration can name the primary spatial relationship, workflow, landmark classes, transform, validation rule, error ceiling, exclusion and independent unit. Exploratory overlays can be shown separately. This prevents tuning a registration until the preferred organelle sits under the fluorescence spot.

Quality control should include a dual-modality standard or patterned substrate, fluorescence retention, dimensional change, landmark annotation, transform residual and withheld-point error. Plot these over time. A new resin batch, detector calibration or software version can alter one side of the correlation.

Data stewardship should preserve raw light and electron images, overview maps, stage and grid coordinates, timestamps, preparation records, section order, landmarks, transforms, deformation fields, validation points, masks, code, software versions and final overlays. Every overlay should be reproducible from unwarped inputs.

Figures should show both original modalities, an overview-to-target relocation chain, fiducials, validation error, deformation map and a transparent overlay at more than one magnification. Avoid opaque colours that hide electron structure. Give uncertainty rings or contours when a point estimate suggests false precision.

Negative correlation can be informative when label retention, relocation, registration and detection limits pass. Calculate the smallest separation or association the workflow could resolve. If the fluorescent target repeatedly falls outside a bounded ultrastructural compartment, that can constrain a localisation model without proving a mechanism.

Follow-up should attack the leading alternative: a second label for specificity, faster fixation for movement, cryogenic preservation for deformation, tomography for depth, immunogold for molecular identity or perturbation for causation. CLEM’s strength is not that two images agree, but that each modality tests what the other cannot.

Safety spans both workflows. Chemical fixatives, heavy-metal stains, resins, ultramicrotomes, electron microscopes, cryogens and biological specimens require trained staff and approved disposal. Classroom learning should use printed images and public datasets. Registration reasoning can be authentic without exposure to hazardous preparation.

Reproducibility depends on stable identifiers across modalities. Connect specimen, live event, light image, section, electron field, landmark set, transform and final claim in one manifest. When a target is lost, record the stage at which it disappeared. This lineage reveals whether failure arose from biology, preparation, relocation or registration and prevents a successful overlay from becoming detached from all the unsuccessful attempts around it.

Contents · Previous section · Continue to the Science Learning Hub