Why Science? | STitch3D, Multiple Slices, Single-Cell Atlases and 3D Evidence

Section 1 of 36

1. Begin with measured slices, not a magical volume

Most spatial-transcriptomics experiments measure thin two-dimensional sections. STitch3D combines several sections to study three-dimensional organisation. Every reconstruction should begin by showing the measured planes, their order and the gaps between them.

Contents · Next section

Section 2 of 36

2. Align sections before modelling

The documented workflow aligns multiple tissue slices before building the model. Iterative closest point or PASTE can be used. This registration step contributes its own uncertainty and must not disappear behind the later neural network.

Contents · Previous section · Next section

Section 3 of 36

3. Add a paired single-cell reference

A single-cell RNA-sequencing atlas supplies cell-type-specific expression profiles. Reference quality, tissue match, dissociation bias and annotation affect inferred proportions. A reference from another condition or stage may omit genuine states.

Contents · Previous section · Next section

Section 4 of 36

4. Build a three-dimensional spatial graph

Aligned spots become nodes in a graph that includes within-slice and cross-slice neighbourhoods. Slice spacing and centre-to-centre distance shape those edges. The graph is a model of proximity, not direct evidence of cellular contact.

Contents · Previous section · Next section

Section 5 of 36

5. Learn a shared latent representation

Deep learning combines spatial-transcriptomics measurements, graph structure and reference information. Latent features can support domains and denoising. They are mathematical summaries rather than directly observed genes, pathways or anatomy.

Contents · Previous section · Next section

Section 6 of 36

6. Identify three-dimensional spatial regions

STitch3D seeks regions with coherent expression across the reconstructed volume. These domains may reveal layers or compartments. Their continuity depends on registration, graph scale, section spacing and missing tissue.

Contents · Previous section · Next section

Section 7 of 36

7. Infer three-dimensional cell-type distributions

The method estimates cell-type proportions from the reference across spots and slices. Proportions are model outputs, not counted cells. Fine-grained labels need markers, morphology and independent imaging.

Contents · Previous section · Next section

Section 8 of 36

8. Create a section manifest at the cryostat

Record specimen, section order, thickness, interval, orientation marks, staining, damage and collection time. State whether slices are adjacent, near-adjacent or separated. Computational smoothness cannot recover laboratory metadata that were never recorded.

Contents · Previous section · Next section

Section 9 of 36

9. Preserve images, coordinates and masks

Keep uncropped histology, raw coordinates, tissue masks and transformed coordinates. Show excluded tissue and tears. A 3D graph can bridge missing regions unless masks and boundaries are explicit.

Contents · Previous section · Next section

Section 10 of 36

10. Audit the single-cell reference

Document specimens, batches, filtering, doublets, gene identifiers, cell labels and missing states. Perform leave-one-cell-type-out or alternative-reference checks. Confident mapping to the nearest available type can be wrong when the reference is incomplete.

Contents · Previous section · Next section

Section 11 of 36

11. Harmonise genes without erasing specificity

Record gene intersection, normalisation and variable-feature selection for the spatial and reference data. Withhold some markers for validation. If a cell-type distribution depends on markers used to define the reference, say so.

Contents · Previous section · Next section

Section 12 of 36

12. Set slice distances in physical units

The documentation accepts distances between adjacent slices in micrometres. Use laboratory values where available and test plausible uncertainty. Treating unknown distances as negligible changes the 3D graph and the apparent continuity of domains.

Contents · Previous section · Next section

Section 13 of 36

13. Choose alignment with evidence

Compare ICP, PASTE or another justified route using independent landmarks and tissue masks. Do not select the method solely because it produces a compact stack. Preserve local residuals and regions that cannot be aligned.

Contents · Previous section · Next section

Section 14 of 36

14. Archive the full reconstruction pipeline

Save raw slices, order, spacing, masks, transforms, graph edges, reference profiles, features, model settings, seeds, checkpoints, domains, proportions, software versions, hardware and validation. A final 3D render is not a reproducibility package.

Contents · Previous section · Next section

Section 15 of 36

15. Practise with an invented 3D audit table

This fictional table teaches interpretation, not performance.

Contents · Previous section · Next section

Section 16 of 36

16. Read domains as reconstructed regions

A 3D domain joins evidence across measured planes through a model. Show per-slice assignments and uncertainty before rendering a surface. Continuity in the image should not imply measurement between sections.

Contents · Previous section · Next section

Section 17 of 36

17. Read proportions as estimates

Plot posterior or model uncertainty where available and compare raw marker expression. A proportion can be influenced by reference similarity, capture resolution and cell density. It is not a direct cell count.

Contents · Previous section · Next section

Section 18 of 36

18. Inspect cross-slice edges

Map edges that connect slices and summarise their distances. Check whether they cross missing tissue, cavities or damaged regions. A few implausible edges can make separate structures appear continuous.

Contents · Previous section · Next section

Section 19 of 36

19. Distinguish denoising from discovery

Borrowing information across slices can stabilise sparse genes. It can also propagate a pattern. Validate denoised expression with held-out genes or imaging and retain raw values beside reconstructed maps.

Contents · Previous section · Next section

Section 20 of 36

20. Keep virtual slices visibly virtual

The repository describes downstream use including newly generated virtual slices. Label them as predictions, show neighbouring measured sections and quantify interpolation uncertainty. Never mix virtual and measured planes without a clear visual distinction.

Contents · Previous section · Next section

Section 21 of 36

21. Replicate the whole laboratory pipeline

A second specimen should repeat sectioning, registration, graph construction and modelling. Thousands of spots in one stack do not establish population generality. Report specimens that cannot be reconstructed cleanly.

Contents · Previous section · Next section

Section 22 of 36

22. Audit section order

Deliberately swap a pair or reverse a subset as a negative control. A model that still produces the same smooth anatomy may be driven by the reference or within-slice structure rather than correct three-dimensional order.

Contents · Previous section · Next section

Section 23 of 36

23. Audit spacing and graph scale

Vary slice distances and cross-slice neighbourhood thresholds over plausible ranges. Track domains, cell-type gradients and local continuity. Report conclusions that survive and regions that remain sensitive.

Contents · Previous section · Next section

Section 24 of 36

24. Audit reference sensitivity

Use an alternative reference, leave out a cell type and perturb annotations. Observe whether omitted biology becomes diffuse or is reassigned confidently. This reveals how reference incompleteness appears in the final atlas.

Contents · Previous section · Next section

Section 25 of 36

25. Audit registration sensitivity

Compare candidate alignments, landmarks and masks. Quantify residual vectors and local distortion. If a domain exists only under one registration, the claim is about that alignment choice, not established 3D anatomy.

Contents · Previous section · Next section

Section 26 of 36

26. Validate with independent three-dimensional evidence

Use histology, protein imaging, known landmarks, cleared-tissue imaging or another modality not fitted by the model. Validate stable, ambiguous and negative structures. A prettier volume is not stronger evidence by itself.

Contents · Previous section · Next section

Section 27 of 36

27. Compare with slice-by-slice and simpler baselines

Show what each measured section says alone, plus a simple stack or alternative deconvolution. STitch3D earns complexity when it adds reproducible cross-slice structure without erasing real variation or inventing continuity.

Contents · Previous section · Next section

Section 28 of 36

28. Report gaps and failures

Missing sections, damaged tissue, poor reference match or uncertain alignment may make a local 3D claim impossible. Leave gaps visible and archive failed reconstructions. Refusing false continuity is a scientific success.

Contents · Previous section · Next section

Section 29 of 36

29. Use uncertainty to plan the next section

Identify the plane or region where another measured slice, stain or targeted panel would most reduce uncertainty. Reconstruction becomes especially valuable when it tells the laboratory where to collect the next piece of evidence.

Contents · Previous section · Next section

Section 30 of 36

30. Start with Primary Science habits

Primary Science already teaches the habits beneath STitch3D: observe carefully, compare fairly, record consistently and keep conclusions within the experiment. A simple map of plants, light or water can show why location changes interpretation without advanced mathematics.

Contents · Previous section · Next section

Section 31 of 36

31. Use PSLE Science to practise claim limits

PSLE Science connects observations to processes and explanations. STitch3D adds a modern reminder that a statistical pattern is not automatically a cause. Students can ask what changed, what was measured, what remained uncontrolled and which new observation would separate competing explanations.

Contents · Previous section · Next section

Section 32 of 36

32. Connect Secondary and O-Level Science

Secondary Science and O-Level Biology develop cells, organisation, variation, experimental design and evaluation. STitch3D provides a contemporary case where molecular measurements, tissue position and computing meet. The learning goal is disciplined reasoning, not memorising a package name.

Contents · Previous section · Next section

Section 33 of 36

33. Let Computing support the science

Computing contributes data structures, graphs, algorithms, optimisation, visualisation and reproducibility. Biology supplies the specimen, mechanism and independent validation. STitch3D shows why correct code is necessary but insufficient: an algorithm can run perfectly on mislabelled, confounded or incomplete data.

Contents · Previous section · Next section

Section 34 of 36

34. Choose school opportunities by fit

Families should verify Biology, Computing, mathematics, microscopy, laboratory research, data visualisation and interdisciplinary mentoring on current official school and MOE pages. A school mention of genomics, AI or data science does not guarantee programme depth, admission or a career outcome. Look for sustained inquiry, careful teaching, accessible mentoring and time to explain evidence.

Contents · Previous section · Next section

Section 35 of 36

35. See the connected career families

Reasoning used in STitch3D appears in three-dimensional genomics, developmental biology, pathology, computational anatomy, biomedical imaging, bioinformatics and scientific software. Routes can pass through polytechnic, junior college, university or continuing education with different blends of biology, mathematics, statistics, computing and communication. Current course requirements must be checked directly; one project cannot guarantee entry or employment.

Contents · Previous section · Next section

Section 36 of 36

36. Finish with a bounded scientific claim

A defensible conclusion states exactly what STitch3D estimated and under which inputs, settings and specimens. STitch3D reconstructs domains and cell-type distributions from aligned slices and a reference; it does not make virtual planes measured tissue, inferred proportions direct counts or a smooth volume continuous observed anatomy. Report uncertainty, sensitivity, replication, negative results and orthogonal validation together. That boundary is not weakness; it makes the result testable and reusable.

A publication-ready STitch3D project starts at sectioning. Record specimen, block orientation, section order, thickness, interval, stains, missing sections, folds and damaged regions. Photograph orientation marks and retain uncropped images. A neural model cannot recover whether two tubes were swapped or whether an unrecorded gap contains important anatomy.

Build a slice manifest with barcodes, spots, genes, coordinate units, scaling, mask area, quality summaries and expected overlap. State whether sections are adjacent, near-adjacent or widely spaced. Enter slice distances from laboratory records rather than choosing values because the final volume looks smooth.

Treat registration as an independent analysis. Compare the documented ICP or PASTE routes, show landmarks and residual vectors, and retain unmatched tissue. Registration error becomes input to every 3D graph edge and domain. A successful downstream model cannot retroactively prove that the initial transform was correct.

Audit the single-cell reference with the same care as the spatial data. Record tissue, condition, developmental stage, specimens, dissociation, filtering, doublets, gene identifiers and annotation. Use alternative references or leave-one-cell-type-out tests. When a state is missing, the model may assign it confidently to the nearest available type.

Inspect the 3D graph directly. Plot within-slice and cross-slice edge lengths, degrees and edges crossing masks. Vary slice distance and cross-slice radius. A graph that bridges a cavity or missing section can create a continuous domain that no measured tissue supports.

Keep measured, denoised and virtual outputs visually distinct. Measured planes should be obvious; denoised genes should link back to raw counts; virtual slices should carry prediction labels and uncertainty. Never use one colour scale that makes measured and inferred tissue indistinguishable.

Validate cell-type proportions with held-out markers, protein imaging or another reference. Compare abundance at the specimen level and acknowledge spot resolution. Proportions do not identify exact cell positions inside a spot, and reference annotation uncertainty should propagate into interpretation.

Validate 3D domains with independent anatomy: histological layers, cleared-tissue imaging, vessels, known shapes or another modality not used to tune the model. Include difficult and negative regions. A reconstruction that succeeds only on familiar broad anatomy may still fail at fine boundaries.

Run order and spacing negative controls. Swap adjacent slices, reverse a subset or perturb spacing within plausible uncertainty. If the same volume appears, the result may be driven mainly by within-slice structure or the reference. If everything collapses, section metadata deserve special emphasis.

Reproducibility materials include raw and processed matrices, images, masks, section manifest, alignment choices, transforms, spacing, reference profiles, gene lists, graph edges, model configuration, seeds, checkpoints, domains, proportions, denoised values, virtual slices, validation, software commit, hardware and checksums. Archive failed stacks.

Did you know? A beautiful 3D surface may contain mostly inference between thin measured planes. The surface can be scientifically useful, but only when readers can see where measurements stop and interpolation begins. Good visualisation shows uncertainty instead of polishing it away.

A bounded conclusion might say: ‘STitch3D reconstructed three-dimensional domains and cell-type distributions that were stable across registration, spacing, graph and reference checks, repeated in an independent specimen and agreed with orthogonal anatomy.’ It should not call virtual slices measured tissue, inferred proportions counted cells or a smooth volume continuous observation.

Use an evidence ladder from measured slice, to stable registration, to robust graph, to held-out molecular support, to independent 3D anatomy, to replicated specimen. The ladder turns a striking volume into a testable argument and shows exactly which new section or assay would strengthen the weakest link.

STitch3D should include a section-density experiment. Reconstruct the volume using all measured slices, then repeat after removing every second slice or omitting a difficult region. Compare domains, cell-type gradients and landmark positions. This estimates how much the claimed structure depends on sampling density and identifies the spacing at which interpolation begins to dominate observation.

Registration and deconvolution errors can reinforce each other. A misplaced slice may bring a region near the wrong neighbours, and the reference may then assign a plausible cell mixture that makes the placement look convincing. Review alignment with image or anatomical evidence that is independent of the reference expression profiles, and review cell-type estimates without using the reconstructed geometry as their only support.

Build a provenance layer into every 3D visual. A reader should be able to select a voxel, spot or surface patch and see its nearest measured slices, interpolation distance, contributing observations, reference support, local graph edges and uncertainty. The same rendering can then serve exploration and audit without making virtual space look like direct microscopy.

When comparing groups, the specimen—not the reconstructed voxel—is the unit of biological replication. Summarise volume, domain position and composition per specimen, preserve laterality and orientation, and use an analysis appropriate to the number of independent specimens. Millions of rendered points do not create millions of biological replicates.

End the project with a measurement-return plan. Identify the region where another section, targeted marker, protein stain or higher-resolution assay would most reduce uncertainty. State how that observation would change the current claim. A responsible 3D reconstruction is not merely a finished picture; it is an instrument for choosing the next experiment.

Keep the measured-plane labels visible in every export.

Contents · Previous section · Continue to the Science Learning Hub