WHY MATHEMATICS?

Why Mathematics? | SEDR Spatial Representation: Masked Self-Supervision, Deep Autoencoders, Variational Graph Embeddings and Soft Clustering

How SEDR combines corrupted-expression recovery and probabilistic graph learning, then reconstructs tissue profiles without treating denoised predictions as measurements.

Why mathematics is important becomes easier to see when a real decision depends on it. This guide follows the mechanism from definitions and dimensions to a worked calculation, validation, practical learning and responsible use. It keeps factual evidence separate from interpretation and treats every illustrative scenario as an example rather than a promise of results.

Choose a reading route

Mask complete spot vectors to create a self-supervised task
Learn expression and spatial codes through separate encoders
Regularize probabilistic graph embeddings with KL divergence
Use soft clustering while checking self-reinforcing errors

Chapter map

1. Spot-by-gene matrix
2. Masked spot sampling
3. Learnable mask vector
4. Deep expression encoder
5. Expression decoder
6. Mean-squared reconstruction
7. Spatial adjacency
8. Graph convolution
9. Variational mean
10. Variational scale
11. Reparameterized sample
12. KL regularization
13. Latent concatenation
14. Soft assignment
15. Sharpened target distribution
16. Grouped validation

Visit the How Mathematics Works hub for the wider series and its connections to education, reasoning, science, computing, design and careers.

1. Spot-by-gene matrix

This is where modelling choices begin to affect evidence. Rows are tissue spots and columns are selected genes.

A concrete situation

X has n by m entries.

What the mathematics does not license

Filtering changes the learning target. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Publish the feature set. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

2. Masked spot sampling

The central mathematical idea is concrete. A random subset of spot rows is chosen for corruption.

A concrete situation

Fifteen percent of spots are masked.

What the mathematics does not license

Uniform masking may not mimic dropout. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Compare structured masks. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

3. Learnable mask vector

Here the geometry becomes operational. Chosen expression rows are replaced by one trainable token.

A concrete situation

xi becomes x-tilde.

What the mathematics does not license

One token may oversimplify missingness. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Audit recovery by tissue region. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

4. Deep expression encoder

This step deserves a careful pause. Nonlinear layers compress the masked matrix.

A concrete situation

Zf has thirty-two dimensions.

What the mathematics does not license

The encoder may learn batch. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Colour embeddings by donor. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

Did You Know? In this chapter, the most important number is useful only together with its construction rule. Changing scaling, sample membership or validation design can change the numerical answer without changing the arithmetic.

5. Expression decoder

The useful question is not only how to calculate, but what the calculation assumes. A fully connected or graph decoder reconstructs gene values.

A concrete situation

X-hat has the same n by m shape.

What the mathematics does not license

Reconstruction can oversmooth rare genes. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Report gene-wise errors. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

6. Mean-squared reconstruction

A small example makes the mechanism visible. Squared differences reward accurate recovery.

A concrete situation

Residuals 1 and 2 contribute 1 and 4.

What the mathematics does not license

Large-scale genes dominate. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Check scaling and robust loss. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

7. Spatial adjacency

This is where modelling choices begin to affect evidence. Coordinate neighbours form a graph.

A concrete situation

Each spot connects to six nearby spots.

What the mathematics does not license

Graph errors become representation errors. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Compare radii and kNN. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

8. Graph convolution

The central mathematical idea is concrete. Neighbour-normalized messages build spatial features.

A concrete situation

A node averages transformed local vectors.

What the mathematics does not license

Deep layers can erase boundaries. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Measure oversmoothing. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

9. Variational mean

Here the geometry becomes operational. The graph encoder outputs a latent mean for each spot.

A concrete situation

mu-g has sixteen values.

What the mathematics does not license

Mean alone hides uncertainty. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Keep variance outputs. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

10. Variational scale

This step deserves a careful pause. A second output defines spread of the approximate posterior.

A concrete situation

log sigma squared equals minus one.

What the mathematics does not license

Very small variance can signal posterior collapse. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Plot KL by spot. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

11. Reparameterized sample

The useful question is not only how to calculate, but what the calculation assumes. Noise is scaled and shifted so gradients pass through sampling.

A concrete situation

zg equals mu plus sigma times epsilon.

What the mathematics does not license

One sample adds stochasticity. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Repeat samples. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

12. KL regularization

A small example makes the mechanism visible. The approximate posterior is penalized for departing from its prior.

A concrete situation

KL equals 0.8 for one spot.

What the mathematics does not license

Too much KL can erase information. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Sweep its weight. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

13. Latent concatenation

This is where modelling choices begin to affect evidence. Expression and spatial codes are joined.

A concrete situation

Z contains Zf beside Zg.

What the mathematics does not license

One branch may dominate by scale. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Standardize and ablate. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

14. Soft assignment

The central mathematical idea is concrete. Distances to cluster centres become probabilities.

A concrete situation

A spot has probabilities 0.7, 0.2 and 0.1.

What the mathematics does not license

Uncertainty can be hidden by hard labels. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Map entropy. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

15. Sharpened target distribution

Here the geometry becomes operational. High-confidence assignments receive more weight during refinement.

A concrete situation

A 0.7 probability increases in the target.

What the mathematics does not license

Early errors can reinforce themselves. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Freeze an untouched evaluation set. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

16. Grouped validation

This step deserves a careful pause. Whole tissue sections remain outside tuning and imputation checks.

A concrete situation

One DLPFC section is held out.

What the mathematics does not license

Random spots leak local structure. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Use section and platform transfer. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

A seven-day practice lab

Day 1: define the decision, observation unit, X variables and Y or reference condition.
Day 2: reproduce the smallest worked example and check every dimension, unit and sign.
Day 3: build a transparent baseline and write what it assumes.
Day 4: fit the main method inside a training-only pipeline; record every parameter.
Day 5: challenge one assumption with a deliberately difficult dataset or operating regime.
Day 6: compare validation metrics with the practical workload created by errors.
Day 7: present one supported conclusion, one unresolved uncertainty and one safe next experiment.

Questions for a student, parent or mentor

Which part of the model is a mathematical definition, and which part is a choice?
What small calculation would catch a sign, scale or dimension error?
What evidence comes from training, validation and the untouched test?
What benign condition could look like a fault or prediction failure?
How would the answer change for another response, asset, school project or time period?
What should a human review before acting on the score?

A method-specific audit before deployment

SEDR joins two representation-learning routes. A deep autoencoder compresses the spot-by-gene matrix, while a variational graph autoencoder embeds the same masked features through a spatial adjacency graph. Their expression and spatial codes are concatenated. A decoder reconstructs expression, and the final representation supports clustering, visualization, trajectory inference, batch correction and cautious imputation.

The self-supervised task randomly selects spots and replaces each selected expression vector with a learnable mask vector. The network must recover information that is missing from the input by using regularities in the remaining expression and graph context. This is a pretext task: success shows that the model learned predictable structure, but it does not prove that every reconstructed zero is a biological transcript.

A variational graph autoencoder represents every spot with a distribution rather than a single fixed vector. Graph-convolution encoders produce parameters for a latent distribution; sampling is made trainable through reparameterization. Reconstruction rewards embeddings that explain the graph or expression structure, while a Kullback-Leibler term regularizes the approximate posterior toward its prior. Concatenation exposes both spatial and expression codes to the decoder.

SEDR can add an iterative deep-clustering objective that creates soft cluster probabilities and a sharpened target distribution. Reconstruction and clustering then influence one another. This can create compact groups, but an early mistaken assignment can become self-reinforcing. The responsible workflow preserves an unclustered embedding, runs ablations, varies the graph and masks, repeats seeds, withholds whole sections and labels denoised values as predictions.

Frequently asked questions

What is SEDR?

SEDR is an unsupervised method that combines a masked deep expression autoencoder with a variational graph autoencoder for spatial transcriptomic representation learning.

What is masked self-supervision?

Selected spot vectors are replaced by a learnable mask vector, and the network learns to reconstruct informative structure without external labels.

Why is the graph autoencoder variational?

It represents each spot with parameters of an approximate latent distribution and regularizes that distribution toward a prior.

How are expression and spatial information joined?

The expression code and graph-derived spatial code are concatenated into one latent vector used by the decoder.

What is soft clustering?

Each spot receives probabilities over clusters; an iterative target distribution can sharpen compactness.

Are reconstructed counts observations?

No. They are model outputs whose usefulness must be checked on hidden genes, sections and independent assays.

Primary source and fact-check note

Hang Xu, Huazhu Fu, Yahui Long et al., “Unsupervised spatially embedded deep representation of spatial transcriptomics,” Genome Medicine 16, 12 (2024), DOI 10.1186/s13073-024-01283-x. Definitions and core claims in this article were checked against this primary publication. Calculations shown here were independently recomputed. Examples are educational illustrations; they are not reported industrial performance results.

Useful next reading

Keep the options open. Mathematics education supports clearer questions, transferable problem-solving skills and informed choices across science, engineering, computing, AI, statistics, design and many other careers. It does not guarantee intelligence, income, admission or employment; those outcomes depend on far more than one subject or model.

Extended laboratory: topic-specific validation and responsible use

Mask audit

Publish masked fractions, patterns and mask-token behavior.

Expression audit

Keep encoder dimensions and reconstruction residuals.

Graph audit

Record adjacency and graph-convolution settings.

Variational audit

Inspect means, variances, KL scale and posterior collapse.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Fusion audit

Compare branch scales and single-branch ablations.

Clustering audit

Map soft probabilities and entropy before hard labels.

Transfer audit

Withhold sections, donors and platforms.

Final claim

Denoised values remain predictions under the model.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Mechanism workshop: calculate, challenge and transfer

Workshop 1: Spot-by-gene matrix

Begin with this mechanism: Rows are tissue spots and columns are selected genes. Recreate the concrete case: X has n by m entries. Write mask indicators, mask-token values, expression codes, graph degrees, variational means, log variances, noise samples, KL terms, reconstructed genes, soft probabilities, sharpened targets and grouped errors with direction, dimensions, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Filtering changes the learning target. Complete the transfer step: Publish the feature set. Change exactly one gene filter, PC dimension, cluster count, prior, gamma, degrees of freedom, lattice edge, initialization, chain length, burn-in, jitter scale, subspot layout, radius, graph degree, attention score, alpha mixture, latent width, tied parameter, mask rate, mask pattern, KL weight, cluster target, corruption rule, negative sample, pairwise distance normalization, Leiden resolution, pseudotime root, tissue section or random seed. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep observation, preprocessing, graph, prior, objective, optimization, posterior or embedding, clustering, biological interpretation and independent validation on separate notebook lines. Reserve whole donors, slides, tissues, patients or platforms outside tuning when testing transfer.

Workshop 2: Masked spot sampling

Begin with this mechanism: A random subset of spot rows is chosen for corruption. Recreate the concrete case: Fifteen percent of spots are masked. Write mask indicators, mask-token values, expression codes, graph degrees, variational means, log variances, noise samples, KL terms, reconstructed genes, soft probabilities, sharpened targets and grouped errors with direction, dimensions, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Uniform masking may not mimic dropout. Complete the transfer step: Compare structured masks. Change exactly one gene filter, PC dimension, cluster count, prior, gamma, degrees of freedom, lattice edge, initialization, chain length, burn-in, jitter scale, subspot layout, radius, graph degree, attention score, alpha mixture, latent width, tied parameter, mask rate, mask pattern, KL weight, cluster target, corruption rule, negative sample, pairwise distance normalization, Leiden resolution, pseudotime root, tissue section or random seed. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep observation, preprocessing, graph, prior, objective, optimization, posterior or embedding, clustering, biological interpretation and independent validation on separate notebook lines. Reserve whole donors, slides, tissues, patients or platforms outside tuning when testing transfer.

Workshop 3: Learnable mask vector

Begin with this mechanism: Chosen expression rows are replaced by one trainable token. Recreate the concrete case: xi becomes x-tilde. Write mask indicators, mask-token values, expression codes, graph degrees, variational means, log variances, noise samples, KL terms, reconstructed genes, soft probabilities, sharpened targets and grouped errors with direction, dimensions, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: One token may oversimplify missingness. Complete the transfer step: Audit recovery by tissue region. Change exactly one gene filter, PC dimension, cluster count, prior, gamma, degrees of freedom, lattice edge, initialization, chain length, burn-in, jitter scale, subspot layout, radius, graph degree, attention score, alpha mixture, latent width, tied parameter, mask rate, mask pattern, KL weight, cluster target, corruption rule, negative sample, pairwise distance normalization, Leiden resolution, pseudotime root, tissue section or random seed. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep observation, preprocessing, graph, prior, objective, optimization, posterior or embedding, clustering, biological interpretation and independent validation on separate notebook lines. Reserve whole donors, slides, tissues, patients or platforms outside tuning when testing transfer.

Workshop 4: Deep expression encoder

Begin with this mechanism: Nonlinear layers compress the masked matrix. Recreate the concrete case: Zf has thirty-two dimensions. Write mask indicators, mask-token values, expression codes, graph degrees, variational means, log variances, noise samples, KL terms, reconstructed genes, soft probabilities, sharpened targets and grouped errors with direction, dimensions, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: The encoder may learn batch. Complete the transfer step: Colour embeddings by donor. Change exactly one gene filter, PC dimension, cluster count, prior, gamma, degrees of freedom, lattice edge, initialization, chain length, burn-in, jitter scale, subspot layout, radius, graph degree, attention score, alpha mixture, latent width, tied parameter, mask rate, mask pattern, KL weight, cluster target, corruption rule, negative sample, pairwise distance normalization, Leiden resolution, pseudotime root, tissue section or random seed. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep observation, preprocessing, graph, prior, objective, optimization, posterior or embedding, clustering, biological interpretation and independent validation on separate notebook lines. Reserve whole donors, slides, tissues, patients or platforms outside tuning when testing transfer.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Workshop 5: Expression decoder

Begin with this mechanism: A fully connected or graph decoder reconstructs gene values. Recreate the concrete case: X-hat has the same n by m shape. Write mask indicators, mask-token values, expression codes, graph degrees, variational means, log variances, noise samples, KL terms, reconstructed genes, soft probabilities, sharpened targets and grouped errors with direction, dimensions, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Reconstruction can oversmooth rare genes. Complete the transfer step: Report gene-wise errors. Change exactly one gene filter, PC dimension, cluster count, prior, gamma, degrees of freedom, lattice edge, initialization, chain length, burn-in, jitter scale, subspot layout, radius, graph degree, attention score, alpha mixture, latent width, tied parameter, mask rate, mask pattern, KL weight, cluster target, corruption rule, negative sample, pairwise distance normalization, Leiden resolution, pseudotime root, tissue section or random seed. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep observation, preprocessing, graph, prior, objective, optimization, posterior or embedding, clustering, biological interpretation and independent validation on separate notebook lines. Reserve whole donors, slides, tissues, patients or platforms outside tuning when testing transfer.

Workshop 6: Mean-squared reconstruction

Begin with this mechanism: Squared differences reward accurate recovery. Recreate the concrete case: Residuals 1 and 2 contribute 1 and 4. Write mask indicators, mask-token values, expression codes, graph degrees, variational means, log variances, noise samples, KL terms, reconstructed genes, soft probabilities, sharpened targets and grouped errors with direction, dimensions, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Large-scale genes dominate. Complete the transfer step: Check scaling and robust loss. Change exactly one gene filter, PC dimension, cluster count, prior, gamma, degrees of freedom, lattice edge, initialization, chain length, burn-in, jitter scale, subspot layout, radius, graph degree, attention score, alpha mixture, latent width, tied parameter, mask rate, mask pattern, KL weight, cluster target, corruption rule, negative sample, pairwise distance normalization, Leiden resolution, pseudotime root, tissue section or random seed. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep observation, preprocessing, graph, prior, objective, optimization, posterior or embedding, clustering, biological interpretation and independent validation on separate notebook lines. Reserve whole donors, slides, tissues, patients or platforms outside tuning when testing transfer.

Workshop 7: Spatial adjacency

Begin with this mechanism: Coordinate neighbours form a graph. Recreate the concrete case: Each spot connects to six nearby spots. Write mask indicators, mask-token values, expression codes, graph degrees, variational means, log variances, noise samples, KL terms, reconstructed genes, soft probabilities, sharpened targets and grouped errors with direction, dimensions, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Graph errors become representation errors. Complete the transfer step: Compare radii and kNN. Change exactly one gene filter, PC dimension, cluster count, prior, gamma, degrees of freedom, lattice edge, initialization, chain length, burn-in, jitter scale, subspot layout, radius, graph degree, attention score, alpha mixture, latent width, tied parameter, mask rate, mask pattern, KL weight, cluster target, corruption rule, negative sample, pairwise distance normalization, Leiden resolution, pseudotime root, tissue section or random seed. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep observation, preprocessing, graph, prior, objective, optimization, posterior or embedding, clustering, biological interpretation and independent validation on separate notebook lines. Reserve whole donors, slides, tissues, patients or platforms outside tuning when testing transfer.

Workshop 8: Graph convolution

Begin with this mechanism: Neighbour-normalized messages build spatial features. Recreate the concrete case: A node averages transformed local vectors. Write mask indicators, mask-token values, expression codes, graph degrees, variational means, log variances, noise samples, KL terms, reconstructed genes, soft probabilities, sharpened targets and grouped errors with direction, dimensions, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Deep layers can erase boundaries. Complete the transfer step: Measure oversmoothing. Change exactly one gene filter, PC dimension, cluster count, prior, gamma, degrees of freedom, lattice edge, initialization, chain length, burn-in, jitter scale, subspot layout, radius, graph degree, attention score, alpha mixture, latent width, tied parameter, mask rate, mask pattern, KL weight, cluster target, corruption rule, negative sample, pairwise distance normalization, Leiden resolution, pseudotime root, tissue section or random seed. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep observation, preprocessing, graph, prior, objective, optimization, posterior or embedding, clustering, biological interpretation and independent validation on separate notebook lines. Reserve whole donors, slides, tissues, patients or platforms outside tuning when testing transfer.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Workshop 9: Variational mean

Begin with this mechanism: The graph encoder outputs a latent mean for each spot. Recreate the concrete case: mu-g has sixteen values. Write mask indicators, mask-token values, expression codes, graph degrees, variational means, log variances, noise samples, KL terms, reconstructed genes, soft probabilities, sharpened targets and grouped errors with direction, dimensions, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Mean alone hides uncertainty. Complete the transfer step: Keep variance outputs. Change exactly one gene filter, PC dimension, cluster count, prior, gamma, degrees of freedom, lattice edge, initialization, chain length, burn-in, jitter scale, subspot layout, radius, graph degree, attention score, alpha mixture, latent width, tied parameter, mask rate, mask pattern, KL weight, cluster target, corruption rule, negative sample, pairwise distance normalization, Leiden resolution, pseudotime root, tissue section or random seed. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep observation, preprocessing, graph, prior, objective, optimization, posterior or embedding, clustering, biological interpretation and independent validation on separate notebook lines. Reserve whole donors, slides, tissues, patients or platforms outside tuning when testing transfer.

Workshop 10: Variational scale

Begin with this mechanism: A second output defines spread of the approximate posterior. Recreate the concrete case: log sigma squared equals minus one. Write mask indicators, mask-token values, expression codes, graph degrees, variational means, log variances, noise samples, KL terms, reconstructed genes, soft probabilities, sharpened targets and grouped errors with direction, dimensions, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Very small variance can signal posterior collapse. Complete the transfer step: Plot KL by spot. Change exactly one gene filter, PC dimension, cluster count, prior, gamma, degrees of freedom, lattice edge, initialization, chain length, burn-in, jitter scale, subspot layout, radius, graph degree, attention score, alpha mixture, latent width, tied parameter, mask rate, mask pattern, KL weight, cluster target, corruption rule, negative sample, pairwise distance normalization, Leiden resolution, pseudotime root, tissue section or random seed. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep observation, preprocessing, graph, prior, objective, optimization, posterior or embedding, clustering, biological interpretation and independent validation on separate notebook lines. Reserve whole donors, slides, tissues, patients or platforms outside tuning when testing transfer.

Workshop 11: Reparameterized sample

Begin with this mechanism: Noise is scaled and shifted so gradients pass through sampling. Recreate the concrete case: zg equals mu plus sigma times epsilon. Write mask indicators, mask-token values, expression codes, graph degrees, variational means, log variances, noise samples, KL terms, reconstructed genes, soft probabilities, sharpened targets and grouped errors with direction, dimensions, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: One sample adds stochasticity. Complete the transfer step: Repeat samples. Change exactly one gene filter, PC dimension, cluster count, prior, gamma, degrees of freedom, lattice edge, initialization, chain length, burn-in, jitter scale, subspot layout, radius, graph degree, attention score, alpha mixture, latent width, tied parameter, mask rate, mask pattern, KL weight, cluster target, corruption rule, negative sample, pairwise distance normalization, Leiden resolution, pseudotime root, tissue section or random seed. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep observation, preprocessing, graph, prior, objective, optimization, posterior or embedding, clustering, biological interpretation and independent validation on separate notebook lines. Reserve whole donors, slides, tissues, patients or platforms outside tuning when testing transfer.

Workshop 12: KL regularization

Begin with this mechanism: The approximate posterior is penalized for departing from its prior. Recreate the concrete case: KL equals 0.8 for one spot. Write mask indicators, mask-token values, expression codes, graph degrees, variational means, log variances, noise samples, KL terms, reconstructed genes, soft probabilities, sharpened targets and grouped errors with direction, dimensions, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Too much KL can erase information. Complete the transfer step: Sweep its weight. Change exactly one gene filter, PC dimension, cluster count, prior, gamma, degrees of freedom, lattice edge, initialization, chain length, burn-in, jitter scale, subspot layout, radius, graph degree, attention score, alpha mixture, latent width, tied parameter, mask rate, mask pattern, KL weight, cluster target, corruption rule, negative sample, pairwise distance normalization, Leiden resolution, pseudotime root, tissue section or random seed. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep observation, preprocessing, graph, prior, objective, optimization, posterior or embedding, clustering, biological interpretation and independent validation on separate notebook lines. Reserve whole donors, slides, tissues, patients or platforms outside tuning when testing transfer.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Workshop 13: Latent concatenation

Begin with this mechanism: Expression and spatial codes are joined. Recreate the concrete case: Z contains Zf beside Zg. Write mask indicators, mask-token values, expression codes, graph degrees, variational means, log variances, noise samples, KL terms, reconstructed genes, soft probabilities, sharpened targets and grouped errors with direction, dimensions, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: One branch may dominate by scale. Complete the transfer step: Standardize and ablate. Change exactly one gene filter, PC dimension, cluster count, prior, gamma, degrees of freedom, lattice edge, initialization, chain length, burn-in, jitter scale, subspot layout, radius, graph degree, attention score, alpha mixture, latent width, tied parameter, mask rate, mask pattern, KL weight, cluster target, corruption rule, negative sample, pairwise distance normalization, Leiden resolution, pseudotime root, tissue section or random seed. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep observation, preprocessing, graph, prior, objective, optimization, posterior or embedding, clustering, biological interpretation and independent validation on separate notebook lines. Reserve whole donors, slides, tissues, patients or platforms outside tuning when testing transfer.

Workshop 14: Soft assignment

Begin with this mechanism: Distances to cluster centres become probabilities. Recreate the concrete case: A spot has probabilities 0.7, 0.2 and 0.1. Write mask indicators, mask-token values, expression codes, graph degrees, variational means, log variances, noise samples, KL terms, reconstructed genes, soft probabilities, sharpened targets and grouped errors with direction, dimensions, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Uncertainty can be hidden by hard labels. Complete the transfer step: Map entropy. Change exactly one gene filter, PC dimension, cluster count, prior, gamma, degrees of freedom, lattice edge, initialization, chain length, burn-in, jitter scale, subspot layout, radius, graph degree, attention score, alpha mixture, latent width, tied parameter, mask rate, mask pattern, KL weight, cluster target, corruption rule, negative sample, pairwise distance normalization, Leiden resolution, pseudotime root, tissue section or random seed. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep observation, preprocessing, graph, prior, objective, optimization, posterior or embedding, clustering, biological interpretation and independent validation on separate notebook lines. Reserve whole donors, slides, tissues, patients or platforms outside tuning when testing transfer.

Workshop 15: Sharpened target distribution

Begin with this mechanism: High-confidence assignments receive more weight during refinement. Recreate the concrete case: A 0.7 probability increases in the target. Write mask indicators, mask-token values, expression codes, graph degrees, variational means, log variances, noise samples, KL terms, reconstructed genes, soft probabilities, sharpened targets and grouped errors with direction, dimensions, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Early errors can reinforce themselves. Complete the transfer step: Freeze an untouched evaluation set. Change exactly one gene filter, PC dimension, cluster count, prior, gamma, degrees of freedom, lattice edge, initialization, chain length, burn-in, jitter scale, subspot layout, radius, graph degree, attention score, alpha mixture, latent width, tied parameter, mask rate, mask pattern, KL weight, cluster target, corruption rule, negative sample, pairwise distance normalization, Leiden resolution, pseudotime root, tissue section or random seed. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep observation, preprocessing, graph, prior, objective, optimization, posterior or embedding, clustering, biological interpretation and independent validation on separate notebook lines. Reserve whole donors, slides, tissues, patients or platforms outside tuning when testing transfer.

Workshop 16: Grouped validation

Begin with this mechanism: Whole tissue sections remain outside tuning and imputation checks. Recreate the concrete case: One DLPFC section is held out. Write mask indicators, mask-token values, expression codes, graph degrees, variational means, log variances, noise samples, KL terms, reconstructed genes, soft probabilities, sharpened targets and grouped errors with direction, dimensions, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Random spots leak local structure. Complete the transfer step: Use section and platform transfer. Change exactly one gene filter, PC dimension, cluster count, prior, gamma, degrees of freedom, lattice edge, initialization, chain length, burn-in, jitter scale, subspot layout, radius, graph degree, attention score, alpha mixture, latent width, tied parameter, mask rate, mask pattern, KL weight, cluster target, corruption rule, negative sample, pairwise distance normalization, Leiden resolution, pseudotime root, tissue section or random seed. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep observation, preprocessing, graph, prior, objective, optimization, posterior or embedding, clustering, biological interpretation and independent validation on separate notebook lines. Reserve whole donors, slides, tissues, patients or platforms outside tuning when testing transfer.