WHY MATHEMATICS?

Why Mathematics? | SoptSC Similarity Matrices: Low-Rank Representation, NMF, Graph Laplacians, Shortest Paths and Minimum Spanning Trees

How SoptSC uses one learned cell-cell similarity structure to connect clustering, marker genes, pseudotime, lineage and pathway-specific communication without confusing inferred geometry with observed time.

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

Learn a sparse low-rank representation of cells from expression
Factor a symmetric nonnegative similarity matrix for clustering
Use graph spectra, shortest paths and spanning trees for structure
Combine ligand, receptor and target activity into communication probabilities

Chapter map

1. Expression matrix X
2. Self-representation
3. Coefficient matrix Z
4. Low-rank penalty
5. Sparse penalty
6. Symmetrization
7. Symmetric NMF
8. Membership weights
9. Consensus graph
10. Graph Laplacian
11. Eigengap
12. Similarity distance
13. Rooted shortest path
14. Cluster graph
15. Minimum spanning tree
16. Communication probability

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

1. Expression matrix X

Here the geometry becomes operational. Genes form rows and cells form columns.

A concrete situation

X has 10,000 genes and 2,000 cells.

What the mathematics does not license

Normalization changes geometry. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Freeze preprocessing. 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. Self-representation

This step deserves a careful pause. Each cell is reconstructed from other cells.

A concrete situation

Cell j equals a weighted combination of neighbours.

What the mathematics does not license

Linear combinations may miss nonlinear structure. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Inspect residuals. 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. Coefficient matrix Z

The useful question is not only how to calculate, but what the calculation assumes. Z stores cell-to-cell representation weights.

A concrete situation

Zij equals 0.4.

What the mathematics does not license

Coefficients are not biological transitions. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Name them representation weights. 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. Low-rank penalty

A small example makes the mechanism visible. A nuclear-norm-like constraint favours shared structure.

A concrete situation

Many singular values shrink.

What the mathematics does not license

Too much shrinkage merges states. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Sweep regularization. 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. Sparse penalty

This is where modelling choices begin to affect evidence. An L1-style term favours few neighbours.

A concrete situation

Most Zij become zero.

What the mathematics does not license

Sparsity can disconnect rare cells. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Audit graph components. 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. Symmetrization

The central mathematical idea is concrete. S takes the maximum of absolute Zij and Zji.

A concrete situation

0.4 and -0.2 become similarity 0.4.

What the mathematics does not license

Direction is discarded. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Keep Z for audit. 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. Symmetric NMF

Here the geometry becomes operational. S is approximated by H H transpose.

A concrete situation

H has k nonnegative columns.

What the mathematics does not license

Factorization can have local minima. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Use repeated starts. 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. Membership weights

This step deserves a careful pause. A row of H gives soft cluster affiliation.

A concrete situation

Cell i is 0.8 in cluster one.

What the mathematics does not license

Hard labels discard ambiguity. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Report entropy or margins. 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. Consensus graph

The useful question is not only how to calculate, but what the calculation assumes. Repeated factorization supports stable co-membership edges.

A concrete situation

Two cells co-cluster in 90% of runs.

What the mathematics does not license

Repetition shares the same data. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Do not call it external replication. 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. Graph Laplacian

A small example makes the mechanism visible. Degree minus adjacency encodes connectivity.

A concrete situation

L equals D minus A.

What the mathematics does not license

Thresholding changes the spectrum. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Sweep graph truncation. 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. Eigengap

This is where modelling choices begin to affect evidence. A jump between ordered eigenvalues suggests connected groups.

A concrete situation

lambda3 to lambda4 is large.

What the mathematics does not license

No universal gap guarantees k. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Compare biological plausibility. 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. Similarity distance

The central mathematical idea is concrete. Strong similarity becomes short edge length.

A concrete situation

Length equals one over weight.

What the mathematics does not license

The transform changes paths. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

State the formula. 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. Rooted shortest path

Here the geometry becomes operational. Distance from a chosen root orders cells.

A concrete situation

Cell B is path length 2.4 from root A.

What the mathematics does not license

Root choice sets direction. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Test alternative roots. 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. Cluster graph

This step deserves a careful pause. Cell edges aggregate into cluster connections.

A concrete situation

Cluster one connects to two and three.

What the mathematics does not license

Aggregation hides cell heterogeneity. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Retain edge distributions. 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. Minimum spanning tree

The useful question is not only how to calculate, but what the calculation assumes. A tree connects all clusters with minimum total weight.

A concrete situation

One branch point links three states.

What the mathematics does not license

Trees cannot express cycles. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Check alternative near-optimal trees. 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. Communication probability

A small example makes the mechanism visible. Ligand, receptor and signed target activity score ordered pairs.

A concrete situation

Sender L and receiver R plus target evidence yield 0.7.

What the mathematics does not license

Inputs are inferred and dependent. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Validate with perturbation and time. 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

SoptSC starts with a gene-by-cell matrix X and asks how well each cell can be represented as a linear combination of other cells. It learns a coefficient matrix Z under low-rank and sparse constraints, then symmetrizes it with S equals the elementwise maximum of the absolute values of Z and its transpose. S is a cell-cell similarity matrix: each edge weight records how strongly two cells help represent one another.

Clustering uses symmetric nonnegative matrix factorization, approximating S by H times H transpose. Each row of H gives a cell’s nonnegative membership weights across candidate clusters. The number of clusters can be informed by eigenvalues of a graph Laplacian derived from a truncated consensus matrix. Eigengaps summarize graph connectivity, but a clear gap is evidence about one representation rather than proof of a uniquely correct biological taxonomy.

The same similarity graph supports temporal reasoning. Edge lengths are derived from similarities, shortest-path distance from a chosen root orders cells along pseudotime, and a minimum spanning tree over cluster-level connections proposes lineage branches. Pseudotime is an inferred coordinate, not clock time. Root choice, graph sparsification and distance conversion can reverse or reroute interpretations, so alternative roots and graph settings must be tested.

SoptSC also builds pathway-specific communication probabilities from ligand expression in sender cells, receptor expression in receiver cells and signed target-gene activity. Aggregating single-cell probabilities creates cluster-to-cluster summaries that can be placed beside the inferred lineage to suggest feedback or feedforward patterns. The mathematics provides coherence across clustering, markers, paths and communication, but shared inputs also create dependence. External time-course or perturbation evidence is needed before calling a route regulatory control.

Frequently asked questions

What is SoptSC?

SoptSC is similarity matrix-based optimization for single-cell analysis, integrating clustering, marker genes, pseudotime, lineage and communication.

How is the similarity matrix built?

A sparse low-rank coefficient matrix represents each cell from other cells; absolute coefficients are symmetrized with an elementwise maximum.

What does symmetric NMF do?

It approximates the similarity matrix as H times H transpose with nonnegative H, producing soft cluster membership weights.

How can a graph Laplacian suggest cluster number?

Eigenvalue gaps summarize connected structure in a truncated consensus graph; a gap can support a candidate k under that representation.

Why use shortest paths and a minimum spanning tree?

Shortest paths from a root provide a pseudotime coordinate, while a spanning tree proposes a parsimonious connection among cluster states.

Does inferred pseudotime equal real time?

No. It is an ordering along a data-derived graph and needs experimental time-course or lineage evidence for temporal interpretation.

Primary source and fact-check note

Shuxiong Wang, Matthew Karikomi, Adam L. MacLean and Qing Nie, “Cell lineage and communication network inference via optimization for single-cell transcriptomics,” Nucleic Acids Research 47(11), e66 (2019), DOI 10.1093/nar/gkz204. 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

Matrix audit

Version X preprocessing and dimensions.

Optimization audit

Publish low-rank, sparse and convergence settings.

Factor audit

Use repeated NMF starts and keep soft memberships.

Spectrum audit

Show graph thresholds and eigenvalues.

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.

Path audit

Declare similarity-to-distance conversion and root.

Tree audit

Inspect near-optimal lineages and possible cycles.

Communication audit

Separate ligand, receptor and target contributions.

Final claim

A coherent inferred geometry still needs independent temporal and perturbation evidence.

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: Expression matrix X

Begin with this mechanism: Genes form rows and cells form columns. Recreate the concrete case: X has 10,000 genes and 2,000 cells. Write matrix dimensions, coefficients, singular values, sparsity, similarities, factor weights, Laplacian eigenvalues, path lengths, tree costs and communication terms with direction, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Normalization changes geometry. Complete the transfer step: Freeze preprocessing. Change exactly one reference profile, normalization, regularization value, graph radius, spatial edge, shuffle unit, pair sample, anchor, bin size, feature scale, random seed, matrix rank, similarity threshold, root cell or pathway target set. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep measurement, transformation, optimization, null comparison, graph result, biological interpretation and validation on separate notebook lines. Reserve whole donors, tissues or conditions outside tuning when testing transfer.

Workshop 2: Self-representation

Begin with this mechanism: Each cell is reconstructed from other cells. Recreate the concrete case: Cell j equals a weighted combination of neighbours. Write matrix dimensions, coefficients, singular values, sparsity, similarities, factor weights, Laplacian eigenvalues, path lengths, tree costs and communication terms with direction, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Linear combinations may miss nonlinear structure. Complete the transfer step: Inspect residuals. Change exactly one reference profile, normalization, regularization value, graph radius, spatial edge, shuffle unit, pair sample, anchor, bin size, feature scale, random seed, matrix rank, similarity threshold, root cell or pathway target set. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep measurement, transformation, optimization, null comparison, graph result, biological interpretation and validation on separate notebook lines. Reserve whole donors, tissues or conditions outside tuning when testing transfer.

Workshop 3: Coefficient matrix Z

Begin with this mechanism: Z stores cell-to-cell representation weights. Recreate the concrete case: Zij equals 0.4. Write matrix dimensions, coefficients, singular values, sparsity, similarities, factor weights, Laplacian eigenvalues, path lengths, tree costs and communication terms with direction, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Coefficients are not biological transitions. Complete the transfer step: Name them representation weights. Change exactly one reference profile, normalization, regularization value, graph radius, spatial edge, shuffle unit, pair sample, anchor, bin size, feature scale, random seed, matrix rank, similarity threshold, root cell or pathway target set. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep measurement, transformation, optimization, null comparison, graph result, biological interpretation and validation on separate notebook lines. Reserve whole donors, tissues or conditions outside tuning when testing transfer.

Workshop 4: Low-rank penalty

Begin with this mechanism: A nuclear-norm-like constraint favours shared structure. Recreate the concrete case: Many singular values shrink. Write matrix dimensions, coefficients, singular values, sparsity, similarities, factor weights, Laplacian eigenvalues, path lengths, tree costs and communication terms with direction, 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 shrinkage merges states. Complete the transfer step: Sweep regularization. Change exactly one reference profile, normalization, regularization value, graph radius, spatial edge, shuffle unit, pair sample, anchor, bin size, feature scale, random seed, matrix rank, similarity threshold, root cell or pathway target set. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep measurement, transformation, optimization, null comparison, graph result, biological interpretation and validation on separate notebook lines. Reserve whole donors, tissues or conditions 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: Sparse penalty

Begin with this mechanism: An L1-style term favours few neighbours. Recreate the concrete case: Most Zij become zero. Write matrix dimensions, coefficients, singular values, sparsity, similarities, factor weights, Laplacian eigenvalues, path lengths, tree costs and communication terms with direction, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Sparsity can disconnect rare cells. Complete the transfer step: Audit graph components. Change exactly one reference profile, normalization, regularization value, graph radius, spatial edge, shuffle unit, pair sample, anchor, bin size, feature scale, random seed, matrix rank, similarity threshold, root cell or pathway target set. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep measurement, transformation, optimization, null comparison, graph result, biological interpretation and validation on separate notebook lines. Reserve whole donors, tissues or conditions outside tuning when testing transfer.

Workshop 6: Symmetrization

Begin with this mechanism: S takes the maximum of absolute Zij and Zji. Recreate the concrete case: 0.4 and -0.2 become similarity 0.4. Write matrix dimensions, coefficients, singular values, sparsity, similarities, factor weights, Laplacian eigenvalues, path lengths, tree costs and communication terms with direction, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Direction is discarded. Complete the transfer step: Keep Z for audit. Change exactly one reference profile, normalization, regularization value, graph radius, spatial edge, shuffle unit, pair sample, anchor, bin size, feature scale, random seed, matrix rank, similarity threshold, root cell or pathway target set. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep measurement, transformation, optimization, null comparison, graph result, biological interpretation and validation on separate notebook lines. Reserve whole donors, tissues or conditions outside tuning when testing transfer.

Workshop 7: Symmetric NMF

Begin with this mechanism: S is approximated by H H transpose. Recreate the concrete case: H has k nonnegative columns. Write matrix dimensions, coefficients, singular values, sparsity, similarities, factor weights, Laplacian eigenvalues, path lengths, tree costs and communication terms with direction, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Factorization can have local minima. Complete the transfer step: Use repeated starts. Change exactly one reference profile, normalization, regularization value, graph radius, spatial edge, shuffle unit, pair sample, anchor, bin size, feature scale, random seed, matrix rank, similarity threshold, root cell or pathway target set. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep measurement, transformation, optimization, null comparison, graph result, biological interpretation and validation on separate notebook lines. Reserve whole donors, tissues or conditions outside tuning when testing transfer.

Workshop 8: Membership weights

Begin with this mechanism: A row of H gives soft cluster affiliation. Recreate the concrete case: Cell i is 0.8 in cluster one. Write matrix dimensions, coefficients, singular values, sparsity, similarities, factor weights, Laplacian eigenvalues, path lengths, tree costs and communication terms with direction, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Hard labels discard ambiguity. Complete the transfer step: Report entropy or margins. Change exactly one reference profile, normalization, regularization value, graph radius, spatial edge, shuffle unit, pair sample, anchor, bin size, feature scale, random seed, matrix rank, similarity threshold, root cell or pathway target set. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep measurement, transformation, optimization, null comparison, graph result, biological interpretation and validation on separate notebook lines. Reserve whole donors, tissues or conditions 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: Consensus graph

Begin with this mechanism: Repeated factorization supports stable co-membership edges. Recreate the concrete case: Two cells co-cluster in 90% of runs. Write matrix dimensions, coefficients, singular values, sparsity, similarities, factor weights, Laplacian eigenvalues, path lengths, tree costs and communication terms with direction, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Repetition shares the same data. Complete the transfer step: Do not call it external replication. Change exactly one reference profile, normalization, regularization value, graph radius, spatial edge, shuffle unit, pair sample, anchor, bin size, feature scale, random seed, matrix rank, similarity threshold, root cell or pathway target set. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep measurement, transformation, optimization, null comparison, graph result, biological interpretation and validation on separate notebook lines. Reserve whole donors, tissues or conditions outside tuning when testing transfer.

Workshop 10: Graph Laplacian

Begin with this mechanism: Degree minus adjacency encodes connectivity. Recreate the concrete case: L equals D minus A. Write matrix dimensions, coefficients, singular values, sparsity, similarities, factor weights, Laplacian eigenvalues, path lengths, tree costs and communication terms with direction, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Thresholding changes the spectrum. Complete the transfer step: Sweep graph truncation. Change exactly one reference profile, normalization, regularization value, graph radius, spatial edge, shuffle unit, pair sample, anchor, bin size, feature scale, random seed, matrix rank, similarity threshold, root cell or pathway target set. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep measurement, transformation, optimization, null comparison, graph result, biological interpretation and validation on separate notebook lines. Reserve whole donors, tissues or conditions outside tuning when testing transfer.

Workshop 11: Eigengap

Begin with this mechanism: A jump between ordered eigenvalues suggests connected groups. Recreate the concrete case: lambda3 to lambda4 is large. Write matrix dimensions, coefficients, singular values, sparsity, similarities, factor weights, Laplacian eigenvalues, path lengths, tree costs and communication terms with direction, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: No universal gap guarantees k. Complete the transfer step: Compare biological plausibility. Change exactly one reference profile, normalization, regularization value, graph radius, spatial edge, shuffle unit, pair sample, anchor, bin size, feature scale, random seed, matrix rank, similarity threshold, root cell or pathway target set. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep measurement, transformation, optimization, null comparison, graph result, biological interpretation and validation on separate notebook lines. Reserve whole donors, tissues or conditions outside tuning when testing transfer.

Workshop 12: Similarity distance

Begin with this mechanism: Strong similarity becomes short edge length. Recreate the concrete case: Length equals one over weight. Write matrix dimensions, coefficients, singular values, sparsity, similarities, factor weights, Laplacian eigenvalues, path lengths, tree costs and communication terms with direction, 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 transform changes paths. Complete the transfer step: State the formula. Change exactly one reference profile, normalization, regularization value, graph radius, spatial edge, shuffle unit, pair sample, anchor, bin size, feature scale, random seed, matrix rank, similarity threshold, root cell or pathway target set. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep measurement, transformation, optimization, null comparison, graph result, biological interpretation and validation on separate notebook lines. Reserve whole donors, tissues or conditions 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: Rooted shortest path

Begin with this mechanism: Distance from a chosen root orders cells. Recreate the concrete case: Cell B is path length 2.4 from root A. Write matrix dimensions, coefficients, singular values, sparsity, similarities, factor weights, Laplacian eigenvalues, path lengths, tree costs and communication terms with direction, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Root choice sets direction. Complete the transfer step: Test alternative roots. Change exactly one reference profile, normalization, regularization value, graph radius, spatial edge, shuffle unit, pair sample, anchor, bin size, feature scale, random seed, matrix rank, similarity threshold, root cell or pathway target set. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep measurement, transformation, optimization, null comparison, graph result, biological interpretation and validation on separate notebook lines. Reserve whole donors, tissues or conditions outside tuning when testing transfer.

Workshop 14: Cluster graph

Begin with this mechanism: Cell edges aggregate into cluster connections. Recreate the concrete case: Cluster one connects to two and three. Write matrix dimensions, coefficients, singular values, sparsity, similarities, factor weights, Laplacian eigenvalues, path lengths, tree costs and communication terms with direction, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Aggregation hides cell heterogeneity. Complete the transfer step: Retain edge distributions. Change exactly one reference profile, normalization, regularization value, graph radius, spatial edge, shuffle unit, pair sample, anchor, bin size, feature scale, random seed, matrix rank, similarity threshold, root cell or pathway target set. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep measurement, transformation, optimization, null comparison, graph result, biological interpretation and validation on separate notebook lines. Reserve whole donors, tissues or conditions outside tuning when testing transfer.

Workshop 15: Minimum spanning tree

Begin with this mechanism: A tree connects all clusters with minimum total weight. Recreate the concrete case: One branch point links three states. Write matrix dimensions, coefficients, singular values, sparsity, similarities, factor weights, Laplacian eigenvalues, path lengths, tree costs and communication terms with direction, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Trees cannot express cycles. Complete the transfer step: Check alternative near-optimal trees. Change exactly one reference profile, normalization, regularization value, graph radius, spatial edge, shuffle unit, pair sample, anchor, bin size, feature scale, random seed, matrix rank, similarity threshold, root cell or pathway target set. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep measurement, transformation, optimization, null comparison, graph result, biological interpretation and validation on separate notebook lines. Reserve whole donors, tissues or conditions outside tuning when testing transfer.

Workshop 16: Communication probability

Begin with this mechanism: Ligand, receptor and signed target activity score ordered pairs. Recreate the concrete case: Sender L and receiver R plus target evidence yield 0.7. Write matrix dimensions, coefficients, singular values, sparsity, similarities, factor weights, Laplacian eigenvalues, path lengths, tree costs and communication terms with direction, units, denominators and assumptions. Calculate the smallest non-trivial example by hand, then reproduce it in a spreadsheet or short program. Challenge this limitation: Inputs are inferred and dependent. Complete the transfer step: Validate with perturbation and time. Change exactly one reference profile, normalization, regularization value, graph radius, spatial edge, shuffle unit, pair sample, anchor, bin size, feature scale, random seed, matrix rank, similarity threshold, root cell or pathway target set. Explain which mathematical object changes, which conclusion survives and which claim must weaken. Keep measurement, transformation, optimization, null comparison, graph result, biological interpretation and validation on separate notebook lines. Reserve whole donors, tissues or conditions outside tuning when testing transfer.