WHY MATHEMATICS?

Why Mathematics? | SingleCellNet for Cross-Platform Classification: Gene-Pair Comparisons, Binary Features, Random Forests and Unknown Cells

How SingleCellNet replaces fragile expression magnitudes with within-cell gene-order comparisons, trains a random-forest ensemble and adds random profiles to resist coercing every query into a known type.

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

Select genes and ordered pairs that distinguish reference types
Turn every pair comparison into a binary feature
Train a multiclass random forest with an unknown category
Audit votes across platforms, species and absent cell types

Chapter map

1. Why order can outlast scale
2. Balanced reference sampling
3. Candidate genes
4. Ordered gene pairs
5. Binary transformation
6. Template matching for pairs
7. Randomised unknown profiles
8. Decision trees
9. Random forest ensemble
10. Class vote scores
11. Unknown-category decisions
12. Cross-platform transfer
13. Cross-species mapping
14. Held-out performance
15. A student forest project
16. Governance and careers

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

1. Why order can outlast scale

This step deserves a careful pause. A platform may change expression magnitudes while preserving whether one gene exceeds another within a cell.

A concrete situation

Gene A remains above gene B after rescaling.

What the mathematics does not license

Dropout can flip the order. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Test pair stability across platforms. 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. Balanced reference sampling

The useful question is not only how to calculate, but what the calculation assumes. Training cells are sampled by type so abundant labels do not dominate every tree.

A concrete situation

Each pancreas type contributes the same training count.

What the mathematics does not license

Aggressive balancing discards diversity. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Report sampled and available cells. 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. Candidate genes

A small example makes the mechanism visible. Genes enriched and under-expressed by type define a manageable search space for useful pairs.

A concrete situation

A beta-cell gene and an immune gene become candidates.

What the mathematics does not license

Differential expression can capture batch. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Select within biological folds. 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. Ordered gene pairs

This is where modelling choices begin to affect evidence. Every candidate pair asks a directional question, not a distance. Reversing the pair reverses the binary meaning.

A concrete situation

Feature AB equals one when A is greater than B.

What the mathematics does not license

Ties and missing genes need a declared rule. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Calculate ten cell-wise comparisons. 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. Binary transformation

The central mathematical idea is concrete. Selected pairs convert the expression matrix into zeros and ones.

A concrete situation

Thousands of counts become a compact pattern.

What the mathematics does not license

Binarisation removes magnitude information. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Compare rank robustness and information 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.

6. Template matching for pairs

Here the geometry becomes operational. Pair patterns are scored against ideal class templates to retain discriminating features.

A concrete situation

A pair is usually one for one type and zero elsewhere.

What the mathematics does not license

Many correlated pairs can duplicate evidence. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Measure pair redundancy and stability. 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. Randomised unknown profiles

This step deserves a careful pause. Random sampling of pair values creates profiles unlike the observed known classes and supplies an extra category.

A concrete situation

A scrambled pattern trains the forest to reject some queries.

What the mathematics does not license

Synthetic unknowns cannot represent every real novelty. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Challenge with withheld real types. 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. Decision trees

The useful question is not only how to calculate, but what the calculation assumes. Each tree repeatedly splits on binary gene-pair questions until leaves accumulate class evidence.

A concrete situation

One path asks AB, then CD, then EF.

What the mathematics does not license

Deep trees can learn reference noise. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Inspect depth and out-of-bag error. 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. Random forest ensemble

A small example makes the mechanism visible. Bootstrap samples and random feature subsets create diverse trees whose votes are combined.

A concrete situation

One thousand trees stabilise a class score in the documented implementation.

What the mathematics does not license

More trees do not fix biased labels. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Plot error and vote stability versus tree count. 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. Class vote scores

This is where modelling choices begin to affect evidence. The fraction of trees voting for each class forms a comparative score vector.

A concrete situation

A query receives 0.72 beta, 0.18 alpha, 0.10 random.

What the mathematics does not license

Vote fractions are not automatically calibrated probabilities. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Build reliability curves on held-out studies. 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. Unknown-category decisions

The central mathematical idea is concrete. A high random-class score or weak known scores can flag cells unsupported by the reference.

A concrete situation

A withheld ductal type avoids a forced endocrine label.

What the mathematics does not license

Novel states can still resemble known pair patterns. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Report rejection precision and recall. 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. Cross-platform transfer

Here the geometry becomes operational. Because features are within-cell comparisons, the same classifier can sometimes transfer across protocols.

A concrete situation

Droplet reference classifies plate-based queries.

What the mathematics does not license

Gene detection limits can change many pairs. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Stratify missing and flipped pairs. 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. Cross-species mapping

This step deserves a careful pause. Ortholog tables align genes before the same pair transform is applied.

A concrete situation

Mouse gene pairs classify a related human programme.

What the mathematics does not license

One-to-many orthology and species biology add uncertainty. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Publish the mapping and lost pairs. 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. Held-out performance

The useful question is not only how to calculate, but what the calculation assumes. Cohen’s kappa measures agreement beyond chance while class-wise precision–recall handles imbalance.

A concrete situation

Rare classes receive their own AUPR.

What the mathematics does not license

A single average hides failures. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Report every class and confidence interval. 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. A student forest project

A small example makes the mechanism visible. Convert ranked measurements into binary pairs, train a small forest and add random profiles.

A concrete situation

Known synthetic labels reveal pair and vote behaviour.

What the mathematics does not license

Real data add dropout, missing genes and species mapping. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Withhold one class as unknown. 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. Governance and careers

This is where modelling choices begin to affect evidence. Version genes, pairs, transforms, random profiles, forest seeds, votes and mappings. This joins combinatorics, trees and computational biology.

A concrete situation

Students can begin with binary comparisons.

What the mathematics does not license

A high vote proves no function or diagnosis. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Write a pair-feature model card. 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

A SingleCellNet audit begins with labelled reference cells and shared genes. It selects genes enriched or depleted by cell type, compares selected genes in pairs and converts every cell into binary questions such as “is gene A greater than gene B?” Template matching selects discriminating pairs, and a multiclass random forest learns the resulting binary patterns.

Record label balance, selected genes, ordered gene pairs, binary transforms, random-profile generation, forest settings, seeds, class vote scores, unknown-category scores and orthology mapping. Pairwise ranks can be more portable across platforms than raw expression magnitudes, but missing genes, changed biology and bad reference labels still matter.

Validate with donor-, study-, platform- and species-held-out data, per-class precision–recall, Cohen’s kappa, gene-pair stability, forest calibration and truly absent cell types. A forest vote is comparative evidence within the trained class system, not a direct measurement of identity or function.

Frequently asked questions

What is SingleCellNet?

It is a supervised single-cell classifier that converts selected within-cell gene comparisons into binary features and trains a multiclass random forest.

What is a top gene-pair feature?

For an ordered pair A and B, the feature is 1 when expression of A exceeds B in that cell and 0 otherwise.

Why can gene pairs help transfer?

Within-cell order is often less sensitive to platform scale than raw expression magnitude, though missing genes and biological shift still matter.

What does the random forest output?

Its trees vote across candidate classes, producing class scores or vote proportions for each query cell.

Why include random profiles?

Randomised binary profiles create an additional category that can absorb queries unlike the known training classes.

How should SingleCellNet be validated?

Use donor-, platform-, study- and species-held-out sets, per-class precision–recall, kappa, pair stability and absent-type challenges.

Primary source and fact-check note

Yuqi Tan and Patrick Cahan, “SingleCellNet: A Computational Tool to Classify Single Cell RNA-Seq Data Across Platforms and Across Species,” Cell Systems 9, 207–213.e2 (2019), DOI 10.1016/j.cels.2019.06.004. 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

Gene-pair provenance

Store each pair, direction, source genes, class template score and selection fold. Review whether a few batches or genes dominate the feature set.

Binary-flip audit

Measure how often every selected pair flips across technical replicates, platforms and normalisation choices. Rank robustness must be demonstrated, not assumed.

Unknown-profile audit

Record how random profiles were generated and compare them with real withheld classes. Synthetic randomness covers only part of biological novelty.

Forest stability

Repeat training with declared seeds. Compare selected pairs, out-of-bag error, per-class votes and feature importance across forests.

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.

Vote calibration

Relate class vote fractions to observed correctness in held-out studies. A 0.8 vote share need not mean 80% probability.

Cross-platform test

Train on one platform and test another without tuning on the final labels. Report missing genes, unusable pairs and performance by type.

Cross-species test

Publish ortholog rules, one-to-many choices and lost pairs. Validate species separately; conserved order does not guarantee conserved cell state.

Final claim

Name references, pairs, random class, forest settings, scores and held-out metrics. A quantitative comparison remains conditional on the training classes.

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: Why order can outlast scale

Start with the mechanism: A platform may change expression magnitudes while preserving whether one gene exceeds another within a cell. The worked case is concrete: Gene A remains above gene B after rescaling. Expose the boundary: Dropout can flip the order. Complete this evidence step: Test pair stability across platforms. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare correlations, binary pairs, margins, probabilities, paths or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 2: Balanced reference sampling

Start with the mechanism: Training cells are sampled by type so abundant labels do not dominate every tree. The worked case is concrete: Each pancreas type contributes the same training count. Expose the boundary: Aggressive balancing discards diversity. Complete this evidence step: Report sampled and available cells. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare correlations, binary pairs, margins, probabilities, paths or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 3: Candidate genes

Start with the mechanism: Genes enriched and under-expressed by type define a manageable search space for useful pairs. The worked case is concrete: A beta-cell gene and an immune gene become candidates. Expose the boundary: Differential expression can capture batch. Complete this evidence step: Select within biological folds. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare correlations, binary pairs, margins, probabilities, paths or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 4: Ordered gene pairs

Start with the mechanism: Every candidate pair asks a directional question, not a distance. Reversing the pair reverses the binary meaning. The worked case is concrete: Feature AB equals one when A is greater than B. Expose the boundary: Ties and missing genes need a declared rule. Complete this evidence step: Calculate ten cell-wise comparisons. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare correlations, binary pairs, margins, probabilities, paths or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

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: Binary transformation

Start with the mechanism: Selected pairs convert the expression matrix into zeros and ones. The worked case is concrete: Thousands of counts become a compact pattern. Expose the boundary: Binarisation removes magnitude information. Complete this evidence step: Compare rank robustness and information loss. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare correlations, binary pairs, margins, probabilities, paths or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 6: Template matching for pairs

Start with the mechanism: Pair patterns are scored against ideal class templates to retain discriminating features. The worked case is concrete: A pair is usually one for one type and zero elsewhere. Expose the boundary: Many correlated pairs can duplicate evidence. Complete this evidence step: Measure pair redundancy and stability. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare correlations, binary pairs, margins, probabilities, paths or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 7: Randomised unknown profiles

Start with the mechanism: Random sampling of pair values creates profiles unlike the observed known classes and supplies an extra category. The worked case is concrete: A scrambled pattern trains the forest to reject some queries. Expose the boundary: Synthetic unknowns cannot represent every real novelty. Complete this evidence step: Challenge with withheld real types. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare correlations, binary pairs, margins, probabilities, paths or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 8: Decision trees

Start with the mechanism: Each tree repeatedly splits on binary gene-pair questions until leaves accumulate class evidence. The worked case is concrete: One path asks AB, then CD, then EF. Expose the boundary: Deep trees can learn reference noise. Complete this evidence step: Inspect depth and out-of-bag error. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare correlations, binary pairs, margins, probabilities, paths or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

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: Random forest ensemble

Start with the mechanism: Bootstrap samples and random feature subsets create diverse trees whose votes are combined. The worked case is concrete: One thousand trees stabilise a class score in the documented implementation. Expose the boundary: More trees do not fix biased labels. Complete this evidence step: Plot error and vote stability versus tree count. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare correlations, binary pairs, margins, probabilities, paths or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 10: Class vote scores

Start with the mechanism: The fraction of trees voting for each class forms a comparative score vector. The worked case is concrete: A query receives 0.72 beta, 0.18 alpha, 0.10 random. Expose the boundary: Vote fractions are not automatically calibrated probabilities. Complete this evidence step: Build reliability curves on held-out studies. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare correlations, binary pairs, margins, probabilities, paths or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 11: Unknown-category decisions

Start with the mechanism: A high random-class score or weak known scores can flag cells unsupported by the reference. The worked case is concrete: A withheld ductal type avoids a forced endocrine label. Expose the boundary: Novel states can still resemble known pair patterns. Complete this evidence step: Report rejection precision and recall. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare correlations, binary pairs, margins, probabilities, paths or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 12: Cross-platform transfer

Start with the mechanism: Because features are within-cell comparisons, the same classifier can sometimes transfer across protocols. The worked case is concrete: Droplet reference classifies plate-based queries. Expose the boundary: Gene detection limits can change many pairs. Complete this evidence step: Stratify missing and flipped pairs. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare correlations, binary pairs, margins, probabilities, paths or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

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: Cross-species mapping

Start with the mechanism: Ortholog tables align genes before the same pair transform is applied. The worked case is concrete: Mouse gene pairs classify a related human programme. Expose the boundary: One-to-many orthology and species biology add uncertainty. Complete this evidence step: Publish the mapping and lost pairs. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare correlations, binary pairs, margins, probabilities, paths or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 14: Held-out performance

Start with the mechanism: Cohen’s kappa measures agreement beyond chance while class-wise precision–recall handles imbalance. The worked case is concrete: Rare classes receive their own AUPR. Expose the boundary: A single average hides failures. Complete this evidence step: Report every class and confidence interval. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare correlations, binary pairs, margins, probabilities, paths or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 15: A student forest project

Start with the mechanism: Convert ranked measurements into binary pairs, train a small forest and add random profiles. The worked case is concrete: Known synthetic labels reveal pair and vote behaviour. Expose the boundary: Real data add dropout, missing genes and species mapping. Complete this evidence step: Withhold one class as unknown. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare correlations, binary pairs, margins, probabilities, paths or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.

Workshop 16: Governance and careers

Start with the mechanism: Version genes, pairs, transforms, random profiles, forest seeds, votes and mappings. This joins combinatorics, trees and computational biology. The worked case is concrete: Students can begin with binary comparisons. Expose the boundary: A high vote proves no function or diagnosis. Complete this evidence step: Write a pair-feature model card. Write inputs, the mathematical operation, expected output, uncertainty and one observation that would contradict the claim. Repeat under a nearby defensible setting and compare correlations, binary pairs, margins, probabilities, paths or labels. Use untouched donors, studies or simulated truth because random cells can share hidden structure with training data. Keep measurement, model estimate, interpretation and decision on separate notebook lines so the conclusion remains auditable.