WHY MATHEMATICS?

Why Mathematics? | UCE for Universal Cell Embeddings: Expression-Weighted Gene Sampling, Protein Language Vectors, Chromosome Positions and Zero-Shot Transfer

How Universal Cell Embedding turns sparse RNA measurements into a reusable cross-species coordinate system by sampling genes according to expression, describing their protein products with sequence vectors, adding genomic positions and learning cell context with a large Transformer.

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

Convert expression counts into transparent sampling probabilities
Represent genes through protein sequence and genomic position
Build a contextual cell embedding with Transformer attention
Test zero-shot transfer, sampling variation and biological limits

Chapter map

1. From counts to a probability distribution
2. Expression-weighted sampling
3. Token multiplicity as abundance
4. Protein language vector
5. Unseen-gene construction
6. Chromosome metadata
7. Species and cell context
8. Transformer input sequence
9. Scaled self-attention
10. Many layers and parameters
11. Cell-vector pooling
12. Distances in universal space
13. Zero-shot label transfer
14. Worked six-gene sample
15. Cross-species validation
16. Governance and learning

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

1. From counts to a probability distribution

A small example makes the mechanism visible. For every expressed gene compute log(x_i + 1), then divide by the sum of those log values to obtain sampling probability p_i.

A concrete situation

Counts 6, 3 and 1 become weights log 7, log 4 and log 2 before normalization.

What the mathematics does not license

Library preparation and dropout affect the distribution. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Check that probabilities sum to one. 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. Expression-weighted sampling

This is where modelling choices begin to affect evidence. Draw gene tokens from the categorical distribution so abundance influences how often each gene appears.

A concrete situation

In ten draws the first gene may appear six times.

What the mathematics does not license

A finite sample has Monte Carlo noise. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Repeat seeds and report variation. 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. Token multiplicity as abundance

The central mathematical idea is concrete. Repeated appearances allow the sequence to carry quantitative information without attaching a raw count to each token.

A concrete situation

A highly expressed ribosomal gene can occupy several positions.

What the mathematics does not license

Common genes can crowd out rare signals. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Measure inclusion probability by abundance. 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. Protein language vector

Here the geometry becomes operational. A pretrained protein model converts the amino-acid sequence of each gene product into a numeric embedding.

A concrete situation

A 300-residue protein becomes one fixed-length vector.

What the mathematics does not license

Isoforms and non-protein-coding genes need explicit handling. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Record protein accession and pooling. 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. Unseen-gene construction

This step deserves a careful pause. Because the starting vector is computed from sequence, a new protein-coding gene can enter without retraining a lookup table.

A concrete situation

A gene from a new species receives an ESM-derived vector.

What the mathematics does not license

Sequence similarity does not capture every regulatory role. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Compare homologous and divergent genes. 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. Chromosome metadata

The useful question is not only how to calculate, but what the calculation assumes. Genomic location is encoded alongside sequence-derived information to distinguish positional context.

A concrete situation

Two proteins with similar sequence occur on different chromosomes.

What the mathematics does not license

Genome builds and annotations change coordinates. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Version assembly and gene annotation. 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. Species and cell context

A small example makes the mechanism visible. Metadata and the collection of sampled genes identify the biological context in which a gene vector is interpreted.

A concrete situation

The same conserved protein appears in human and mouse cells.

What the mathematics does not license

Species tokens can also reveal dataset shortcuts. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Probe species predictability after alignment. 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. Transformer input sequence

This is where modelling choices begin to affect evidence. The 1,024 sampled genes are sorted by genomic position within chromosome groups, separated by species-specific chromosome start and end tokens, and processed by stacked attention and feed-forward blocks.

A concrete situation

A CLS token begins one cell sentence.

What the mathematics does not license

Sampling and the finite sequence change which evidence enters. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Publish sequence length, seed and retained mass. 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. Scaled self-attention

The central mathematical idea is concrete. Query–key dot products divided by the square root of dimension become softmax weights over value vectors.

A concrete situation

A transcription factor token attends to a pathway partner.

What the mathematics does not license

Attention is computation, not a regulatory edge. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Compare weights with perturbation evidence. 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. Many layers and parameters

Here the geometry becomes operational. The reported UCE architecture uses a deep Transformer with roughly 650 million parameters to refine contextual gene representations.

A concrete situation

Later layers integrate wider gene context.

What the mathematics does not license

Capacity can memorize corpus structure. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Use species and study holdouts. 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. Cell-vector pooling

This step deserves a careful pause. The final hidden state of the special CLS token, decoded through an MLP, supplies the reported 1,280-dimensional cell embedding.

A concrete situation

A cell becomes one point in the universal space.

What the mathematics does not license

A single vector can hide conflicting programmes. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Compare CLS geometry with token-level evidence. 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. Distances in universal space

The useful question is not only how to calculate, but what the calculation assumes. Cosine similarity or Euclidean distance quantifies neighbourhoods among cell vectors.

A concrete situation

A new fish immune cell lies near related vertebrate cells.

What the mathematics does not license

Distance depends on checkpoint and preprocessing. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Report metric and neighbour 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.

13. Zero-shot label transfer

A small example makes the mechanism visible. Reference neighbours can supply candidate labels without updating model weights.

A concrete situation

A new-species cell inherits a broad immune label.

What the mathematics does not license

Reference coverage constrains the vocabulary. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Allow rejection and broad labels. 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. Worked six-gene sample

This is where modelling choices begin to affect evidence. Normalize six counts, calculate draw probabilities, perform twenty seeded draws and compare two pooled vectors.

A concrete situation

The most abundant gene appears often but not exactly at expectation.

What the mathematics does not license

One random draw is unstable. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Compute confidence intervals over seeds. 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. Cross-species validation

The central mathematical idea is concrete. Keep complete species and tissues outside training and compare with orthologue-based and count-based baselines.

A concrete situation

A held-out organism tests genuine zero-shot transfer.

What the mathematics does not license

Shared technologies can still leak structure. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Stratify by assay and study. 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 learning

Here the geometry becomes operational. Version counts, probabilities, draws, sequences, coordinates, checkpoint and distances. This joins probability, genomics and AI.

A concrete situation

Students can start with categorical sampling.

What the mathematics does not license

A universal vector remains a model estimate. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Write a UCE 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 UCE audit begins with the expression vector but does not end there. Expression-weighted sampling chooses gene tokens, a protein language model supplies sequence-derived gene vectors, chromosome metadata locates genes in the genome, and a large Transformer turns the sampled list into a cell embedding. This combination is designed to let a previously unseen protein-coding gene receive a vector from its amino-acid sequence rather than requiring a learned gene-ID lookup.

Record the gene identifiers, expression units, sampling probabilities, random seed, number of sampled tokens, protein sequence release, protein-model checkpoint, chromosome coordinates, species metadata, Transformer checkpoint and pooling rule. Sampling converts abundance into a probability distribution, so repeated runs can differ even for the same cell unless the seed and draw count are fixed.

Validate zero-shot transfer with species, tissues and technologies kept completely outside training; compare sampling with deterministic rankings and count-based baselines; and report uncertainty over repeated draws. A universal coordinate system is an ambitious statistical representation, not proof that every species, rare state or regulatory mechanism has been captured.

During pretraining, UCE masks some expressed genes from the sampled cell sentence. It then pairs the resulting CLS cell vector with compressed protein embeddings for balanced sets of expressed and unexpressed query genes. A multilayer perceptron predicts whether each query gene was expressed, and binary cross-entropy supplies the training loss. This objective learns cell context without requiring a cell-type label.

Frequently asked questions

What is UCE?

UCE, Universal Cell Embedding, is a foundation model that represents single cells from different species in one learned vector space. The Nature study reports pretraining on a 36-million-cell, eight-species atlas and zero-shot use on new datasets.

Why sample genes instead of sorting them once?

UCE samples 1,024 expressed genes with replacement using probabilities proportional to log(count + 1). Highly expressed genes are more likely to appear, yet lower-expression genes can still enter, and repeated draws preserve abundance information statistically.

How can UCE represent a gene it never saw in training?

It derives the starting gene vector from the amino-acid sequence of the encoded protein using a pretrained protein language model. A new protein-coding gene can therefore receive a sequence-based representation without a new learned ID row.

Why include chromosome position?

Chromosome and genomic-location metadata distinguish where genes occur and add structural context that protein sequence alone does not contain.

What does zero-shot mean here?

It means applying the pretrained model to new cells, genes or species without task-specific parameter updates. Zero-shot is an evaluation condition, not a promise of equal accuracy everywhere.

Does a universal embedding prove evolutionary or regulatory equivalence?

No. Nearby vectors show similarity under the model and its training objective. Evolutionary homology, regulation and causal function require independent biological evidence.

Primary source and fact-check note

Yanay Rosen, Yusuf Roohani, Ayush Agrawal et al., “Universal cell embedding provides a foundation model for cell biology,” Nature 656, 183–191 (2026), DOI 10.1038/s41586-026-10689-z. 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

Sampling audit

Store probabilities, number of draws and seeds; repeat embeddings to quantify Monte Carlo variation.

Protein-reference audit

Version protein sequences, isoform rules, protein-language checkpoint and pooling for every gene.

Genome-position audit

Record assembly, chromosome labels and coordinate encoding; test genes with revised annotations.

Unseen-species firewall

Hold whole species and tissues out of all tuning, threshold selection and model choice.

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.

Deterministic comparison

Compare weighted sampling with ranked tokens and count-based baselines under identical splits.

Neighbour stability

Measure how nearest cells change across seeds, metrics, reference composition and preprocessing.

Biological validation

Check zero-shot labels with markers, independent assays and perturbations rather than geometry alone.

Final claim

Name sampling, sequence model, genome build and checkpoint. Universal means reusable representation, not universal truth.

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: From counts to a probability distribution

Start with this article's mechanism: For every expressed gene compute log(x_i + 1), then divide by the sum of those log values to obtain sampling probability p_i. Work through the concrete case: Counts 6, 3 and 1 become weights log 7, log 4 and log 2 before normalization. Make the boundary visible: Library preparation and dropout affect the distribution. Complete the evidence step: Check that probabilities sum to one. Write the input table or token list, transformation, output, units, uncertainty and one observation that would contradict the conclusion. Repeat the calculation under a nearby defensible choice and compare sampling distributions, prior embeddings, similarities, attention weights, masks, losses, distances, neighbourhoods, proportions, logits or calibration as relevant. Keep complete donors, studies, tissues, species or technologies outside training when the claim concerns transfer. Random cells from the same biological source are not an independent test. Place measured facts, model estimates, interpretation and decisions on separate notebook lines so a student, parent or reviewer can see exactly what the mathematics establishes.

Workshop 2: Expression-weighted sampling

Start with this article's mechanism: Draw gene tokens from the categorical distribution so abundance influences how often each gene appears. Work through the concrete case: In ten draws the first gene may appear six times. Make the boundary visible: A finite sample has Monte Carlo noise. Complete the evidence step: Repeat seeds and report variation. Write the input table or token list, transformation, output, units, uncertainty and one observation that would contradict the conclusion. Repeat the calculation under a nearby defensible choice and compare sampling distributions, prior embeddings, similarities, attention weights, masks, losses, distances, neighbourhoods, proportions, logits or calibration as relevant. Keep complete donors, studies, tissues, species or technologies outside training when the claim concerns transfer. Random cells from the same biological source are not an independent test. Place measured facts, model estimates, interpretation and decisions on separate notebook lines so a student, parent or reviewer can see exactly what the mathematics establishes.

Workshop 3: Token multiplicity as abundance

Start with this article's mechanism: Repeated appearances allow the sequence to carry quantitative information without attaching a raw count to each token. Work through the concrete case: A highly expressed ribosomal gene can occupy several positions. Make the boundary visible: Common genes can crowd out rare signals. Complete the evidence step: Measure inclusion probability by abundance. Write the input table or token list, transformation, output, units, uncertainty and one observation that would contradict the conclusion. Repeat the calculation under a nearby defensible choice and compare sampling distributions, prior embeddings, similarities, attention weights, masks, losses, distances, neighbourhoods, proportions, logits or calibration as relevant. Keep complete donors, studies, tissues, species or technologies outside training when the claim concerns transfer. Random cells from the same biological source are not an independent test. Place measured facts, model estimates, interpretation and decisions on separate notebook lines so a student, parent or reviewer can see exactly what the mathematics establishes.

Workshop 4: Protein language vector

Start with this article's mechanism: A pretrained protein model converts the amino-acid sequence of each gene product into a numeric embedding. Work through the concrete case: A 300-residue protein becomes one fixed-length vector. Make the boundary visible: Isoforms and non-protein-coding genes need explicit handling. Complete the evidence step: Record protein accession and pooling. Write the input table or token list, transformation, output, units, uncertainty and one observation that would contradict the conclusion. Repeat the calculation under a nearby defensible choice and compare sampling distributions, prior embeddings, similarities, attention weights, masks, losses, distances, neighbourhoods, proportions, logits or calibration as relevant. Keep complete donors, studies, tissues, species or technologies outside training when the claim concerns transfer. Random cells from the same biological source are not an independent test. Place measured facts, model estimates, interpretation and decisions on separate notebook lines so a student, parent or reviewer can see exactly what the mathematics establishes.

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: Unseen-gene construction

Start with this article's mechanism: Because the starting vector is computed from sequence, a new protein-coding gene can enter without retraining a lookup table. Work through the concrete case: A gene from a new species receives an ESM-derived vector. Make the boundary visible: Sequence similarity does not capture every regulatory role. Complete the evidence step: Compare homologous and divergent genes. Write the input table or token list, transformation, output, units, uncertainty and one observation that would contradict the conclusion. Repeat the calculation under a nearby defensible choice and compare sampling distributions, prior embeddings, similarities, attention weights, masks, losses, distances, neighbourhoods, proportions, logits or calibration as relevant. Keep complete donors, studies, tissues, species or technologies outside training when the claim concerns transfer. Random cells from the same biological source are not an independent test. Place measured facts, model estimates, interpretation and decisions on separate notebook lines so a student, parent or reviewer can see exactly what the mathematics establishes.

Workshop 6: Chromosome metadata

Start with this article's mechanism: Genomic location is encoded alongside sequence-derived information to distinguish positional context. Work through the concrete case: Two proteins with similar sequence occur on different chromosomes. Make the boundary visible: Genome builds and annotations change coordinates. Complete the evidence step: Version assembly and gene annotation. Write the input table or token list, transformation, output, units, uncertainty and one observation that would contradict the conclusion. Repeat the calculation under a nearby defensible choice and compare sampling distributions, prior embeddings, similarities, attention weights, masks, losses, distances, neighbourhoods, proportions, logits or calibration as relevant. Keep complete donors, studies, tissues, species or technologies outside training when the claim concerns transfer. Random cells from the same biological source are not an independent test. Place measured facts, model estimates, interpretation and decisions on separate notebook lines so a student, parent or reviewer can see exactly what the mathematics establishes.

Workshop 7: Species and cell context

Start with this article's mechanism: Metadata and the collection of sampled genes identify the biological context in which a gene vector is interpreted. Work through the concrete case: The same conserved protein appears in human and mouse cells. Make the boundary visible: Species tokens can also reveal dataset shortcuts. Complete the evidence step: Probe species predictability after alignment. Write the input table or token list, transformation, output, units, uncertainty and one observation that would contradict the conclusion. Repeat the calculation under a nearby defensible choice and compare sampling distributions, prior embeddings, similarities, attention weights, masks, losses, distances, neighbourhoods, proportions, logits or calibration as relevant. Keep complete donors, studies, tissues, species or technologies outside training when the claim concerns transfer. Random cells from the same biological source are not an independent test. Place measured facts, model estimates, interpretation and decisions on separate notebook lines so a student, parent or reviewer can see exactly what the mathematics establishes.

Workshop 8: Transformer input sequence

Start with this article's mechanism: The 1,024 sampled genes are sorted by genomic position within chromosome groups, separated by species-specific chromosome start and end tokens, and processed by stacked attention and feed-forward blocks. Work through the concrete case: A CLS token begins one cell sentence. Make the boundary visible: Sampling and the finite sequence change which evidence enters. Complete the evidence step: Publish sequence length, seed and retained mass. Write the input table or token list, transformation, output, units, uncertainty and one observation that would contradict the conclusion. Repeat the calculation under a nearby defensible choice and compare sampling distributions, prior embeddings, similarities, attention weights, masks, losses, distances, neighbourhoods, proportions, logits or calibration as relevant. Keep complete donors, studies, tissues, species or technologies outside training when the claim concerns transfer. Random cells from the same biological source are not an independent test. Place measured facts, model estimates, interpretation and decisions on separate notebook lines so a student, parent or reviewer can see exactly what the mathematics establishes.

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: Scaled self-attention

Start with this article's mechanism: Query–key dot products divided by the square root of dimension become softmax weights over value vectors. Work through the concrete case: A transcription factor token attends to a pathway partner. Make the boundary visible: Attention is computation, not a regulatory edge. Complete the evidence step: Compare weights with perturbation evidence. Write the input table or token list, transformation, output, units, uncertainty and one observation that would contradict the conclusion. Repeat the calculation under a nearby defensible choice and compare sampling distributions, prior embeddings, similarities, attention weights, masks, losses, distances, neighbourhoods, proportions, logits or calibration as relevant. Keep complete donors, studies, tissues, species or technologies outside training when the claim concerns transfer. Random cells from the same biological source are not an independent test. Place measured facts, model estimates, interpretation and decisions on separate notebook lines so a student, parent or reviewer can see exactly what the mathematics establishes.

Workshop 10: Many layers and parameters

Start with this article's mechanism: The reported UCE architecture uses a deep Transformer with roughly 650 million parameters to refine contextual gene representations. Work through the concrete case: Later layers integrate wider gene context. Make the boundary visible: Capacity can memorize corpus structure. Complete the evidence step: Use species and study holdouts. Write the input table or token list, transformation, output, units, uncertainty and one observation that would contradict the conclusion. Repeat the calculation under a nearby defensible choice and compare sampling distributions, prior embeddings, similarities, attention weights, masks, losses, distances, neighbourhoods, proportions, logits or calibration as relevant. Keep complete donors, studies, tissues, species or technologies outside training when the claim concerns transfer. Random cells from the same biological source are not an independent test. Place measured facts, model estimates, interpretation and decisions on separate notebook lines so a student, parent or reviewer can see exactly what the mathematics establishes.

Workshop 11: Cell-vector pooling

Start with this article's mechanism: The final hidden state of the special CLS token, decoded through an MLP, supplies the reported 1,280-dimensional cell embedding. Work through the concrete case: A cell becomes one point in the universal space. Make the boundary visible: A single vector can hide conflicting programmes. Complete the evidence step: Compare CLS geometry with token-level evidence. Write the input table or token list, transformation, output, units, uncertainty and one observation that would contradict the conclusion. Repeat the calculation under a nearby defensible choice and compare sampling distributions, prior embeddings, similarities, attention weights, masks, losses, distances, neighbourhoods, proportions, logits or calibration as relevant. Keep complete donors, studies, tissues, species or technologies outside training when the claim concerns transfer. Random cells from the same biological source are not an independent test. Place measured facts, model estimates, interpretation and decisions on separate notebook lines so a student, parent or reviewer can see exactly what the mathematics establishes.

Workshop 12: Distances in universal space

Start with this article's mechanism: Cosine similarity or Euclidean distance quantifies neighbourhoods among cell vectors. Work through the concrete case: A new fish immune cell lies near related vertebrate cells. Make the boundary visible: Distance depends on checkpoint and preprocessing. Complete the evidence step: Report metric and neighbour stability. Write the input table or token list, transformation, output, units, uncertainty and one observation that would contradict the conclusion. Repeat the calculation under a nearby defensible choice and compare sampling distributions, prior embeddings, similarities, attention weights, masks, losses, distances, neighbourhoods, proportions, logits or calibration as relevant. Keep complete donors, studies, tissues, species or technologies outside training when the claim concerns transfer. Random cells from the same biological source are not an independent test. Place measured facts, model estimates, interpretation and decisions on separate notebook lines so a student, parent or reviewer can see exactly what the mathematics establishes.

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: Zero-shot label transfer

Start with this article's mechanism: Reference neighbours can supply candidate labels without updating model weights. Work through the concrete case: A new-species cell inherits a broad immune label. Make the boundary visible: Reference coverage constrains the vocabulary. Complete the evidence step: Allow rejection and broad labels. Write the input table or token list, transformation, output, units, uncertainty and one observation that would contradict the conclusion. Repeat the calculation under a nearby defensible choice and compare sampling distributions, prior embeddings, similarities, attention weights, masks, losses, distances, neighbourhoods, proportions, logits or calibration as relevant. Keep complete donors, studies, tissues, species or technologies outside training when the claim concerns transfer. Random cells from the same biological source are not an independent test. Place measured facts, model estimates, interpretation and decisions on separate notebook lines so a student, parent or reviewer can see exactly what the mathematics establishes.

Workshop 14: Worked six-gene sample

Start with this article's mechanism: Normalize six counts, calculate draw probabilities, perform twenty seeded draws and compare two pooled vectors. Work through the concrete case: The most abundant gene appears often but not exactly at expectation. Make the boundary visible: One random draw is unstable. Complete the evidence step: Compute confidence intervals over seeds. Write the input table or token list, transformation, output, units, uncertainty and one observation that would contradict the conclusion. Repeat the calculation under a nearby defensible choice and compare sampling distributions, prior embeddings, similarities, attention weights, masks, losses, distances, neighbourhoods, proportions, logits or calibration as relevant. Keep complete donors, studies, tissues, species or technologies outside training when the claim concerns transfer. Random cells from the same biological source are not an independent test. Place measured facts, model estimates, interpretation and decisions on separate notebook lines so a student, parent or reviewer can see exactly what the mathematics establishes.

Workshop 15: Cross-species validation

Start with this article's mechanism: Keep complete species and tissues outside training and compare with orthologue-based and count-based baselines. Work through the concrete case: A held-out organism tests genuine zero-shot transfer. Make the boundary visible: Shared technologies can still leak structure. Complete the evidence step: Stratify by assay and study. Write the input table or token list, transformation, output, units, uncertainty and one observation that would contradict the conclusion. Repeat the calculation under a nearby defensible choice and compare sampling distributions, prior embeddings, similarities, attention weights, masks, losses, distances, neighbourhoods, proportions, logits or calibration as relevant. Keep complete donors, studies, tissues, species or technologies outside training when the claim concerns transfer. Random cells from the same biological source are not an independent test. Place measured facts, model estimates, interpretation and decisions on separate notebook lines so a student, parent or reviewer can see exactly what the mathematics establishes.

Workshop 16: Governance and learning

Start with this article's mechanism: Version counts, probabilities, draws, sequences, coordinates, checkpoint and distances. This joins probability, genomics and AI. Work through the concrete case: Students can start with categorical sampling. Make the boundary visible: A universal vector remains a model estimate. Complete the evidence step: Write a UCE model card. Write the input table or token list, transformation, output, units, uncertainty and one observation that would contradict the conclusion. Repeat the calculation under a nearby defensible choice and compare sampling distributions, prior embeddings, similarities, attention weights, masks, losses, distances, neighbourhoods, proportions, logits or calibration as relevant. Keep complete donors, studies, tissues, species or technologies outside training when the claim concerns transfer. Random cells from the same biological source are not an independent test. Place measured facts, model estimates, interpretation and decisions on separate notebook lines so a student, parent or reviewer can see exactly what the mathematics establishes.