WHY MATHEMATICS?

Why Mathematics? | LangCell for Language–Cell Learning: Transcriptome Encoders, Contrastive Alignment, Text Matching and Zero-Shot Identity

How LangCell connects single-cell transcriptomes with natural-language descriptions: separate encoders create cell and text vectors, contrastive losses align matched pairs, and a matching classifier supports retrieval, few-shot learning and zero-shot identity scoring.

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

Encode cells and identity descriptions as vectors
Align cell–cell and cell–text pairs with contrastive loss
Train a direct cell–text matching decision
Test prompts, omitted labels, retrieval and calibrated rejection

Chapter map

1. Why language can carry identity
2. Cell sequence construction
3. Cell encoder
4. Text tokenization
5. Text encoder
6. Cosine similarity
7. Cell–cell contrastive learning
8. Momentum encoder
9. Negative queue or batch
10. Cell–text contrastive loss
11. Symmetric alignment
12. Cell–text matching head
13. Zero-shot ranking
14. Worked three-by-three matrix
15. Few-shot and fine-tuned use
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. Why language can carry identity

The useful question is not only how to calculate, but what the calculation assumes. Cell names, tissues, functions and ontology descriptions contain structured knowledge that labels alone discard.

A concrete situation

“CD4-positive memory T cell” conveys lineage and state.

What the mathematics does not license

Text can contain ambiguity or outdated claims. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Version every description source. 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. Cell sequence construction

A small example makes the mechanism visible. A transcriptome is converted into an ordered or otherwise bounded gene-token representation compatible with the cell encoder.

A concrete situation

Highly informative expressed genes enter the sequence.

What the mathematics does not license

Preprocessing controls which genes appear. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Record vocabulary and 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.

3. Cell encoder

This is where modelling choices begin to affect evidence. Transformer layers map gene context to a single normalized cell vector.

A concrete situation

One transcriptome becomes z_c in a shared dimension.

What the mathematics does not license

The vector can retain batch shortcuts. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Probe donor and study information. 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. Text tokenization

The central mathematical idea is concrete. A language tokenizer breaks a description into subword IDs with attention masks and length limits.

A concrete situation

“alveolar macrophage” becomes several tokens.

What the mathematics does not license

Rare biological terms may fragment oddly. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Inspect vocabulary and 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.

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. Text encoder

Here the geometry becomes operational. A pretrained language model maps the description into a normalized vector z_t.

A concrete situation

Two paraphrases occupy nearby points.

What the mathematics does not license

General language pretraining may miss specialist meaning. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Test domain synonyms and definitions. 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. Cosine similarity

This step deserves a careful pause. Normalize both vectors and take their dot product, producing a score from minus one to one.

A concrete situation

A matched pair scores 0.82 and a mismatch 0.15.

What the mathematics does not license

High similarity is relative, not calibrated truth. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Publish candidate set and margin. 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. Cell–cell contrastive learning

The useful question is not only how to calculate, but what the calculation assumes. Two views or related representations of a cell are pulled together while other cells act as negatives.

A concrete situation

Augmented views of one cell form a positive pair.

What the mathematics does not license

False negatives may share the same identity. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Use metadata-aware negatives. 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. Momentum encoder

A small example makes the mechanism visible. A slowly updated encoder supplies stable comparison vectors while the main encoder learns by gradient descent.

A concrete situation

Theta_m becomes m theta_m plus one-minus-m theta.

What the mathematics does not license

Momentum can lag rapid improvements. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Report coefficient and queue age. 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. Negative queue or batch

This is where modelling choices begin to affect evidence. Many comparison vectors make the contrastive denominator informative without encoding all examples at once.

A concrete situation

One positive competes with thousands of negatives.

What the mathematics does not license

Stale or biased negatives distort learning. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Audit class and tissue composition. 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. Cell–text contrastive loss

The central mathematical idea is concrete. For each cell, exponentiated matched similarity divided by the sum over candidate texts becomes a probability-like term.

A concrete situation

A positive logit dominates mismatches after temperature scaling.

What the mathematics does not license

Prompt duplication changes the denominator. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Deduplicate and stratify texts. 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. Symmetric alignment

Here the geometry becomes operational. Text-to-cell and cell-to-text directions can both contribute so neither modality is merely an output label.

A concrete situation

A description retrieves cells while a cell retrieves descriptions.

What the mathematics does not license

Retrieval asymmetry can remain. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Report both recall directions. 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. Cell–text matching head

This step deserves a careful pause. A joint classifier predicts matched versus mismatched after encoders supply representations.

A concrete situation

A hard negative sibling type receives careful comparison.

What the mathematics does not license

Easy negatives inflate accuracy. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Mine biologically close negatives. 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 ranking

The useful question is not only how to calculate, but what the calculation assumes. Encode candidate descriptions once, score a new cell against each and choose or reject based on similarity and margin.

A concrete situation

A cell ranks three immune descriptions.

What the mathematics does not license

The true identity may be absent. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Use coverage and rejection tests. 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 three-by-three matrix

A small example makes the mechanism visible. Build three normalized cell vectors and three text vectors, calculate nine dot products, apply temperature softmax and compute one cross-entropy term.

A concrete situation

Matched diagonal scores win but one sibling remains close.

What the mathematics does not license

Tiny vectors hide encoder uncertainty. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Paraphrase one text and recompute. 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. Few-shot and fine-tuned use

This is where modelling choices begin to affect evidence. A small labelled set can train a head or update selected encoder layers after multimodal pretraining.

A concrete situation

Ten examples per class adapt the representation.

What the mathematics does not license

Few-shot splits have high variance. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Repeat seeds and donor splits. 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

The central mathematical idea is concrete. Version cell tokens, text, positives, negatives, temperature, momentum and prompts. This links geometry, language and biology.

A concrete situation

Students can start with cosine similarity.

What the mathematics does not license

Semantic alignment remains a hypothesis. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Write a LangCell 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 LangCell audit places two data types in one geometric space. A cell encoder represents a ranked transcriptome, a text encoder represents natural-language descriptions of cell identity, and contrastive objectives pull matched cells and descriptions together while pushing mismatches apart. A cell–text matching task adds a direct compatibility classifier, allowing retrieval and zero-shot identity scoring.

Record gene vocabulary, rank encoding, cell encoder checkpoint, tokenizer, text encoder, description source, positive pairs, negative sampling, temperature, momentum update, queue or batch construction, matching head and label prompts. The words used to describe an identity are part of the model input, so prompt wording and ontology granularity can change a score.

Validate with synonym and paraphrase tests, omitted cell types, unrelated descriptions, donor and study holdouts, calibrated rejection and strong expression-only baselines. Language alignment can make cell representations searchable and transferable, but semantic similarity is not a substitute for molecular or experimental confirmation.

Frequently asked questions

What is LangCell?

LangCell is a multimodal pretraining framework that aligns single-cell transcriptomes with natural-language descriptions of cell identity. It supports cell and text retrieval as well as zero-shot, few-shot and fine-tuned tasks.

What is contrastive learning?

It gives matched cell–description pairs high similarity and unmatched pairs lower similarity. A softmax-style loss compares each positive pair with many alternatives.

Why use a temperature?

Temperature rescales similarity logits before softmax. A small value makes the distribution sharper; a larger value spreads probability more evenly across candidates.

What is cell–text matching?

A classifier examines a cell and a description jointly and predicts whether they belong together, complementing the global geometry learned by contrastive alignment.

How can LangCell perform zero-shot identification?

Encode a new cell and candidate descriptions, calculate similarities, and rank the descriptions without training a new classifier on labelled examples for those exact candidates.

Can wording change a zero-shot result?

Yes. Synonyms, detail level and unsupported claims can move a text vector. Responsible evaluation uses several documented prompts and permits an uncertain or unassigned result.

Primary source and fact-check note

Suyuan Zhao et al., “LangCell: Language-Cell Pre-training for Cell Identity Understanding,” Proceedings of the 41st International Conference on Machine Learning, PMLR 235, 61159–61185 (2024). 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

Prompt ledger

Store every identity description, source, tokenizer output, synonym and length truncation.

Paraphrase test

Use several accurate descriptions per type and report ranking and confidence variation.

Hard-negative test

Compare sibling cell types and near-synonymous descriptions rather than relying on easy mismatches.

Queue audit

Report momentum coefficient, queue age, class balance, tissues and false-negative risk.

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.

Bidirectional retrieval

Publish cell-to-text and text-to-cell recall, not only one favourable direction.

Omitted-label challenge

Remove true identities from the candidate set and evaluate rejection instead of forced selection.

Expression baseline

Compare identical splits with logistic regression and expression-only pretrained encoders.

Final claim

Name cell encoding, text, positives, negatives, temperature and prompts. Language similarity needs biological confirmation.

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 language can carry identity

Start with this article's mechanism: Cell names, tissues, functions and ontology descriptions contain structured knowledge that labels alone discard. Work through the concrete case: “CD4-positive memory T cell” conveys lineage and state. Make the boundary visible: Text can contain ambiguity or outdated claims. Complete the evidence step: Version every description source. 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: Cell sequence construction

Start with this article's mechanism: A transcriptome is converted into an ordered or otherwise bounded gene-token representation compatible with the cell encoder. Work through the concrete case: Highly informative expressed genes enter the sequence. Make the boundary visible: Preprocessing controls which genes appear. Complete the evidence step: Record vocabulary and truncation. 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: Cell encoder

Start with this article's mechanism: Transformer layers map gene context to a single normalized cell vector. Work through the concrete case: One transcriptome becomes z_c in a shared dimension. Make the boundary visible: The vector can retain batch shortcuts. Complete the evidence step: Probe donor and study information. 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: Text tokenization

Start with this article's mechanism: A language tokenizer breaks a description into subword IDs with attention masks and length limits. Work through the concrete case: “alveolar macrophage” becomes several tokens. Make the boundary visible: Rare biological terms may fragment oddly. Complete the evidence step: Inspect vocabulary and truncation. 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: Text encoder

Start with this article's mechanism: A pretrained language model maps the description into a normalized vector z_t. Work through the concrete case: Two paraphrases occupy nearby points. Make the boundary visible: General language pretraining may miss specialist meaning. Complete the evidence step: Test domain synonyms and definitions. 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: Cosine similarity

Start with this article's mechanism: Normalize both vectors and take their dot product, producing a score from minus one to one. Work through the concrete case: A matched pair scores 0.82 and a mismatch 0.15. Make the boundary visible: High similarity is relative, not calibrated truth. Complete the evidence step: Publish candidate set and margin. 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: Cell–cell contrastive learning

Start with this article's mechanism: Two views or related representations of a cell are pulled together while other cells act as negatives. Work through the concrete case: Augmented views of one cell form a positive pair. Make the boundary visible: False negatives may share the same identity. Complete the evidence step: Use metadata-aware negatives. 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: Momentum encoder

Start with this article's mechanism: A slowly updated encoder supplies stable comparison vectors while the main encoder learns by gradient descent. Work through the concrete case: Theta_m becomes m theta_m plus one-minus-m theta. Make the boundary visible: Momentum can lag rapid improvements. Complete the evidence step: Report coefficient and queue age. 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: Negative queue or batch

Start with this article's mechanism: Many comparison vectors make the contrastive denominator informative without encoding all examples at once. Work through the concrete case: One positive competes with thousands of negatives. Make the boundary visible: Stale or biased negatives distort learning. Complete the evidence step: Audit class and tissue composition. 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: Cell–text contrastive loss

Start with this article's mechanism: For each cell, exponentiated matched similarity divided by the sum over candidate texts becomes a probability-like term. Work through the concrete case: A positive logit dominates mismatches after temperature scaling. Make the boundary visible: Prompt duplication changes the denominator. Complete the evidence step: Deduplicate and stratify texts. 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: Symmetric alignment

Start with this article's mechanism: Text-to-cell and cell-to-text directions can both contribute so neither modality is merely an output label. Work through the concrete case: A description retrieves cells while a cell retrieves descriptions. Make the boundary visible: Retrieval asymmetry can remain. Complete the evidence step: Report both recall directions. 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: Cell–text matching head

Start with this article's mechanism: A joint classifier predicts matched versus mismatched after encoders supply representations. Work through the concrete case: A hard negative sibling type receives careful comparison. Make the boundary visible: Easy negatives inflate accuracy. Complete the evidence step: Mine biologically close negatives. 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 ranking

Start with this article's mechanism: Encode candidate descriptions once, score a new cell against each and choose or reject based on similarity and margin. Work through the concrete case: A cell ranks three immune descriptions. Make the boundary visible: The true identity may be absent. Complete the evidence step: Use coverage and rejection tests. 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 three-by-three matrix

Start with this article's mechanism: Build three normalized cell vectors and three text vectors, calculate nine dot products, apply temperature softmax and compute one cross-entropy term. Work through the concrete case: Matched diagonal scores win but one sibling remains close. Make the boundary visible: Tiny vectors hide encoder uncertainty. Complete the evidence step: Paraphrase one text and recompute. 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: Few-shot and fine-tuned use

Start with this article's mechanism: A small labelled set can train a head or update selected encoder layers after multimodal pretraining. Work through the concrete case: Ten examples per class adapt the representation. Make the boundary visible: Few-shot splits have high variance. Complete the evidence step: Repeat seeds and donor splits. 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 cell tokens, text, positives, negatives, temperature, momentum and prompts. This links geometry, language and biology. Work through the concrete case: Students can start with cosine similarity. Make the boundary visible: Semantic alignment remains a hypothesis. Complete the evidence step: Write a LangCell 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.