WHY MATHEMATICS?

Why Mathematics? | scPRINT for Gene Networks: Expression and Protein Embeddings, Denoising, Bottleneck Learning and Attention Graphs

How scPRINT builds a bidirectional cell model from gene identity, expression, genomic location and protein sequence, then combines three pretraining losses and selected attention heads to propose cell-specific gene networks.

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

Build multi-part gene tokens with explicit biological provenance
Understand denoising, bottleneck and hierarchical-label losses
Construct attention-derived gene-by-gene matrices
Validate networks, read-depth recovery and zero-shot outputs

Chapter map

1. Weighted corpus sampling
2. Gene identity representation
3. Expression embedding
4. Genomic-location encoding
5. Matched ESM2 protein embedding
6. Bidirectional Transformer
7. Purposeful downsampling
8. Denoising objective
9. Cell bottleneck
10. Hierarchical label prediction
11. Joint loss arithmetic
12. Attention-head selection
13. Gene-by-gene network matrix
14. Worked four-gene network
15. Network benchmarking
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. Weighted corpus sampling

This step deserves a careful pause. A custom sampler draws training cells across species, diseases and populations rather than treating every stored cell identically.

A concrete situation

Rare groups receive planned representation.

What the mathematics does not license

Weights encode editorial choices. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Publish probabilities and effective counts. 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. Gene identity representation

The useful question is not only how to calculate, but what the calculation assumes. Every recognized gene receives an ID embedding shared across cells.

A concrete situation

The same gene starts from the same identity vector.

What the mathematics does not license

Aliases and missing genes need mapping rules. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Version the vocabulary. 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. Expression embedding

A small example makes the mechanism visible. Observed count or normalized expression is projected into a vector that changes with the cell.

A concrete situation

Gene A at count 5 differs from count 1.

What the mathematics does not license

Counts depend on sequencing depth. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Record preprocessing and total counts. 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. Genomic-location encoding

This is where modelling choices begin to affect evidence. Gene position supplies chromosome and coordinate context to the token.

A concrete situation

Two genes near one another share positional structure.

What the mathematics does not license

Genome builds and rearrangements complicate location. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Version assembly and coordinates. 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. Matched ESM2 protein embedding

The central mathematical idea is concrete. A protein-language vector represents sequence information associated with a gene product.

A concrete situation

Related proteins begin with related numerical features.

What the mathematics does not license

Isoforms and noncoding genes need explicit rules. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Store accessions 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.

6. Bidirectional Transformer

Here the geometry becomes operational. Each retained gene can attend to genes on either side of the input ordering because the encoder is not causal.

A concrete situation

A masked gene uses the whole observed context.

What the mathematics does not license

Attention can learn batch co-occurrence. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Probe technical 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.

7. Purposeful downsampling

This step deserves a careful pause. Training reduces transcript counts to create a noisy source profile with a known fuller target.

A concrete situation

A 10,000-count cell is sampled to 2,000 counts.

What the mathematics does not license

Random thinning is only one noise model. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Compare with real technical replicates. 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. Denoising objective

The useful question is not only how to calculate, but what the calculation assumes. The decoder predicts the original profile from the downsampled input and requested output depth.

A concrete situation

Predicted counts are compared with known counts.

What the mathematics does not license

Recovery can smooth rare transcripts. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Report abundance-stratified 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. Cell bottleneck

A small example makes the mechanism visible. Requested genes are reconstructed using only a cell embedding, forcing that compact vector to retain global information.

A concrete situation

One embedding predicts a panel of genes.

What the mathematics does not license

Compression can discard subtle states. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Compare bottleneck sizes. 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. Hierarchical label prediction

This is where modelling choices begin to affect evidence. Separate disentangled embeddings predict cell type, tissue, disease, sex, sequencer, ethnicity and organism under a hierarchy-aware classifier.

A concrete situation

One vector specializes in disease information.

What the mathematics does not license

Metadata can be confounded and sensitive. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Audit privacy, balance and leakage. 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. Joint loss arithmetic

The central mathematical idea is concrete. Denoising, bottleneck and label losses are weighted and added before gradient updates.

A concrete situation

Three losses 0.4, 0.2 and 0.3 combine under declared weights.

What the mathematics does not license

One task can dominate training. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Plot every term and gradient scale. 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. Attention-head selection

Here the geometry becomes operational. Heads judged useful for gene-network inference are selected rather than averaging every head blindly.

A concrete situation

A subset yields a sharper gene-by-gene matrix.

What the mathematics does not license

Selection on the test set leaks evidence. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Choose heads on separate validation data. 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. Gene-by-gene network matrix

This step deserves a careful pause. Attention entries across selected heads and layers are combined into weighted candidate edges.

A concrete situation

Edge A to B receives score 0.74.

What the mathematics does not license

Scores are not calibrated causal probabilities. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Compare with null and known networks. 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 four-gene network

The useful question is not only how to calculate, but what the calculation assumes. Build a four-by-four attention matrix, combine two heads, remove the diagonal and rank candidate edges.

A concrete situation

A–C becomes the strongest proposed link.

What the mathematics does not license

A tiny example ignores layer and cell variation. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Shuffle genes and repeat. 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. Network benchmarking

A small example makes the mechanism visible. Use perturbation, transcription-factor binding and trusted pathway edges, plus simple correlation and dedicated baselines.

A concrete situation

A held-out perturbation tests predicted targets.

What the mathematics does not license

Gold standards are incomplete. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Report precision–recall and uncertainty. 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

This is where modelling choices begin to affect evidence. Version corpus weights, token parts, depths, losses, heads and network. This joins graphs, probability and biology.

A concrete situation

Students can begin with matrices and downsampling.

What the mathematics does not license

An attention graph is a hypothesis generator. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Write a scPRINT network 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 scPRINT audit adds three views of each gene before bidirectional attention: identity, expression and genomic-location information, with matched ESM2 protein embeddings supporting gene representation. Its pretraining combines denoising, bottleneck reconstruction and hierarchical label prediction so the model learns cell embeddings, read-depth-aware expression recovery and several disentangled label views.

Record cellxgene corpus version, weighted sampling, species and metadata balance, gene IDs, positions, protein sequences, downsampling depth, requested reconstruction depth, bottleneck genes, hierarchical labels, attention heads selected for a network and every loss weight. The published models span roughly 2 to 100 million parameters and were trained on more than 50 million cells.

Validate inferred gene networks against perturbation and binding evidence, repeat head selection, compare simple correlation and dedicated network methods, and keep atlases outside tuning. Attention-derived edges are model associations; they become regulatory claims only after independent causal evidence.

Frequently asked questions

What is scPRINT?

scPRINT, single-cell PRe-trained Inference of Networks with Transformers, is a bidirectional foundation model pretrained on more than 50 million cellxgene cells for cell-specific gene-network inference and other zero-shot tasks.

What information enters a gene token?

The model combines gene identity, expression and genomic-location representations, alongside matched ESM2 protein embeddings and cell context.

Which pretraining tasks are combined?

The paper adds three losses: denoising a deliberately downsampled profile, reconstructing requested genes through a cell bottleneck, and predicting hierarchical metadata labels.

What is read-depth-aware recovery?

The model receives a reduced-depth profile and a requested output depth, then estimates expression at that target total count. The estimate is not a new sequencing measurement.

How does scPRINT form a gene network?

Selected attention heads are combined into a gene-by-gene matrix. This yields model-weighted associations that require independent network validation.

Can attention prove regulation?

No. Directional or causal gene regulation needs perturbation, binding, temporal or other experimental evidence beyond an attention score.

Primary source and fact-check note

Jérémie Kalfon, Jules Samaran, Gabriel Peyré et al., “scPRINT: pre-training on 50 million cells allows robust gene network predictions,” Nature Communications 16, 3607 (2025), DOI 10.1038/s41467-025-58699-1. 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

Corpus-sampling audit

Store weighted sampling probabilities and effective cells by species, disease and technology.

Token-provenance audit

Version gene IDs, coordinates, protein accessions, expression transformation and missing cases.

Downsampling reality check

Compare simulated thinning with real lower-depth replicates and technology-specific noise.

Loss ledger

Report denoising, bottleneck and each hierarchical-label loss separately.

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.

Head-selection firewall

Choose attention heads without test-network labels and repeat selection across seeds.

Network nulls

Compare shuffled cells, shuffled genes, correlation and dedicated GRN methods.

Causal validation

Use held-out perturbation and binding evidence before describing an edge as regulatory.

Final claim

Name corpus, token parts, depths, losses and heads. Attention networks are candidate models.

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: Weighted corpus sampling

Start with this article's mechanism: A custom sampler draws training cells across species, diseases and populations rather than treating every stored cell identically. Work through the concrete case: Rare groups receive planned representation. Make the boundary visible: Weights encode editorial choices. Complete the evidence step: Publish probabilities and effective counts. Write the input table or token list, transformation, output, dimensions, uncertainty and one observation that would contradict the conclusion. Repeat under a nearby defensible setting and compare masks, ranks, value vectors, retention outputs, logits, losses, embeddings, generated profiles, attention networks or calibration as relevant. Use complete donors, studies, tissues or technologies outside training when testing transfer, because random cells from one biological source are not independent evidence. Keep measurement, model estimate, interpretation and decision on separate notebook lines so a student, parent or reviewer can see what the mathematics establishes.

Workshop 2: Gene identity representation

Start with this article's mechanism: Every recognized gene receives an ID embedding shared across cells. Work through the concrete case: The same gene starts from the same identity vector. Make the boundary visible: Aliases and missing genes need mapping rules. Complete the evidence step: Version the vocabulary. Write the input table or token list, transformation, output, dimensions, uncertainty and one observation that would contradict the conclusion. Repeat under a nearby defensible setting and compare masks, ranks, value vectors, retention outputs, logits, losses, embeddings, generated profiles, attention networks or calibration as relevant. Use complete donors, studies, tissues or technologies outside training when testing transfer, because random cells from one biological source are not independent evidence. Keep measurement, model estimate, interpretation and decision on separate notebook lines so a student, parent or reviewer can see what the mathematics establishes.

Workshop 3: Expression embedding

Start with this article's mechanism: Observed count or normalized expression is projected into a vector that changes with the cell. Work through the concrete case: Gene A at count 5 differs from count 1. Make the boundary visible: Counts depend on sequencing depth. Complete the evidence step: Record preprocessing and total counts. Write the input table or token list, transformation, output, dimensions, uncertainty and one observation that would contradict the conclusion. Repeat under a nearby defensible setting and compare masks, ranks, value vectors, retention outputs, logits, losses, embeddings, generated profiles, attention networks or calibration as relevant. Use complete donors, studies, tissues or technologies outside training when testing transfer, because random cells from one biological source are not independent evidence. Keep measurement, model estimate, interpretation and decision on separate notebook lines so a student, parent or reviewer can see what the mathematics establishes.

Workshop 4: Genomic-location encoding

Start with this article's mechanism: Gene position supplies chromosome and coordinate context to the token. Work through the concrete case: Two genes near one another share positional structure. Make the boundary visible: Genome builds and rearrangements complicate location. Complete the evidence step: Version assembly and coordinates. Write the input table or token list, transformation, output, dimensions, uncertainty and one observation that would contradict the conclusion. Repeat under a nearby defensible setting and compare masks, ranks, value vectors, retention outputs, logits, losses, embeddings, generated profiles, attention networks or calibration as relevant. Use complete donors, studies, tissues or technologies outside training when testing transfer, because random cells from one biological source are not independent evidence. Keep measurement, model estimate, interpretation and decision on separate notebook lines so a student, parent or reviewer can see 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: Matched ESM2 protein embedding

Start with this article's mechanism: A protein-language vector represents sequence information associated with a gene product. Work through the concrete case: Related proteins begin with related numerical features. Make the boundary visible: Isoforms and noncoding genes need explicit rules. Complete the evidence step: Store accessions and pooling. Write the input table or token list, transformation, output, dimensions, uncertainty and one observation that would contradict the conclusion. Repeat under a nearby defensible setting and compare masks, ranks, value vectors, retention outputs, logits, losses, embeddings, generated profiles, attention networks or calibration as relevant. Use complete donors, studies, tissues or technologies outside training when testing transfer, because random cells from one biological source are not independent evidence. Keep measurement, model estimate, interpretation and decision on separate notebook lines so a student, parent or reviewer can see what the mathematics establishes.

Workshop 6: Bidirectional Transformer

Start with this article's mechanism: Each retained gene can attend to genes on either side of the input ordering because the encoder is not causal. Work through the concrete case: A masked gene uses the whole observed context. Make the boundary visible: Attention can learn batch co-occurrence. Complete the evidence step: Probe technical labels. Write the input table or token list, transformation, output, dimensions, uncertainty and one observation that would contradict the conclusion. Repeat under a nearby defensible setting and compare masks, ranks, value vectors, retention outputs, logits, losses, embeddings, generated profiles, attention networks or calibration as relevant. Use complete donors, studies, tissues or technologies outside training when testing transfer, because random cells from one biological source are not independent evidence. Keep measurement, model estimate, interpretation and decision on separate notebook lines so a student, parent or reviewer can see what the mathematics establishes.

Workshop 7: Purposeful downsampling

Start with this article's mechanism: Training reduces transcript counts to create a noisy source profile with a known fuller target. Work through the concrete case: A 10,000-count cell is sampled to 2,000 counts. Make the boundary visible: Random thinning is only one noise model. Complete the evidence step: Compare with real technical replicates. Write the input table or token list, transformation, output, dimensions, uncertainty and one observation that would contradict the conclusion. Repeat under a nearby defensible setting and compare masks, ranks, value vectors, retention outputs, logits, losses, embeddings, generated profiles, attention networks or calibration as relevant. Use complete donors, studies, tissues or technologies outside training when testing transfer, because random cells from one biological source are not independent evidence. Keep measurement, model estimate, interpretation and decision on separate notebook lines so a student, parent or reviewer can see what the mathematics establishes.

Workshop 8: Denoising objective

Start with this article's mechanism: The decoder predicts the original profile from the downsampled input and requested output depth. Work through the concrete case: Predicted counts are compared with known counts. Make the boundary visible: Recovery can smooth rare transcripts. Complete the evidence step: Report abundance-stratified error. Write the input table or token list, transformation, output, dimensions, uncertainty and one observation that would contradict the conclusion. Repeat under a nearby defensible setting and compare masks, ranks, value vectors, retention outputs, logits, losses, embeddings, generated profiles, attention networks or calibration as relevant. Use complete donors, studies, tissues or technologies outside training when testing transfer, because random cells from one biological source are not independent evidence. Keep measurement, model estimate, interpretation and decision on separate notebook lines so a student, parent or reviewer can see 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: Cell bottleneck

Start with this article's mechanism: Requested genes are reconstructed using only a cell embedding, forcing that compact vector to retain global information. Work through the concrete case: One embedding predicts a panel of genes. Make the boundary visible: Compression can discard subtle states. Complete the evidence step: Compare bottleneck sizes. Write the input table or token list, transformation, output, dimensions, uncertainty and one observation that would contradict the conclusion. Repeat under a nearby defensible setting and compare masks, ranks, value vectors, retention outputs, logits, losses, embeddings, generated profiles, attention networks or calibration as relevant. Use complete donors, studies, tissues or technologies outside training when testing transfer, because random cells from one biological source are not independent evidence. Keep measurement, model estimate, interpretation and decision on separate notebook lines so a student, parent or reviewer can see what the mathematics establishes.

Workshop 10: Hierarchical label prediction

Start with this article's mechanism: Separate disentangled embeddings predict cell type, tissue, disease, sex, sequencer, ethnicity and organism under a hierarchy-aware classifier. Work through the concrete case: One vector specializes in disease information. Make the boundary visible: Metadata can be confounded and sensitive. Complete the evidence step: Audit privacy, balance and leakage. Write the input table or token list, transformation, output, dimensions, uncertainty and one observation that would contradict the conclusion. Repeat under a nearby defensible setting and compare masks, ranks, value vectors, retention outputs, logits, losses, embeddings, generated profiles, attention networks or calibration as relevant. Use complete donors, studies, tissues or technologies outside training when testing transfer, because random cells from one biological source are not independent evidence. Keep measurement, model estimate, interpretation and decision on separate notebook lines so a student, parent or reviewer can see what the mathematics establishes.

Workshop 11: Joint loss arithmetic

Start with this article's mechanism: Denoising, bottleneck and label losses are weighted and added before gradient updates. Work through the concrete case: Three losses 0.4, 0.2 and 0.3 combine under declared weights. Make the boundary visible: One task can dominate training. Complete the evidence step: Plot every term and gradient scale. Write the input table or token list, transformation, output, dimensions, uncertainty and one observation that would contradict the conclusion. Repeat under a nearby defensible setting and compare masks, ranks, value vectors, retention outputs, logits, losses, embeddings, generated profiles, attention networks or calibration as relevant. Use complete donors, studies, tissues or technologies outside training when testing transfer, because random cells from one biological source are not independent evidence. Keep measurement, model estimate, interpretation and decision on separate notebook lines so a student, parent or reviewer can see what the mathematics establishes.

Workshop 12: Attention-head selection

Start with this article's mechanism: Heads judged useful for gene-network inference are selected rather than averaging every head blindly. Work through the concrete case: A subset yields a sharper gene-by-gene matrix. Make the boundary visible: Selection on the test set leaks evidence. Complete the evidence step: Choose heads on separate validation data. Write the input table or token list, transformation, output, dimensions, uncertainty and one observation that would contradict the conclusion. Repeat under a nearby defensible setting and compare masks, ranks, value vectors, retention outputs, logits, losses, embeddings, generated profiles, attention networks or calibration as relevant. Use complete donors, studies, tissues or technologies outside training when testing transfer, because random cells from one biological source are not independent evidence. Keep measurement, model estimate, interpretation and decision on separate notebook lines so a student, parent or reviewer can see 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: Gene-by-gene network matrix

Start with this article's mechanism: Attention entries across selected heads and layers are combined into weighted candidate edges. Work through the concrete case: Edge A to B receives score 0.74. Make the boundary visible: Scores are not calibrated causal probabilities. Complete the evidence step: Compare with null and known networks. Write the input table or token list, transformation, output, dimensions, uncertainty and one observation that would contradict the conclusion. Repeat under a nearby defensible setting and compare masks, ranks, value vectors, retention outputs, logits, losses, embeddings, generated profiles, attention networks or calibration as relevant. Use complete donors, studies, tissues or technologies outside training when testing transfer, because random cells from one biological source are not independent evidence. Keep measurement, model estimate, interpretation and decision on separate notebook lines so a student, parent or reviewer can see what the mathematics establishes.

Workshop 14: Worked four-gene network

Start with this article's mechanism: Build a four-by-four attention matrix, combine two heads, remove the diagonal and rank candidate edges. Work through the concrete case: A–C becomes the strongest proposed link. Make the boundary visible: A tiny example ignores layer and cell variation. Complete the evidence step: Shuffle genes and repeat. Write the input table or token list, transformation, output, dimensions, uncertainty and one observation that would contradict the conclusion. Repeat under a nearby defensible setting and compare masks, ranks, value vectors, retention outputs, logits, losses, embeddings, generated profiles, attention networks or calibration as relevant. Use complete donors, studies, tissues or technologies outside training when testing transfer, because random cells from one biological source are not independent evidence. Keep measurement, model estimate, interpretation and decision on separate notebook lines so a student, parent or reviewer can see what the mathematics establishes.

Workshop 15: Network benchmarking

Start with this article's mechanism: Use perturbation, transcription-factor binding and trusted pathway edges, plus simple correlation and dedicated baselines. Work through the concrete case: A held-out perturbation tests predicted targets. Make the boundary visible: Gold standards are incomplete. Complete the evidence step: Report precision–recall and uncertainty. Write the input table or token list, transformation, output, dimensions, uncertainty and one observation that would contradict the conclusion. Repeat under a nearby defensible setting and compare masks, ranks, value vectors, retention outputs, logits, losses, embeddings, generated profiles, attention networks or calibration as relevant. Use complete donors, studies, tissues or technologies outside training when testing transfer, because random cells from one biological source are not independent evidence. Keep measurement, model estimate, interpretation and decision on separate notebook lines so a student, parent or reviewer can see what the mathematics establishes.

Workshop 16: Governance and learning

Start with this article's mechanism: Version corpus weights, token parts, depths, losses, heads and network. This joins graphs, probability and biology. Work through the concrete case: Students can begin with matrices and downsampling. Make the boundary visible: An attention graph is a hypothesis generator. Complete the evidence step: Write a scPRINT network card. Write the input table or token list, transformation, output, dimensions, uncertainty and one observation that would contradict the conclusion. Repeat under a nearby defensible setting and compare masks, ranks, value vectors, retention outputs, logits, losses, embeddings, generated profiles, attention networks or calibration as relevant. Use complete donors, studies, tissues or technologies outside training when testing transfer, because random cells from one biological source are not independent evidence. Keep measurement, model estimate, interpretation and decision on separate notebook lines so a student, parent or reviewer can see what the mathematics establishes.