eduKateSG · Why Science?
Find the sample that breaks its cohort pattern—and prove it is an RNA outlier before telling a clinical story
Choose the closest route, then return to the full index whenever you need the wider evidence chain.
Science learning becomes powerful when students can tell a measured pattern from a model, a statistical decision from a mechanism, and a useful lead from a finished conclusion. FRASER—Find RAre Splicing Events in RNA-seq—detects sample-specific aberrant splicing by modelling splice ratios across a cohort. It uses the splice metrics ψ5 and ψ3 for alternative donor or acceptor use and θ for intron retention, learns latent confounders with a denoising autoencoder, and evaluates residual deviations with a beta-binomial model that includes junction-specific dispersion. It can support rare-disease RNA analysis, but an outlier is evidence for follow-up, not a diagnosis or proof that a DNA variant caused the event. This guide supports Primary Science, PSLE Science, Secondary Science, O-Level Science and STEM education as a longform evidence-reading route; it is not medical advice, a laboratory protocol, an admission promise or a career guarantee.
Related eduKate reading: IRFinder and intron-retention detection; LeafCutter and annotation-free intron clusters; JunctionSeq and differential junction usage; Science Learning Hub; Education Hub; STEM Education; Careers by Subject Capability.
Primary and official sources checked on 11 October 2026: Official Bioconductor FRASER package, release 3.23, version 2.8.0; Official FRASER vignette; Official FRASER repository; FRASER primary Nature Communications study; FRASER primary study full text; 2026 Singapore–Cambridge O-Level Biology syllabus; MOE G2/G3 Lower Secondary Science teaching and learning syllabus; 2026 MOE G2 Computing syllabus. Package interfaces and recommendations can change, so reproduce an analysis only with the documentation for the installed release.
| Stage | Question | Evidence | Responsible output |
|---|---|---|---|
| Count | What junction and intron evidence exists? | Splice-aware alignments and sample metadata | Versioned count matrices |
| Normalise | What cohort pattern is expected? | ψ5, ψ3 and θ ratios plus latent factors | A confounder-aware expectation |
| Test | Which sample-event pairs deviate unusually? | Beta-binomial residual evidence and multiplicity control | A ranked outlier set |
| Validate | Is the outlier real and relevant? | Genome browser, RNA assays, genotype and phenotype review | A bounded candidate, not a diagnosis |
Inside this guide
1–12 · Foundations and design
- 1. Define a rare outlier question
- 2. Spell out FRASER
- 3. Choose an informative tissue
- 4. Build a comparable cohort
- 5. Name independent samples
- 6. Align reads consistently
- 7. Count split reads
- 8. Measure psi5
- 9. Measure psi3
- 10. Measure theta
- 11. Filter low-information events
- 12. Learn latent structure
13–24 · Models and estimation
- 13. Select the latent dimension
- 14. Use a beta-binomial model
- 15. Calculate residual evidence
- 16. Control multiplicity
- 17. Rank without declaring diagnosis
- 18. Inspect the genome browser
- 19. Check sample-wide burden
- 20. Check cohort recurrence
- 21. Connect to DNA variants carefully
- 22. Look for allelic evidence
- 23. Interpret intron retention separately
- 24. Test leave-one-out influence
25–36 · Diagnostics and interpretation
- 25. Test cohort sensitivity
- 26. Test alignment sensitivity
- 27. Validate the RNA event
- 28. Seek functional follow-up
- 29. Connect to school science
- 30. Connect to statistics
- 31. Connect to computing
- 32. Connect to pathways
- 33. Use current package facts
- 34. Did you know?
- 35. Protect privacy
- 36. Finish with a reproducible bundle
Section 1 of 36
1. Define a rare outlier question
FRASER looks for an event that is unusual for one sample relative to a cohort, not a conventional average difference between two large groups. The comparison unit is therefore sample-event against learned cohort expectation.
Working checkpoint — rare splicing outlier: Write who the focal sample is and what defines its reference cohort. Keep the independent sample, design, annotation and software version beside this decision so the evidence can be reconstructed.
For a classroom model, use invented counts and clearly label the activity as an analogy rather than a sequencing experiment.
Section 2 of 36
2. Spell out FRASER
The name expands to Find RAre Splicing Events in RNA-seq. The capitals are memorable, but the scientific object remains a read-count ratio tied to a junction, splice site or intron.
Working checkpoint — psi5 metric: Use coordinates and metric names beside every reported event. Read the numerical result together with sample-level plots, effect size and a sensitivity check capable of weakening the headline.
For future study, connect the idea to biology, mathematics and computing while checking current official pathway information separately.
Section 3 of 36
3. Choose an informative tissue
RNA evidence can reveal aberrant splicing only when the relevant gene and event are expressed in the sampled tissue. A clean negative result in an uninformative tissue is weak evidence against a defect.
Working checkpoint — psi3 metric: Report expression and tissue relevance before interpreting absence. Preserve the complete tested family and every excluded feature; a polished shortlist is not a substitute for provenance.
For scientific writing, prefer “supports”, “is associated with” and “is predicted to” when stronger causal verbs are not justified.
Section 4 of 36
4. Build a comparable cohort
The model learns typical behaviour from available samples. Library preparation, tissue, ancestry, age, RNA quality and batch can create structure that looks like biology if the cohort is poorly matched.
Working checkpoint — theta metric: Create a sample table before counting junctions. Separate what the model estimates from the biological story proposed afterward, then name the extra evidence that story would require.
For students, the transferable habit is to ask: What was measured? Compared with what? What alternative explanation remains?
Section 5 of 36
5. Name independent samples
Each person or independently collected specimen contributes one biological unit. Technical repeats can improve measurement but should not inflate the apparent cohort size.
Working checkpoint — denoising autoencoder: Mark related, repeated and paired specimens explicitly. Treat identifiers, denominators and factor levels as scientific quantities: a silent mismatch can reverse an otherwise correct calculation.
For parents, a calm checkpoint is to ask how many independent samples support the claim and which assumption matters most.
Section 6 of 36
6. Align reads consistently
Splice-junction discovery depends on genome build, annotation, aligner and parameters. Mixed pipelines can create sample-specific junction differences that a downstream model cannot distinguish from biology.
Working checkpoint — beta-binomial model: Use one versioned alignment policy or model the unavoidable batch. Keep the independent sample, design, annotation and software version beside this decision so the evidence can be reconstructed.
For a classroom model, use invented counts and clearly label the activity as an analogy rather than a sequencing experiment.
Section 7 of 36
7. Count split reads
Reads spanning exon–exon junctions identify donor and acceptor connections. Raw counts remain important because a dramatic ratio based on two reads carries different weight from one supported by hundreds.
Working checkpoint — rare splicing outlier: Keep component counts in every review table. Read the numerical result together with sample-level plots, effect size and a sensitivity check capable of weakening the headline.
For future study, connect the idea to biology, mathematics and computing while checking current official pathway information separately.
Section 8 of 36
8. Measure psi5
The ψ5 metric describes how reads leaving a donor are distributed among acceptors. An outlying ψ5 can flag unusual donor-centred choice, but the interpretation depends on the local junction network.
Working checkpoint — psi5 metric: Draw all competing junctions from the donor. Preserve the complete tested family and every excluded feature; a polished shortlist is not a substitute for provenance.
For scientific writing, prefer “supports”, “is associated with” and “is predicted to” when stronger causal verbs are not justified.
Section 9 of 36
9. Measure psi3
The ψ3 metric describes how reads arriving at an acceptor are distributed among donors. It provides a complementary acceptor-centred view of alternative splicing.
Working checkpoint — psi3 metric: Check whether the same junction is unusual from both perspectives. Separate what the model estimates from the biological story proposed afterward, then name the extra evidence that story would require.
For students, the transferable habit is to ask: What was measured? Compared with what? What alternative explanation remains?
Section 10 of 36
10. Measure theta
FRASER uses θ to capture intron-retention evidence relative to splice support. Intronic reads can also reflect nascent RNA, DNA contamination or overlapping transcription.
Working checkpoint — theta metric: Inspect coverage shape and library protocol before calling retention. Treat identifiers, denominators and factor levels as scientific quantities: a silent mismatch can reverse an otherwise correct calculation.
For parents, a calm checkpoint is to ask how many independent samples support the claim and which assumption matters most.
Section 11 of 36
11. Filter low-information events
Very low counts produce unstable ratios and a large sparse testing problem. Coverage thresholds improve reliability but also define the tested family.
Working checkpoint — denoising autoencoder: Publish thresholds and retained event counts. Keep the independent sample, design, annotation and software version beside this decision so the evidence can be reconstructed.
For a classroom model, use invented counts and clearly label the activity as an analogy rather than a sequencing experiment.
Section 12 of 36
12. Learn latent structure
A denoising autoencoder models systematic variation across samples so expected splice ratios reflect hidden technical or biological factors. Latent correction helps only when the learned structure is sensible.
Working checkpoint — beta-binomial model: Inspect factor associations with known batches and quality metrics. Read the numerical result together with sample-level plots, effect size and a sensitivity check capable of weakening the headline.
For future study, connect the idea to biology, mathematics and computing while checking current official pathway information separately.
Section 13 of 36
13. Select the latent dimension
Too little correction can leave confounding; too much can absorb rare biological signal. The FRASER workflow includes data-driven dimension selection, but the selected value still deserves documentation.
Working checkpoint — rare splicing outlier: Save the optimisation curve and chosen dimension. Preserve the complete tested family and every excluded feature; a polished shortlist is not a substitute for provenance.
For scientific writing, prefer “supports”, “is associated with” and “is predicted to” when stronger causal verbs are not justified.
Section 14 of 36
14. Use a beta-binomial model
Given expected ratios, FRASER treats count variation with a beta-binomial distribution and junction-specific dispersion. This allows more variability than a simple binomial model.
Working checkpoint — psi5 metric: Explain dispersion as biological and technical variability, not noise erased. Separate what the model estimates from the biological story proposed afterward, then name the extra evidence that story would require.
For students, the transferable habit is to ask: What was measured? Compared with what? What alternative explanation remains?
Section 15 of 36
15. Calculate residual evidence
An outlier is a sample-event pair whose observed ratio deviates strongly from expectation under the fitted model. Residuals and z-scores help show direction and magnitude.
Working checkpoint — psi3 metric: Report observed ratio, expected ratio and residual together. Treat identifiers, denominators and factor levels as scientific quantities: a silent mismatch can reverse an otherwise correct calculation.
For parents, a calm checkpoint is to ask how many independent samples support the claim and which assumption matters most.
Section 16 of 36
16. Control multiplicity
A cohort contains many samples and events, creating millions of chances for extreme values. Adjusted evidence must match the tested family and the intended sample-level claim.
Working checkpoint — theta metric: Preserve all tested events and the correction scope. Keep the independent sample, design, annotation and software version beside this decision so the evidence can be reconstructed.
For a classroom model, use invented counts and clearly label the activity as an analogy rather than a sequencing experiment.
Section 17 of 36
17. Rank without declaring diagnosis
A small adjusted value prioritises an unusual RNA pattern. It does not identify the causal DNA change, prove clinical significance or establish that the event explains a person’s phenotype.
Working checkpoint — denoising autoencoder: Use “candidate aberrant-splicing outlier” until validation supports more. Read the numerical result together with sample-level plots, effect size and a sensitivity check capable of weakening the headline.
For future study, connect the idea to biology, mathematics and computing while checking current official pathway information separately.
Section 18 of 36
18. Inspect the genome browser
Coverage, junction arcs and neighbouring exons can expose mapping artefacts, unannotated exons, low support or broader transcriptional disruption.
Working checkpoint — beta-binomial model: Review focal and control samples at one shared scale. Preserve the complete tested family and every excluded feature; a polished shortlist is not a substitute for provenance.
For scientific writing, prefer “supports”, “is associated with” and “is predicted to” when stronger causal verbs are not justified.
Section 19 of 36
19. Check sample-wide burden
A specimen with thousands of outliers may have RNA degradation, contamination, extreme batch effects or unusual cell composition. A globally problematic sample needs different handling from one local event.
Working checkpoint — rare splicing outlier: Plot outlier counts and quality metrics per sample. Separate what the model estimates from the biological story proposed afterward, then name the extra evidence that story would require.
For students, the transferable habit is to ask: What was measured? Compared with what? What alternative explanation remains?
Section 20 of 36
20. Check cohort recurrence
An event rare in one cohort may be common in another tissue, ancestry group or protocol. Public and internal recurrence evidence helps distinguish technical habits from unusual biology.
Working checkpoint — psi5 metric: Record cohort composition and recurrence counts. Treat identifiers, denominators and factor levels as scientific quantities: a silent mismatch can reverse an otherwise correct calculation.
For parents, a calm checkpoint is to ask how many independent samples support the claim and which assumption matters most.
Section 21 of 36
21. Connect to DNA variants carefully
A nearby rare variant can strengthen a causal hypothesis, but proximity is not proof. Phasing, inheritance, predicted splice effect and RNA pattern should tell a coherent story.
Working checkpoint — psi3 metric: Keep variant and RNA evidence in separate columns before integration. Keep the independent sample, design, annotation and software version beside this decision so the evidence can be reconstructed.
For a classroom model, use invented counts and clearly label the activity as an analogy rather than a sequencing experiment.
Section 22 of 36
22. Look for allelic evidence
When heterozygous sites are informative, allele-specific expression or junction reads may link an RNA defect to one haplotype. Mapping bias can imitate imbalance.
Working checkpoint — theta metric: Use bias-aware alignments and explicit phasing where possible. Read the numerical result together with sample-level plots, effect size and a sensitivity check capable of weakening the headline.
For future study, connect the idea to biology, mathematics and computing while checking current official pathway information separately.
Section 23 of 36
23. Interpret intron retention separately
Retention has distinct confounders from alternative junction choice. Nuclear RNA, incomplete processing and DNA carry-over deserve direct checks.
Working checkpoint — denoising autoencoder: Do not treat every θ outlier as a mature transcript. Preserve the complete tested family and every excluded feature; a polished shortlist is not a substitute for provenance.
For scientific writing, prefer “supports”, “is associated with” and “is predicted to” when stronger causal verbs are not justified.
Section 24 of 36
24. Test leave-one-out influence
Because expectations are cohort-derived, removing the focal sample or influential neighbours can reveal whether its score depends on its own contribution or a small subgroup.
Working checkpoint — beta-binomial model: Repeat priority events under a leave-one-out fit. Separate what the model estimates from the biological story proposed afterward, then name the extra evidence that story would require.
For students, the transferable habit is to ask: What was measured? Compared with what? What alternative explanation remains?
Section 25 of 36
25. Test cohort sensitivity
A justified alternative reference cohort can change latent structure and dispersion. Robust events remain unusual across reasonable cohort definitions.
Working checkpoint — rare splicing outlier: Predefine alternative cohorts rather than shopping for significance. Treat identifiers, denominators and factor levels as scientific quantities: a silent mismatch can reverse an otherwise correct calculation.
For parents, a calm checkpoint is to ask how many independent samples support the claim and which assumption matters most.
Section 26 of 36
26. Test alignment sensitivity
A decisive novel junction should survive a reasonable remapping or independent junction counter. Disagreement may identify reference or parameter dependence.
Working checkpoint — psi5 metric: Compare coordinates and component counts, not only adjusted values. Keep the independent sample, design, annotation and software version beside this decision so the evidence can be reconstructed.
For a classroom model, use invented counts and clearly label the activity as an analogy rather than a sequencing experiment.
Section 27 of 36
27. Validate the RNA event
RT-PCR, targeted sequencing or an independent RNA sample can confirm the aberrant product. Assays should capture both abnormal and normal forms when their ratio matters.
Working checkpoint — psi3 metric: Design primers around the predicted junction structure. Read the numerical result together with sample-level plots, effect size and a sensitivity check capable of weakening the headline.
For future study, connect the idea to biology, mathematics and computing while checking current official pathway information separately.
Section 28 of 36
28. Seek functional follow-up
A confirmed splice change may affect coding sequence, stability or protein, but each consequence requires new evidence. Nonsense-mediated decay can even hide the altered transcript.
Working checkpoint — theta metric: Build a validation ladder matched to the proposed consequence. Preserve the complete tested family and every excluded feature; a polished shortlist is not a substitute for provenance.
For scientific writing, prefer “supports”, “is associated with” and “is predicted to” when stronger causal verbs are not justified.
Section 29 of 36
29. Connect to school science
FRASER turns “unusual” into a fair comparison: define the reference, measure consistently and ask whether alternative explanations remain. That habit spans Primary Science, PSLE Science and Secondary Science.
Working checkpoint — denoising autoencoder: Let students compare one observation with a distribution, not one favourite control. Separate what the model estimates from the biological story proposed afterward, then name the extra evidence that story would require.
For students, the transferable habit is to ask: What was measured? Compared with what? What alternative explanation remains?
Section 30 of 36
30. Connect to statistics
An outlier is relative to an expected distribution. Sample size, dispersion and multiplicity determine whether a difference is truly surprising.
Working checkpoint — beta-binomial model: Show how the same deviation changes meaning under narrow and wide distributions. Treat identifiers, denominators and factor levels as scientific quantities: a silent mismatch can reverse an otherwise correct calculation.
For parents, a calm checkpoint is to ask how many independent samples support the claim and which assumption matters most.
Section 31 of 36
31. Connect to computing
Sparse matrices, latent models and coordinate joins make reproducible software practice essential. A sample-label swap can create convincing rare outliers everywhere.
Working checkpoint — rare splicing outlier: Use identity checks and immutable sample manifests. Keep the independent sample, design, annotation and software version beside this decision so the evidence can be reconstructed.
For a classroom model, use invented counts and clearly label the activity as an analogy rather than a sequencing experiment.
Section 32 of 36
32. Connect to pathways
Rare-splicing work joins molecular biology, statistics, computing and clinical research. It can illuminate education pathways without promising admission, diagnosis or employment.
Working checkpoint — psi5 metric: Use official institution pages for current pathway choices. Read the numerical result together with sample-level plots, effect size and a sensitivity check capable of weakening the headline.
For future study, connect the idea to biology, mathematics and computing while checking current official pathway information separately.
Section 33 of 36
33. Use current package facts
On 11 October 2026 the official Bioconductor release page lists FRASER 2.8.0 for Bioconductor 3.23. A reproducible report should cite the version actually installed.
Working checkpoint — psi3 metric: Save session information and package sources. Preserve the complete tested family and every excluded feature; a polished shortlist is not a substitute for provenance.
For scientific writing, prefer “supports”, “is associated with” and “is predicted to” when stronger causal verbs are not justified.
Section 34 of 36
34. Did you know?
The same junction can be viewed from its donor side with ψ5 and its acceptor side with ψ3. Those two neighbourhoods can reveal different kinds of competition around one observed connection.
Working checkpoint — theta metric: Inspect both local denominators before narrating the event. Separate what the model estimates from the biological story proposed afterward, then name the extra evidence that story would require.
For students, the transferable habit is to ask: What was measured? Compared with what? What alternative explanation remains?
Section 35 of 36
35. Protect privacy
Rare-disease genomic and transcriptomic data can be identifying. Reproducibility should use controlled access, de-identification and governance rather than publishing sensitive raw files indiscriminately.
Working checkpoint — denoising autoencoder: Document access conditions and share only what consent permits. Treat identifiers, denominators and factor levels as scientific quantities: a silent mismatch can reverse an otherwise correct calculation.
For parents, a calm checkpoint is to ask how many independent samples support the claim and which assumption matters most.
Section 36 of 36
36. Finish with a reproducible bundle
Deliver controlled-access read identifiers, sample metadata, tissue and library protocol, genome and annotation, alignments, counts, FRASER version, latent-dimension evidence, complete results, plots, variant integration and validation status.
Working checkpoint — beta-binomial model: Make every outlier traceable without overstating it. Keep the independent sample, design, annotation and software version beside this decision so the evidence can be reconstructed.
For a classroom model, use invented counts and clearly label the activity as an analogy rather than a sequencing experiment.
Contents · Previous section · Continue to the Science Learning Hub