eduKateSG · Why Science?
Follow the differential diagnosis one observation at a time: see which finding raises a candidate, which lowers it and how the prior becomes a post-test probability
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. LIRICAL—LIkelihood Ratio Interpretation of Clinical AbnormaLities—is an open-source application for phenotype-driven prioritisation of candidate diseases and genes in genomic diagnostics. Its official documentation describes HPO-based analysis in phenotype-only mode or with genotype findings from panel, exome or genome sequencing. The method reports likelihood-ratio contributions and post-test probabilities so users can inspect how evidence moved a differential. Those probabilities remain model outputs conditioned on priors, annotations and assumptions; they are not diagnoses or guarantees. 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: Human Phenotype Ontology and structured clinical phenotypes; Gene2Phenotype and structured gene–disease models; DECIPHER and phenotype-linked variant sharing; Science Learning Hub; Education Hub; STEM Education; Careers by Subject Capability.
Primary and official sources checked on 12 October 2026: Official LIRICAL 2.4.1 documentation observed 12 October 2026; Official LIRICAL repository; Interpretable Clinical Genomics with a Likelihood Ratio Paradigm; 2026 Singapore–Cambridge O-Level Biology syllabus; MOE G2/G3 Lower Secondary Science teaching and learning syllabus; 2026 MOE G2 Computing syllabus. Resource interfaces and recommendations can change, so preserve the release, access date and documentation used.
| Stage | Question | Evidence | Responsible output |
|---|---|---|---|
| Prior | How plausible is each candidate before the current findings? | Declared prior model and disease knowledge | A starting probability distribution |
| Update | How does each observed or excluded feature change support? | HPO-specific likelihood ratios | Per-feature evidence contributions |
| Integrate | What changes when compatible genotype evidence is added? | Variant and inheritance likelihood evidence | An updated candidate differential |
| Validate | Does independent review support the leading explanation? | Clinical assessment, variant evidence, segregation and follow-up | A bounded interpretation, not a probability-only diagnosis |
Inside this guide
1–12 · Foundations and design
- 1. Start with a differential
- 2. Understand the likelihood ratio
- 3. Separate priors from evidence
- 4. Read post-test probability conditionally
- 5. Encode observed features with HPO
- 6. Represent excluded features carefully
- 7. Add onset and age context
- 8. Use phenotype-only mode
- 9. Add genotype evidence when appropriate
- 10. Read inheritance compatibility
- 11. Inspect every phenotype contribution
- 12. Do not overcount correlated findings
13–24 · Models and estimation
- 13. Respect annotation frequency
- 14. Understand information content
- 15. Read evidence above one and below one
- 16. Keep rank and probability distinct
- 17. Inspect the full differential
- 18. Use Phenopackets reproducibly
- 19. Check genotype representation
- 20. Challenge the prior
- 21. Challenge annotation bias
- 22. Challenge noisy phenotyping
- 23. Challenge circular evaluation
- 24. Pin data and software versions
25–36 · Diagnostics and interpretation
- 25. Use bounded clinical language
- 26. Protect the case profile
- 27. Connect to Primary Science
- 28. Connect to Secondary Science
- 29. Connect to mathematics
- 30. Connect to computing
- 31. Connect to pathways
- 32. Did you know?
- 33. Write a reader-ready differential
- 34. Build a probability audit
- 35. Build a validation ladder
- 36. Finish with a reproducible bundle
Section 1 of 36
1. Start with a differential
LIRICAL compares multiple candidate diseases rather than asking a model to approve one favourite diagnosis.
Working checkpoint — likelihood ratio: Define the candidate space and intended use before reading probabilities. Keep the phenotype profile, prior model, optional genotype data, ontology, disease data and software version together.
For a classroom model, use fictional clues and candidates rather than personal health histories.
Section 2 of 36
2. Understand the likelihood ratio
A likelihood ratio compares how probable a finding is under a candidate disorder with how probable it is outside that disorder.
Working checkpoint — prior probability: State the numerator, denominator and evidence source. Read each post-test probability beside the likelihood-ratio contributions that moved it.
For future study, connect likelihood ratios to biology, mathematics and computing while checking official pathways separately.
Section 3 of 36
3. Separate priors from evidence
The prior describes the starting plausibility; likelihood evidence updates it. Mixing these roles hides why a candidate rose.
Working checkpoint — post-test probability: Record the prior model explicitly. Preserve the full differential and near-ties; a top diagnosis list is not the complete model output.
For scientific writing, say “post-test probability under these inputs” rather than “chance the person has the disease”.
Section 4 of 36
4. Read post-test probability conditionally
A post-test probability is conditional on the supplied data, annotations, model and candidate set.
Working checkpoint — phenotype contribution: Report those conditions beside the value. Separate probability under the model from clinical truth and independent validation.
For students, ask which observation raised a candidate, which lowered it and which was never assessed.
Section 5 of 36
5. Encode observed features with HPO
HPO terms make clinical observations computable, while original notes keep nuance that an ontology code may omit.
Working checkpoint — differential diagnosis: Preserve the mapping and source wording. Treat negation, onset, correlated terms and priors as scientific inputs: a silent change can transform the differential.
For parents, ask what assumptions and prior information sit behind the displayed probability.
Section 6 of 36
6. Represent excluded features carefully
A genuinely absent feature can lower some candidates, but unassessed or age-dependent features should not be treated as absent.
Working checkpoint — genotype evidence: Require a defensible assessment before adding a negated term. Keep the phenotype profile, prior model, optional genotype data, ontology, disease data and software version together.
For a classroom model, use fictional clues and candidates rather than personal health histories.
Section 7 of 36
7. Add onset and age context
The probability of observing a feature can depend on when the person was examined and when the feature usually appears.
Working checkpoint — likelihood ratio: Keep onset and observation age with the phenotype profile. Read each post-test probability beside the likelihood-ratio contributions that moved it.
For future study, connect likelihood ratios to biology, mathematics and computing while checking official pathways separately.
Section 8 of 36
8. Use phenotype-only mode
LIRICAL can rank diseases and genes from HPO observations alone, which is useful when genotype evidence is unavailable or intentionally separated.
Working checkpoint — prior probability: Call the result a phenotype-driven differential. Preserve the full differential and near-ties; a top diagnosis list is not the complete model output.
For scientific writing, say “post-test probability under these inputs” rather than “chance the person has the disease”.
Section 9 of 36
9. Add genotype evidence when appropriate
With sequencing data, LIRICAL can combine phenotypic and genotypic evidence under candidate disease and inheritance models.
Working checkpoint — post-test probability: Verify assembly, VCF samples and pedigree before integration. Separate probability under the model from clinical truth and independent validation.
For students, ask which observation raised a candidate, which lowered it and which was never assessed.
Section 10 of 36
10. Read inheritance compatibility
A genotype contributes differently under dominant, recessive, X-linked or other models.
Working checkpoint — phenotype contribution: Explain which allelic configuration was evaluated. Treat negation, onset, correlated terms and priors as scientific inputs: a silent change can transform the differential.
For parents, ask what assumptions and prior information sit behind the displayed probability.
Section 11 of 36
11. Inspect every phenotype contribution
One strength of the framework is that individual observations can show how they raised or lowered a candidate.
Working checkpoint — differential diagnosis: List the decisive positive and negative contributions. Keep the phenotype profile, prior model, optional genotype data, ontology, disease data and software version together.
For a classroom model, use fictional clues and candidates rather than personal health histories.
Section 12 of 36
12. Do not overcount correlated findings
Several closely related HPO terms may describe one biological feature, making independence assumptions optimistic.
Working checkpoint — genotype evidence: Collapse redundant descriptions in a sensitivity analysis. Read each post-test probability beside the likelihood-ratio contributions that moved it.
For future study, connect likelihood ratios to biology, mathematics and computing while checking official pathways separately.
Section 13 of 36
13. Respect annotation frequency
Likelihood calculations depend on how often features are annotated for diseases, and rare-disease records can be incomplete.
Working checkpoint — likelihood ratio: Treat missing annotation as an evidence limitation. Preserve the full differential and near-ties; a top diagnosis list is not the complete model output.
For scientific writing, say “post-test probability under these inputs” rather than “chance the person has the disease”.
Section 14 of 36
14. Understand information content
Specific phenotypes can be more discriminating than broad ancestors in an ontology, but specificity must remain clinically accurate.
Working checkpoint — prior probability: Prefer the narrowest justified term. Separate probability under the model from clinical truth and independent validation.
For students, ask which observation raised a candidate, which lowered it and which was never assessed.
Section 15 of 36
15. Read evidence above one and below one
A likelihood ratio above one supports a candidate; below one weighs against it under the model. Neither is absolute proof.
Working checkpoint — post-test probability: Translate direction and magnitude in plain language. Treat negation, onset, correlated terms and priors as scientific inputs: a silent change can transform the differential.
For parents, ask what assumptions and prior information sit behind the displayed probability.
Section 16 of 36
16. Keep rank and probability distinct
The first-ranked candidate can still have a low probability, while several close candidates may remain credible.
Working checkpoint — phenotype contribution: Report both values and the gap. Keep the phenotype profile, prior model, optional genotype data, ontology, disease data and software version together.
For a classroom model, use fictional clues and candidates rather than personal health histories.
Section 17 of 36
17. Inspect the full differential
A short top list can hide near-ties, broad syndromes or candidates penalised by one uncertain feature.
Working checkpoint — differential diagnosis: Archive all results and filter only for presentation. Read each post-test probability beside the likelihood-ratio contributions that moved it.
For future study, connect likelihood ratios to biology, mathematics and computing while checking official pathways separately.
Section 18 of 36
18. Use Phenopackets reproducibly
Structured case representations can carry phenotypes and related context, but schema and identifier versions still matter.
Working checkpoint — genotype evidence: Pin the Phenopacket and ontology versions. Preserve the full differential and near-ties; a top diagnosis list is not the complete model output.
For scientific writing, say “post-test probability under these inputs” rather than “chance the person has the disease”.
Section 19 of 36
19. Check genotype representation
The same allele can be encoded differently, and transcript or consequence choices can affect compatibility.
Working checkpoint — likelihood ratio: Normalise variants and retain the source form. Separate probability under the model from clinical truth and independent validation.
For students, ask which observation raised a candidate, which lowered it and which was never assessed.
Section 20 of 36
20. Challenge the prior
Geographic, ancestry, referral and ascertainment context can make a generic prior poorly calibrated for one case.
Working checkpoint — prior probability: Run a justified alternative prior analysis. Treat negation, onset, correlated terms and priors as scientific inputs: a silent change can transform the differential.
For parents, ask what assumptions and prior information sit behind the displayed probability.
Section 21 of 36
21. Challenge annotation bias
Well-studied disorders usually have richer phenotype descriptions than newly recognised ones.
Working checkpoint — post-test probability: Consider whether knowledge depth favours familiar diseases. Keep the phenotype profile, prior model, optional genotype data, ontology, disease data and software version together.
For a classroom model, use fictional clues and candidates rather than personal health histories.
Section 22 of 36
22. Challenge noisy phenotyping
Incorrect, broad or copied HPO terms can redirect a differential with false precision.
Working checkpoint — phenotype contribution: Trace every decisive term to a documented observation. Read each post-test probability beside the likelihood-ratio contributions that moved it.
For future study, connect likelihood ratios to biology, mathematics and computing while checking official pathways separately.
Section 23 of 36
23. Challenge circular evaluation
Benchmark cases whose publications informed disease annotations can inflate apparent performance.
Working checkpoint — differential diagnosis: Use time-split or held-out knowledge where possible. Preserve the full differential and near-ties; a top diagnosis list is not the complete model output.
For scientific writing, say “post-test probability under these inputs” rather than “chance the person has the disease”.
Section 24 of 36
24. Pin data and software versions
Disease annotations, ontology structure and software releases change the same case result over time.
Working checkpoint — genotype evidence: Archive the executable, data releases and command. Separate probability under the model from clinical truth and independent validation.
For students, ask which observation raised a candidate, which lowered it and which was never assessed.
Section 25 of 36
25. Use bounded clinical language
LIRICAL supports an interpretable differential; it does not diagnose, prescribe or replace clinical review.
Working checkpoint — likelihood ratio: Write “modelled probability under these inputs”. Treat negation, onset, correlated terms and priors as scientific inputs: a silent change can transform the differential.
For parents, ask what assumptions and prior information sit behind the displayed probability.
Section 26 of 36
26. Protect the case profile
Rare combinations of features and variants can identify a person even without direct identifiers.
Working checkpoint — prior probability: Minimise and govern every shared input. Keep the phenotype profile, prior model, optional genotype data, ontology, disease data and software version together.
For a classroom model, use fictional clues and candidates rather than personal health histories.
Section 27 of 36
27. Connect to Primary Science
A fictional clue game shows that one observation can support one explanation and weaken another without settling the case.
Working checkpoint — post-test probability: Use invented traits and transparent rules. Read each post-test probability beside the likelihood-ratio contributions that moved it.
For future study, connect likelihood ratios to biology, mathematics and computing while checking official pathways separately.
Section 28 of 36
28. Connect to Secondary Science
Genetics and phenotype become evidence that updates competing biological explanations.
Working checkpoint — phenotype contribution: Separate observation, likelihood and mechanism. Preserve the full differential and near-ties; a top diagnosis list is not the complete model output.
For scientific writing, say “post-test probability under these inputs” rather than “chance the person has the disease”.
Section 29 of 36
29. Connect to mathematics
Prior odds, likelihood ratios and post-test odds make Bayesian updating visible and checkable.
Working checkpoint — differential diagnosis: Work a small example by hand before using software. Separate probability under the model from clinical truth and independent validation.
For students, ask which observation raised a candidate, which lowered it and which was never assessed.
Section 30 of 36
30. Connect to computing
Ontologies, structured files, disease databases and reproducible commands make the calculation possible.
Working checkpoint — genotype evidence: Validate identifiers and stop on unmapped terms. Treat negation, onset, correlated terms and priors as scientific inputs: a silent change can transform the differential.
For parents, ask what assumptions and prior information sit behind the displayed probability.
Section 31 of 36
31. Connect to pathways
Interpretable genomic reasoning links biology, medicine, statistics and software and can illuminate STEM pathways.
Working checkpoint — likelihood ratio: Check official study and professional requirements separately. Keep the phenotype profile, prior model, optional genotype data, ontology, disease data and software version together.
For a classroom model, use fictional clues and candidates rather than personal health histories.
Section 32 of 36
32. Did you know?
LIRICAL is designed to show how individual phenotypic observations contribute to a candidate rather than giving only a mysterious rank.
Working checkpoint — prior probability: Read the per-feature contribution table. Read each post-test probability beside the likelihood-ratio contributions that moved it.
For future study, connect likelihood ratios to biology, mathematics and computing while checking official pathways separately.
Section 33 of 36
33. Write a reader-ready differential
Name the prior, case profile, genotype scope, leading candidates, probabilities, decisive contributions and unresolved alternatives.
Working checkpoint — post-test probability: Do not compress uncertainty into one label. Preserve the full differential and near-ties; a top diagnosis list is not the complete model output.
For scientific writing, say “post-test probability under these inputs” rather than “chance the person has the disease”.
Section 34 of 36
34. Build a probability audit
Recalculate one candidate from prior through major likelihood contributions and compare it with the software output.
Working checkpoint — phenotype contribution: Investigate every unexplained discrepancy. Separate probability under the model from clinical truth and independent validation.
For students, ask which observation raised a candidate, which lowered it and which was never assessed.
Section 35 of 36
35. Build a validation ladder
Verify phenotype mapping, test correlated terms and priors, inspect genotype evidence, then seek clinical and laboratory confirmation.
Working checkpoint — differential diagnosis: Strengthen the evidence that moves the result most. Treat negation, onset, correlated terms and priors as scientific inputs: a silent change can transform the differential.
For parents, ask what assumptions and prior information sit behind the displayed probability.
Section 36 of 36
36. Finish with a reproducible bundle
Deliver permitted Phenopacket or HPO input, VCF references, priors, versions, command, complete outputs, audit table and validation status.
Working checkpoint — genotype evidence: Make the probability path reconstructable. Keep the phenotype profile, prior model, optional genotype data, ontology, disease data and software version together.
For a classroom model, use fictional clues and candidates rather than personal health histories.
Contents · Previous section · Continue to the Science Learning Hub