eduKateSG · Why Science?
Turn sequence clues into an interpretable priority score—then return to RNA evidence before making a clinical claim
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. SQUIRLS—Super-quick Information Content and Random Forest Learning for Splice Variants—is an interpretable machine-learning tool for prioritising genomic variants by their predicted effect on mRNA splicing. It computes hand-designed sequence and annotation features, uses models such as random forests and logistic regression, and can annotate positions, CSV files or VCFs. Its score helps order candidates for review; it is not a diagnosis and does not replace transcript evidence, phenotype matching or clinical interpretation. Because the model exposes feature-level explanations, readers can examine why a variant was prioritised instead of treating one probability-like value as an oracle. 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: SpliceAI and sequence-based splice-variant prediction; MMSplice and modular splice-variant effects; FRASER and rare aberrant-splicing detection; Science Learning Hub; Education Hub; STEM Education; Careers by Subject Capability.
Primary and official sources checked on 11 October 2026: Official SQUIRLS repository; Official SQUIRLS 2.0 documentation; SQUIRLS primary American Journal of Human Genetics study; Jackson Laboratory overview of SQUIRLS; 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 |
|---|---|---|---|
| Locate | Where is the variant relative to transcript features? | Genome build, transcript annotation and allele | A defined variant–transcript context |
| Describe | Which interpretable splicing features change? | Information content, motif and positional features | A feature vector with biological labels |
| Score | How strongly does the trained model prioritise it? | Random-forest or logistic estimator | A model score for ranking |
| Validate | Does RNA and phenotype evidence agree? | Independent transcript and clinical evidence | A reviewed candidate, not an automated diagnosis |
Inside this guide
1–12 · Foundations and design
- 1. Expand the name
- 2. Define prioritisation
- 3. Choose the reference assembly
- 4. Map transcript context
- 5. Understand information content
- 6. Use positional features
- 7. Inspect motif-related features
- 8. Understand the random forest
- 9. Understand logistic regression
- 10. Use supported input modes
- 11. Read HTML reports
- 12. Read VCF annotations
13–24 · Models and estimation
- 13. Read variant categories
- 14. Check training labels
- 15. Avoid score transplantation
- 16. Compare models on matched cases
- 17. Study disagreements
- 18. Challenge canonical-site confidence
- 19. Challenge deep intronic confidence
- 20. Challenge exonic variants
- 21. Consider tissue expression
- 22. Consider incomplete annotation
- 23. Inspect false positives
- 24. Inspect false negatives
25–36 · Diagnostics and interpretation
- 25. Use orthogonal RNA evidence
- 26. Integrate phenotype cautiously
- 27. Use bounded diagnostic language
- 28. Connect to Primary Science
- 29. Connect to Secondary Science
- 30. Connect to mathematics
- 31. Connect to computing
- 32. Connect to pathways
- 33. Did you know?
- 34. Audit one candidate end to end
- 35. Build a validation ladder
- 36. Finish with a reproducible bundle
Section 1 of 36
1. Expand the name
SQUIRLS stands for Super-quick Information Content and Random Forest Learning for Splice Variants. The name signals both interpretable sequence features and a trained classifier.
Working checkpoint — information-content feature: Use the expanded name once and then describe the actual prediction task. 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. Define prioritisation
The tool ranks variants by predicted splice impact so reviewers can inspect the most plausible candidates earlier. Prioritisation changes review order; it does not determine truth automatically.
Working checkpoint — random forest estimator: Write the downstream decision and acceptable false-positive burden. 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 the reference assembly
Variant coordinates only have meaning against a specific genome assembly. Resources and transcripts used by SQUIRLS must match that assembly.
Working checkpoint — splice-variant score: Record build and verify reference alleles before annotation. 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. Map transcript context
A variant can affect different transcripts differently because exon boundaries and strands differ. The same genomic coordinate may therefore generate several annotations.
Working checkpoint — variant category: Retain transcript-specific rows until the review rule is explicit. 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. Understand information content
Splice-site motifs can be described by how informative each base is at each position relative to observed splice sites. A variant may weaken or strengthen that pattern.
Working checkpoint — VCF annotation: Treat the feature as a modelled motif signal, not direct spliceosome binding. 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. Use positional features
Distance to donor, acceptor, exon or intron boundaries helps distinguish canonical-site variants from nearby and deeper changes. Position contributes context but does not prove effect.
Working checkpoint — orthogonal validation: Report distance and transcript orientation beside the score. 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. Inspect motif-related features
SQUIRLS includes features designed to represent sequence signals linked to splicing. Each is an approximation of biological recognition within a trained model.
Working checkpoint — information-content feature: Use the HTML explanation to see which features influenced a candidate. 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. Understand the random forest
A random forest combines decisions from many trees trained on labelled variants. Nonlinear interactions can improve prediction while feature views preserve some interpretability.
Working checkpoint — random forest estimator: Record model and resource versions, not only the executable name. 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. Understand logistic regression
The documentation also describes logistic-regression modelling. A linear combination is easier to trace but still depends on training labels and feature scaling.
Working checkpoint — splice-variant score: Do not compare scores from different estimators as if they share one calibration. 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. Use supported input modes
SQUIRLS can annotate individual positions, CSV records and VCF files. Input mode changes workflow convenience, not the biological evidence behind a score.
Working checkpoint — variant category: Validate parsing, allele order and sample fields on a small test set. 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. Read HTML reports
The HTML output can display scores, variant categories and explanatory figures. A readable report supports review but can still conceal an upstream coordinate mismatch.
Working checkpoint — VCF annotation: Keep machine-readable output and provenance beside the visual report. 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. Read VCF annotations
Adding scores to VCF records makes them portable through pipelines. Header definitions, transcript multiplicity and field versions must travel with the file.
Working checkpoint — orthogonal validation: Preserve the original VCF and a new provenance-rich annotated copy. 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. Read variant categories
The documentation groups variants by splice-relevant location or interpretation. Categories help organise review but should not be mistaken for clinical classifications.
Working checkpoint — information-content feature: Use “category” and “predicted impact” rather than “pathogenic class”. 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. Check training labels
A model learns from available labelled variants and negatives. Bias in genes, variant positions, assays or reporting can shape what the model recognises best.
Working checkpoint — random forest estimator: Describe the benchmark population before generalising performance. 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. Avoid score transplantation
A threshold tuned on one dataset may not retain the same sensitivity, specificity or predictive value in a new diagnostic or research cohort.
Working checkpoint — splice-variant score: Recalibrate or report ranking performance on representative data. 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. Compare models on matched cases
SpliceAI, MMSplice and SQUIRLS predict related but distinct targets using different features and outputs. Raw scores cannot be placed on one universal scale.
Working checkpoint — variant category: Use the same measured variant set and outcome definition. 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. Study disagreements
When predictors disagree, the disagreement can reveal different sequence windows, transcript choices or learned assumptions. It is evidence for review, not a reason to select the preferred score.
Working checkpoint — VCF annotation: Inspect features and measured RNA rather than majority voting blindly. 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. Challenge canonical-site confidence
Variants at invariant splice dinucleotides often look compelling, but the expressed transcript, rescue sites and measured product still matter.
Working checkpoint — orthogonal validation: Confirm which transcript is expressed in the relevant tissue. 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. Challenge deep intronic confidence
Far-from-boundary variants are numerous, making false positives costly. A predicted cryptic site needs evidence that it is used and changes transcript structure.
Working checkpoint — information-content feature: Review local sequence, junction reads and an appropriate assay. 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. Challenge exonic variants
An exonic change may affect coding, splicing or both. Synonymous status does not rule out splice impact, while a splice prediction does not erase a protein-level consequence.
Working checkpoint — random forest estimator: Keep coding and splicing evidence as separate columns. 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. Consider tissue expression
No genomic predictor directly observes whether a gene is expressed in the sampled tissue or whether regulatory factors favour a predicted event.
Working checkpoint — splice-variant score: Pair sequence prediction with tissue-aware RNA evidence. 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. Consider incomplete annotation
A missed transcript boundary or novel exon can alter feature calculation and interpretation. Annotation is a model of known transcripts, not the complete transcriptome.
Working checkpoint — variant category: Inspect decisive loci against current annotations and RNA tracks. 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. Inspect false positives
A candidate can score highly yet show no relevant RNA change. Mapping errors, uninformative tissue and imperfect model calibration are different explanations.
Working checkpoint — VCF annotation: Classify the reason before calling the predictor wrong. 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. Inspect false negatives
A low score does not prove absence of splice effect, especially outside well-represented training patterns or with complex variants.
Working checkpoint — orthogonal validation: Keep measured aberrant RNA evidence even when prediction is low. 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. Use orthogonal RNA evidence
Patient RNA-seq, targeted RT-PCR, minigene assays and long reads answer different aspects of a splice hypothesis.
Working checkpoint — information-content feature: Choose an assay that distinguishes the predicted products. 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. Integrate phenotype cautiously
Phenotype matching can raise or lower the priority of a gene, but it should not retroactively change what the sequence model measured.
Working checkpoint — random forest estimator: Keep phenotype, inheritance, RNA and predictor evidence separable. 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. Use bounded diagnostic language
A SQUIRLS score supports computational prioritisation of splice-impact candidates. It does not alone establish variant pathogenicity or a molecular diagnosis.
Working checkpoint — splice-variant score: Follow current professional interpretation frameworks where clinical use is intended. 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. Connect to Primary Science
Classification and fair comparison begin with visible features and agreed rules. A model adds weights, but students can still ask which observations drive the decision.
Working checkpoint — variant category: Use harmless card-sorting examples to teach prioritisation. 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 Secondary Science
DNA variants, RNA splicing and protein-coding consequences form a chain of hypotheses. Evidence at one link does not automatically prove all later links.
Working checkpoint — VCF annotation: Draw the chain and label observed versus predicted steps. 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 mathematics
Probabilities, ranks, sensitivity and specificity show why a score is not certainty. Predictive value also depends on how common true effects are in the tested set.
Working checkpoint — orthogonal validation: Use a confusion matrix with invented 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 31 of 36
31. Connect to computing
Command-line tools, VCF schemas, resource databases and model files make the prediction reproducible. Interfaces can change even when the scientific idea remains.
Working checkpoint — information-content feature: Pin the release and save configuration plus checksums. 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
Interpretable genomics combines biology, information theory, statistics and software engineering and can illuminate STEM education and career pathways.
Working checkpoint — random forest estimator: Use current official education sources for actual course decisions. 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. Did you know?
SQUIRLS was designed to make prioritisation interpretable: its name explicitly mentions information content and random forests, and its reports expose feature-level reasoning.
Working checkpoint — splice-variant score: Interpretability helps audit a model; it does not guarantee correctness. 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. Audit one candidate end to end
Trace coordinate, allele, transcript, feature values, estimator, score, category, phenotype and RNA evidence for one variant.
Working checkpoint — variant category: An auditable row is more valuable than an unexplained top-ten list. 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. Build a validation ladder
Begin with assembly and transcript checks, compare explanatory features, review independent predictors and then seek measured RNA and clinical evidence.
Working checkpoint — VCF annotation: Escalate evidence in proportion to the decision stakes. 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 original and annotated variants, assembly and transcript resources, SQUIRLS version, estimator and resource versions, complete outputs, review notes, RNA evidence and final bounded interpretation.
Working checkpoint — orthogonal validation: Ensure every prioritised variant can be reconstructed without the HTML alone. 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