eduKateSG · Why Science?
Relate changes in single-cell expression distributions across measured time windows—without overstating temporal prediction as direct molecular causation
Reading routes
Science learning becomes useful when a familiar object or observation is turned into a system of quantities, mechanisms and claim limits. This guide owns one applied evidence-reading job inside eduKateSG’s wider Science estate. It connects naturally to Why Science Inferelator 3 0 Regularised Regression Context Specific Network Evidence; Why Science Genie3 Tree Ensembles Gene Regulatory Network Inference Evidence; Why Science Pidc Partial Information Decomposition Single Cell Network Evidence; Why Science Panda Message Passing Multi Omic Networks Regulatory Evidence; Education Hub; Singapore Secondary School Directory; Career Adulthood Hub. It also keeps current school and public claims traceable to visible primary sources: SINCERITIES primary study; SINCERITIES full text at PMC; Official SINCERITIES repository; 2026 Singapore–Cambridge O-Level Biology syllabus; MOE G2/G3 Lower Secondary Science syllabus; 2026 MOE G2 Computing syllabus. The sources describe the scientific scope; this article translates that scope into a calm route for Primary Science, PSLE Science, Secondary Science, O-Level Science, STEM exploration, school choices and career pathways without inventing admission or employment outcomes.
Papili Gao and colleagues introduced SINCERITIES in Bioinformatics in 2018 for gene-regulatory-network inference from time-stamped, cross-sectional single-cell expression profiles. The method measures changes in each gene’s expression distribution between consecutive time points, then uses regularised linear regression to infer directed relationships; partial correlation supplies proposed activation or repression signs. The paper evaluated simulated data and single-cell studies including THP-1 cells. Temporal structure strengthens a hypothesis, but it does not remove hidden causes or establish direct binding.
Inside this guide
1–12 · Foundations and models
- 1. Begin with repeated population snapshots
- 2. Look beyond the mean
- 3. Understand KS distance
- 4. Form the temporal regression
- 5. See why ridge regularisation appears
- 6. Treat edge signs as a separate layer
- 7. Read the primary evidence in scope
- 8. Define stimulation and response
- 9. Create a time-stamped manifest
- 10. Balance cells and specimens across time
- 11. Inspect distributions directly
- 12. Handle zeros and detection limits
13–24 · Evidence, testing and applications
- 13. Choose time windows biologically
- 14. Tune regularisation inside training evidence
- 15. Practise with a fictional SINCERITIES table
- 16. Read weight, direction and sign separately
- 17. Audit the KS time series
- 18. Check alternative distribution distances
- 19. Test the number of time points
- 20. Audit the ridge penalty
- 21. Use specimen-level temporal controls
- 22. Construct matched temporal nulls
- 23. Compare with trajectory methods
- 24. Examine indirect paths
25–36 · Learning, decisions and pathways
- 25. Build a temporal evidence card
- 26. Add direct molecular evidence
- 27. Design intervention around the window
- 28. Report uncertainty in activation and repression
- 29. Connect SINCERITIES to school science
- 30. Try a distribution-change activity
- 31. Connect to mathematics
- 32. Connect to computing and AI literacy
- 33. Connect to school and career choices
- 34. Create a family evidence habit
- 35. Use precise temporal language
- 36. Finish with a time-window-to-test bundle
Section 1 of 36
1. Begin with repeated population snapshots
Time-stamped single-cell studies usually measure different cells at each collection. SINCERITIES tracks how every gene’s expression distribution changes between consecutive times and asks whether an earlier change helps explain a later change in another gene.
SINCERITIES uses time-stamped cross-sectional cells and measures how each gene’s expression distribution changes between adjacent collection times.
Archive the time-stamped manifest, cell counts, preprocessing, distance metric, penalty selection, software version and full edge table.
Section 2 of 36
2. Look beyond the mean
A population can change while its average stays similar because variability, tails or subpopulations move. Comparing distributions captures this richness, although each gene is still summarised separately before network fitting.
Kolmogorov–Smirnov distance sees more than a mean but still compresses a full distribution into one change statistic.
Plot cumulative curves and distributions; a KS value cannot reveal whether mean, spread or a small subpopulation moved.
Section 3 of 36
3. Understand KS distance
The published implementation uses the Kolmogorov–Smirnov distance, the largest absolute gap between two cumulative distribution functions. A large distance shows change but not which molecular event caused it.
Ridge regression stabilises an underdetermined temporal model, and its penalty and scaling affect every directed weight.
Leave out each independent biological series and refit; thousands of cells do not replace donor or culture replication.
Section 4 of 36
4. Form the temporal regression
Changes of candidate regulators in one window are related to changes of each target in the next. Earlier features predicting later responses create direction, while hidden causes and indirect paths can still produce edges.
Directed weights and proposed activation or repression signs arise from distinct calculations and need separate uncertainty.
Use temporal nulls preserving counts, smooth trends and specimen blocks rather than an unrealistically random matrix.
Section 5 of 36
5. See why ridge regularisation appears
Genes can greatly outnumber time windows, making ordinary regression underdetermined. Ridge regression shrinks coefficients and stabilises estimation instead of selecting a tiny subset, so the penalty must be reported.
Measured time strengthens temporal reasoning but cannot remove hidden causes, indirect cascades or batch–time confounding.
Keep binding, perturbation and rescue beside temporal regression instead of letting order become direct molecular proof.
Section 6 of 36
6. Treat edge signs as a separate layer
Proposed activation or repression signs come from partial correlation, while directed strengths come from temporal regularised regression. An edge can have stable direction and uncertain sign.
The 2018 simulated and biological applications establish method evidence in those settings, not universal causal certainty.
Use time-stamped distributional change, regularised directed hypothesis and proposed sign until experiments justify stronger language.
Section 7 of 36
7. Read the primary evidence in scope
Papili Gao and colleagues evaluated SINCERITIES on simulated data and profiles including THP-1 cells, then applied it to chicken erythrocyte development. These applications demonstrate the method without making every edge causal.
SINCERITIES uses time-stamped cross-sectional cells and measures how each gene’s expression distribution changes between adjacent collection times.
Archive the time-stamped manifest, cell counts, preprocessing, distance metric, penalty selection, software version and full edge table.
Section 8 of 36
8. Define stimulation and response
State the intervention or developmental trigger, measured times, expected response window and population. A network following one stimulation should not be presented as a permanent cell-type network.
Kolmogorov–Smirnov distance sees more than a mean but still compresses a full distribution into one change statistic.
Plot cumulative curves and distributions; a KS value cannot reveal whether mean, spread or a small subpopulation moved.
Section 9 of 36
9. Create a time-stamped manifest
Record specimen or culture, collection time, batch, treatment, assay, cell count, quality thresholds, identifiers and preprocessing. Note whether time points belong to independent or paired biological series.
Ridge regression stabilises an underdetermined temporal model, and its penalty and scaling affect every directed weight.
Leave out each independent biological series and refit; thousands of cells do not replace donor or culture replication.
Section 10 of 36
10. Balance cells and specimens across time
Unequal cell counts affect distribution precision, while missing specimens can confound time with donor. Plot counts by time and specimen before computing change statistics.
Directed weights and proposed activation or repression signs arise from distinct calculations and need separate uncertainty.
Use temporal nulls preserving counts, smooth trends and specimen blocks rather than an unrealistically random matrix.
Section 11 of 36
11. Inspect distributions directly
For priority genes, show cumulative distributions, density or quantile plots at every time. One KS number hides whether change spans the whole population or only a tail.
Measured time strengthens temporal reasoning but cannot remove hidden causes, indirect cascades or batch–time confounding.
Keep binding, perturbation and rescue beside temporal regression instead of letting order become direct molecular proof.
Section 12 of 36
12. Handle zeros and detection limits
Single-cell expression contains biological zeros, technical dropouts and detection limits. Transformation or imputation can change distributions and partial correlations; imputed values are not directly observed molecules.
The 2018 simulated and biological applications establish method evidence in those settings, not universal causal certainty.
Use time-stamped distributional change, regularised directed hypothesis and proposed sign until experiments justify stronger language.
Section 13 of 36
13. Choose time windows biologically
Closely spaced samples may capture early regulation but show small distribution changes; wide gaps can mix many indirect events. Unequal intervals require explicit treatment and sensitivity checks.
SINCERITIES uses time-stamped cross-sectional cells and measures how each gene’s expression distribution changes between adjacent collection times.
Archive the time-stamped manifest, cell counts, preprocessing, distance metric, penalty selection, software version and full edge table.
Section 14 of 36
14. Tune regularisation inside training evidence
Choose ridge strength through a predeclared or cross-validated procedure that respects specimen and temporal structure. Random cell folds leak donor and time information.
Kolmogorov–Smirnov distance sees more than a mean but still compresses a full distribution into one change statistic.
Plot cumulative curves and distributions; a KS value cannot reveal whether mean, spread or a small subpopulation moved.
Section 15 of 36
15. Practise with a fictional SINCERITIES table
This classroom table is invented and is not output from the SINCERITIES paper.
| Fictional edge | Temporal weight | Proposed sign | Specimen recurrence | Next check |
|---|---|---|---|---|
| TF-A → Gene-B | 0.67 | Positive | 86% | Early knockdown |
| TF-C → Gene-D | 0.59 | Uncertain | 48% | Add time points |
| TF-E → Gene-F | 0.31 | Negative | 82% | Binding and rescue |
Ridge regression stabilises an underdetermined temporal model, and its penalty and scaling affect every directed weight.
Leave out each independent biological series and refit; thousands of cells do not replace donor or culture replication.
Section 16 of 36
16. Read weight, direction and sign separately
Regression weight ranks temporal predictive influence, direction comes from earlier-to-later windows and sign comes from partial correlation. None alone proves direct regulation.
Directed weights and proposed activation or repression signs arise from distinct calculations and need separate uncertainty.
Use temporal nulls preserving counts, smooth trends and specimen blocks rather than an unrealistically random matrix.
Section 17 of 36
17. Audit the KS time series
Plot every gene’s distributional-change statistic across windows. A single shock affecting many genes can create broad predictive relationships, so compare cell viability and global response metrics.
Measured time strengthens temporal reasoning but cannot remove hidden causes, indirect cascades or batch–time confounding.
Keep binding, perturbation and rescue beside temporal regression instead of letting order become direct molecular proof.
Section 18 of 36
18. Check alternative distribution distances
Compare KS with mean difference or another justified distance. Agreement supports robustness; disagreement reveals whether tails, means or other features drive an edge.
The 2018 simulated and biological applications establish method evidence in those settings, not universal causal certainty.
Use time-stamped distributional change, regularised directed hypothesis and proposed sign until experiments justify stronger language.
Section 19 of 36
19. Test the number of time points
Few windows provide little information for high-dimensional regression. Subsample or simulate the design to learn whether the next experiment needs denser early sampling.
SINCERITIES uses time-stamped cross-sectional cells and measures how each gene’s expression distribution changes between adjacent collection times.
Archive the time-stamped manifest, cell counts, preprocessing, distance metric, penalty selection, software version and full edge table.
Section 20 of 36
20. Audit the ridge penalty
Trace coefficients and ranks across a plausible penalty grid. Heavy shrinkage flattens edges; weak shrinkage amplifies unstable combinations. Report tiers when exact rank moves.
Kolmogorov–Smirnov distance sees more than a mean but still compresses a full distribution into one change statistic.
Plot cumulative curves and distributions; a KS value cannot reveal whether mean, spread or a small subpopulation moved.
Section 21 of 36
21. Use specimen-level temporal controls
Repeat inference after leaving out each biological series. If one time point was processed in one batch, time and batch cannot be separated without matched controls.
Ridge regression stabilises an underdetermined temporal model, and its penalty and scaling affect every directed weight.
Leave out each independent biological series and refit; thousands of cells do not replace donor or culture replication.
Section 22 of 36
22. Construct matched temporal nulls
Permute time labels within valid blocks, circularly shift series when appropriate or simulate independent genes with similar distributional change. Preserve cell counts and smoothness.
Directed weights and proposed activation or repression signs arise from distinct calculations and need separate uncertainty.
Use temporal nulls preserving counts, smooth trends and specimen blocks rather than an unrealistically random matrix.
Section 23 of 36
23. Compare with trajectory methods
Contrast SINCERITIES measured-time evidence with SCODE or LEAP pseudotime and with static GENIE3 or PIDC. Agreement prioritises edges; disagreement identifies assumptions.
Measured time strengthens temporal reasoning but cannot remove hidden causes, indirect cascades or batch–time confounding.
Keep binding, perturbation and rescue beside temporal regression instead of letting order become direct molecular proof.
Section 24 of 36
24. Examine indirect paths
An early regulator can affect an intermediate that later affects the target. The temporal model may rank the first gene without resolving the chain; denser sampling and perturbation are needed.
The 2018 simulated and biological applications establish method evidence in those settings, not universal causal certainty.
Use time-stamped distributional change, regularised directed hypothesis and proposed sign until experiments justify stronger language.
Section 25 of 36
25. Build a temporal evidence card
Record time windows, distributions, KS series, ridge path, directed weight, sign evidence, specimen recurrence, null comparison, binding, perturbation and alternative pathways.
SINCERITIES uses time-stamped cross-sectional cells and measures how each gene’s expression distribution changes between adjacent collection times.
Archive the time-stamped manifest, cell counts, preprocessing, distance metric, penalty selection, software version and full edge table.
Section 26 of 36
26. Add direct molecular evidence
Motif, chromatin and binding assays can test whether the regulator could act at the target. Perform them in the relevant time window and cell state.
Kolmogorov–Smirnov distance sees more than a mean but still compresses a full distribution into one change statistic.
Plot cumulative curves and distributions; a KS value cannot reveal whether mean, spread or a small subpopulation moved.
Section 27 of 36
27. Design intervention around the window
Perturb the regulator before its predicted influence window, sample several times and measure the target distribution plus its mean. Include rescue, negatives and an intermediate candidate.
Ridge regression stabilises an underdetermined temporal model, and its penalty and scaling affect every directed weight.
Leave out each independent biological series and refit; thousands of cells do not replace donor or culture replication.
Section 28 of 36
28. Report uncertainty in activation and repression
Partial-correlation signs can change under confounding, scaling or omitted variables. Compare signs across specimens, preprocessing and conditioning sets, leaving unstable edges unsigned.
Directed weights and proposed activation or repression signs arise from distinct calculations and need separate uncertainty.
Use temporal nulls preserving counts, smooth trends and specimen blocks rather than an unrealistically random matrix.
Section 29 of 36
29. Connect SINCERITIES to school science
The method shows why averages can hide variation, why time order matters and why controls must match the design. Students ask what changed and what claim is justified.
Measured time strengthens temporal reasoning but cannot remove hidden causes, indirect cascades or batch–time confounding.
Keep binding, perturbation and rescue beside temporal regression instead of letting order become direct molecular proof.
Section 30 of 36
30. Try a distribution-change activity
Give groups dot plots of fictional cell measurements at four times. Ask whether mean, spread or a subpopulation changed, then draw cumulative curves and locate their largest gap.
The 2018 simulated and biological applications establish method evidence in those settings, not universal causal certainty.
Use time-stamped distributional change, regularised directed hypothesis and proposed sign until experiments justify stronger language.
Section 31 of 36
31. Connect to mathematics
Cumulative distributions, maximum distance, regression, regularisation and partial correlation link biology with mathematics. A hand-drawn KS example shows how a metric compresses a picture.
SINCERITIES uses time-stamped cross-sectional cells and measures how each gene’s expression distribution changes between adjacent collection times.
Archive the time-stamped manifest, cell counts, preprocessing, distance metric, penalty selection, software version and full edge table.
Section 32 of 36
32. Connect to computing and AI literacy
The pipeline transforms distributions into temporal features, regularised fits and graphs. Students learn that predicting what comes next does not automatically identify the real cause.
Kolmogorov–Smirnov distance sees more than a mean but still compresses a full distribution into one change statistic.
Plot cumulative curves and distributions; a KS value cannot reveal whether mean, spread or a small subpopulation moved.
Section 33 of 36
33. Connect to school and career choices
Students excited by populations and time can explore biology, chemistry, mathematics, computing or statistics. Look for opportunities to investigate, code, visualise and communicate evidence.
Ridge regression stabilises an underdetermined temporal model, and its penalty and scaling affect every directed weight.
Leave out each independent biological series and refit; thousands of cells do not replace donor or culture replication.
Section 34 of 36
34. Create a family evidence habit
When an average is reported, ask what individual values looked like, how many independent groups were measured and whether another change happened at the same time.
Directed weights and proposed activation or repression signs arise from distinct calculations and need separate uncertainty.
Use temporal nulls preserving counts, smooth trends and specimen blocks rather than an unrealistically random matrix.
Section 35 of 36
35. Use precise temporal language
Say time-stamped distributional change, regularised directed hypothesis and proposed sign. Reserve causes, activates and represses for experiments supporting those mechanisms.
Measured time strengthens temporal reasoning but cannot remove hidden causes, indirect cascades or batch–time confounding.
Keep binding, perturbation and rescue beside temporal regression instead of letting order become direct molecular proof.
Section 36 of 36
36. Finish with a time-window-to-test bundle
Deliver the sampling manifest, distribution plots, distance sensitivity, penalty paths, specimen leave-one-out results, matched nulls, method comparison, edge cards and timed perturbation queue.
The 2018 simulated and biological applications establish method evidence in those settings, not universal causal certainty.
Use time-stamped distributional change, regularised directed hypothesis and proposed sign until experiments justify stronger language.
Did you know? Two time points can have the same mean expression and still yield a large Kolmogorov–Smirnov distance because their spreads or subpopulations differ. Always show the distributions behind the change statistic so the biology remains visible.
The number of cells at each time affects how precisely a distribution is estimated. Downsample large time points repeatedly to the size of the smallest credible sample, then inspect edge stability. A conclusion driven by unequal cell counts needs a more cautious label.
KS distance is unsigned: it reports how different two distributions are, not whether expression rose or fell. Direction of distributional shift and sign of a regulatory edge require additional summaries. Keep these concepts separate in tables and figures.
Unequal time intervals can make one window accumulate more biological change than another. Consider interval-aware normalisation or analyse equal intervals where possible. At minimum, report the actual timing beside every change statistic and inspect whether long gaps dominate.
Ridge regression shares weight across correlated predictors. This can stabilise estimation but make exact gene interpretation difficult. Examine coefficient paths and regulator families, and design group-aware perturbations when several candidates carry nearly the same temporal information.
Partial correlation signs depend on which variables are conditioned upon. Compare reasonable conditioning sets and specimen-level estimates. If signs disagree, preserve a directed unsigned edge rather than forcing activation or repression into the figure.
Batch and time are dangerous twins. If all early samples were processed on one day and late samples on another, no algorithm can fully separate biology from processing. Balanced experimental design is the strongest correction; statistical adjustment cannot recreate missing overlap.
Use a leave-one-time-window-out exercise to identify dominant intervals. A network collapsing after removal of one transition may still describe an important acute response, but its scope should be that window rather than the full process.
Simulate distributional changes that alter only the mean, only the variance or only a small subpopulation. Running SINCERITIES on these scenarios helps users learn which patterns the workflow detects and where sign or direction can be misunderstood.
For human data, report donor representation at every time and avoid pooling away demographic imbalance. Keep raw profiles in appropriate controlled environments while sharing code, aggregate distributions and derived networks as governance permits.
Students can compare two cumulative-frequency graphs and locate the largest vertical gap. Then ask what different dot plots could produce the same KS distance. The activity teaches that a metric is concise evidence, not the entire observation.
Prioritise experiments that sample before, during and after the predicted window. Measure distribution shape as well as the mean, because a perturbation may change only a subpopulation. Rescue and intermediate measurements can separate direct response from a cascade.
Contents · Previous section · Continue to the Science Learning Hub