Loading animations
Some things are far easier to watch than to read about. These explainers cover the concepts students get stuck on — how sequencing actually works, why we run FastQC, how a read finds its place in the genome, and how a neural network learns.
Statistics
Test 20,000 genes where nothing is really changed, and watch 1,000 of them come out significant. Then compare Bonferroni and Benjamini–Hochberg on the same data.
Statistics
Build the null hypothesis by shuffling labels rather than by quoting a formula, then watch where your result lands in it. Then run the same study forty times and see the p-value jump about.
Statistics
Two numbers per gene and two lines drawn across them. Drag the lines and watch the gene list you would take to the bench change, with nothing about the biology moving at all.
Statistics
Watch a 200-gene matrix collapse into two axes, then inject a batch effect and see PC1 quietly stop being about your treatment — with nothing in the plot saying so.
Statistics
Four cell types at known, deliberately unequal distances, run through the real t-SNE and UMAP. The neighbourhoods survive; the gaps, the spacing and the blob sizes do not.
Statistics
Run both methods on three real groups, on two elongated bands, and on a single featureless cloud. All three return clusters, with respectable scores, and the pictures look equally convincing.
Statistics
Raise the flexibility of a model and training error falls to zero — including on data where the outcome is pure noise. A training R² of 1.000 on nothing at all.
Statistics
Pick the best genes using all your patients, then cross-validate: 98% accuracy on data where the labels were assigned at random. Move one line inside the loop and it drops to chance.
Statistics
Hold the classifier completely fixed and move only how rare the positives are. The ROC curve does not budge; precision falls from 87% to 4%, and 25 benign variants get flagged for every real one.
Statistics
A real least-squares surface for a dose-response fit. Too high a learning rate blows up and everyone notices; too low goes quietly flat and looks exactly like convergence while the fit is nowhere near.