Alzheimer's gene co-expression · Workflow

Each row is one probe, a single gene measurement on the microarray; each column is one person. The 31 columns run Control (9 people), Incipient (7), Moderate (8) and Severe (7).

Each group’s matrix is standardized with a single mean and standard deviation computed over all its values, rather than one per probe. t-SNE and K-Means then run on each disease stage separately, with the settings below.

Each dot is one probe. For the chosen disease stage, t-SNE arranges all probes so that those with similar expression patterns across that stage's people sit close together, and K-Means cuts the map into clusters. Anchor’s cluster shows the anchor, the candidates and the other probes in the anchor's cluster at this stage. All clusters gives every cluster its own colour; the colours are labels, not an order.

K-Means needs the number of clusters, k, to be set in advance. The curve scores every k that the original notebook tried with the Calinski–Harabasz score, which is higher when clusters are tight and well separated. Both the chosen k and the best-scoring k are marked.

probe (probe set)

one measurement feature on the microarray. A row of the matrix is one probe; its columns are the people it was measured in.

anchor probe

the probe the search starts from — here the ones for the genes PLD2 and TPP2. A candidate is a probe that shares the anchor's cluster in all three disease stages.

global z-score

as used here: every value in one severity matrix has that whole matrix's mean subtracted and is divided by that whole matrix's standard deviation. One number per matrix, not per row — which is why changing it would move every distance on the map.

t-SNE

a way of placing high-dimensional points on a flat map so that points which were close stay close. As used here, close means similar expression pattern across people. The axes, the compass directions and the distances between separate clumps mean nothing at all.

K-Means, and k

a way of cutting a map into k groups. As used here k is 35, chosen where the quality score starts to level off.

Calinski-Harabasz score

a number that rewards clusters for being tight and far apart. Used here to show where the chosen k sits on the curve.

the anchor’s cluster

the K-Means cluster that contains the anchor on one disease stage’s map. It groups probes by expression pattern.

candidates (the shortlist)

the probes that share the anchor's cluster in all three disease stages. A probe that qualifies in two stages out of three does not appear. The control group is not used.

Pearson correlation, r

how tightly two probes' readings rise and fall together across the people in one stage, from −1 to 1. Computed here on the raw readings, not the standardised ones, and shown per stage because a single number would hide the disagreement between stages.