Compound–target analysis

The strip under the plate counts, for each column, how many of the selected compounds carry a published Kᵢ or Kd at that target. The bar beneath each number is that same number and nothing else — one value drawn twice, not a second measurement. Both are counted off the cells above them, so they follow the current selection rather than the whole archive. Columns are ordered by that count, which is why the bars fall away to the right: most of this pharmacology rests on a handful of targets that have been assayed against nearly everything, and a long tail measured once or twice.

Every cell is a published equilibrium constant — K_i or K_d, the strongest recorded for that pair. Functional readouts are not pooled with them. The whole grid is shown: 163 compounds against 80 molecular targets, from 2,351 measurements. It is mostly empty, and that is a true fact about how thinly this pharmacology has been assayed rather than a fact about the compounds. 24 of those targets were recorded under more than one name — ChEMBL writes “Cannabinoid receptor 1” where the PDSP ingest writes “CB1” — and are folded onto one column by the table in src/lib/targets.ts. Point at a column heading to see the other spellings; the fault and every fold are set out on the integrity report.

Each compound is a vector over all 80 targets of this source, carrying 1/K_i at each target and zero where there is no interaction. A cell is cos θ between two such vectors: 1 is the same direction, 0 is orthogonal. The diagonal is 1 by construction, not by finding. Cosine normalizes each vector by its own length, so the units cancel and the axis needs none.

The reciprocal is what makes the zero honest. 1/K_i tends to zero as K_i tends to infinity, so an empty cell holds the limit of the quantity at “does not bind” rather than a number standing in for one. Measured in pX instead, the same cell would assert a binding constant of one molar — and because a weak target would then contribute a component of the same order as a strong one, the angle would largely track whether two compounds had been assayed at the same targets. Here a target bound at 10 µM contributes about a hundred-thousandth of what one bound at 100 pM does, so the profile is set by the interactions that dominate it and being co-assayed buys nothing on its own.

What the reciprocal does not fix. An empty cell here still means “nobody ran this assay”, which is not the same fact as “this compound does not bind”. Zero is the right value for the second and says nothing about the first, and no complete-vector method can tell them apart. The drug–target plate above is the one that refuses to conflate them — it never fills a gap, and pays for it by comparing a pair only over targets they share. The two orderings disagreeing is informative rather than a fault.

Because every vector is complete, every pair here is comparable and the clustering carries no fallback distance at all. Distances are 1 − cos θ, average linkage; a group is shaded where its members joined within the closest quartile of observed distances.

μ = 4.82, σ = 5.48 over 163 compounds; the sample standard deviation, with n − 1 in the denominator. Both are computed over every compound of this source and over its whole target space, so neither moves when you change the selection or hide a column above.

This is why there is no p-value beside the Z. The distribution is strongly right-skewed and bounded below at zero — most compounds have been assayed at a handful of targets, a few at scores of them. A Z-score needs no distributional assumption to be computed and reads as exactly what it is: how far from typical a compound’s coverage is, in standard deviations. Turning it into a two-tailed p would assume a normal distribution, and this histogram is what that assumption would have to hold for. It also matters what is being standardized — how thoroughly a compound has been studied, which is a fact about the literature rather than about the molecule, and not a thing a hypothesis test has a null for.