What a UMAP of chemical space can't tell you

SARVault renders a map of chemical space with every molecule as a point using structural fingerprints and coloring by potency. It is a useful way to explore a large collection of molecules, but it is also easy to read more into the picture than it can actually tell you. This note provides some context.

What the map is good for

The map is good at showing which molecules are similar to each other. Molecules that sit together on the map tend to be genuinely similar according to the fingerprint used to build it.

That makes the map useful for three things: spotting groups of related molecules, finding lonely outliers that might be candidates for closer inspection and navigating a large SAR table. Click on a cluster and you can jump straight to the corresponding series.

As a way to explore nearly two thousand compounds, it earns its screen space.

What the map cannot say

If you want a mental model, think of a subway map. It tells you which stations connect, but the distances and directions on paper do not correspond to the real world. UMAP does something similar for chemical similarity.

There are three things you should not read too much into.

The distance between distant clusters. Two clusters being far apart does not necessarily mean that they are very different. UMAP is mainly concerned with keeping nearby molecules together, not with making the distances between faraway groups meaningful.

The size and density of clusters. A large cluster is not necessarily a large or important chemical family. The algorithm can stretch some regions and compress others as it creates a readable layout.

The axes. The x and y axes have no useful units or direct interpretation, which is why they are left unlabeled. Moving to the right, for example, does not mean that a particular chemical property is increasing.

There is another, quieter caveat: the map can only be as meaningful as the definition of similarity behind it. SARVault uses structural fingerprints, which are one particular way of deciding whether two molecules are similar. Change that definition and you can get a different map.