Chemical space(s), a concept still evolving

⚡ An interesting preprint, Growth vs. Diversity: A Time-Evolution Analysis of the Chemical Space, by a group that has done a lot of work on the concept of chemical space or "chemical multiverse" (a nice review by some of the same authors in Open Access! link), is revisiting the age-old (well, perhaps Lipinski-old) question of "diversity" in chemical libraries. A commendable effort that also highlights the limits of the concept itself and the need for its continued exploration.

😵‍💫 As a non-chemist, I have always enjoyed the concept of chemical space, with its implicit promise of bringing order to the messy world of molecules.

🪠 Of course that promise is broken upon even cursory inspection because there are so many ways to define chemical space: you first have to decide on the underlying topology. I like the Wikipedia definition (which is itself extracted from a 2017 article link): "Chemical space is a concept in cheminformatics referring to the property space spanned by all possible molecules and chemical compounds adhering to a given set of construction principles and boundary conditions." The construction principles define the topology, how to "go" from one molecule to another, and neighboring molecules are those that can be obtained by applying a minimal operator. And there you have the first challenge: different construction principles can lead to very different neighborhoods and therefore distances (neighbor molecules are at distance 1) which can lead to very different clusters as clustering or dimension-reduction algorithms rely on distance.

💰 The second challenge is how to connect the topology to a relevant drug-related question: what do molecules do? So now we are moving from chemical space to CHEMICAL LANDSCAPE, because each point in the space has a value vector attached to it. And that's a very high-dimensional vector! In practice we use only a small subset of the values: selected in silico assays, such as qSAR (quantitative structure-activity relationship) and docking models. Some of these models are used for all molecules (e.g., solubility, metabolic stability) but others are specific to the target. Two molecules may have the same activity profile for one target and very different for another. That may even be true for two neighboring molecules, in which case, if there is a greater than 100-fold affinity difference between them, there is an ACTIVITY CLIFF (see for example the cool paper by some of the authors of the preprint: link).

💡 So, a structurally diverse library might cluster in narrow regions of bioactivity space, limiting therapeutic utility. But the preprint suggests that structural diversity has not increased very much since 2011, certainly not as would be expected from the library size increases.