Chemical space(s), a concept still evolving, part II

💡 Or, evolving on chemical space(s)!

The appeal of the concept for me comes from the fact that the topology of a chemical space (the "construction operators") and its boundary conditions are so natural to map to an Evolutionary Algorithm: construction operators map to evolutionary operators (mutation, crossover) and boundary conditions to constraints on the solutions.

1️⃣ Swapping R groups or "mutating" moieties from large libraries of acceptable groups and moieties plus a grammar of combinations,

2️⃣ under calculated constraints of, say, ranging from simple molecular weight to synthetic feasibility (some constraints are easy to compute, some are hard, some need to be estimated).

3️⃣ The next piece required to complete the picture is a fitness (or objective) function to transform the space into a landscape. Some of the soft constraints can be moved into the objective by adding a penalty to how far a molecule deviates from the boundary. The core of the fitness, though, is a high-dimensional objective: the "optimization" problem is therefore a (many-)multi-objective one.

4️⃣ Expanding on this last point, we (humans) are not great at navigating a Pareto surface even with only two objectives (a Pareto line in this case), so now imagine 5 or 10 or 10,000. To cope with the complexity of the issue we tend to optimize for one thing at a time, usually starting with the star of the show, potency, then moving down the list of objectives. That usually does not end well. So in order to navigate this nightmarish landscape, we need (1) the right operators, starting with the right representation(s), and (2) the ability to interact with the navigator.

I mentioned it in a previous post, in 2006 🦕 , we used a Markush representation to navigate various (small molecule) chemical spaces, which allowed for navigation and interaction -adding the 2008 paper here for a refresh. But the work on representations of small molecules has made a lot of progress since then. I have been out of the game for a long time and I am curious to see/know how EAs have been used in the last 10 years with these new representations.

🫠 On that note, I am a superfan of the brilliant (well, I am a groupie) Alán Aspuru-Guzik who has designed interesting representations over the years, i.e., SELFIES. I just found their work accepted for ICLR 2025: Efficient Evolutionary Search Over Chemical Space with Large Language Models

(link). From the abstract: "we redesign crossover and mutation operations in EAs using LLMs trained on large corpora of chemical information". What's not to love?? ❤️