Chemist Insight Is First-Class Data

By Blaise AI Team

Medicinal chemistry runs on judgement.

Not the vague kind. The sharp, hard-earned kind that shows up as a throwaway comment in a design meeting:

  • “That substitution will spike clearance.”
  • “This looks like a P-gp magnet.”
  • “We’re about to grow a rotatable bond tax.”
  • “This series is going to fall apart on solubility.”

Most software discards that signal.

Blaise treats it as first-class data.

The key idea: feedback is an asset, not an annotation

In a real program, the scarce resource isn’t compute. It’s experienced attention.

Chemists see nuance instantly because they’ve internalised:

  • series-specific structure–property trade-offs
  • project history and priors
  • the unspoken constraints (synthesis, IP, tox, developability)

If your system can’t capture that, it will keep repeating the same mistakes with better typography.

What “first-class” means in practice

In Blaise, comments and critique are not sticky notes.

They are structured signals that plug back into the design loop:

  • the critique is linked to the exact molecules/series it refers to
  • the system learns what a given team calls “good” or “bad” in context
  • the scoring and proposal logic adapts immediately

That closes a loop most “AI for chemistry” products leave open.

Why this beats static scoring functions

Static property predictors are brittle:

  • they miss project-specific constraints
  • they over-generalise across chemical space
  • they ignore the implicit multi-objective nature of optimisation

Chemist feedback is the missing term in the objective.

Capture it, and the system stops being a demo and starts becoming a colleague.

The compounding effect

The first time you capture critique, you get a better next suggestion.

The hundredth time, you’ve built something more important:

  • a project memory of what your team learned
  • an auditable design trail
  • a reusable set of “house rules” grounded in actual outcomes

That’s how you make AI useful in drug discovery: not by pretending models are omniscient, but by wiring expert judgement into the loop.