liveBook · Manning

Threshold-based reward

As explained in Build AI Drug Discovery Pipelines

A threshold-based reward assigns optimization feedback only when candidate properties meet specified minimum or maximum thresholds.

A threshold-based reward gives a molecule optimization feedback only after it meets specified minimum or maximum property values. This turns several drug-design requirements into joint gates rather than allowing one excellent property to compensate for a failure elsewhere. section 13.3.1

Drug-design thresholds

A candidate receives zero reward unless IC50 is below 100 nM, solubility exceeds 10 µM, and synthetic accessibility exceeds 0.5. The approach guarantees these constraints but provides no gradient signal while properties remain far from their thresholds, which can slow early learning. section 13.3.1

Use threshold-based rewards when:

  • Hard requirements matter more than trade-offs among normalized property scores. section 13.3.1
  • Structural constraints such as molecular weight, chiral-center count, or scaffold presence must be enforced separately. section 13.3.1

How Build AI Drug Discovery Pipelines covers it

How deep it goes
Defines threshold gating, explains its learning-signal limitation, and contrasts it with weighted rewards and hard constraints.
Examples
Uses IC50 below 100 nM, solubility above 10 µM, and synthetic accessibility above 0.5 as simultaneous gates.
What you can do after
Apply it via reinforcement-learning reward functions when candidate molecules must satisfy multiple minimum property requirements.
How it connects
Connects reinforcement learning, reward functions, property prediction, constraint satisfaction, and scaffold-constrained generation.

This explainer was generated by Manning's knowledge graph from the text of Build AI Drug Discovery Pipelines. The section links point to the parts of the book it draws on.