liveBook · Manning

Bounding box method

As explained in Build AI Drug Discovery Pipelines

The bounding box method defines a region by the minimum and maximum value of each feature or descriptor dimension.

The bounding box method estimates a model’s applicability domain by enclosing training examples within an n-dimensional hyperrectangle defined by each descriptor’s minimum and maximum. section 4.1.3

Apply the method by:

  • Recording the minimum and maximum value of every descriptor. section 4.1.3
  • Treating new points inside every descriptor range as inside the applicability domain; treat points beyond any range as unreliable or at the model’s edge. section 4.1.3

Solubility prediction

For a solubility model, descriptors such as molecular weight, hydrogen-bond donors, TPSA, logP, ring count, and aromatic proportion define the bounding ranges for new molecules.

How Build AI Drug Discovery Pipelines covers it

How deep it goes
Defines the hyperrectangle and its membership rule, but stops short of handling empty regions or descriptor correlations.
Examples
Uses solubility descriptors including molecular weight, TPSA, logP, hydrogen-bond donors, rings, and aromatic proportion.
What it assumes
Assumes descriptor values and distinguishes interpolation inside observed ranges from extrapolation beyond them.
What you can do after
Apply it to identify new molecules whose descriptor values exceed the training-data ranges.
How it connects
Connects applicability domain with interpolation, extrapolation, descriptor correlation, and convex hull methods.

Related concepts

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.