liveBook · Manning

Continuity under perturbation

As explained in Build AI Drug Discovery Pipelines

Continuity under perturbation is an evaluation technique that measures how smoothly decoded outputs change when noise is added to a latent representation.

Continuity under perturbation measures whether decoded molecules change smoothly when increasing noise is added to a molecule’s latent representation. A well-structured latent space shows gradual chemical-similarity loss rather than an abrupt transition to invalid molecules. section 10.2.4

A practical evaluation workflow:

  • Encode a molecule into its latent representation. section 10.2.4
  • Systematically increase noise and decode each perturbed representation. section 10.2.4
  • Measure structural similarity between the original and decoded molecules, while monitoring invalid outputs. section 10.2.4

MOSES benchmark evaluation

The evaluate_continuity method tests whether a BasicAutoencoder degrades gracefully as latent noise increases. Interpret a gradual similarity decline as stronger continuity; a sharp drop into invalidity indicates a poorly structured latent space. section 10.2.4

How Build AI Drug Discovery Pipelines covers it

How deep it goes
Defines systematic noise perturbation and decoding, but does not provide the method’s implementation details or results.
Examples
Uses a BasicAutoencoder evaluated on the MOSES benchmark with progressively noised molecular latent representations.
What it assumes
Assumes encoded molecules, latent representations, decoding, and a structural-similarity measure.
What you can do after
Evaluate whether a molecular latent space changes gradually or collapses into dissimilar or invalid decoded structures.
How it connects
Connects latent space, molecular decoding, chemical similarity, interpolation smoothness, and molecular 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.