liveBook · Manning

Underrepresentation

As explained in Build AI Drug Discovery Pipelines

Underrepresentation is insufficient inclusion of a population, class, structure, or condition in a dataset relative to its relevant real-world distribution.

Underrepresentation occurs when training data includes a population, structure, or condition less often than its relevant real-world distribution. This imbalance can reduce model accuracy and limit generalization in drug discovery and clinical prediction. section 13.4.3

Drug-discovery datasets show underrepresentation in several forms:

  • Protein structure data contains 91% soluble proteins but only 3–4% membrane proteins, although membrane proteins comprise about 30% of human proteins. section 13.4.3
  • Small-molecule datasets emphasize compounds within Lipinski’s Rule of Five, leaving larger molecules and macrocycles less represented. section 13.4.3
  • Clinical prediction data underrepresents some racial and ethnic groups, geographic regions, and age groups, weakening generalization across patients. section 13.4.3

How Build AI Drug Discovery Pipelines covers it

How deep it goes
Compares dataset proportions with real-world distributions and traces effects on accuracy and generalization.
Examples
Uses membrane proteins, macrocycles, larger molecules, and underrepresented patient populations as concrete cases.
What you can do after
Evaluate how skewed training data can limit prediction accuracy and generalization across targets and patients.
How it connects
Links training data bias to model accuracy, generalization, clinical prediction, and drug-target classes.

Also explained in

Compare how 2 books explain underrepresentation →

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.