liveBook · Manning

Sensitivity analysis

As explained in Build AI Drug Discovery Pipelines

Sensitivity analysis evaluates how changes in inputs or assumptions affect a model's outputs.

Sensitivity analysis measures how changing a molecular descriptor changes a model’s prediction. Large output changes indicate greater influence on the target prediction. section 6.4.3

Use two complementary perturbation strategies:

  • Global sensitivity analysis permutes a feature or masks it to simulate removing the descriptor, then compares predictions on the original and modified test sets. section 6.4.3
  • Local sensitivity analysis estimates how small descriptor changes affect one input, using partial derivatives to approximate prediction uncertainty. section 6.4.3

Finite-difference analysis of an RNA binder

For each descriptor of the first ROBIN RNA binder, perturb the value by delta = 0.05 in both directions, predict both variants, and compute the central difference. Plot descriptors whose estimated partial derivative exceeds 0.01.

How Build AI Drug Discovery Pipelines covers it

How deep it goes
Defines global and local procedures, then implements local derivatives with central finite differences.
Examples
Analyzes descriptors for a single RNA binder from the ROBIN data set and visualizes derivatives above 0.01.
What it assumes
Assumes a trained predictive model and descriptor matrix while defining the comparison procedure from first principles.
What you can do after
Diagnose descriptor influence by perturbing features and plotting estimated prediction sensitivities.
How it connects
Connects model-agnostic interpretability with descriptors, partial derivative analysis, and SHAP.

Also explained in

Compare how 2 books explain sensitivity analysis →

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.