Global models and applicability domain

Global models are pretrained models provided by deepmirror for predicting molecular properties. They are built from curated experimental datasets and can be used without training a model on your own data. Each model targets a specific property; its predictions should be interpreted in the context of that property's units and assay conditions.

Applicability domain

The applicability domain indicates how closely a molecule resembles the reference chemistry associated with a model. deepmirror compares its structure with a public reference set and applies a model-specific threshold:

  • In domain: the molecule meets the similarity threshold and has closer structural support in the reference set.
  • Out of domain: the molecule falls below the threshold and its prediction warrants additional caution.

Thresholds are selected using validation results before evaluating performance on separate test data. The label describes structural similarity: an in-domain prediction can still be wrong, and an out-of-domain prediction can still be useful.

Reading the performance metrics

Where sufficient test data is available, deepmirror reports performance separately for in-domain and out-of-domain molecules:

  • MAE: the average prediction error. Lower is better. Check the displayed units, which may use a logarithmic scale.
  • Spearman's ρ: how well predictions rank molecules. Higher is better; values near zero indicate little ranking agreement.

These metrics come from held-out test molecules. Results are calculated within each eligible dataset and then averaged, giving each dataset equal weight. Cross-validation results are assessed separately.

The two groups may contain different datasets and numbers of observations, so in-domain performance will not always appear better. Unavailable metrics indicate insufficient support for reporting them.

Use these scores as evidence about model performance across groups of molecules, rather than as an error estimate for an individual prediction.