Science & Validation
Rigorous evaluation across seven photophysical properties against hold-out synthetic and experimental sets.
Trained on 20,796 points
The underlying GIN is trained on a highly curated dataset of over 20,000 distinct chromophore–solvent measurement points. Our evaluation methodology separates identical scaffolds into the training block to prevent data leakage, ensuring test-set performance reflects genuine generalisation to novel compounds.
Model performance
Wavelength targeting and lifetime are the two strongest predictors - the properties that matter most for first-stage screening. We publish detailed metrics on all seven properties.
Appropriate role in screening
PyLabs is built as a pre-screening funnel ahead of TD-DFT or synthesis. It is exceptionally accurate at rapidly discarding poor candidates (e.g., incorrect absorption regions, negligible quantum yield) across vast virtual libraries.
While correlation is very high within the applicability domain, the model does not replace high-fidelity computational methods for final characterisation. Treat PyLabs as your compass, not the destination.