UV Absorption Wavelength Predictor (λmax)

This model predicts the wavelength of maximum UV absorption (λmax) of a molecule, expressed in nanometers (nm). λmax indicates at which part of the UV/visible light spectrum a molecule absorbs most strongly — a key property for evaluating and designing UV filters.

Training Data

The model was trained on a combined dataset of two sources:

• Computational data: absorption wavelengths calculated using quantum chemistry simulations (ORCA software) for a broad and diverse set of organic molecules (source: https://doi.org/10.1038/s41597-019-0306-0).
• Experimental data: measured λmax values for known UV filter compounds, sourced from the ECHA (European Chemicals Agency) database.

To ensure that the model performs well for UV filter-like molecules — which are the primary target of this application — the experimental ECHA data was given higher weight during training. Only molecules with λmax ≤ 750 nm were included, focusing the model on the UV and visible light range relevant to sun protection.

The dataset was split into a training set (80%) and a held-out test set (20%) to evaluate how well the model generalises to new, unseen molecules.

How Molecules Are Represented

Molecules are represented using molecular descriptors — a set of numerical values that capture structural and physicochemical properties of a molecule, such as the number and type of rings, the presence of specific chemical groups, and electronic surface properties. These descriptors are computed automatically from the molecular structure (provided as a SMILES string) using the RDKit cheminformatics library.

Feature Selection

Starting from over 200 molecular descriptors, a multi-step selection process was applied to identify the most informative ones:

1. Descriptors with no variation across the dataset were removed.
2. Highly redundant descriptors (those carrying nearly identical information) were eliminated.
3. A recursive feature elimination procedure was used to rank descriptors by their contribution to prediction accuracy, and cross-validation was used to determine the optimal number to retain.

This process resulted in 25 descriptors being selected for the final model:

BertzCT, HallKierAlpha, PEOE_VSA12, PEOE_VSA14, SMR_VSA10, SMR_VSA3, SlogP_VSA3, SlogP_VSA7, SlogP_VSA8, NumAliphaticCarbocycles, NumAliphaticHeterocycles, NumAliphaticRings, NumAromaticRings, NumAtomStereoCenters, NumBridgeheadAtoms, NumHAcceptors, fr_Ar_N, fr_C_S, fr_aldehyde, fr_allylic_oxid, fr_azo, fr_ketone, fr_nitro, fr_nitroso, fr_unbrch_alkane

Algorithm

The model is based on XGBoost (Extreme Gradient Boosting).

Performance