Smiles2Class predicts which molecular classes a structure belongs to, directly from its SMILES string. A structure can belong to several classes at once.

Input and output

• Input: a SMILES string.
• Output: a present/absent probability for each of 426 classes. The classes are independent binary decisions, not a single exclusive assignment.

Class vocabulary

The 426 classes are ClassyFire terms taken from the Alternative Parents annotation of HMDB. The vocabulary and its order are fixed with the model, so output column k always refers to the same class.

Model

• Primary model: a 15-layer, 512-dimensional Uni-Mol-style encoder with a 3-layer Transformer decoder carrying 426 class queries. Each query is projected to a present/absent softmax.
• The encoder used here is UniMol-Lite, a lightweight reimplementation of the Uni-Mol architecture (MIT licence).
• A second implementation trains one binary AutoGluon classifier per class on the frozen 512-dimensional embedding.

Training and evaluation

Training, validation and test partitions contain 86,685 / 28,935 / 28,935 molecules and are assigned per class. Each class is evaluated as an independent one-versus-rest problem: AUROC and average precision from continuous scores; precision, recall, F1 and balanced accuracy at a fixed 0.5 threshold.

Comparison with NPClassifier

The two tools use different class systems (ClassyFire terms here, a natural-product taxonomy in NPClassifier) and share 22 class names. The comparison is therefore restricted to those 22 classes and to a common cohort of 28,935 molecules. Only NPClassifier V1 can be distributed publicly, and its scores are not on the same probability scale, so it is reported at its legacy class-score threshold of 1e-7 - the most favourable setting available for V1.

Where to use it

• Web platform: AI Tools module.
• Source code: smiles2class (MIT licence).