Softmax converts a vector of logits—raw scores produced by a model—into nonnegative probabilities that sum to one. Each class has its own feature weights, and prediction selects the class with the highest probability. section 5.4.1 section 10.4.3
Use the equation to see how each class score becomes a probability relative to the exponentiated scores of all classes. section 5.4.1
Softmax appears in two different molecular-machine-learning workflows:
- In CYP inhibition classification, it provides probabilities across five cytochrome P450 classes and supports cross-entropy training. This setup incorrectly forces every molecule into exactly one class, even though a molecule may inhibit none or several CYPs. section 5.4.1
- In a molecule-generating seq2seq decoder, it converts vocabulary-sized logits into next-character probabilities for SMILES generation. A sampling method then selects the next character. section 10.4.3
Temperature sampling
A decoder can generate molecules from a latent vector by repeatedly applying softmax to output logits and sampling characters until the SMILES sequence ends. Temperature sampling controls this generation process.

























