Navigating latent space of natural language processing models to explain the galaxy of chiral moleculesBaimacheva, Natalia (Avtor) Podlipnik, Črtomir (Mentor) latent space arithmeticsQSARlanguage processing modelsExploring effectiveness of representing molecular structural features in latent space of molecular heteroencoders, namely chirality. Latent space vectors are numerical vectors that can encode SMILES strings. Latent space vectors (LSV) showed ability to encode chirality of a molecule in order to classify enantiomers by order of their elution from the ADH chromatographic column. Additionaly delta LSV of the two enantiomers can highlight the chiral structural difference in order to improve the classification results. Delta LSV were calculated in 2 ways: difference of the opposite enantiomers and difference of one enantiomer and its non-stereo representation. Random forest models for LSV and DLSV reach prediction accuracies 0.753, 0.763 respectively.20252025-05-13 08:40:00Magistrsko delo/naloga169112VisID: 25437COBISS_ID: 238557443sl