Fig. 4
From: Positional embeddings and zero-shot learning using BERT for molecular-property prediction
Analysis of the similarity between the pretraining and downstream (i.e., Clintox) SMILES dataset samples (i.e., 100 randomly selected samples) based on the molecular structures using BERT with “Absolute” PE. a Tanimoto similarity heatmap, b Embedding similarity heatmap, and c PCA Visualization of molecular embeddings in latent space