Anatol Presents his poster titled “Bio-Isostere Guided Molecular Property Prediction” at the 13th International Conference on Chemical Structures in Noordwijkerhout, The Netherlands.

The poster proposes a novel idea to pre-train a GNN by creating bio-isosteres from a large unlabeled dataset. These bio-isosteres are dynamically sampled during training to form triplets, where the anchor and positive instances are bio-isosteres and the negative instance is a randomly chosen molecule. We show that pre-training a GNN in this way improves the embeddings when compared to a randomly initialized GNN.