Xiangru (Robert) Tang: Most DL models in chemistry work with text or 2D images/diagrams. Let's jointly embed text + 2D *and* 3D molecular representations. Or at least an encoding of euclidean distance, for the last item. Use contrastive learning - have training loss for both positive and negative conditions. With this model we can make text to (chemical) graph. Could include MoFlow too, to learn mappings between molecular graphs and their latent representations.
#ismbbeccb2023
#textmining