... molecules with comparable affinity. Contrastive learning methods such as ConPLex (Singh et al., 2023) represent an important step toward grounded representations, demonstrating that protein language model embeddings can be aligned with chemical representations through contrastive objectives...