• Comprises two tasks: classifying protein sequences by EC number and retrieving enzyme sequences based on chemical reactions
• Proposes Contrastive Reaction-EnzymE Pretraining (CREEP) leveraging multimodal contrastive learning for Task 2 (Next Thread)

• Rxnfp: A BERT-style language model trained on reactions represented as SMILES/SMARTS strings
• ProtT5 as protein sequence encoder
• SciBERT for additional textual descriptions based on gene ontology
• Adopts EBM-NCE to maximize the mutual information between all three modalities