An Open MetaGenomic Corpus for Mixed-Modality Genomic Language Modeling

Explore a comprehensive talk on the Open MetaGenomic (OMG) corpus, a groundbreaking genomic pretraining dataset, and its applications in mixed-modality genomic language modeling. Delve into the creation of this massive dataset, comprising 3.1T base pairs and 3.3B protein coding sequences, derived from JGI's IMG and EMBL's MGnify repositories. Learn about the quality filtering process, the dataset's composition, and its unique mixed-modality approach combining translated amino acids for protein coding sequences and nucleic acids for intergenic sequences. Discover the development of gLM2, the first mixed-modality genomic language model, and its ability to leverage genomic context for robust functional representations and coevolutionary signals in protein-protein interfaces. Gain insights into embedding space deduplication techniques for corpus balancing and improved downstream task performance. Explore unsupervised protein-protein interaction analysis and future directions in this field. Access the OMG dataset and gLM2 model through provided Hugging Face Hub links, and engage in a Q&A session to deepen your understanding of this innovative approach to genomic language modeling.