A new frontier in generative genomics with Omnii (www.radicalnumerics.ai)

Omnii has emerged as a groundbreaking genome language model (gLM) that signifies a pivotal advancement in generative genomics, aiming to enhance our understanding of the human genome by predicting, designing, and interpreting genomic sequences. By utilizing a sophisticated architecture featuring a multi-hybrid backbone and dynamic sparse attention, Omnii processes a context window of 2 million base pairs and seamlessly integrates genomic annotations with raw DNA sequence data. This multimodal approach allows Omnii to outperform previous models, especially in noncoding variant effect prediction, making significant advancements in areas such as complex trait classification and clinical variant analysis, including a notable ability to identify key Alzheimer's-related variants. The significance of Omnii lies in its ability to combine scaling, alignment, and modality fusion, which were historically pursued separately, to create a model that can directly reason over DNA sequences and associated biological signals. This integration marks a departure from earlier gLMs that were often unimodal and unaligned, thereby enhancing practical usability for researchers. As a state-of-the-art solution, Omnii not only eclipses traditional bioinformatics models but also establishes a new performance frontier for gLMs in genomic research, providing a more accessible tool for researchers to interpret complex genomic information without needing extensive task-specific training.