Together AI Releases Biological Foundational Model Evo

San Francisco-based AI research startup Together AI has introduced an advanced new biological foundational model called Evo,  to understand and create sequences from DNA, RNA, and proteins. Evo is trained with a large dataset of prokaryotic genomic sequences, covering 2.7 million whole genomes. 

The main goal of Evo is to address the challenges of modelling entire genomes, given their long length and the intricate changes happening at the level of individual building blocks, or nucleotides.

Unlike previous AI models in biology that focused on specific tasks, Evo is designed as a foundational model. It integrates information across long genomic sequences while being sensitive to individual nucleotide changes. To overcome challenges related to long sequences and precise resolution, Evo uses the StripedHyena architecture, employing a hybrid design of rotary attention and hyena operators.

Evo-1 exhibits several notable capabilities, such as predicting essential genes for an organism's survival based on small DNA mutations without prior training (zero-shot gene essentiality testing). It also excels in predicting functions across DNA, RNA, and proteins, outperforming other models in protein function prediction. Evo goes a step further by generating novel CRISPR systems, showcasing its ability to design complex molecular structures involving proteins and RNA simultaneously.

The model can generate sequences at the scale of entire genomes, up to 650,000 characters, using a single GPU. 

The researchers see Evo as a groundbreaking technology with the potential to accelerate discoveries in various scientific fields, including biology, chemistry, and material science. Its applications extend to practical challenges such as drug discovery, agriculture, and sustainability. 

In December last year, Together AI received a $102.5 million Series A investment from NVIDIA, Kleiner Perkins, and Emergence Capital.