“The improved scaling performance of the ESM Cambrian model was achieved by increasing its training dataset from 50 million to 2.8 billion protein sequences, primarily by incorporating metagenomic data.”