Scaling ProGen

ProGen in Practice Generative artificial intelligence has officially mastered the language of biology. By treating amino acid sequences like text sentences, the ProGen family of models—originally developed by Salesforce Research and advanced by organizations like Profluent Bio—is fundamentally reshaping protein engineering. Instead of relying on slow, trial-and-error directed evolution, scientists are actively deploying ProGen in practice to generate functional, custom-tailored proteins from scratch in a matter of weeks. The Architecture: From Words to Amino Acids

Traditional protein design requires mapping complex 3D structures or executing thousands of random mutations to find a working variant. ProGen sidesteps these roadblocks by framing biology as a sequence generation problem.

Billion-Scale Training: Early iterations utilized a 1.2-billion-parameter neural network trained on 280 million sequences. The advanced ProGen3 family scales this process up to 46 billion parameters, digesting over 1.5 trillion amino-acid tokens.

Next-Token Prediction: Just as standard large language models predict the next word in a sentence, ProGen predicts the next amino acid in a molecular chain.

Conditional Control Tags: Scientists guide the generation by inputting specific tags. These tags dictate desired properties like taxonomic origin, cellular environment, or exact molecular function. Real-World Applications: ProGen in the Lab

ProGen’s theoretical capabilities have translated directly into viable, real-world laboratory breakthroughs. 1. De Novo Enzyme Design ProGen: Language Modeling for Protein Generation – bioRxiv