Salesforce's language model just designed working proteins from scratch

While most people are still using large language models to write code and draft emails, Salesforce just used one to leapfrog millions of years of natural evolution.

Researchers from Salesforce Research, UC San Francisco, and UC Berkeley published a milestone paper in Nature Biotechnology. They demonstrated that a language model called ProGen can design entirely new, functional proteins from scratch.

Treating biology like syntax

Proteins are basically strings of 20 different amino acids folded into intricate functional shapes. The exact sequence determines how the protein behaves in a living organism.

Instead of treating protein design as a brutally complex 3D physics problem, the team treated it as a language. Words string together to form sentences, and amino acids string together to form proteins.

Salesforce trained ProGen on 280 million known protein sequences. The model simply learned the grammar of life by predicting the next amino acid in line, just like ChatGPT predicts the next word in a chat window.

To keep the AI from outputting biological gibberish, researchers set up guardrails and prompted it to generate specific families of molecules, specifically targeting antimicrobial proteins.

It walks like a duck and kills bacteria

Generating digital sequences is one thing, but making them work in the real world is another. The researchers went into the wet lab and physically synthesized 100 molecules designed by ProGen.

Out of that batch, 66 molecules demonstrated enzymatic reactions similar to hen egg white lysozymes. Overall, Salesforce estimated that 73% of ProGen-generated proteins were fully functional, compared to just 59% in natural protein baselines.

When they took the five strongest performers and introduced them to live Escherichia coli bacteria, two of the synthetic proteins successfully tore through the bacteria and destroyed them.

X-ray crystallography revealed something even wilder: although their amino acid sequences differed from any known natural protein by up to 30%, their 3D shapes almost perfectly matched natural molecules.

As UCSF researcher James Fraser put it, it was a classic case of looking like a duck and quacking like a duck, but the X-rays proved it actually walked like one too.

Why it matters

Evolution takes millions of years of random genetic mutations to stumble across an effective enzyme. ProGen demonstrates that neural networks can learn the underlying rules of biological grammar in a fraction of that time.

We think the real brilliance here is the dimensional hack. Instead of wrestling with atomic force fields and heavy physics simulations, Salesforce turned protein generation into a standard next-token prediction task, and the model somehow inferred evolutionary logic on its own.

This paves an entirely new path for targeted therapeutics. The team is already looking at treatments for rheumatoid arthritis and multiple sclerosis, and synthetic enzymes could offer brand new ammunition against antibiotic-resistant bacteria.

The real catch is the validation bottleneck. As biotech startup founder Ali Madani pointed out, an AI can dream up millions of molecules in an afternoon, but physically synthesizing and testing them in wet labs is still painstakingly slow.

For the biotech industry, the primary bottleneck just shifted from molecule discovery to high-throughput experimental verification.