Researchers detail how ProGen, an LLM trained on 280M proteins, designed proteins with anti-microbial properties that were tested in real life and shown to work
An AI was tasked with creating proteins with anti-microbial properties. Researchers then created a subset of the proteins and found some did the job
Context & Ripple Effects
This closes a loop opened three years earlier, when MIT researchers used AI to discover potent new antibiotics against drug-resistant diseases. Those efforts mined existing compounds with machine learning; ProGen instead generates entirely new protein sequences, trained on 280M known proteins — and crucially, the outputs were actually synthesized and shown to work, not just scored in silico.
That generate-then-validate result is what makes the follow-on work possible: researchers went on to design functional antibodies from scratch without animal testing (AI-designed antibodies) and to produce sixteen viable AI-designed viruses that infect bacteria. ProGen is the proof point that sequence-generating models produce biology that survives contact with the lab.
First-order effects
- Antimicrobial development gets a new starting point: rather than screening natural compounds for activity against drug-resistant bacteria, labs can specify a property and synthesize candidate proteins directly, compressing the discovery step into generation plus testing.
- Synthesis and wet-lab validation capacity becomes the bottleneck — the model produces candidates cheaply, so the scarce resource shifts to physically making and testing them.
Second-order effects
- The same generate-and-test pipeline extends beyond antimicrobials into adjacent targets — antibody discovery and bacteria-infecting viruses — pulling biotech funders toward generative platforms over traditional screening shops.
- Dual-use risk enters regulators' scope: a method proven to yield working antimicrobials demonstrably generalizes to designing viable pathogens, forcing biosecurity review of protein- and sequence-generation models alongside their therapeutic upside.
Third-order effects
- If generation-plus-validation keeps replacing screening across proteins, antibodies, and viruses, drug discovery restructures around foundation models of biological sequences, with access controls on training data and synthesis becoming the industry's compliance layer.
- Wet-lab verification becomes the trust anchor for AI-designed biology — 'proof-carrying' claims where every generated candidate is backed by a real test — separating validated platforms from paper results.
The trend: Biology is shifting from mining nature's molecules to generating them, with LLM-style sequence models producing functional proteins, antibodies, and viruses faster than safety frameworks are being built around them.