Google's new science workbench hunts for antibiotics inside protein 'dark matter'
Antigravity pairs a biologist with an AI agent to sift millions of unlabelled protein sequences for antimicrobial candidates. The first lead is now headed to the lab bench.
Key points
- Google DeepMind says a biologist using its new Antigravity workbench has pulled a candidate antimicrobial peptide out of the unlabelled portion of the UniProt protein database, and the molecule is now going to wet-lab testing.
- The search targeted Streptococcus pseudopneumoniae, a bacterium linked to respiratory tract infections, and used ProtNLM, a model that guesses what a protein does from its raw amino-acid sequence.
- Antimicrobial resistance could kill an estimated 10 million people a year globally by 2050, according to figures cited by the DeepMind team.
- The workbench embeds a 3D structure viewer that pulls predictions from the AlphaFold database, so human and agent can point at the same loop on the same molecule.
- One candidate has been flagged for wet-lab validation; no results are back yet.
Keep this straight: it's a software demo, not a drug. What Google DeepMind actually showed is a biologist and an AI agent working together inside one browser tab to shortlist a molecule worth testing.
The target was Streptococcus pseudopneumoniae, a close cousin of the bug behind many chest infections. Antimicrobial peptides are short proteins that kill bacteria by punching holes in them or jamming their internal machinery. Nisin, used for decades as a food preservative, is the classic example.
The trick is where they looked. UniProt, the main public catalogue of protein sequences, holds millions of entries that have never been given a function. Genome sequencers deposit them and move on. DeepMind calls this the "dark matter" of the protein world.
How does the workbench actually work?
A researcher types a goal into Antigravity, Google's agentic development platform. Database queries run, candidate sequences get pulled, ProtNLM predicts their function, and 3D shapes arrive from the AlphaFold database, DeepMind's public library of predicted protein structures.
Those structures load into a Biomolecular Structure Viewer built into the same window. A scientist can pin a note on a specific pocket of the molecule and ask the agent a question about that exact spot, because the agent sees the same view. That shared pointer is the main new piece of plumbing here.
Without it, this kind of work normally means custom scripts, exported files, a separate modelling program and a chain of emails. Antigravity compresses that into one session. We've tracked DeepMind's drive to embed AI into biological workflows since July, including its watermark tool for lab-grown proteins published on 30 September.
Why does this matter for patients?
Because the drugs we have are losing. The pipeline of genuinely new antibiotics has been thin for decades, and resistance keeps climbing. Resistance could kill an estimated 10 million people a year by 2050, a figure the DeepMind team cites to frame the urgency.
Peptides like the one DeepMind flagged are attractive because bacteria find them harder to shrug off. Genome-mined finds such as darobactin and lariocidin have revived interest in the approach. A faster way to triage candidates could matter a lot, if the lab results hold up.
| Fact | Figure |
|---|---|
| Projected global AMR deaths by 2050 | ~10 million per year |
| Candidates sent to wet-lab validation | 1 (results pending) |
What is new here, really?
The parts aren't new. ProtNLM has been annotating UniProt entries for a while. AlphaFold structures have been free to download since 2021.
What's new is the stitching. Antigravity puts the database search, function prediction and the shared structure viewer in one place, with a chat interface woven through. That's a workflow story, not a model story.
Honest take: this tool probably saves a graduate student a week of file wrangling, and nothing more until a wet lab confirms a hit. The entire case rests on that first validation coming back positive. Watch for the follow-up paper, not the walkthrough video.
Common questions
Is this an actual new antibiotic?
No. It's a computational prediction of a peptide that might kill a specific respiratory bug. It still has to be made in a lab, tested against live bacteria, assessed for toxicity, and then put through years of trials before any patient sees it.
Can my doctor use this?
Not directly. Antigravity is a research tool for biologists, not a clinical system. The practical message for patients is the older one: finish prescribed courses, don't pressure a GP for antibiotics for a cold, and expect prescribing rules to keep tightening.
