Google is pointing AI at one of medicine’s biggest problems: drug-resistant bacteria. According to Google DeepMind, the company is using its ProtNLM model and its agentic development platform, Google Antigravity, to search huge protein databases for antimicrobial peptides. These are small molecules that bacteria use to kill each other. They could become the base for new treatments at a time when our current antibiotics keep failing.
The stakes are high. Google DeepMind cites estimates that antimicrobial resistance could cause 10 million deaths a year worldwide by 2050. Old antibiotics are losing ground, and the pipeline of new ones is thin.
🧬 Why antimicrobial peptides?
The research focuses on antimicrobial peptides (AMPs), especially a class called bacteriocins. Bacteria have evolved these molecules over millions of years to attack their competitors. They hit basic weak points such as:
- Cell envelopes, the protective outer layers of bacteria
- Essential intracellular machinery that bacteria need to survive
What stands out is the resistance angle. These peptides target such basic parts of bacterial biology that it’s hard for pathogens to evolve around them. That’s the weakness of many traditional antibiotics, and it’s the reason AMPs are getting so much attention.
The approach also has a track record. Nisin, a bacteriocin, is already used around the world. Recent compounds found by mining genomes, including Darobactin, Dynobactin and Lariocidin, show that natural sequence data still holds new drug candidates.
🔍 The “dark matter” problem
This is where AI comes in. Genome and metagenome sequencing has filled global databases with billions of natural amino acid sequences. Most of them have never been studied.
Google DeepMind calls this the unannotated “dark matter” of protein databases. These sequences sit there with no known function because they don’t look like any protein scientists have already studied. Traditional tools mostly find function by similarity: if a new sequence resembles a known one, you can guess what it does. If it resembles nothing, it stays a blank.
UniProt, the world’s leading public resource for protein sequences and function data, holds millions of these unknown entries. Many came straight from genome sequencing projects and were never labeled.
🤖 How ProtNLM fills the gaps
Google built ProtNLM (Protein Natural Language Model) to tackle this. It’s a deep learning model trained to read what Google DeepMind calls the evolutionary “grammar” of amino acid sequences.
The idea is close to how large language models learn patterns in text:
- The model learns the structure and patterns inside raw protein sequences.
- It predicts biological function straight from the sequence, without needing a close match to a known protein.
- It gives previously unnamed entries high-confidence functional labels.
The result so far: millions of UniProt entries that had no name now carry predicted functions. That turns a pile of unknown sequences into a list researchers can actually search for antimicrobial candidates.
⚠️ What to keep in mind
The excerpt Google DeepMind shared covers the problem and the ProtNLM foundation. It doesn’t include benchmark numbers or lab validation results. It also doesn’t explain in detail how Antigravity’s agent workflows speed up the search. So a few caveats:
- Predicted function isn’t proven function. “High-confidence” labels still need lab work to confirm.
- A candidate peptide isn’t a drug. Toxicity, stability, delivery and clinical trials all come later and take years.
- No performance data is public in this excerpt, so it’s too early to judge how ProtNLM’s hit rate compares with other discovery methods.
💡 Why this matters beyond biology
This is significant because it shows a pattern that keeps repeating across AI research: language-model methods moving into fields where the “text” isn’t words. Protein sequences, genomes and chemical structures all have a grammar of their own. Models that learn that grammar can open up datasets that rule-based tools can’t handle.
The Antigravity angle is worth watching too. Google is presenting its agentic coding platform as a scientific research tool, not just a developer product. If AI agents can handle the data wrangling, pipeline building and analysis that slow down computational biology, small research teams could cover far more ground than before.
🧭 Strategic takeaway
If you work in biotech, bioinformatics or applied AI, the move here is practical. Look at the unlabeled data you already have before you go hunting for new data. ProtNLM’s annotations are already in UniProt, so researchers can start filtering those newly labeled sequences for antimicrobial candidates today. For AI teams in other fields, the lesson carries over: the biggest wins often hide in data everyone has written off as noise.
The full write-up on the project is available from Google DeepMind.