Google applies AI to antimicrobial peptide search

Google DeepMind has published a blog post on using AI to speed up the search for antimicrobial peptides (AMPs). The work centres on ProtNLM (Protein Natural Language Model), and the post's title also names Google Antigravity.

DeepMind cites an estimate that antimicrobial resistance could cause 10 million deaths a year worldwide by 2050. The post argues that AMPs, especially bacteriocins, are a promising alternative to conventional antibiotics. They evolved to attack basic weaknesses in competing bacteria, from cell envelopes to essential intracellular machinery, and DeepMind says this makes resistance harder for pathogens to develop.

Genomic and metagenomic sequencing has filled public databases with billions of natural amino acid sequences. Mining these has already produced the widely used bacteriocin Nisin and, more recently, Darobactin, Dynobactin and Lariocidin. But DeepMind notes that many candidates sit in the "dark matter" of protein databases: sequences with no known function because they don't resemble any characterised protein. UniProt alone holds millions of such entries, transferred directly from sequencing projects.

ProtNLM is a deep learning model trained to read the evolutionary "grammar" of amino acid sequences and predict function directly from raw sequence. According to Google, it has added high-confidence functional labels to millions of previously unnamed UniProt entries.

The available excerpt doesn't explain Antigravity's role in the AMP search. It also gives no measured results, such as the number of candidate peptides found, hit rates, or any lab validation of new antimicrobial activity.