On October 15, 2026, Google DeepMind published a walkthrough of researchers using Google Antigravity and ProtNLM to computationally prioritize a possible antimicrobial peptide in Streptococcus pseudopneumoniae. DeepMind predicts the candidate may act against pathogens involved in respiratory tract infections and says wet-lab validation is underway. The walkthrough describes a research lead, not a measured antimicrobial result. Google DeepMind’s account gives the work’s status and scope.
How Antigravity narrowed the search
ProtNLM predicts protein function from amino-acid sequences. In the walkthrough, a researcher sets the biological question while Antigravity coordinates work across protein databases, sequence analysis, structure comparisons, genomic context and literature. UniProt describes ProtNLM’s sequence-based function predictions in its model help page.
Unite.AI identifies the candidate as UniProt A0A0U0D591, a 97-amino-acid peptide in the Lactococcin 972 family. The outlet also reports that a sequence comparison found no significant homologs in the reviewed set, that structural analysis suggested similarity to a characterized bacteriocin, and that the candidate sits 6 base pairs upstream of an operon with predicted transporter-related genes and other annotations. Those are computational clues reported by Unite.AI, not laboratory measurements of the candidate’s activity.
DeepMind also announced the Biomolecular Structure Viewer, an integrated 3D tool in which an agent can highlight parts of a structure and researchers can send annotations back to the agent. The English-language demonstration follows the Antigravity workflow from protein screening through structural and genomic analysis to a proposed wet-lab plan.
What laboratory evidence looks like in another study
A separate ProteoGPT study, published in Nature Microbiology on October 3, 2025, tested its own peptide candidates. In a pilot assay, 18 of 20 generated PT-series peptides inhibited at least one tested strain at a minimum inhibitory concentration (MIC) of 512 μg/mL or less; 12 of 20 reached 256 μg/mL or less. MIC is the concentration used to assess inhibition of microbial growth under the assay conditions. These measurements belong to the ProteoGPT study’s peptides, not DeepMind’s candidate. The paper also reports in-vitro testing and experiments in a mouse thigh-infection model for selected peptides.
The study assessed more than antimicrobial activity: its reported tests included hemolysis, mammalian-cell cytotoxicity, degradation by serum proteases and resistance during repeated bacterial passages. Each addresses a different question—whether a peptide inhibits bacteria, harms other cells, degrades, or is followed by resistant growth. A promising result in one assay does not answer all four.
Why no single hurdle dominates
A cross-study synthesis found no consensus that toxicity, stability or resistance is invariably the hardest obstacle in antimicrobial-peptide research; toxicity and limited data recur as concerns. The separate ProteoGPT experiments illustrate several ways researchers assess these challenges, but do not rank one as universally decisive. (AI Methods for Antimicrobial Peptides; Nature Reviews Bioengineering)
