Accelerating antimicrobial peptide discovery with Google Antigravity

Author: Google Antigravity · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Accelerating antimicrobial peptide discovery with Google Antigravity · Channel: Google Antigravity · 05:56 · Watch on YouTube

Topics: AI agents, Health and science

Summary

Fei, a research scientist at Google DeepMind, demonstrates how Google Antigravity can support antimicrobial peptide discovery. The agent searches the Streptococcus pneumoniae proteome, identifies a candidate predicted to have antimicrobial activity, and assesses its novelty and predicted structure. Further analysis finds a similar 3D fold, a biosynthetic gene cluster, and a possible mechanism involving lipid II. The workflow concludes with a proposed wet-lab validation plan, including expression, susceptibility, and safety testing.

Key points

The demonstration uses Google Antigravity as an agentic science workbench for protein discovery.
The agent searches UniProt for uncharacterized Streptococcus pneumoniae proteins 90–100 amino acids long.
Prot-ALM predicts antimicrobial activity for a candidate described as a Lactococcin 972 family bacteriocin.
A BLAST search against Swiss-Prot finds no known homologs for the candidate.
The AlphaFold prediction gives the mature peptide a global pLDDT score of 88.5 and a beta-sheet-rich fold.
A Protein Data Bank search finds a structurally similar antimicrobial peptide despite only 27% sequence identity.
The candidate’s genomic neighborhood includes genes predicted to support bacteriocin export and expression.
The agent proposes lipid II binding as a possible mechanism and predicts activity against respiratory pathogens.
The proposed experimental roadmap covers recombinant expression, antimicrobial susceptibility profiling, and preclinical safety assays.

Chapters

Start	Chapter	Summary
00:00	Antimicrobial peptides and the discovery workflow	Fei introduces antimicrobial peptides as a potential response to antimicrobial resistance and outlines the Antigravity demonstration.
01:02	Finding and checking a candidate	The agent searches the pneumococcal proteome, identifies a predicted bacteriocin, and checks its novelty against Swiss-Prot.
02:42	Structure and genomic evidence	The agent analyzes the candidate’s predicted structure, compares it with a known antimicrobial peptide, and examines its genomic context.
04:20	Proposed mechanism and experimental plan	The agent infers a possible lipid II-targeting mechanism, predicts potential pathogen activity, and lays out wet-lab validation steps.
05:22	From analysis to validation	Fei summarizes the proposed experimental roadmap and the broader potential of using genetic workflows to explore protein molecules.

Quotes

Today, I will show you how we can use Google anti-gravity as our agentic science workbench to accelerate this protein discovery process. — 00:30

This confirms our candidate lacks known homologs and is a genuinely novel sequence. — 02:06

the core beta-sheet domains align almost perfectly between the two structures with an RMSD of around one angstrom. — 03:14

The agent generated a multi-step wet-lab experimental roadmap. — 05:22

Concepts and entities

Google Antigravity, Fei, Google DeepMind, UniProt, Prot-ALM, BLAST, Swiss-Prot, AlphaFold, Protein Data Bank, Lactococcin 972, lipid II, pLDDT, RMSD

Full transcript (original language)

00:00 Hello. I’m Fei, a research scientist at Google DeepMind working on understanding the relationship between protein sequence and function. Antimicrobial resistance is one of today’s most urgent global health threats, projected to claim millions of lives annually by 2050. To protect global health, we need to discover new therapeutics that the pathogens cannot easily bypass. That is where antimicrobial peptides come in. As nature’s frontline defense, these small proteins attack the fundamental

00:30 structure of pathogen membranes. This makes it harder for antimicrobial resistance to evolve. Today, I will show you how we can use Google anti-gravity as our agentic science workbench to accelerate this protein discovery process. In this demo, we will find a potential antimicrobial peptide, validate its novelty, analyze its 3D structure in silico, and generate an end-to-end wet-lab experimental plan, all in one workbench. I have launched anti-gravity with the

01:02 science skills installed. We start our exploration by searching the UniProt database. We instruct the agent to screen the Streptococcus pneumoniae proteome for short peptides with length between 90 to 100 amino acids. We specifically want to look for uncharacterized proteins that our latest protein function model, Prot-ALM, predicts have antimicrobial activity. The agent identifies a prime candidate

01:35 predicted to have antimicrobial activity. Our model predicts this sequence is a Lactococcin 972 family bacteriocin. Characterized by a 28-residue N-terminal signal sequence and a 69-residue mature peptide domain. Next, we need to confirm that this is a truly novel candidate. I will ask the agent to run a BLAST search against the Swiss-Prot, the gold standard for experimentally validated proteins. The top sequence match is an unrelated

02:06 mitochondrial peptidase with only a 26% sequence identity with a statistically insignificant E-value of 0.03. All subsequent hits present are even less similar. This confirms our candidate lacks known homologs and is a genuinely novel sequence. Next, let’s look at the its predicted 3D structure. The agent retrieves the AlphaFold prediction. It shows that the mature peptide, which spans residues 29 through 97, folds into a rigid singular domain

02:42 with a high-confidence global pLDDT score of 88.5. Structurally, the N-terminus forms a short alpha-helix with lower confidence, likely representing a flexible signal peptide. While the C-terminal domain adopts a highly stable beta-sheet-rich conformation. Next, the agent runs a full six search against the Protein Data Bank to find the similar structures. It identifies a high-probability match with the structure of a known AMP from a different bacteria species. Even though

03:14 their protein sequences are very different with a low sequence identity of only 27%, their 3D folds are strikingly similar. Let’s visualize that the similarity. The agent uh superimposes our AlphaFold prediction onto the known experimental structure and the renders the result in the 3D viewer. As you can see, the core beta-sheet domains align almost perfectly between the two structures with an RMSD of around one angstrom. This a strong structure match confirms

03:45 our candidate adopts the precise fold of an antimicrobial peptide. The agent has annotated the signal peptide with a different color directly on the structure. To back up our Prot ARM predictions with additional biological evidence, let’s look at the genomic context. The agent analyzes the neighboring genes and identifies a tightly organized biosynthetic gene cluster directly downstream of our candidate. The operon is predicted to encode transporter proteins and ATP-binding proteins

04:20 essential for peptide export. Along with a secondary bacteriocin and a dedicated transcriptional regulator, this highly conserved genomic neighborhood provides compelling biological evidence of a complete functional bacteriocin expression system. Next, we want to know how our candidate works and what it targets. The agent synthesizes insights from existing literature on lacticin-like peptides to deduce the candidate’s molecular mechanism. It predicts that

04:52 this peptide family selectively bind to the cell wall precursor lipid II, effectively inhibiting septum synthesis during cellular division. Because ecologically this peptide likely evolved for niche competition within the respiratory tract, the agent predicts potent activity against respiratory pathogens, including Streptococcus pneumoniae and Streptococcus pyogenes. Finally, let’s translate these insights into an experimental validation plan.

05:22 The agent generated a multi-step wet-lab experimental roadmap. It outlines construct design for recombinant expression and antimicrobial susceptibility profiling against the drug-resistant strains. It also details preclinical safety assays to establish therapeutic index. In just minutes, we went from a massive database to a novel, structurally and the corroborated candidate complete with an action plan for wet lab testing. Nature has spent billions of years evolving an incredible universe of

05:56 protein molecules. With a genetic workflows, we can explore and understand this universe in much greater depth.