Can AI Predict New Antibiotics Before Lab Tests? DeepMind's Antigrav

TL;DR: Google DeepMind’s Antigravity platform identified a novel 97-amino-acid bacteriocin from Streptococcus pseudopneumoniae in hours, demonstrating AI’s speed in scanning uncharacterized protein sequences. However, the candidate remains unverified by wet-lab tests, highlighting that while AI accelerates discovery, biological confirmation is still essential for drug development.

Key facts

• Google DeepMind announced its Antigravity platform on October 15, 2026, claiming AI can identify novel antibiotic candidates before wet-lab validation.
• The system identified a 97-amino-acid bacteriocin from Streptococcus pseudopneumoniae in hours, with no known homologs in the Swiss-Prot database.
• Antigravity utilizes ProtNLM to infer function from uncharacterized protein sequences and AlphaFold for structural modeling of candidates.
• The platform integrates Gemini models with AlphaGenome and operates as an agentic science workbench connecting over 30 life-science databases.
• Pushmeet Kohli stated Antigravity evolved over a year, combining multiple breakthroughs to allow AI agents to propose mechanisms interactively.
• No antimicrobial activity or toxicity data has been published for the candidate; it remains unverified pending experimental validation.

Can AI Actually Discover New Antibiotics Before Lab Tests?

The question on every researcher’s mind after Google DeepMind’s October 15, 2026 announcement: can artificial intelligence really find a new antibiotic before any lab test confirms it works? The short answer is yes—but only in theory. Google Antigravity identified a promising antimicrobial peptide candidate from Streptococcus pseudopneumoniae in a matter of hours, but the platform has not yet proven it kills bacteria in real-world conditions.

This distinction is critical. While AI can rapidly scan millions of uncharacterized protein sequences using ProtNLM and predict biological function from amino acid patterns [1][2], actual drug efficacy requires wet-lab experiments. The candidate peptide—97 amino acids long, with no known homologs in Swiss-Prot—is described as biologically plausible based on structural modeling via AlphaFold and Foldseek [3]. Yet Google explicitly states it is still undergoing experimental validation [1][2]. This gap between prediction and proof underscores a growing trend: AI accelerates discovery, but biology still demands physical testing.

The Power of Agentic Science: From Data to Hypothesis in Hours

Google Antigravity functions as an agentic science workbench, integrating tools like the Biomolecular Structure Viewer to let researchers and AI agents collaborate on 3D models in real time [5][6]. Unlike traditional pipelines that require months of manual screening across fragmented databases, Antigravity compresses searches into single sessions. It uses ProtNLM—a deep learning model trained to infer function from sequence—to analyze the “dark matter” of protein databases like UniProt, where most sequences lack functional labels [1][2].

In this case, the agent scanned millions of uncharacterized entries and flagged a 97-amino-acid bacteriocin from S. pseudopneumoniae with no meaningful matches in manually curated Swiss-Prot data, suggesting novelty [3]. Structural analysis further supported its plausibility: an AlphaFold model of the candidate closely aligned with known bacteriocin structures in Foldseek searches [3]. These results are compelling, but they remain computational—no antimicrobial activity or toxicity data has been published yet [3].

Why This Isn’t a Breakthrough Yet—And Why It Still Matters

The real story here isn’t the candidate itself, but what it reveals about the evolving role of AI in drug discovery. While other teams, like Cesar de la Fuente’s at UPenn, have used AI to reduce antibiotic discovery timelines from years to hours [7], DeepMind’s approach is different: it leverages agentic workflows where AI doesn’t just predict but proposes mechanisms and explores structural contexts interactively [3]. Pushmeet Kohli noted that Antigravity evolved over a year by combining Gemini models with breakthroughs like AlphaFold and AlphaGenome [6].

Still, the absence of published code, agent transcripts, or a public repository limits reproducibility [3]. And while Google launched an ‘Antigravity for Science’ credits program to support academic adoption [6], access remains gated. The platform’s claim of scanning over 30 life-science databases via plugins is unverified in the public record [4]. Even the assertion that the candidate lies upstream of a tightly coupled operon lacks independent confirmation [4].

The Bigger Picture: AI as a Co-Pilot, Not a Replacement

Antimicrobial resistance (AMR) could cause 10 million annual deaths globally by 2050 if new treatments aren’t found [1][2]. Existing bacteriocins like Nisin and genome-mined candidates such as Darobactin have entered clinical use, but billions of sequences remain unexplored [1][2]. Antigravity’s strength lies in navigating this vast, uncharted territory faster than ever before.

Yet the core thesis remains: AI predicts. Biology confirms. The candidate from S. pseudopneumoniae may be a real lead—but only time, and lab work, will tell. For now, Antigravity is less a magic bullet and more a powerful co-pilot in the race against drug-resistant bacteria.

Sources

1. Google Antigravity: Accelerating Antimicrobial Peptide Search (deepmind.google) — 2026-10-15
2. Google Antigravity: Accelerating Antimicrobial Peptide Search (deepmind.google) — 2026-10-15
3. DeepMind Antigravity Antimicrobial Peptide: Real Lead? | explainx.ai Blog — 2026-10-15
4. HyperAI (hyper.ai) — 2026-10-15
5. Pushmeet Kohli (@pushmeet) on X (x.com) — 2026-10-16
6. Artificial Intelligence to Accelerate Antibiotic Discovery (beblogarchive.engineering.upenn.edu) — 2024-06-07
7. Google DeepMind Walkthrough Details Antimicrobial Peptide Discovery (www.unite.ai) — 2026-10-15