Machine learning reveals bacteria carry far more antiviral defenses than scientists assumed

Overview

Updated May 31

Every major tool in genetic engineering — from the enzymes that cut DNA in the 1970s to CRISPR gene editing — started as a defense weapon bacteria use against viruses. Two teams revealed bacteria carry three times more weapons than anyone realized, identifying millions of antiviral proteins from tens of thousands of genomes using machine-learning models that flag a new defense system in five minutes.

The findings were published simultaneously in Science in April 2026. The known catalog of bacterial immune defenses has expanded from around 100 system families to a searchable database of more than 44,000 predicted antiviral systems and 2.39 million candidate defense proteins. Each one is a potential starting point for a new biotechnology — a molecular tool that evolution spent billions of years optimizing and that scientists can now, for the first time, actually find.

Why it matters

Every previous bacterial defense tool scientists repurposed — restriction enzymes, CRISPR — launched a multibillion-dollar industry. Scientists now have thousands more to explore.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

2.39M

Predicted antiviral proteins identified

The Pasteur team's models flagged 2.39 million candidate defense proteins across more than 30,000 bacterial genomes.

3×

Increase over previous estimates

Roughly 1.5% of a bacterium's genes serve antiviral functions — three times more than previously thought.

624

Defense proteins found in E. coli alone

The MIT team's DefensePredictor tool identified 624 defense-related proteins across 69 E. coli strains, over 100 of them previously unknown.

85%

Proteins with no known link to immunity

The vast majority of the Pasteur team's predicted defense proteins had never been connected to antiviral function before.

45%

Experimental validation rate

When the MIT team cloned 94 predicted defense systems into bacteria and exposed them to phages, nearly half protected against infection.

Voices

Curated perspectives — historical figures and your fellow readers.

Ever wondered what historical figures would say about today's headlines?

People Involved

Peter DeWeirdt

Corresponding author, DefensePredictor study (MIT / Broad Institute)

Active researcher at the Broad Institute of MIT and Harvard

Aude Bernheim

Senior author, protein language model study (Institut Pasteur)

Active researcher at the Molecular Diversity of Microbes Lab, Institut Pasteur

Ernest Mordret

First author, protein language model study (Institut Pasteur)

Researcher at the Molecular Diversity of Microbes Lab, Institut Pasteur

Feng Zhang

Co-author on DefensePredictor; pioneer in gene-editing tool discovery

Investigator at the Broad Institute, MIT, and Howard Hughes Medical Institute

Michael Laub

MIT biologist and co-discoverer of the SNIPE defense system

Professor at MIT; lab characterized the SNIPE membrane-bound defense system

Rotem Sorek

Pioneer of systematic bacterial defense system discovery (Weizmann Institute)

Won the 2025 Gruber Genetics Prize ($500,000) for bacterial immunity discoveries

Organizations Involved

Broad Institute of MIT and Harvard

Genomics research institute

Home institution of the DefensePredictor team

A biomedical and genomic research center in Cambridge, Massachusetts, jointly operated by MIT and Harvard, where DefensePredictor was developed.

Institut Pasteur

Research Institute

Home institution of the protein language model team

A French biomedical research institute in Paris whose Molecular Diversity of Microbes Lab developed the deep-learning models that predicted 2.39 million antiphage proteins.

Weizmann Institute of Science

Research Institute

Pioneered the computational approach to bacterial defense discovery

An Israeli research institute in Rehovot where Rotem Sorek's lab established the defense-island scanning approach that the new machine-learning tools now automate at scale.

Timeline

January 1968 April 2026

11 events Latest: April 16th, 2026 · 5 months ago Showing 8 of 11

Tap a bar to jump to that date

April 2026
Both studies featured in Nature's April 16 issue
Latest Publication

Nature's Volume 652, Issue 8110 featured coverage of the two simultaneously published Science papers, bringing the findings to the journal's broad readership and cementing the moment as a milestone in microbiology.
April 16th, 2026
SNIPE defense system characterized at MIT
Discovery

Michael Laub's lab at MIT published the characterization of SNIPE, a membrane-bound nuclease system that rapidly degrades invading phage DNA — one of the first newly predicted systems to receive detailed structural and mechanistic analysis.
April 9th, 2026
Nature covers both studies as a 'treasure trove'
Publication

Nature published a news feature highlighting both teams' findings, quoting researchers who said the field had been 'massively underestimating' the number of bacterial defense systems.
April 2nd, 2026
January 2025
Both machine-learning studies posted as preprints
Publication

The DeWeirdt (MIT/Broad) and Mordret (Pasteur) teams posted their studies to bioRxiv on the same day, signaling their simultaneous development of AI-powered approaches to bacterial defense discovery.
January 8th, 2025
October 2022
Known defense system count passes 100 families
Research

After successive rounds of discovery by multiple labs, the catalog of experimentally validated bacterial defense system families exceeded 100 — a tenfold increase from the roughly 10 families known before 2018.
October 1st, 2022
March 2018
Sorek discovers 10 new antiphage defense systems
Discovery

Rotem Sorek's lab at the Weizmann Institute published a systematic scan of 50,000 bacterial genomes using the defense-island strategy, discovering 10 previously unknown immune systems. This paper opened the current era of rapid defense system discovery.
March 2nd, 2018
June 2012
CRISPR-Cas9 repurposed as a gene-editing tool
Discovery

Jennifer Doudna and Emmanuelle Charpentier published their landmark paper showing CRISPR-Cas9 could be programmed to cut any DNA sequence, transforming a bacterial defense mechanism into the most powerful gene-editing tool in history.
June 28th, 2012
June 2011
Defense islands formally defined in bacterial genomes
Research

Eugene Koonin's team at the National Center for Biotechnology Information published the foundational paper describing 'defense islands' — genomic regions where antiviral genes cluster — establishing the search strategy later teams would automate.
June 1st, 2011
March 2007
CRISPR proven as a bacterial immune system
Discovery

Rodolphe Barrangou and Philippe Horvath at Danisco demonstrated experimentally that bacteria use CRISPR to acquire resistance to viruses, confirming it as an adaptive immune system.
March 1st, 2007
January 1987
CRISPR sequences first noticed in E. coli
Discovery

Yoshizumi Ishino and colleagues at Osaka University observed unusual repeated DNA sequences in E. coli. Their function would remain unknown for two decades.
January 1st, 1987
January 1968
Restriction enzymes discovered in bacteria
Discovery

Werner Arber, Hamilton Smith, and Daniel Nathans identified bacterial enzymes that cut foreign DNA at specific sequences — the first characterized bacterial defense mechanism, later foundational to all of genetic engineering.
January 1st, 1968

#timeline [x-show="expanded"] { display: revert !important; opacity: 1 !important; } #timeline .timeline-toggle-btn { display: none; }

Scenarios

Next-generation gene-editing tools emerge from the new protein catalog

Likely

Discussed by: Feng Zhang (Broad Institute), Singularity Hub, Nature news coverage

The pattern holds: restriction enzymes became cloning tools, CRISPR became a gene editor, and now researchers systematically screen the new catalog for programmable nucleases, RNA-guided systems, or novel enzymatic activities suitable for biotechnology. Within two to five years, one or more of the newly discovered defense proteins is adapted into a molecular tool that complements or improves on CRISPR-Cas9 — potentially smaller, more specific, or targeting previously inaccessible genome features. Zhang's lab, which already discovered Fanzor and OMEGA systems through similar mining efforts, is actively pursuing this path.

Defense protein databases accelerate antiviral drug development

Possible

Discussed by: Nature news, News-Medical, Phys.org coverage

Pharmaceutical researchers use the databases to identify bacterial antiviral mechanisms that could be adapted against human viral pathogens. The proteins' billions of years of evolutionary refinement against viral attack make them candidates for novel antiviral therapies, diagnostics, or prophylactics — particularly for viruses where current treatments are limited. This path is longer than tool development, likely requiring five to ten years for clinical applications, but the sheer scale of the candidate pool improves the odds of finding something therapeutically useful.

Machine-learning discovery becomes the default method for finding biological mechanisms

Likely

Discussed by: Science editors, computational biologists quoted in coverage

The success of DefensePredictor and the Pasteur team's models — finding in minutes what took years via traditional methods — establishes AI-powered protein function prediction as a standard first step in microbiology. Labs apply similar approaches to other poorly understood functional categories in bacterial genomes, such as signaling systems, metabolic pathways, and symbiotic mechanisms. The five-minute screening capability effectively ends the era of discovering bacterial systems one at a time.

Most predicted defense proteins turn out to be false positives or biochemically intractable

Possible

Discussed by: Implicit in the 45% experimental validation rate; standard caution in computational biology

The 45% validation rate for the MIT team's predictions means over half of the candidates did not protect bacteria in lab conditions. At the scale of 2.39 million predicted proteins, even a high false-positive rate leaves hundreds of thousands of genuine defense systems — but the practical challenge of experimentally characterizing each one could slow the translation to usable tools. The field may spend years sorting signal from noise in the expanded databases.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

1968–1978

Restriction enzymes launch the biotech industry (1968–1978)

Werner Arber, Hamilton Smith, and Daniel Nathans discovered that bacteria use restriction enzymes to chop up foreign viral DNA at specific sequences while protecting their own DNA through chemical modification. Herb Boyer and Stan Cohen used these enzymes in 1973 to cut and paste DNA from different organisms — the birth of recombinant DNA technology.

Then

The trio shared the 1978 Nobel Prize in Physiology or Medicine. Boyer co-founded Genentech in 1976, launching the modern biotechnology industry.

Now

Restriction enzymes became the foundational toolkit for molecular biology, enabling DNA cloning, sequencing, forensics, and genetically engineered medicines — a multi-hundred-billion-dollar industry built on a bacterial defense mechanism.

Why this matters now

The pattern is identical: scientists discover how bacteria fight viruses, then repurpose the mechanism as a laboratory tool. The 2026 findings represent the largest-ever expansion of that source material.

1987–2020

CRISPR: from curiosity to Nobel Prize (1987–2020)

Yoshizumi Ishino noticed strange repeated DNA sequences in E. coli in 1987. It took 20 years for Rodolphe Barrangou to prove these sequences were an adaptive immune system. In 2012, Jennifer Doudna and Emmanuelle Charpentier showed the system could be programmed to edit any gene. Within a year, Feng Zhang demonstrated it worked in human cells.

Then

Doudna and Charpentier won the 2020 Nobel Prize in Chemistry. CRISPR-based therapies reached patients by 2023, when the first CRISPR medicine (Casgevy for sickle cell disease) was approved.

Now

CRISPR gene editing is now used across agriculture, medicine, and basic research worldwide. The $30 billion gene-editing market traces directly to an obscure bacterial defense mechanism that sat uncharacterized for two decades.

Why this matters now

CRISPR proves that a single bacterial defense system can reshape an entire industry. The new studies suggest there are thousands of undiscovered systems of comparable biochemical sophistication — any one of which could be the next CRISPR.

2020–2022

AlphaFold transforms protein science (2020–2022)

DeepMind's AlphaFold2 solved the protein-folding problem in 2020, then released predicted structures for nearly every known protein in 2022 — over 200 million structures. The achievement earned Demis Hassabis and John Jumper the 2024 Nobel Prize in Chemistry.

Then

Structural biologists gained instant access to protein shapes that previously required years of laboratory work to determine.

Now

AlphaFold established machine learning as a first-class tool in biology, creating the infrastructure and expectations that made protein language models like ESM2 — the backbone of DefensePredictor — possible.

Why this matters now

DefensePredictor and the Pasteur team's models are direct descendants of the AI-for-biology revolution that AlphaFold ignited. The same protein language models that predict structure now predict function, enabling the bacterial defense discovery at scale.