MIT has been harvesting bacteria for the next CRISPR—and found hundreds of potential new tools

CRISPR is a breakthrough technology with humble origins. For the first time, scientists have discovered a powerful gene editor in bacteria that used it as a weapon against invading viruses, so-called phages. In a day, phages can destroy up to a quarter of the bacterial population. In a never-ending arms race, bacteria under attack have developed a large arsenal of defenses.

These bacterial immune systems often digest the DNA or RNA of invading viruses and are relatively easy to produce, making them attractive targets for scientists developing genetic engineering tools. CRISPR is just one example. There are many more of them. But traditional methods of finding them are slow and time-consuming, leaving most CRISPR-like proteins unexplored.

Now MIT scientists have released an AI called DefensePredictor that can eradicate new bacterial defense systems in five minutes, not weeks or months. As a proof of concept, DefensePredictor analyzed hundreds of thousands of proteins in different strains Escherichia coli (Escherichia coli). More than 600 proteins previously not associated with immune defense were discovered. In addition to the vulnerable strain of bacteria, some protect them from attack.

“Escherichia coli contains a much broader landscape of antiphage defense than previously thought, increasing the likely number of systems by several orders of magnitude,” the team wrote.

These systems may hold secrets to how immunity evolved. And because proteins can work in different ways, they could be a goldmine for next-generation precision molecular tools.

Unsurpassed success

About three decades ago, Japanese scientists discovered an interesting, repetitive DNA sequence Art Escherichia coli. Other researchers soon realized it was widespread among bacterial species and matched it with viral DNA sequences, suggesting it might be part of the bacteria’s immunity against phages.

The system, now known as CRISPR, stores fragments of DNA from past infections and uses protein “scissors” to cut the corresponding viral DNA during reinfection. Intrigued by its precision, scientists repurposed CRISPR into a variety of gene-editing tools and revolutionized gene therapy.

CRISPR is the best known, but a whole host of bacterial defense systems have been revolutionized by genetic engineering. One that contains enzyme which cuts specific sequences of foreign DNA, is widely used to add genetic material to cells. Another encodes a balance of toxins and antitoxins that can cause bacterial death after phage infection. This one was adapted into a switch to prevent the uncontrolled spread of engineered microbes or genetically modified cultures.

Researchers are also studying the use of newly discovered systems – with names similar to video games, for example Zorya and Terris— as molecular sensors and programmable signals in synthetic biology.

There are likely more undiscovered tools in the bacterial defense universe, and scientists have ways to hunt them down. There are some protection genes grouped close togethertherefore, a known gene could guide the discovery of others. Researchers too genes found by screening libraries of free-floating circular genome fragments in bacterial populations.

Finished 250 systems were thoroughly tested. But much more can escape modern detection methods if, for example, their components are distributed throughout the genome.

“The full repertoire of antiphage defense systems in bacteria remains unknown,” the team wrote. “We currently lack the tools to systematically identify systems with high speed, sensitivity, and specificity.”

AI Discoverer

DefensePredictor’s new algorithm closes this gap.

At its core is a protein language model called ESM-2. Proteins are made up of 20 molecular “letters” that join together into threads and fold into complex three-dimensional shapes. Like large language models, algorithms like ESM-2 learn the language of proteins and can predict their structure and function based on sequence alone.

ESM-2 and other similar algorithms have already helped scientists decipher mysterious proteins in bacteria, viruses and other microorganisms previously unknown to science. The researchers hope that their unique shapes can provide inspiration antibiotics, biofuelor even get used to it build synthetic organisms.

To build their AI, the team first created a training ground. With the previous model, DefenseFinderthey screened approximately 17,000 microbial genomes for genes related and unrelated to defense systems. They translated these genes into the corresponding proteins and created a database of about 15,000 antiphage proteins and 186,000 proteins not related to defense.

These numbers are too staggering for a human, but artificial intelligence took the job calmly. Together with ESM-2, the model used several algorithms to distinguish between defensive and non-defensive proteins. Eventually DefensePredictor learned some common characteristics that make a protein more likely to be part of the immune system. (As with other language models, it’s hard to fully understand the reasoning behind the system, which the team is still trying to unpack.)

When tested on 69 strains Escherichia coliDefensePredictor discovered a treasure trove of more than 600 new defense-related proteins, including more than 100 that differed from anything previously discovered. While some were encoded next to each other or in circular DNA, like previous findings, nearly half were not. Instead, they were scattered throughout the genome, but still able to work together.

To test the results, the team designed a highly vulnerable Escherichia coli strain to express candidate defense proteins predicted to function individually or as part of a system and exposed them to two dozen aggressive phages. Almost 45 percent of the proteins provided protection against at least one phage.

Outside the borders Escherichia coliscientists expanded their search to 1,000 more microorganisms and discovered thousands of potential protective proteins unlike anything seen before. “New immune mechanisms remain to be discovered,” the team wrote.

The race is on. Also published this week by the Pasteur Institute team combined several models of artificial intelligence to search for antiphage systems in protein sequences. In more than 32,000 bacterial genomes, the model predicted nearly 2.4 million antiphage proteins—most of which were previously unknown. Released Atlas AI-predicted bacterial immunity proteins for others to study.

“The diversity of phage defense systems is vast and largely unexploited,” they wrote.

Microorganisms harbor a colossal array of biological tools that we are only beginning to uncover at scale. More species are being found all the time, thriving in a variety of environments, from pond scum to boiling sulfur springs and the crushing pressure of the Mariana Trench. Each new genome that scientists now discover and dissect with the help of artificial intelligence may harbor the next CRISPR.