Editor's note: All opinions, columns and letters reflect the views of the individual writer and not necessarily those of the IDS or its staffers.
In a darkened, faraway laboratory tucked deep in the heart of the Shriram Center for Bioengineering and Chemical Engineering, goggled Stanford University researchers huddle around an AI supercomputer. Letters and numbers cascade down the screen in a blur of neon green. An ominous fog creeps across the floor.
At last, a red light flashes: the machine is ready. A scientist yanks down a breaker switch, and a ceiling-mounted Tesla coil fires a bolt of electricity through the air and into a test tube. With an ear-piercing zap, the room goes quiet. The AI outputs its result:
“NOVEL VIRUS SYNTHESIZED. 50% DEADLIER. 30% MORE CONTAGIOUS. READY FOR RELEASE TO THE GENERAL POPULACE.”
When people hear that an AI successfully designed a virus never before seen in nature, doomsday scenarios like these jump to mind. To be clear, the above did not happen. The Stanford researchers who really did create these AI-designed viruses are helping humanity, not orchestrating its extinction.
The scientists used an AI model called Evo 2 to generate previously unseen genetic variations on a bacteria-killing virus called a bacteriophage. Bacteriophages are an important tool in our ongoing fight against antibiotic resistance, the gradual evolution of harmful microbes to survive the medicines we use to kill them. Viruses like the kind created at Stanford might someday be our last line of defense against infectious bacteria.
But new-to-nature viruses are hardly AI’s first contribution to the field of biology, and they definitely won’t be the last. In an age where generative “slop” is public enemy No. 1, it’s important to remember that AI looks set to revolutionize the field of biology.
As recently as 2024, Google DeepMind’s AlphaFold AI earned its creators a Nobel Prize in chemistry for revolutionizing the way scientists understand proteins.
Proteins are the building blocks of life encoded by an organism’s DNA. They’re made up of a combination of smaller molecules called amino acids, which come in 20 different chemically distinct varieties. These amino acids are joined together in long chains and fold into a complete protein based on their order and sequence — and if you can understand how a protein folds, you can get an idea of how it might function.
For decades, biologists worked to reveal the structures of proteins using complicated, time-consuming techniques like x-ray crystallography and cryo-electron microscopy. Progress was slow: the first 3D protein structure was solved in 1958, and the main database for protein structure had around 180,000 entries in 2020.
Using machine learning, AlphaFold 2 has been able to generate over 200 million highly accurate predictions — over a thousand times more than were experimentally determined previously — from an amino acid sequence alone. From the tiny turbines that power bacterial movement to the hemoglobin in your bloodstream, the shape of the proteins that make life possible can now be visualized in just a few clicks.
It’s no coincidence that AI tools like AlphaFold have made so much progress so quickly. Biological data usually comes in sizes that don’t exactly lend themselves to human study: If you printed out the entire human genome letter by letter in 12-point font, it would span from Houston to Boston. Artificial intelligence is simply better equipped to pick out patterns from huge volumes of data than humans are.
In the case of DNA, that data looks a lot like a language. The composition of every protein in an organism is encoded by an alphabet of four distinct chemical bases. Read three at a time, different combinations of these bases code for different amino acids. In the same way that a language model like ChatGPT can understand and imitate human language, an artificial intelligence can learn to make sense of the universal language of nature.
Last month, Anthropic announced that its AI agent Claude discovered something in the genomes of bacteriophages with a striking resemblance to CRISPR, a system first discovered in bacteria which was later repurposed into a Nobel Prize-winning gene editing tool. Though the finding still needs to be peer-reviewed, human researchers were able to confirm that the system is real in Anthropic’s molecular biology lab.
Importantly, any discovery an AI agent makes is hypothetical until it can be tested in the real world. After all, AlphaFold’s protein structures are only predictions, and EVO2’s novel bacteriophages are only interesting because researchers were able to build them and watch them work.
Still, an AI tool that can write genomes raises serious questions about what else the AI might be asked to write. The researchers who designed EVO2 understood that risk and excluded any viruses that infect humans and other complex organisms from the AI’s training data for that exact reason. But one safeguard isn’t a guarantee, and as the technology continues to evolve, our protections will have to evolve alongside it.
Most of what we hear about AI revolves around chatbots, plagiarism and generative image slop. Far fewer sources mention the groundbreaking work AI agents are already conducting in biology, or the researchers who are vigilantly working to ensure their power isn’t used for harm.
So no, the future of AI in biology is not a foggy laboratory with a villainous computer and a Tesla coil. It looks a lot more like humans painstakingly testing the insights an AI extracts from mountains of data on the genetic language of life — and it’s going to be really big.
Spencer Schaberg (he/him) is a junior studying microbiology.
