You’ve probably seen the headlines. AI is learning to “speak DNA.” Foundation models like Evo and Genomics-FM are treating our genetic code like just another language, ready to be parsed, embedded, and generated. We’re on the brink of inventing new proteins, predicting evolutionary trajectories, and maybe even designing entirely new creatures in a biological sandbox.
It sounds like science fiction becoming reality. But there’s a terrifying catch hiding in the code.
We’ve spent decades convincing ourselves that DNA is a linear blueprint. A static string of TCAG base pairs that you can read like a book. But biology doesn’t work like that.
DNA isn’t a blueprint; it’s a chaotic, feedback-driven mosh pit where context is everything.
When you map discrete, deterministic genetic sequences into the continuous, probabilistic latent space of an AI embedding, you flatten the very thing that makes life work. You ignore the 3D folding of the helix. You ignore epigenetics. You ignore the dynamic, environment-interacting regulatory elements that actually govern whether a gene wakes up or stays asleep.
Sure, you can train a model on all the DNA in the world. You can ask it to generate a novel sequence. It might even spit out something that looks perfectly valid to a sequencer.
But an AI that only reads the letters will happily design a genome that is biologically inert—or worse, actively harmful.
Think about it. If you model a C. elegans worm as a flat array of vectors, you might be able to play with it in a sandbox. But when you try to synthesize that AI-generated DNA in a lab, the lack of real-world context will bite you. The regulatory elements won’t trigger. The protein folding will collapse. The “creature” won’t live.
This is the fundamental tension of the bleeding edge of AI and biology. We want the awe of wielding AI to invent new life. We want to test drugs on synthetic humans modeled in silicon. But we are treating a highly non-linear system like a static text file.
Foundation models can absolutely learn patterns in DNA. They can help us understand evolution. But until we figure out how to embed the 3D structure, the environmental triggers, and the regulatory chaos, we aren’t playing god. We’re just playing with matches.
Life isn’t a sequence to be read. It’s a system to be experienced. Any AI that forgets that will build dead code.
FAQ
Q: Can't AI just learn the context if we feed it enough data?
A: No. You can't solve a 3D, time-dependent, environmentally reactive system by just feeding it more 1D text. The missing dimensions aren't a data problem; they're a fundamental representational problem.
Q: What's the practical implication for drug discovery?
A: It means AI-generated protein targets might look great on a screen but fail instantly in a living cell. You can't skip the biological context when designing therapeutics.
Q: Are foundation models like Evo completely useless then?
A: Not at all. They're incredible for pattern recognition and designing novel proteins. But treating them as a 'biology sandbox' that can predict future states of humans is a massive, dangerous overreach.