AI takes a step toward designing synthetic life
Researchers report an AI system that can draft entire genome sequences, marking a new move for synthetic biology. The study in Nature describes a DNA language model, Evo2, trained on trillions of genetic letters spanning much of life's diversity. Using Evo2, scientists generated multiple genome designs, including one inspired by Mycoplasma genitalium.
This builds on past milestones. In 2008, teams chemically synthesized the M. genitalium genome, and later "rebooted" similar genomes inside cells-early proof that synthetic genomes can run living systems.
Even so, the gap between a digital sequence and a working organism is still large. "It's cool, but it's not there yet," said Nico Claassens, a synthetic biologist at Wageningen University.
What the AI actually did
Evo2 treats DNA like language, learning patterns from vast genomic datasets and proposing plausible sequences. Among its outputs: complete bacterial-scale genome drafts, including one guided by M. genitalium. These are hypotheses, not validated organisms.
Earlier Evo models produced viral genomes that could infect bacteria-orders of magnitude simpler than bacterial genomes. With cells, small design errors can block replication, break regulation, or kill fitness. As Claassens put it: "You cannot design life 70%. You can do that on a computer, but it will not be functional."
Why this matters for working scientists
AI-generated drafts can compress the early design phase, letting teams explore many genome architectures before touching the bench. Paired with build-test-learn cycles and automation, this could speed hypothesis testing and surface non-obvious designs worth trying.
- Treat model output as a starting point. Check synthesis constraints (GC windows, repeats, homopolymers), assembly strategy, and propagation hosts.
- Plan a validation pipeline: essential gene sets, operon structure, regulatory elements (promoters, RBS, terminators), codon usage, and genome minimization assumptions.
- Decide on a chassis and build route (transplantation, recombineering, yeast assembly). Align with robotics for high-throughput screening and phenotyping.
- Risk and compliance: biosafety levels, dual-use review, and institutional approvals come first-especially for any infectious potential.
If you're mapping AI to lab automation and data infrastructure, see our internal primer: AI for Science & Research.
Limits to keep in mind
- Genotype-to-phenotype mapping remains incomplete. Regulatory networks, chromosome topology, and noncoding RNA effects are hard to predict.
- Epistasis is everywhere. "Silent" edits may shift fitness, timing, or toxicity in unexpected ways.
- Synthesis isn't trivial. Error rates, repetitive regions, and toxic fragments can stall cloning and assembly.
- Cost and time still matter. Even with automation, end-to-end validation can take months per design cycle.
What to watch next
- Tighter closed-loop systems: AI design linked to robotic build, automated QC, multiplex assays, and model retraining.
- Stronger objective functions: viability and essentiality constraints embedded directly into sequence generation.
- Safer testbeds: minimal cells and cell-free platforms for early screening before whole-cell trials.
- Standards and benchmarks for AI-generated genomes, plus transparent data sharing for reproducibility.
Expert voices
Patrick Yizhi Cai calls these models the "ChatGPT moment" for synthetic genomics: "You can start writing things that never existed in nature." It's a bold vision-tempered by the current reality that in silico success doesn't guarantee a living cell.
Further reading
- Nature - coverage and research on synthetic genomics and genome design.
- Creation of a bacterial cell controlled by a chemically synthesized genome (Science, 2010) - a landmark in synthetic genome booting.
Your membership also unlocks:
