This Week in Evolution TWiEVO 121: AI imagines viruses
Jan 8, 2026
Nels and Vincent dive into the fascinating world of AI-designed viruses, focusing on bacteriophage phiX174. They discuss the potential and risks of AI in science, including recent NIH funding turmoil. The hosts clarify misconceptions about the paper's findings and the use of genomic language models. They explore the challenges of phage therapy, the competitive edge of synthetic phages, and the ethical considerations of bioengineering. They also emphasize the importance of robust science policy and community involvement in shaping the future of research.
AI Snips
Chapters
Books
Transcript
Episode notes
Human Curation Drives Viability
- Model outputs are often non-viral until heavily filtered and steered with human knowledge and constraints like genome size and GC content.
- The human-supervised fine-tuning (SFT) dominates viability, revealing substantial human intervention behind 'AI-designed' sequences.
Most Variants Look Like Modified PhiX174
- The AI-generated PHIX variants were ~95% identical to the original, representing modest divergence rather than de novo genomes.
- Such divergence (60–392 mutations) is comparable to what simple lab evolution would produce in weeks.
Build Ethics Into Experimental Design
- Pair technological advances with strong, funded bioethics and biosafety efforts from project start to mitigate risks responsibly.
- Invest in humanities and ethics alongside STEM to embed safety and societal context into research programs.
