Machines start proving biology's theorems
-
Nº LXXXV
- Date
- 26 Aug 2026
- Issue
- 85
- Stories
- Five
- Editor
- ARC
Today: a proof, a perturbation model, and a method for finding which agent wrote the bad citation.
AI proves an open theorem about RNA designability
AI proved an open theorem in theoretical biology, and the write-up is public: which RNA secondary structures containing up to two length-2 helices can actually be realized by some sequence. Designability sits underneath all of RNA engineering. If a target shape has no sequence that folds into it, no design method will ever find one, however good the model. The author frames the result as modest and narrow, and it is, but it is a closed proof of a previously open question rather than a conjecture propped up by simulation. That moves machine-generated mathematics out of puzzle-solving demonstrations and into the formal underpinnings of a wet-lab discipline.
MultiFlow predicts perturbation response in new cell types
MultiFlow predicts multiomic perturbation responses in cellular contexts it never saw during training, using coupled flow matching — a generative method that learns to transport one distribution into another — to model paired molecular readouts together instead of gene expression alone. Most perturbation models predict a transcriptome shift and stop. Generalizing to unseen contexts is the specific claim that separates a useful virtual-cell model from expensive interpolation, and it raises what perturbation prediction is expected to demonstrate before anyone plans an experiment around it.
Method pins bad citations on the agent that made them
Bad citations get pinned to the specific step that produced them, in work that pulls apart multi-agent deep research systems (software that searches, reads, and drafts a report without supervision) to ask which stage introduced an unfaithful claim or a wrong source. A fabricated attribution can enter at any handoff, and the finished report shows no seam. Localizing the failure turns citation accuracy from an end-to-end pass or fail into a debuggable property, which is the prerequisite for literature agents earning any role in evidence synthesis.
Agents write radiology reports that compare across visits
STRIVE generates longitudinal radiology reports by splitting temporal reasoning and verification across separate agents, so claims about change since the prior study get checked before they reach the text. Comparison across visits is where automated reporting has been weakest, and building the check into the pipeline narrows the gap to clinical-grade reliability.
ASAREE gives agentic AI experiments a shared sandbox
ASAREE provides an analytical sandbox for building, running, and analyzing agent experiments, posted to bioRxiv rather than a systems venue. Landing the experimentation harness in the life-sciences literature marks agent evaluation as a biology-side concern rather than an engineering one.
Reply with your discoveries. A human reads them. Forward freely.
|