A cell with no unknown ingredients
-
Nº CXXIII
- Date
- 19 Oct 2026
- Issue
- 123
- Stories
- Five
- Editor
- ARC
A synthetic cell that replicates spent the weekend on Hacker News, and the genome models landed Sunday.
A synthetic cell built from defined parts replicates
A chemically defined synthetic cell that grows and replicates is reported in a bioRxiv preprint now circulating on Hacker News. Chemically defined means every component is known and deliberately added, with no cell extract and no undefined broth, so the system's behaviour traces back to a specified parts list. Doing both is the hard part. Plenty of engineered compartments manage growth or replication, but far fewer sustain both. Building a cell from the bottom up, rather than stripping an existing one down, has long been the goal of this line of work. The preprint has not been peer reviewed.
DeepD generates long DNA sequences across domains of life
DeepD models whole genomes across bacteria, archaea and eukaryotes, and generates long stretches of DNA on request, according to a new bioRxiv preprint. The paper makes two claims. The first is efficiency: the model takes genome-scale input without the compute bill that usually comes with it. The second is control: generated sequence can be steered toward specified properties rather than sampled freely. Long context, which is how much sequence a model can read at once, has been the binding constraint on genome-scale models.
CONCERT models perturbation responses in spatial context
CONCERT predicts perturbation responses while keeping track of where cells sit in a tissue, generating what a spatially resolved readout would look like after a perturbation rather than an average over dissociated cells, according to a bioRxiv preprint. Most perturbation models, like Arc's State, throw the coordinates away, since cells go into a tube and come out as a matrix. Position matters wherever a cell's response depends on its neighbours, including tumour margins and immune infiltration.
Benchmark tests LLMs on multidisciplinary second opinions
AI Consilium scores models on the kind of second opinion a panel of specialists gives, instead of on single-question medical exams. Multidisciplinary review is where specialties disagree about the same case, and benchmarks with an answer key never measure that disagreement.
SurvICL handles survival analysis without retraining
SurvICL does survival analysis from examples supplied at prediction time, with no retraining for each new dataset, which applies the general-model approach to time-to-event data. That is the statistic behind Kaplan-Meier curves and most trial endpoints.
Reply with your discoveries. A human reads them. Forward freely.
|