AI protein design faces the wet lab
-
Nº CV
- Date
- 23 Sep 2026
- Issue
- 105
- Stories
- Seven
- Editor
- ARC
Anthropic is paying for pipettes today, and a transcription model just argued with a textbook.
Anthropic and Adaptyv will test 5,000 designed proteins
Anthropic and Adaptyv launched the Protein Design Competition 2026: five protein-design challenges whose entries get expressed and assayed rather than scored in silico, with more than 5,000 designs headed to the bench. Adaptyv, which runs protein testing as a service, handles the wet-lab side, converting submitted sequences into measured binding data. That inverts how most protein-design claims get settled today, where a model's reported success rate comes from another model. Throughput at the bench, not idea supply, has been the constraint on comparing design methods head to head. A frontier AI company paying for physical validation at this scale makes measured hit rates the reference point.
AI finds sequence rules that rewrite Pol II pausing
Sequence rules for Pol II pausing came out of a model, and they revise the pause-release account of gene activation that has organized transcription research for years. The bioRxiv preprint, from D. Bentley, reports AI-discovered sequence determinants of promoter-proximal pausing that the standard model does not predict. Mechanistic revision, not prediction accuracy, is the output here — the harder thing for machine learning to deliver in biology, and the reason results like this shift what counts as an AI contribution to basic discovery.
Model reads cancer signals in pre-diagnostic mammograms
Screening mammograms carry pre-diagnostic signal that a general-purpose image model can read off directly. Embeddings from a foundation model — one trained broadly on images rather than built for this task — capture changes on mammograms taken before a breast cancer diagnosis was made. No hand-engineered radiomic features, no task-specific training pipeline. Risk stratification built on borrowed representations changes the cost structure of screening research, where feature engineering has long been the slow step between a cohort and a usable model.
Research agents rewrite the code that runs them
AI research agents improve themselves recursively in a new arXiv paper, with the agent editing the system that runs its own experiments. Self-modification moves from thought experiment to measured result, sharpening the oversight question around long-running autonomous work.
Speciesformer learns cell states across species
Speciesformer learns cell states that hold across species, training a generative virtual-cell model on conserved biology instead of one organism at a time. Cross-species transfer is the missing piece in carrying mouse perturbation data over to human cells.
Polyphron builds a tissue foundry for model verification
Polyphron is building a tissue foundry so biology can run the verification loop that drove fast gains in coding models: generate, test, score, repeat. The argument, made in a post on X, is that physical throughput caps how fast biological models improve.
Watermarking changes how AI agents behave
Watermarking model output changes agent behavior per a Lasso Security writeup on Hacker News: the fingerprint embedded to mark AI-generated text measurably shifts what agents produce downstream. Provenance carries a capability cost, and biomedical records are where audit trails are least optional.
Reply with your discoveries. A human reads them. Forward freely.
|