5 min read

An AI company does its own structural biology

An AI company does its own structural biology
Nº 01 · The Lede Anthropic Field report

Anthropic reports first 3D structure of prokaryotic vault

Anthropic reports first 3D structure of prokaryotic vault
Fig. IAnthropic · Filed 05 Nov 2026.

Anthropic's biology research lab reported what it describes as the first 3D structure of a vault particle found in prokaryotic cells. Anthropic, which builds the Claude family of AI models, presents the result as another finding from that in-house lab. The post's title, "Investigating vault particles with Claude," puts the company's own model at the center of the work. If the structure holds, researchers studying prokaryotic vaults now have a first 3D model to test ideas about function against. In the full post, look for how the work was split between Claude and the lab's human researchers.

Read the source →

BEAKER tests the AI that runs robotic chemistry labs
Fig. IIarXiv · Filed 05 Nov 2026.
Nº 02 arXiv Field report

BEAKER tests the AI that runs robotic chemistry labs

Self-driving chemistry labs get a benchmark of their own in BEAKER, described in a new arXiv paper. A benchmark is a shared test for comparing AI models. Self-driving labs pair robotic hardware with software that plans and runs experiments with little human input. BEAKER scores what its authors call the embodied brain, meaning the AI model that decides what the robot does rather than the robot itself. Domain experts curated the test, so rival models for lab automation can be scored against one expert-built standard.

Read more →
PaleoRigor keeps an evidence trail for ancient microbiomes
Fig. IIIbioRxiv · Filed 05 Nov 2026.
Nº 03 bioRxiv Computational biology

PaleoRigor keeps an evidence trail for ancient microbiomes

A research agent for ancient microbiomes records its work as a reproducible evidence trail, according to a new bioRxiv preprint. PaleoRigor is an agent (software that carries out multi-step work on its own) built to process paleomicrobiome data. Its authors call it bounded, meaning it operates within defined limits. The trail is the point. Processing ancient microbial DNA involves many judgment calls, and a record another group can rerun makes an agent's choices checkable instead of taken on trust.

Read more →
Also Filed · Two Briefs from the queue
Nº 04 bioRxiv Field report

XAMPGen pipeline finds new antimicrobial peptides

XAMPGen proposes new antimicrobial peptides using a computational pipeline designed to be interpretable, so researchers can see why a candidate was suggested. The bioRxiv preprint aims to discover novel AMPs.

Read →
Nº 05 arXiv Agents · Infrastructure

FrontierPhysics tests AI agents on full physics research

Physics gets its own agent test in FrontierPhysics, an arXiv benchmark that evaluates AI agents on physics research from start to finish. Like the chemistry-lab benchmark above, it grades agents on doing science, here in a different field.

Read →

Reply with your discoveries. A human reads them. Forward freely.

Agentic Discovery  ·  Nº 136  ·  05 Nov 2026

Editor's Note

One AI company publishes a structure while two new benchmarks ask whether agents can do science at all.

 

Nº 01 · The Lede  —  Anthropic  —  Field report

Anthropic reports first 3D structure of prokaryotic vault

Anthropic reports first 3D structure of prokaryotic vault

Fig. I  Anthropic · Filed 05 Nov 2026.

Anthropic's biology research lab reported what it describes as the first 3D structure of a vault particle found in prokaryotic cells. Anthropic, which builds the Claude family of AI models, presents the result as another finding from that in-house lab. The post's title, "Investigating vault particles with Claude," puts the company's own model at the center of the work. If the structure holds, researchers studying prokaryotic vaults now have a first 3D model to test ideas about function against. In the full post, look for how the work was split between Claude and the lab's human researchers.

Read the source →

Why it matters

An AI developer is now publishing primary structural biology under its own name, not just building the models other labs use.

 

Nº 02  —  arXiv  —  Field report

BEAKER tests the AI that runs robotic chemistry labs

Fig. II  arXiv · Filed 05 Nov 2026.

BEAKER tests the AI that runs robotic chemistry labs

Self-driving chemistry labs get a benchmark of their own in BEAKER, described in a new arXiv paper. A benchmark is a shared test for comparing AI models. Self-driving labs pair robotic hardware with software that plans and runs experiments with little human input. BEAKER scores what its authors call the embodied brain, meaning the AI model that decides what the robot does rather than the robot itself. Domain experts curated the test, so rival models for lab automation can be scored against one expert-built standard.

Read more →

 

Nº 03  —  bioRxiv  —  Computational biology

PaleoRigor keeps an evidence trail for ancient microbiomes

Fig. III  bioRxiv · Filed 05 Nov 2026.

PaleoRigor keeps an evidence trail for ancient microbiomes

A research agent for ancient microbiomes records its work as a reproducible evidence trail, according to a new bioRxiv preprint. PaleoRigor is an agent (software that carries out multi-step work on its own) built to process paleomicrobiome data. Its authors call it bounded, meaning it operates within defined limits. The trail is the point. Processing ancient microbial DNA involves many judgment calls, and a record another group can rerun makes an agent's choices checkable instead of taken on trust.

Read more →

 

Also Filed  ·  Two Briefs from the queue

Nº 04  —  bioRxiv  —  Field report

XAMPGen pipeline finds new antimicrobial peptides

XAMPGen proposes new antimicrobial peptides using a computational pipeline designed to be interpretable, so researchers can see why a candidate was suggested. The bioRxiv preprint aims to discover novel AMPs.

Read →

Nº 05  —  arXiv  —  Agents · Infrastructure

FrontierPhysics tests AI agents on full physics research

Physics gets its own agent test in FrontierPhysics, an arXiv benchmark that evaluates AI agents on physics research from start to finish. Like the chemistry-lab benchmark above, it grades agents on doing science, here in a different field.

Read →

 

· · ·

Reply with your discoveries. A human reads them. Forward freely.