7 min read

Your next method ships as a skill

Your next method ships as a skill
Nº 01 · The Lede bioRxiv Computational biology

A bacterial epigenomics toolkit ships as an AI skill

A bacterial epigenomics toolkit ships as an AI skill
Fig. IbioRxiv · Filed 08 Sep 2026.

MErlin ships as agent skill rather than a command-line package, putting a bacterial epigenomics multi-omics workflow behind plain-language requests. A skill here is a bundle of instructions plus tool calls that an assistant loads on demand, so the analysis logic stays fixed while the interface disappears. The bioRxiv preprint from I. Passeri frames it as a move from code to natural language, with the toolkit itself as the published artifact. That shifts the distribution unit for method development: a group releasing a skill hands over a runnable workflow instead of a repository someone else has to install, configure, and learn. Specialist epigenomics analysis becomes something an agent can execute on request.

Read the source

Karenina scores biomedical AI on expert-written rubrics
Fig. IIbioRxiv · Filed 08 Sep 2026.
Nº 02 bioRxiv Computational biology

Karenina scores biomedical AI on expert-written rubrics

Karenina turns expert rubrics into multi-dimensional evaluations of biomedical AI, letting specialists state what a good answer looks like along several axes instead of collapsing quality into one accuracy figure. The bioRxiv preprint positions domain expertise as an input to evaluation design rather than a manual review step bolted on at the end. That makes expert-defined scoring a buildable component, and raises the floor for what counts as evidence that a biomedical model is fit for use.

Read more
Language models plus ontologies rank rare-disease diagnoses
Fig. IIIarXiv · Filed 08 Sep 2026.
Nº 03 arXiv Field report

Language models plus ontologies rank rare-disease diagnoses

Ontology ranking meets language models in a learned combination aimed at rare-disease diagnosis, where candidates have to arrive as a ranked, phenotype-grounded shortlist rather than free text. The arXiv paper learns the fusion instead of hand-tuning the weight between the two signals, keeping structured ontology evidence in the loop. Narrows the gap between chat-style differential suggestions and the ranked outputs clinical genetics workflows already consume — the place that has been least willing to accept an unranked answer.

Read more
Also Filed · Three Briefs from the queue
Nº 04 arXiv Structural biology · Protein design

A retrospective grades five years of ML protein engineering

Machine-learning directed evolution reviewed in a five-year retrospective from Bruce Wittmann, tallying which sequence-function methods actually improved protein engineering campaigns. Gives the field a shared reference point for what the last wave of protein ML delivered.

Read
Nº 05 X Field report

Open Science Index gathers open tools in one place

Open Science Index launches as one searchable home for open-source tools, models, datasets, and workflows across biology, chemistry, physics, materials, and earth science. Open-tool discovery has lived in GitHub and preprint supplements; one index changes where people and agents look first.

Read
Nº 06 X Field report

An investor makes the case for bench robots

An investor post argues that the best target for physical AI is the bench: robots that pipette, handle stem cells, and run existing instruments. Which experimental bottlenecks get automated follows where this kind of capital lands.

Read

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

Agentic Discovery  ·  Nº 94  ·  08 Sep 2026

Editor's Note

Today the unit for shipping a bioinformatics method quietly changed from a repository to a skill.

 

Nº 01 · The Lede  —  bioRxiv  —  Computational biology

A bacterial epigenomics toolkit ships as an AI skill

A bacterial epigenomics toolkit ships as an AI skill

Fig. I  bioRxiv · Filed 08 Sep 2026.

MErlin ships as agent skill rather than a command-line package, putting a bacterial epigenomics multi-omics workflow behind plain-language requests. A skill here is a bundle of instructions plus tool calls that an assistant loads on demand, so the analysis logic stays fixed while the interface disappears. The bioRxiv preprint from I. Passeri frames it as a move from code to natural language, with the toolkit itself as the published artifact. That shifts the distribution unit for method development: a group releasing a skill hands over a runnable workflow instead of a repository someone else has to install, configure, and learn. Specialist epigenomics analysis becomes something an agent can execute on request.

Read the source →

Why it matters

The repository has been the unit of method-sharing in computational biology for two decades; a published agent skill makes the assistant, not the terminal, the place an analysis gets run, and every methods paper now inherits a packaging question it did not have last year.

The Bench NoteFrom Heureka Labs

Writing a method down as instructions an agent follows is becoming a real way to publish it.

Where skills live. In Bench a skill is an editable file in your workspace or a single project, and ARC reaches for it on its own when a task calls for one.
The registry. Published skills sit in a public registry, searchable from inside the app, reviewed before they go live and checksum-verified on the way in; installing never replaces one you already have.
Who gets the credit. Registry skills are licensed CC BY 4.0 and credited to the author by name, and installs a researcher chose are counted separately from ones the agent made mid-task.

What we’re watching: whether a published skill starts getting cited the way a package is, and what an install count comes to mean

 

Nº 02  —  bioRxiv  —  Computational biology

Karenina scores biomedical AI on expert-written rubrics

Fig. II  bioRxiv · Filed 08 Sep 2026.

Karenina scores biomedical AI on expert-written rubrics

Karenina turns expert rubrics into multi-dimensional evaluations of biomedical AI, letting specialists state what a good answer looks like along several axes instead of collapsing quality into one accuracy figure. The bioRxiv preprint positions domain expertise as an input to evaluation design rather than a manual review step bolted on at the end. That makes expert-defined scoring a buildable component, and raises the floor for what counts as evidence that a biomedical model is fit for use.

Read more →

 

Nº 03  —  arXiv  —  Field report

Language models plus ontologies rank rare-disease diagnoses

Fig. III  arXiv · Filed 08 Sep 2026.

Language models plus ontologies rank rare-disease diagnoses

Ontology ranking meets language models in a learned combination aimed at rare-disease diagnosis, where candidates have to arrive as a ranked, phenotype-grounded shortlist rather than free text. The arXiv paper learns the fusion instead of hand-tuning the weight between the two signals, keeping structured ontology evidence in the loop. Narrows the gap between chat-style differential suggestions and the ranked outputs clinical genetics workflows already consume — the place that has been least willing to accept an unranked answer.

Read more →

 

Also Filed  ·  Three Briefs from the queue

Nº 04  —  arXiv  —  Structural biology · Protein design

A retrospective grades five years of ML protein engineering

Machine-learning directed evolution reviewed in a five-year retrospective from Bruce Wittmann, tallying which sequence-function methods actually improved protein engineering campaigns. Gives the field a shared reference point for what the last wave of protein ML delivered.

Read →

Nº 05  —  X  —  Field report

Open Science Index gathers open tools in one place

Open Science Index launches as one searchable home for open-source tools, models, datasets, and workflows across biology, chemistry, physics, materials, and earth science. Open-tool discovery has lived in GitHub and preprint supplements; one index changes where people and agents look first.

Read →

Nº 06  —  X  —  Field report

An investor makes the case for bench robots

An investor post argues that the best target for physical AI is the bench: robots that pipette, handle stem cells, and run existing instruments. Which experimental bottlenecks get automated follows where this kind of capital lands.

Read →

 

· · ·

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