12 Aug 2026 7 min read

AI just wrote a whole genome

Nº 01 · The Lede X Computational biology

Genome-scale generative design lands

Generative AI designed a whole genome , not a protein and not a gene. Work from Samuel H. King and coauthors, circulating on X over the past week, used Evo 1 and Evo 2 (genome language models trained on raw DNA rather than protein sequence) to generate genomes the authors report as viable. Every generative-biology result until now has operated on a part: a binder, an enzyme. This one operates on the system, pushing the design frontier from single molecules to whole replicating genomes and giving synthetic biology a new answer to what a first draft looks like. The viability claim will draw hard scrutiny, and it should.

Read the source →

Nº 02 X Agents · Infrastructure

MDArena scores simulation agents

MDArena benchmarks coding agents on realistic molecular dynamics workflows, measuring whether an agent can run a simulation end to end rather than answer questions about one. Reliability on multi-step scientific computation has been the soft spot in every research-agent pitch, and physics-based simulation is unforgiving: a wrong thermostat setting yields plausible numbers and useless science. Gives the field a reference score for agent-executed simulation, a domain where capability claims have run well ahead of measurement.

The Bench NoteFrom Heureka Labs

Simulation is a shrewd place to start scoring agents, because it punishes confident wrongness that other tasks let slide.

The second question. A score tells you how often an agent got it right across a test set; the other half is whether you can see how, on the one run sitting in front of you.
What the run leaves behind. ARC's analyses write the script and the figures into the project alongside the summary, so the work can be re-read and re-run months later.
Before it starts. For consequential work ARC drafts a plan and waits for approval, and anything in flight shows as a pill with a stop button.

What we’re watching: whether scores like this begin reporting how often an agent caught and fixed its own mistake, rather than only whether the final answer landed

Nº 03 arXiv Agents · Infrastructure

Agent picks the next experiment

LLM-assisted Bayesian experiment design drives a new agent that proposes which experiment to run next, aiming at mechanistic models rather than curve fits. The Model Discovery Agent, from a group including Kevin Murphy, pairs language-model hypothesis generation with Bayesian optimal experimental design so each round of data buys more model than the last. Data efficiency is the whole ballgame where a measurement costs an animal or a patient visit, and this closes the agent loop around experiment selection instead of analysis alone.

Also Filed · Four Briefs from the queue

Nº 04 Hacker News Field report

Materials discovery gets its startup

Discovered Materials launched on Hacker News with AI agents pointed at finding new materials, out of Y Combinator's current batch. Materials search shares its structure with molecular discovery, so the agent-driven-discovery company now has a second market testing whether autonomous search beats trained intuition.

Read →

Nº 05 arXiv Field report

Auto-research as fuzz testing

Autonomous research reframed as fuzz testing in a new arXiv paper, arguing that agents probing hypothesis space behave like fuzzers probing a program for crashes. The analogy imports decades of coverage and reproducibility tooling into how automated discovery gets judged.

Read →

Nº 06 bioRxiv Field report

Interactome embeddings get a map

A map of protein-protein interaction embeddings turns human interactome structure into a searchable coordinate space for functional discovery. Extends embedding-based inference from sequence to interaction networks, widening what can be predicted for the large share of human proteins still thinly annotated.

Read →

Nº 07 bioRxiv Benchmarks · Evaluation

scROMA benchmarks pathway activity

scROMA infers pathway activity across batches in single-cell transcriptomics and ships a ground-truth simulation framework with it. Simulated ground truth has been the missing piece in pathway-activity benchmarking, where methods get compared without a known answer.

Read →

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

Agentic Discovery · Nº 75 · 12 Aug 2026

Editor's Note

Generative biology jumped from proteins to entire genomes this week, and the benchmarks are scrambling to catch up.

Nº 01 · The Lede — X — Computational biology

Genome-scale generative design lands

Fig. I X · Filed 12 Aug 2026.

Read the source →

Why it matters

Genome-scale generation lifts the ceiling of generative biology from molecules to organisms, and every design tool pitched on "we generate proteins" now sits one level below the frontier.

Nº 02 — X — Agents · Infrastructure

Fig. II X · Filed 12 Aug 2026.

MDArena scores simulation agents

The Bench Note

From Heureka Labs

Simulation is a shrewd place to start scoring agents, because it punishes confident wrongness that other tasks let slide.

▪

The second question. A score tells you how often an agent got it right across a test set; the other half is whether you can see how, on the one run sitting in front of you.

▪

What the run leaves behind. ARC's analyses write the script and the figures into the project alongside the summary, so the work can be re-read and re-run months later.

▪

Before it starts. For consequential work ARC drafts a plan and waits for approval, and anything in flight shows as a pill with a stop button.

What we’re watching: whether scores like this begin reporting how often an agent caught and fixed its own mistake, rather than only whether the final answer landed

Nº 03 — arXiv — Agents · Infrastructure

Fig. III arXiv · Filed 12 Aug 2026.