7 min read

AI-designed drug finally has human data

AI-designed drug finally has human data
Nº 01 · The Lede X Drug discovery · Computational

Insilico's AI-designed drug moves proteomic aging clocks

Insilico's AI-designed drug moves proteomic aging clocks
Fig. IX · Filed 11 Sep 2026.

Insilico Medicine reported that rentosertib, an AI-designed molecule that has already reached patients, showed shifts toward younger predicted biological age across six proteomic aging clocks. Caveats travel with it: a company-reported readout, and proteomic clocks are still contested as surrogate measures of anything. But aging clocks have mostly lived in observational cohorts and supplement studies, and seeing six of them read out on a molecule that started as an AI design puts two literatures in the same room. It also moves the AI-discovery argument off in silico scores and onto human proteomics — the evidence base the field has been thinnest on.

Read the source

DeepMind ships a genome atlas 30x AlphaFold's size
Fig. IIX · Filed 11 Sep 2026.
Nº 02 X Computational biology

DeepMind ships a genome atlas 30x AlphaFold's size

Google DeepMind released AlphaGenome Atlas, a one-petabyte resource the company says is over 30 times the size of the AlphaFold Database, wrapped in a browsing layer that ties the interconnected datasets together instead of leaving them as separate downloads. The AlphaFold database changed which structural questions a researcher could ask without running anything. A regulatory-genome equivalent at that scale puts variant-effect work on the same footing, queryable as a reference resource rather than a compute job.

Read more
De la Fuente lab uses Codex to hunt new antibiotics
Fig. IIIOpenAI · Filed 11 Sep 2026.
Nº 03 OpenAI Computational biology

De la Fuente lab uses Codex to hunt new antibiotics

Codex is scanning extinct genomes for antimicrobial candidates in César de la Fuente's lab, per an OpenAI case study: Codex, OpenAI's coding agent, paired with ChatGPT, searching living and extinct genomes for molecules that might work against drug-resistant infections. What's absent is the point. No purpose-built biology model sits in the loop, so the discovery work runs on general tools already available, shifting the question from whether a foundation model exists for a task to how much scaffolding a general agent needs around it.

Read more
Also Filed · Five Briefs from the queue
Nº 04 arXiv Field report

New method picks which CRISPR hits to test next

Adaptive hit discovery turns CRISPR screens into a loop: a model proposes the next batch of perturbations, the assay answers, and the selection policy carries across screens instead of being refit per experiment. Advances screen automation, where follow-up rounds are still mostly hand-picked.

Read
Nº 05 OpenAI Agents · Infrastructure

OpenAI opens a managed service for cloud agents

OpenAI launched the Agents API , a managed service built on the same machinery that runs Codex, covering orchestration, long-running sessions and tool use. Long-running is the operative phrase for biology, where an analysis routinely outlives a chat session and hosting has been the builder's problem, not the vendor's.

Read
Nº 06 bioRxiv Field report

EukaUTR models and designs eukaryotic 3' UTRs

EukaUTR predicts and designs eukaryotic 3' UTRs as a general model rather than one fit to a single organism or reporter. Extends sequence design into the untranslated regions that set mRNA stability and translation rate, the tunable part of any mRNA therapeutic.

Read
Nº 07 arXiv Agents · Infrastructure

Benchmark tests agents on long multimodal research tasks

Mr.LHDR scores agents on long-horizon research tasks spanning figures, tables and text; the name unpacks to Multimodal Real-World Long-Horizon Deep Research. That combination is where literature-scale biology work actually lives, and it has been the thinnest part of the benchmark shelf.

Read
Nº 08 bioRxiv Computational biology

Genomic language model predicts fungal essential genes

A genomic language model transfers essentiality prediction across fungal species without per-species retraining. Points antifungal target selection at genomes nobody has screened, which matters as resistant Candida and Aspergillus keep spreading.

Read

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

Agentic Discovery  ·  Nº 97  ·  11 Sep 2026

Editor's Note

Aging clocks, a petabyte of genome, and antimicrobials hunted with a coding agent — busy 24 hours.

 

Nº 01 · The Lede  —  X  —  Drug discovery · Computational

Insilico's AI-designed drug moves proteomic aging clocks

Insilico's AI-designed drug moves proteomic aging clocks

Fig. I  X · Filed 11 Sep 2026.

Insilico Medicine reported that rentosertib, an AI-designed molecule that has already reached patients, showed shifts toward younger predicted biological age across six proteomic aging clocks. Caveats travel with it: a company-reported readout, and proteomic clocks are still contested as surrogate measures of anything. But aging clocks have mostly lived in observational cohorts and supplement studies, and seeing six of them read out on a molecule that started as an AI design puts two literatures in the same room. It also moves the AI-discovery argument off in silico scores and onto human proteomics — the evidence base the field has been thinnest on.

Read the source →

Why it matters

Human proteomic data attached to an AI-originated molecule raises the evidence bar for the whole category: "our model found this" now invites the question of what happened in people, not what scored well on a screen.

 

Nº 02  —  X  —  Computational biology

DeepMind ships a genome atlas 30x AlphaFold's size

Fig. II  X · Filed 11 Sep 2026.

DeepMind ships a genome atlas 30x AlphaFold's size

Google DeepMind released AlphaGenome Atlas, a one-petabyte resource the company says is over 30 times the size of the AlphaFold Database, wrapped in a browsing layer that ties the interconnected datasets together instead of leaving them as separate downloads. The AlphaFold database changed which structural questions a researcher could ask without running anything. A regulatory-genome equivalent at that scale puts variant-effect work on the same footing, queryable as a reference resource rather than a compute job.

Read more →

 

Nº 03  —  OpenAI  —  Computational biology

De la Fuente lab uses Codex to hunt new antibiotics

Fig. III  OpenAI · Filed 11 Sep 2026.

De la Fuente lab uses Codex to hunt new antibiotics

Codex is scanning extinct genomes for antimicrobial candidates in César de la Fuente's lab, per an OpenAI case study: Codex, OpenAI's coding agent, paired with ChatGPT, searching living and extinct genomes for molecules that might work against drug-resistant infections. What's absent is the point. No purpose-built biology model sits in the loop, so the discovery work runs on general tools already available, shifting the question from whether a foundation model exists for a task to how much scaffolding a general agent needs around it.

Read more →

The Bench NoteFrom Heureka Labs

Broadly capable tools are now enough to run a real search, which puts the interesting engineering in everything around the model.

What the scaffolding is. In Bench it is unglamorous: a project folder where outputs land, a rationale the agent reads before it starts, and a task list it can see.
Checked against a source. ARC resolves a compound from a name, CAS number or structure and returns what is known about it, so the record carries a lookup.
The trail, as it forms. Provenance is captured while the work runs — make a plot from a table and the link is recorded.

What we’re watching: whether procedure files and tool setups start showing up in methods sections next to the result

 

Also Filed  ·  Five Briefs from the queue

Nº 04  —  arXiv  —  Field report

New method picks which CRISPR hits to test next

Adaptive hit discovery turns CRISPR screens into a loop: a model proposes the next batch of perturbations, the assay answers, and the selection policy carries across screens instead of being refit per experiment. Advances screen automation, where follow-up rounds are still mostly hand-picked.

Read →

Nº 05  —  OpenAI  —  Agents · Infrastructure

OpenAI opens a managed service for cloud agents

OpenAI launched the Agents API , a managed service built on the same machinery that runs Codex, covering orchestration, long-running sessions and tool use. Long-running is the operative phrase for biology, where an analysis routinely outlives a chat session and hosting has been the builder's problem, not the vendor's.

Read →

Nº 06  —  bioRxiv  —  Field report

EukaUTR models and designs eukaryotic 3' UTRs

EukaUTR predicts and designs eukaryotic 3' UTRs as a general model rather than one fit to a single organism or reporter. Extends sequence design into the untranslated regions that set mRNA stability and translation rate, the tunable part of any mRNA therapeutic.

Read →

Nº 07  —  arXiv  —  Agents · Infrastructure

Benchmark tests agents on long multimodal research tasks

Mr.LHDR scores agents on long-horizon research tasks spanning figures, tables and text; the name unpacks to Multimodal Real-World Long-Horizon Deep Research. That combination is where literature-scale biology work actually lives, and it has been the thinnest part of the benchmark shelf.

Read →

Nº 08  —  bioRxiv  —  Computational biology

Genomic language model predicts fungal essential genes

A genomic language model transfers essentiality prediction across fungal species without per-species retraining. Points antifungal target selection at genomes nobody has screened, which matters as resistant Candida and Aspergillus keep spreading.

Read →

 

· · ·

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