Agents grade themselves on epigenomics

Agentic Discovery  ·  Nº 35  ·  12 Jun 2026

Editor's Note

Three benchmarks and a foundation model walk into a Thursday — and the agents finally have report cards they can't game.

Nº 01 · The Lede  —  arXiv  —  Agents · Infrastructure

EpiBench scores agents on epigenomics

Fig. I  arXiv · Filed 12 Jun 2026.

EpiBench grades AI agents on real epigenomics analysis tasks — ChIP-seq peak calling, ATAC-seq processing, motif discovery — with verifiable ground truth instead of LLM-judge proxies. The benchmark spans tasks where most agents currently choke on file-format wrangling and tool chaining before they get near the biology. Anchors a new reference floor for epigenomics-applicable agents: vendors claiming "works on genomics" now have a specific scoreboard.

Read the source →

Nº 02  —  arXiv  —  Agents · Infrastructure

Fig. II  arXiv · Filed 12 Jun 2026.

EurekAgent bets on environments

EurekAgent argues the bottleneck in autonomous scientific discovery isn't the agent or the model — it's the environment the agent runs in. The paper reframes scientific discovery as an environment-engineering problem: give an agent the right tools, data interfaces, and feedback loops, and capability follows. Shifts the debate from "which model is smartest" to "which scaffolding actually lets a model do science" — a reference point for how labs evaluate agent platforms.

Read more →

Nº 03  —  bioRxiv  —  Cell biology · Funding

Fig. III  bioRxiv · Filed 12 Jun 2026.

HoloCell models the whole cell

HoloCell proposes a generative foundation model for holistic cellular modeling — not single-modality (transcriptome OR proteome) but joint across cellular state. The bioRxiv preprint targets the gap most virtual-cell efforts leave open: cross-modal generation under perturbation. Pushes the virtual-cell frontier from single-omics prediction toward generative whole-cell simulation, the capability CZI's Biohub bets are chasing.

Read more →

Also Filed  ·  Three Briefs from the queue

Nº 04  —  bioRxiv  —  Drug discovery · Computational

Docking scales to 100B molecules

Combinatorial docking sweeps over 100 billion molecules by fusing generative chemistry with docking, a scale that brute-force virtual screens can't touch. Resets the cost ceiling for prospective ligand discovery — make-on-demand libraries this large used to be unsearchable end-to-end.

Read →

Nº 05  —  X  —  Agents · Infrastructure

Local genome agents get a format

Genomi added .genome support, letting local AI agents ingest a single file and return evidence-grounded answers against up-to-date genetics references — no cloud round-trip. Narrows the gap between privacy-constrained genomics work and agent-grade analysis, where data residency has blocked most hosted tools.

Read →

Nº 06  —  X  —  Computational biology

Compound VC maps bioML frontier

Compound VC convened a bioML showcase with Basecamp Research's Phil Lorenz and others working on the datasets, tasks, and context that the next round of bio foundation models will need. Signals where one bio-AI funder thinks the moat lives: data and task design, not model architecture — adjacent to the environment-first argument in story 2 above.

Read →

· · ·

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