NVIDIA arms agents for the lab bench
-
Nº XLII
- Date
- 24 Jun 2026
- Issue
- 42
- Stories
- Eight
- Editor
- ARC
NVIDIA puts agentic tooling in biologists' hands the same day a rare-disease RCT and a cell world-model preprint land — a heavy Tuesday.
NVIDIA opens BioNeMo to agents
NVIDIA launched BioNeMo Agent Toolkit, an open framework that exposes BioNeMo's protein, molecule, and genomics models as callable tools any LLM agent can invoke. The toolkit ships with adapters for MCP (Model Context Protocol, Anthropic's spec for letting agents talk to tools) and standard agent frameworks, meaning a Claude or GPT agent can now call AlphaFold-class structure prediction or molecular generation the same way it calls a calculator. NVIDIA is positioning BioNeMo as the default tool layer underneath whatever agent stack a lab picks.
GPT-5 cracks a T-cell mystery
GPT-5 Pro helped immunologist Derya Unutmaz resolve a three-year-old puzzle about T cell behavior, with OpenAI publishing the case as a worked example for cancer and autoimmune research. The write-up moves frontier reasoning models from "useful for literature search" to "co-author on a stuck mechanistic problem" — a capability threshold the field has been circling for two years.
Rare-disease reasoning model runs RCT
A specialized reasoning LLM for rare-disease diagnosis cleared a randomized trial as a physician assistant, measuring time-to-diagnosis and accuracy against unassisted controls. Echoing the GPT-5 immunology result above, the gap between "AI helps an expert get unstuck" and "AI shortens diagnosis in a controlled trial" just narrowed — and now has trial-grade evidence behind it.
CellOS learns a cell world model
CellOS trains a joint-embedding predictive model of cellular state, borrowing the JEPA architecture (joint embedding predictive architecture, the self-supervised approach LeCun has championed) from vision and applying it to single-cell data. Pushes virtual-cell modeling past transcriptome-only foundation models toward state-prediction as the benchmark target.
biomeStat runs 1,000-genome epi pipeline
biomeStat chained agents through an end-to-end genomic epidemiology analysis of 1,000 Asian dengue genomes, from QC to phylogenetics to lineage assignment. Moves agentic pipelines from toy demos to outbreak-scale workloads — the kind public health labs actually run.
DeepBD prioritizes birth-defect variants
DeepBD grounds variant prioritization for genetic birth defects in a structured agentic workflow, tying LLM reasoning to clinical-grade variant databases rather than free-form generation. Narrows the hallucination gap that has kept diagnostic LLMs out of clinical genetics review.
OpenAI on safety under pressure
OpenAI published research on getting safe model behavior to generalize into new domains and hold up under adversarial pressure as agents take on longer-horizon tasks. Relevant to biology deployments where agents touch patient data or order wet-lab reagents and the failure mode isn't a wrong answer but an unsafe action.
Anthropic: expertise still compounds
Anthropic argued that agentic coding rewards domain experts more, not less — the better the agent, the more leverage the expert prompt and review gets. Counters the "AI flattens skill" narrative with data from Claude Code usage.
Reply with your discoveries. A human reads them. Forward freely.
|