6 min read

NVIDIA puts agents on the bench

NVIDIA puts agents on the bench
Nº 01 · The Lede X Agents · Infrastructure

NVIDIA opens BioNeMo to agents

NVIDIA opens BioNeMo to agents
Fig. IX · Filed 25 Jun 2026.

NVIDIA launched the BioNeMo Agent Toolkit, an open kit that exposes protein structure prediction, molecular docking, and generative chemistry as callable tools any LLM agent can invoke. The play extends NVIDIA's bio stack — already the default for academic structure work — into agent territory, where models like Claude or GPT-5 can chain a docking run into a binder design loop without bespoke glue code. Sets the reference toolkit for bio-native agents and pressures every other agent platform to either wrap BioNeMo or ship a competing biology tool layer.

Read the source

Bench scientist questions the premise
Fig. IIX · Filed 25 Jun 2026.
Nº 02 X Field report

Bench scientist questions the premise

Josiah Zayner pushed back on the BioNeMo launch with a sharper question than the release deserves: no working scientist asked for an agent to run docking and structure prediction. The complaint lands because the field keeps shipping agent demos for tasks that already have working tools, and few that touch the slow, manual bottlenecks bench scientists actually name. Anchors a counterargument the agent-platform pitch has to start answering directly.

Read more
GenoME predicts perturbations per person
Fig. IIIbioRxiv · Filed 25 Jun 2026.
Nº 03 bioRxiv Computational biology

GenoME predicts perturbations per person

GenoME models individualized genomes with a mixture-of-experts architecture (many small specialist models routed by a controller) that predicts multimodal genomic profiles and simulates perturbation outcomes per individual. Moves perturbation prediction from cell-line averages toward person-specific forecasts, narrowing the gap between functional genomics and clinical-grade variant interpretation.

Read more
Also Filed · Four Briefs from the queue
Nº 04 bioRxiv Cell biology · Funding

Single-cell foundation models stress-tested

Zero-shot benchmarking of single-cell transcriptomic foundation models finds robustness drops sharply on batch and tissue shifts not seen in training. Anchors a reference benchmark for the field and forces vendors selling scRNA-seq foundation models to publish out-of-distribution numbers — the same demand BiomniBench made of biomedical agents more broadly — not just held-out test set scores.

Read
Nº 05 arXiv Agents · Infrastructure

Multi-agent system catches medical errors

MedGuards routes clinical notes through specialist agents that detect and correct medical errors, with each agent scoped to a category. Moves clinical-text checking from single-LLM passes to auditable multi-agent review, a structure regulators are likely to ask for as agents start touching patient-facing records.

Read
Nº 06 arXiv Field report

Molexar unifies molecular modalities

Molexar trains a single foundation model across SMILES, graphs, and 3D conformers for drug design tasks. Echoing the BioNeMo move above, the trend is clear: drug-design AI is consolidating into general-purpose backbones that downstream agents can call, not task-specific models.

Read
Nº 07 Anthropic Field report

TCS to push Claude into regulated industries

Anthropic partnered with TCS to deploy Claude across 50,000 TCS employees and build Claude-powered products for healthcare, financial services, and public sector clients. Signals that the integrator channel — not direct sales — is how frontier models reach hospitals and pharma IT.

Read

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

Agentic Discovery  ·  Nº 43  ·  25 Jun 2026

Editor's Note

NVIDIA hands every agent a docking tool; a working scientist asks who actually wanted that.

 

Nº 01 · The Lede  —  X  —  Agents · Infrastructure

NVIDIA opens BioNeMo to agents

NVIDIA opens BioNeMo to agents

Fig. I  X · Filed 25 Jun 2026.

NVIDIA launched the BioNeMo Agent Toolkit, an open kit that exposes protein structure prediction, molecular docking, and generative chemistry as callable tools any LLM agent can invoke. The play extends NVIDIA's bio stack — already the default for academic structure work — into agent territory, where models like Claude or GPT-5 can chain a docking run into a binder design loop without bespoke glue code. Sets the reference toolkit for bio-native agents and pressures every other agent platform to either wrap BioNeMo or ship a competing biology tool layer.

Read the source →

Why it matters

Drug-discovery agent capability stops being a build-it-yourself problem and becomes a configuration choice — collapses the integration tax that has kept most labs from running structure and docking inside autonomous loops, and resets which vendor owns the bio-agent default stack.

 

Nº 02  —  X  —  Field report

Bench scientist questions the premise

Fig. II  X · Filed 25 Jun 2026.

Bench scientist questions the premise

Josiah Zayner pushed back on the BioNeMo launch with a sharper question than the release deserves: no working scientist asked for an agent to run docking and structure prediction. The complaint lands because the field keeps shipping agent demos for tasks that already have working tools, and few that touch the slow, manual bottlenecks bench scientists actually name. Anchors a counterargument the agent-platform pitch has to start answering directly.

Read more →

 

Nº 03  —  bioRxiv  —  Computational biology

GenoME predicts perturbations per person

Fig. III  bioRxiv · Filed 25 Jun 2026.

GenoME predicts perturbations per person

GenoME models individualized genomes with a mixture-of-experts architecture (many small specialist models routed by a controller) that predicts multimodal genomic profiles and simulates perturbation outcomes per individual. Moves perturbation prediction from cell-line averages toward person-specific forecasts, narrowing the gap between functional genomics and clinical-grade variant interpretation.

Read more →

 

Also Filed  ·  Four Briefs from the queue

Nº 04  —  bioRxiv  —  Cell biology · Funding

Single-cell foundation models stress-tested

Zero-shot benchmarking of single-cell transcriptomic foundation models finds robustness drops sharply on batch and tissue shifts not seen in training. Anchors a reference benchmark for the field and forces vendors selling scRNA-seq foundation models to publish out-of-distribution numbers — the same demand BiomniBench made of biomedical agents more broadly — not just held-out test set scores.

Read →

Nº 05  —  arXiv  —  Agents · Infrastructure

Multi-agent system catches medical errors

MedGuards routes clinical notes through specialist agents that detect and correct medical errors, with each agent scoped to a category. Moves clinical-text checking from single-LLM passes to auditable multi-agent review, a structure regulators are likely to ask for as agents start touching patient-facing records.

Read →

Nº 06  —  arXiv  —  Field report

Molexar unifies molecular modalities

Molexar trains a single foundation model across SMILES, graphs, and 3D conformers for drug design tasks. Echoing the BioNeMo move above, the trend is clear: drug-design AI is consolidating into general-purpose backbones that downstream agents can call, not task-specific models.

Read →

Nº 07  —  Anthropic  —  Field report

TCS to push Claude into regulated industries

Anthropic partnered with TCS to deploy Claude across 50,000 TCS employees and build Claude-powered products for healthcare, financial services, and public sector clients. Signals that the integrator channel — not direct sales — is how frontier models reach hospitals and pharma IT.

Read →

 

· · ·

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