AI agents reshape drug and cell biology
Good Morning!
Clinical trust, inverse design, and a $25k biosafety bounty — the agent layer isn't spreading uniformly, it's accelerating at the edges.
1. Fact-checking boosts LLM trust in oncology
Atomic fact-checking — breaking LLM recommendations into individually verifiable claims — measurably raised clinician trust in AI-generated oncology decision support, according to a randomized controlled trial published on arXiv. Clinicians shown sourced, granular claims accepted AI guidance at higher rates than those shown the same recommendations as unbroken text. The finding reframes the clinical-AI adoption debate: hallucination mitigation becomes a design requirement, not a safety disclaimer. 'Our model is accurate' stops being enough when a verifiable-claim interface is the new bar. Read More →
Why it matters: First RCT evidence that explainability architecture — not just model accuracy — drives clinical uptake. Oncology AI vendors without an atomic-fact interface now have a direct trial result arguing against them.
2. Inverse design targets single-cell transcriptomics
MolGene-E inverts the standard molecular-design direction: instead of predicting how a compound affects cells, it starts from a desired scRNA-seq (single-cell RNA sequencing) expression state and works backward to candidate molecules. A bioRxiv preprint positions the model at the intersection of generative chemistry and functional genomics. The practical shift: a target cell state, not a target protein, anchors the drug-design loop. Read More →
3. Agent re-scores docking poses with physics
AgenticPosesRanker wraps a physics-grounded scoring step around standard protein-ligand docking. An AI agent (a system that calls external tools in sequence to complete multi-step tasks) re-ranks poses after generation, correcting the ranking errors that pure ML scoring functions inherit from training-set bias. The arXiv preprint moves docking from a single-pass prediction into an iterative, physically-checked pipeline — raising the floor for structure-based hit identification before anything goes to synthesis. Read More →
4. Multi-agent framework tackles spatial transcriptomics
STAT (Spatial Transcriptomics Analysis Tool) chains multiple AI agents to handle the full spatial-transcriptomics analysis loop interactively, per a bioRxiv preprint. Routing tasks across specialist agents rather than one monolithic model addresses the scale and heterogeneity that single models struggle with. Spatial platforms are scaling throughput faster than analytical methods can keep up — STAT raises the production ceiling at a moment when that gap is widening. Read More →
5. ChatGPT clinical edition goes free
OpenAI made ChatGPT free for verified U.S. physicians, nurse practitioners, and pharmacists, covering clinical documentation, care support, and research queries. Removing the cost barrier that kept AI writing tools out of smaller practices puts institutional procurement decisions under new pressure. What was a budget line is now a table-stakes question. Read More →
6. OpenAI runs biosafety red-team bounty
OpenAI launched a bug bounty paying up to $25,000 for researchers who find universal jailbreaks in GPT-5.5's biosafety guardrails. The program frames biosecurity red-teaming (adversarial testing designed to expose safety failures) as a structured, compensated discipline. Ad-hoc concern is giving way to a paid market for finding where the guardrails break. Read More →
Reply with your discoveries. A human reads them. Forward freely.