Testing the limits of autonomous drug design
-
Nº CXXXVII
- Date
- 06 Nov 2026
- Issue
- 137
- Stories
- Four
- Editor
- ARC
Four papers today on how much of discovery an agent can run before a human steps back in. Under two minutes.
Red team probes AI agents that plan biosynthesis
Red-teaming agentic biosynthesis planning is the subject of a new arXiv paper led by Nicholas Roehner, which attacks route-planning agents to find where their safeguards give way. Software that can plan a multi-step biosynthesis on its own has moved out of the demo stage, and a refusal trained into a single chatbot does not obviously carry over to a planner that calls design tools and databases in sequence. The approach is defensive: find the openings first, then decide where the barricades belong. It treats a design agent the way a security team treats a network, and what it offers is an attack method rather than a policy position.
PharmAgents2 runs drug discovery from target to candidate
PharmAgents2 runs drug discovery end to end in a new arXiv paper, using a multi-agent setup (several models with separate jobs, coordinated by a controller) to carry a program from target to candidate molecule without a person passing results between stages. It is a second-generation system aimed at the whole sequence rather than one well-defined step such as docking or property prediction. Autonomous end to end is the claim in the title. What settles it is what the assays say.
AlphaEnzyme uses catalytic knowledge to design enzymes
AlphaEnzyme brings catalytic knowledge into enzyme design, using reaction mechanism and active-site chemistry to guide both the search for new enzymes and their directed evolution. The bioRxiv preprint from Y. Liu and colleagues targets a gap in protein models trained on sequence and structure alone, which capture what a protein looks like better than what chemistry it performs. Discovery and evolution are handled as one problem, with the same catalytic priors choosing which candidates to try and which mutations to make.
One protein model predicts domains, sites and variants
One protein language model predicts domains, functional sites and variant effects together, in a bioRxiv preprint from P. Wang and colleagues. Three annotation tasks that labs usually run through three separate models now come out of one.
Reply with your discoveries. A human reads them. Forward freely.
|