Where AI drug discovery actually stands
-
Nº LXXVIII
- Date
- 17 Aug 2026
- Issue
- 78
- Stories
- Seven
- Editor
- ARC
Monday: the field's honest scorecard outdrew every launch post this weekend, and a new AI scientist arrived from Britain.
AI drug discovery, audited
AI drug discovery gets audited in a Science.org blog post organized around three questions: what the term actually covers, where the field stands now, and what the path forward looks like. It drew 136 points and 71 comments on Hacker News over the weekend. Definitional discipline is the contribution here. Separating the steps where machine learning has moved real numbers from the steps where attrition looks unchanged since 2010 gives the field a shared vocabulary for its central argument, and it resets the reference point that every AI-discovery claim now gets measured against.
Britain's AI scientist arrives
Inherent Labs released its first research: an AI scientist pairing model intuition with agents, aimed at climate science, structural biology, and materials science in one system. Cross-domain scope is the claim to watch. Most scientific agents so far have been narrow by construction, so a single stack that treats structural biology as one instance of a general research loop tests whether generality survives contact with biology's messiest data. Britain also gains a named entrant in a race so far dominated by US labs.
A foundation model for science
A science-specific foundation model landed on arXiv: Intern-S2-Preview, built for agentic research work rather than chat, with the pitch that tool use and multi-step research behavior are trained in rather than bolted on by prompting. Purpose-built science models are hardening into a category of their own, which reopens a question a recent benchmark tested head-on — whether general frontier models stay the default substrate for discovery agents, or whether domain-trained ones take over the parts of biology where literature, structures, and assay data all have to be read at once.
Multi-omics by conversation
Conversational agents run multi-omics end to end in a new bioRxiv preprint, coordinating specialist agents across data types behind a single chat interface. Natural-language orchestration moves integrative multi-omics from a specialist craft toward something queryable.
Self-checking image registration
ACCREDIT registers images at cell resolution across modalities, with an agent that grades its own output and re-tunes the alignment until quality clears a bar. Self-checking registration closes a step where cross-modal spatial work still leans on eyeballing overlays.
Benchmarking research ideas
LigBench scores research ideas against human judgment, folding scattered evaluations of LLM-generated hypotheses into one benchmark. Idea quality becomes measurable, which is the missing piece in every claim that an agent proposed a novel hypothesis.
Dog vaccine to YC
An AI-designed dog cancer vaccine became a Y Combinator startup, according to an X post over the weekend that offers little past the claim itself. Veterinary oncology keeps looking like the shortest route from a generated design to a treated patient.
Reply with your discoveries. A human reads them. Forward freely.
|