Self-evolving agents top drug discovery
-
Nº XLV
- Date
- 29 Jun 2026
- Issue
- 45
- Stories
- Six
- Editor
- ARC
Monday opens with agents that learn from their own past runs, and a sobering reminder that benchmarks aren't real science.
DrugSAGE tops drug-discovery agents
DrugSAGE ranks first among nine state-of-the-art drug-discovery agents by accumulating and reusing experience across tasks, according to a new release from Wengong Jin's group. The self-evolving LLM agent (one that updates its own playbook from prior runs rather than relying on a frozen prompt) transfers learned strategies between assays, hit-finding, and property prediction. Cross-task transfer has been the missing piece in agent-based discovery — most systems forget everything between projects.
Real science breaks AI frameworks
Advanced AI frameworks stumble when tested against published studies in uncertainty quantification, Therapeutic Data Commons ML, and agent-based modeling — not curated benchmarks. The bioRxiv evaluation finds the gap between leaderboard scores and reproducing actual papers remains wide. Anchors a counterargument to benchmark-driven capability claims and forces vendors to show published-study replication, not just TDC numbers.
Agents reviewing agents
An AI agent spends three weeks analyzing protein data and flags a drug target; a second agent decides whether the finding is strong enough to act on. The X thread sparked 30 replies debating where human judgment enters the loop. Moves the agent-oversight debate from theoretical to operational — multi-agent review is becoming the default architecture before anyone has agreed on accountability.
scBench-Long stresses single-cell agents
scBench-Long benchmarks long-horizon single-cell biology tasks with verifiable answers, exposing where agents drift over multi-step scRNA-seq workflows. Establishes a reference benchmark for sustained reasoning in cell biology, where most existing evals stop at single-turn questions.
Variant calling goes client-server
Client-server interfaces enable agent-driven variant calling, per a new bioRxiv preprint pushing heavy compute off the agent and onto dedicated callers. Narrows the gap between conversational agents and production genomics infrastructure — variant calling stops being a wall the agent hits.
Synthetic longitudinal clinical notes
A new pipeline generates longitudinal synthetic clinical notes with LLMs, giving researchers patient trajectories without de-identification headaches. Lowers the data-access ceiling for clinical NLP work that has been gated on real-EHR agreements.
Reply with your discoveries. A human reads them. Forward freely.
|