Bigger models don't always win in drug discovery
Good Morning!
A de novo antibody platform, a reproducibility agent, and a scaling law that cuts against the hype — today's issue runs against the grain.
1. De novo antibody design goes generative
Origin-1 designs antibodies against novel epitopes from scratch — no template, no prior binder required, according to a new bioRxiv preprint. The platform chains epitope conditioning, sequence generation, and structural plausibility checks inside a single generative loop, pushing de novo antibody discovery past the narrow class of well-characterized targets where earlier tools stalled. The starting point is now an epitope map, not a known binder. That expands the tractable target space for programs that have been stuck waiting on hybridoma campaigns. Read More →
Why it matters: Generative antibody design clearing the novel-epitope barrier means the 'undruggable by conventional discovery' list just got shorter. The cold-start problem that blocks early-stage biologics programs becomes a software question rather than a biology one.
2. Scaling law flips in molecular prediction
Larger models lose to smaller, task-matched ones on molecular property and activity prediction benchmarks, according to a new bioRxiv assessment of AI-driven drug discovery. The finding directly challenges the assumption — imported from general LLM scaling — that bigger always wins in computational drug discovery. It also sets a concrete reference dataset for vendors whose pitch rests on model size alone. Read More →
3. Agent checks scientific reproducibility at scale
ARA automates reproducibility checks at submission time: the agent runs computational experiments from a manuscript's methods section and flags discrepancies between reported and reproduced results. The preprint positions ARA (Agentic Reproducibility Assessment) as infrastructure for journals and review boards. That moves reproducibility from a post-publication embarrassment into a submission-time gate — a shift the biomedical literature, where replication rates in high-profile studies remain low, has needed for years. Read More →
4. Healthcare agent gaps mapped empirically
Agent skill gaps in healthcare are now catalogued empirically across clinical reasoning, patient communication, and governance — a new study documents where current agents fail and what oversight structures are absent. The findings raise the floor for what a production-grade clinical agent must demonstrate before deployment conversations can begin. Read More →
5. OpenAI clears federal security bar
OpenAI earned FedRAMP Moderate authorization for ChatGPT Enterprise and its API, opening U.S. federal agencies — NIH, FDA, VA among them — to deploy OpenAI models under government security standards without bespoke compliance negotiations. Read More →
6. OpenAI open-sources agent orchestration spec
Symphony, OpenAI's open-source orchestration spec for Codex (OpenAI's coding agent), turns issue trackers into persistent agent queues. The pattern keeps long-running agents live between human interventions — plumbing that biology workflow builders can adapt for multi-step analysis pipelines without rebuilding task-state management from scratch. Read More →
Reply with your discoveries. A human reads them. Forward freely.