This week in AIxBio (April 30, 2026)
Biosecurity risks from chatbots, $500M for cell-modeling data, AI-designed recombinases hit pharma, FDA streams trial data in real time, and a literature agent that actually reads the figures.
Big week at the intersection of AI and biology. The NYT published a biosecurity investigation that will (and should) make the rounds. CZ Biohub committed half a billion dollars to the datasets needed for predictive cell models. Profluent signed a $2.25B-milestone deal with Lilly on AI-designed recombinases. The FDA started streaming trial endpoints in real time. ASI showed what happens when you let a literature agent actually look at the figures instead of just reading text.
Gabriel J.X. Dance in the New York Times: A.I. Bots Told Scientists How to Make Biological Weapons (gift link). Scientists who red-team leading chatbots shared transcripts with the Times showing that ChatGPT, Gemini, and Claude can produce detailed plans for assembling pathogens, dispersing biological payloads, and evading detection, sometimes volunteering strategic details the prompter hadn’t asked for. Meanwhile, the Trump administration has cut biodefense budget requests by nearly 50%, top NSC biosecurity staff have left without replacement, and older model versions with weaker guardrails remain publicly available even after newer ones are patched.
Biohub Launches the Virtual Biology Initiative. CZ Biohub is committing $500M over five years to generate the large-scale, multimodal biological datasets needed to train predictive models of the cell: $400M for internal technology development (cryo-ET, large-scale microscopy, molecular and tissue engineering) and $100M in external grants to seed a coordinated global data-generation effort. Partners include the Allen Institute, Arc Institute, Broad Institute, Wellcome Sanger, the Human Cell Atlas, the Human Protein Atlas, and NVIDIA, with all Biohub-generated data released openly.
Profluent Announces Strategic Partnership with Lilly to Develop AI-Designed Recombinases for Genetic Medicine. Profluent, which uses protein language models to design novel enzymes, is partnering with Eli Lilly on custom site-specific recombinases for kilobase-scale DNA editing, the kind of large-insert precision work that conventional CRISPR systems can’t reliably do. The pitch is that AI-designed recombinases can be programmed to target arbitrary genomic loci rather than relying on whatever nature happened to evolve.
FDA Announces Major Steps to Implement Real-Time Clinical Trials. The FDA unveiled two proof-of-concept trials (AstraZeneca Phase 2 in mantle cell lymphoma, Amgen Phase 1b in small cell lung carcinoma) that stream endpoints and safety signals to the agency in real time rather than through the usual batch-reporting cycle. The agency also issued an RFI for a broader pilot program launching this summer, with the longer-term goal of eliminating the dead time between discrete trial phases and running “continuous” trials.
Marko Brkic: The Figure Problem in Scientific AI: Building a Multimodal Literature Agent for Biology. Applied Scientific Intelligence introduces Alexandria, a a multimodal literature agent that reads, parses, retrieves, and reasons across millions of research papers rather than looking at text alone. The system uses Nemotron Parse for ingestion, a hybrid retrieval pipeline with contextualized embeddings, and a VLM-driven zoom loop that lets the model crop into specific sub-panels when the answer depends on an axis label or plotted value. On FigQA2 (the figure-understanding slice of LABBench2), Alexandria scores 62.5%, a 4.4-point lead over Edison at 58.1% and 45 points over PaperQA2.