Language models are changing how some biotechnology startups choose and analyze experiments even though they have not removed the slowest step: laboratory work. In a Latent Space case study, Endura Therapeutics describes using AI for software, literature review and disease selection while treating physical measurements as the final constraint.
Endura distinguishes “foundries,” which invest in faster experimental systems, from “navigators,” which use general AI tools to improve everyday decisions. Faster coding lets scientists adapt analysis when protocols change, build narrowly tailored dashboards and examine results without waiting for a separate computational team. These operational gains do not require a proprietary scientific model.
For disease selection, Endura says it used research agents to screen about 500 targets, then produced deeper reviews for roughly 100. The prompts asked for failed programs, reasons for failure and conditions under which Endura’s approach might work. Human reviewers still check primary sources and perform full diligence on selected programs.
The account comes from Endura’s chief executive and is not a controlled evaluation. He acknowledges that reports contained errors and that on-demand analysis can weaken reproducibility when code is not preserved in a versioned pipeline. The useful claim is therefore narrower: AI can make preliminary thinking much cheaper, while experiments and expert verification continue to decide what is scientifically valid.