Geodesic Intelligence launches AI drug discovery platform with agentic design engine and structure prediction

Geodesic Intelligence has unveiled a full-stack AI drug discovery platform built around an agentic design engine called NovaDDE and a preview version of its atomic structure prediction model, NovaAtom-Lite-Preview. The company frames the release as a step toward what it calls "AGI for drug discovery" — a system that finds the shortest path from biology to drugs by combining AI, structural biology and biophysics.

NovaDDE operates as an agentic workspace for discovery teams. It plans, runs and evaluates computational experiments from target definition through to prioritized candidates, documenting every decision along the way. The system offers two distinct workspaces. A "Co-Scientist Workspace" allows free-form chat to describe protein or drug development decisions; the environment plans the run, retrieves structures and reasons with the user. A "Protein Binder Design Studio" skips the chat and lets users define a target surface, constraints and generate binders directly.

NovaAtom-Lite-Preview is a first look at the company's most capable structure prediction model, a lighter and faster build distributed as a preview while the full model remains in training. The model takes a sequence and returns a single-atom-resolution structure for proteins, antibodies and complexes. Geodesic emphasizes that NovaDDE calls models such as NovaAtom as tools inside the workflow, not as a separate black box.

The announcement positions the platform as progress toward "AGI for drug discovery," a term the company uses to describe a system that combines reasoning, planning and autonomous evaluation within a narrow, well-defined domain. In practice, the current product is a stack of task-focused agents grounded in structural biology and biophysics, not general intelligence. The Lite version of NovaAtom is explicitly presented as an interim release; the full model has not finished training, and no benchmarks have been published against standard baselines such as CASP or CAMEO.

Several gaps remain. The company has not published a detailed model card with accuracy metrics, runtimes or known limitations for NovaAtom-Lite, nor has it disclosed architecture, training data or a usage license. For NovaDDE, no details were provided on wet-lab integration, compatibility with standard formats such as PDB and mmCIF, or a pricing model. Until such data appears, practical evaluation of the platform remains limited to the claimed capabilities.