What Is AutoDiscovery?

AutoDiscovery transforms how scientists interact with data by inverting the traditional research workflow. Rather than starting with a hypothesis and testing it, researchers upload a structured dataset and let the system autonomously generate hypotheses in natural language, propose experiments, execute Python code, interpret statistical results, and use findings to generate new leads. The system has already produced publishable results across disciplines—from uncovering trophic relationships in marine ecosystem data to identifying mutual-exclusivity patterns in cancer mutations that could inform treatment decisions.

How It Works

The system combines two key technical innovations:

• Bayesian Surprise: AutoDiscovery quantifies how much experimental evidence shifts its internal beliefs (represented as probability distributions). Rather than simply confirming expected findings, it prioritizes results that meaningfully contradict its prior expectations—reflecting the scientific intuition that surprising results often represent genuine discoveries.
• Monte Carlo Tree Search (MCTS): The system uses intelligent search to navigate the vast space of possible scientific questions, balancing exploration of new hypotheses with deeper investigation of promising leads.

Together, these approaches guide AutoDiscovery toward unexpected patterns and non-obvious research directions that human researchers might overlook.

Real-World Impact

Early partnerships demonstrate concrete value across domains. Researchers at the University of Utah used AutoDiscovery to accelerate exploratory analysis at unprecedented speed. At Fred Hutchinson Cancer Center, analyses that would normally require weeks or months were completed in a single day. A case study with oncologists at Swedish Cancer Institute's Paul G. Allen Research Center uncovered a mutual-exclusivity pattern between PIK3CA and TP53 mutations in breast cancer—a potential finding that could inform treatment strategies.

Several AutoDiscovery discoveries in social science were independently verified and published in a peer-reviewed paper (November 2025), demonstrating that the system's findings meet scientific publication standards.

Availability and Next Steps

AutoDiscovery is now available as an experimental feature in AstaLabs within the Asta platform. Researchers can upload datasets and monitor discovery progress through an intuitive interface displaying hypotheses, surprisal scores, and statistical details. The system supports both quick analyses and extended overnight runs generating hundreds of experiments. Full case studies and technical documentation are available on Allen AI's website.