AtlasAI

AtlasAI is deployed as a sidebar within the Human Protein Atlas to enable natural language querying and analysis. This multi-agent system is purpose-built to overcome the structural, technical, and reliability barriers that general-purpose frontier models face when interacting with the atlas. To solve the opacity and hallucination risks of monolithic language models, AtlasAI relies on explicit task decomposition, step-by-step reasoning traces, external artifact storage, and continuous code-level validation.

The core processing infrastructure is driven by four specialized agents: (i) the Deep Search Agent turns plain-language questions into precise search queries step-by-step, with every filter verified before the search runs; (ii) the Investigator Agent extracts actual numbers and biological measurements directly from the HPA data, for a single gene or thousands at once; (iii) the Reader Agent answers questions about the atlas itself, quoting HPA's own documentation word for word; and (iv) the Autonomous Scientific Orchestrator conducts entire research studies on its own, building a study plan from a single question, dispatching its own subagents in parallel, combining 13 analytical tools that compute directly on the full data tables, and delivering figures and a complete scientific report in which every number is verified against the exact table cells it cites. Every data point is explicitly tracked through each recorded step before it, so any figure or statistic can be traced back through every tool and agent call to the original atlas records; the whole study unfolds live as an interactive map, and every dataset, figure and complete study can be downloaded.

Gene summary infographics

An AI-generated visual summary based solely on data from Human Protein Atlas version 25 is provided for each Gene Summary page. The input data was drawn from the gene summary and from the Tissue, Single Cell, Subcellular, Structure & Interaction, Cancer and Blood resources.

These data were first summarized with the gemini-3.8-flash model into a structured text with four sections: identity and function, expression, localization, and disease and clinical relevance. For each section, the model highlighted the most important findings and assessed how well the different data sources agreed. This structured summary was then used as the prompt for the gemini-3-pro-image model, which generated the final infographic.

The visual summaries provide a fast overview of all the information the Human Protein Atlas holds for a given gene and protein: what it does, where it is expressed and located, whether it is clinically relevant, and how well established this knowledge is. They constitute a starting point for exploring the gene page and the underlying data.