General Purpose AI for Public Data Analysis and Experimental Hypothesis Generation

Biology has entered an era where the most valuable discoveries often lie not in new experiments, but in the vast archives of data already collected. For years, scientists have generated enormous libraries of molecular information, recording the activity of genes and proteins in healthy tissues and diseased ones. These records are like detailed maps of the human body's inner workings, yet they were often created to answer specific, narrow questions. The challenge for modern science is to look at these existing maps and ask new questions that the original explorers never thought to pose. This is where artificial intelligence enters the picture, not as a replacement for human researchers, but as a tool capable of scanning these massive datasets to find hidden patterns and suggest new paths for investigation. The goal is to see if a computer can act as a partner to a biologist, sifting through public records to propose testable ideas about how cells behave, how diseases develop, and what might stop them.

A recent study explored exactly this potential by testing how general-purpose artificial intelligence could help scientists generate and investigate new biological questions using public data. The researchers did not simply ask the computer to find a single answer; instead, they set up a series of complementary investigations to see how well the AI could navigate complex biological landscapes. In one line of inquiry, the team used interactive tools to analyze data from individual cells, hoping to find signals that could help human stem cells turn into specific types of tissue. The AI pointed toward a specific signaling pathway involving a molecule called Apelin as a likely candidate for guiding this transformation. However, when the scientists took this suggestion and tested it in the laboratory under the conditions they chose, the expected effect did not appear. This outcome highlighted a crucial distinction: finding a pattern in data is not the same as proving a biological effect exists in the real world.

The researchers then shifted their focus to cancer, asking the AI to search public single-cell data for potential targets on tumor cells that could be attacked by engineered immune cells. They ran the analysis on data from patients with multiple myeloma, a cancer of the blood, and found that three separate runs of the analysis consistently identified the same fifty candidate targets. Among these, two well-known markers, BCMA and CD38, appeared repeatedly, confirming the AI's ability to recover established knowledge. When the same process was applied to pancreatic ductal adenocarcinoma, a type of pancreatic cancer, the results were less uniform. The three runs produced different exploratory findings and failed to nominate a single target that met their strict selection criteria, illustrating that the AI's success can vary depending on the complexity and nature of the disease data.

To test whether the AI could uncover insights without human bias, the team conducted two separate analyses on anonymized lung data without initially telling the computer what tissue it was looking at or what disease the patients had. Even without this context, the AI successfully identified the different types of cells that make up the lung and inferred that the tissue had undergone remodeling associated with disease. When the researchers later provided the AI with the tissue and disease information, it generated a list of new biological questions. Eight of these questions showed statistical support when checked against an independent set of patient data. These included observations about how different cell types coordinated their antioxidant defenses and a specific increase in a protein called CRYM within fibroblast cells, suggesting these are genuine biological relationships worth studying further.

In a final, broader effort, the team integrated multiple types of data to prioritize potential treatments for pancreatic cancer. They combined information about RNA expression, which shows which genes are active, with data on proteins, results from functional screens that test what happens when genes are disabled, and measurements of how these molecules appear in normal, healthy tissues. This multi-layered approach helped them narrow down a long list of possibilities to 213 potential therapeutic targets. By looking at all these measurements together, the scientists could distinguish between candidates that looked similar based on RNA alone and assess whether they were functionally relevant and safe to target in normal tissues. While this work successfully identified a refined list of candidates, the researchers noted that whether these targets will actually work as therapies remains to be established.

Together, these investigations demonstrate how general-purpose artificial intelligence can extend the reach of experimental biologists, allowing them to extract new hypotheses and identify biological relationships from public data that might otherwise go unnoticed. The study also serves as a clear reminder that completing an analysis and finding a reproducible pattern in a dataset are distinct steps from establishing a real-world experimental or therapeutic effect. The AI proved capable of suggesting directions and highlighting connections, but the ultimate validation of those ideas still requires the careful, grounded work of human scientists in the laboratory.