Cell IO: a database linking single-cell transcriptomes to literature-derived knowledge of cell functions and perturbations

Biology has long relied on a fundamental question: what does a specific type of cell actually do? For decades, scientists could identify a cell by the genes it carries or the proteins on its surface, much like recognizing a person by their name tag. But knowing a name does not explain a person's job, their habits, or how they react when the world around them changes. In the modern era of biology, researchers can now read the genetic instructions of individual cells, creating massive libraries of data that show exactly which genes are active in a single cell at a specific moment. These libraries have grown to include hundreds of millions of cells from humans, mice, and rats. However, the knowledge of what these cells actually accomplish—how they fight disease, how they communicate, or what happens when a gene is broken—remains scattered across thousands of scientific papers, buried in text that computers cannot easily read.

A researcher at the RIKEN Center for Biosystems Dynamics Research in Japan has built a bridge between these two worlds. They created a new public database called Cell IO, which links the raw genetic data of individual cells directly to the written knowledge found in scientific literature. By using a common language to describe cell types, the system allows a scientist to start with a list of genes and instantly see what the world's scientific community has written about that cell's behavior. Conversely, a researcher reading about a specific cell's function can immediately find the genetic data that supports those claims. This tool does not just store information; it organizes it so that the "what" of a cell's genetic makeup is always paired with the "how" and "why" of its function, turning a mountain of disconnected facts into a navigable map of cellular life.

The project began by tackling the sheer volume of genetic data available in public archives. The researcher gathered raw genetic sequences from over 65,000 experiments involving human, mouse, and rat cells. These experiments were not uniform; they came from different labs, used different machines, and were processed in different ways. To make them comparable, they ran every single experiment through the same computer pipeline. This process involved cleaning the data, removing errors, and grouping similar cells together. They then assigned a standard name to each group of cells based on a universal dictionary of cell types, ensuring that a "T cell" in a mouse study was recognized as the same type of cell in a human study. The result is a massive, organized collection containing nearly 83 million cells from nearly 15,000 experiments, all processed in the same way so they can be compared directly.

While the genetic data provided the map of the cells, the researcher knew they also needed the story of what those cells do. To find this, they turned to the scientific literature. They collected hundreds of thousands of open-access research articles and review papers, focusing on the immune system and the nervous system. Using advanced computer models trained to read and understand language, they extracted specific statements about cell behavior. The system learned to identify sentences where a cell acts as an agent, such as a cell that kills a virus or a cell that releases a chemical signal. It also looked for findings related to gene perturbations, which are experiments where scientists deliberately break or change a gene to see how the cell reacts.

The computer did not just read the text; it verified the claims. It checked the sentences to ensure the cell was truly the one performing the action and not just a bystander. It then translated the specific names used in the papers into the same standard dictionary used for the genetic data. This created a second layer of knowledge containing nearly 900,000 distinct claims about cell functions. Some of these claims come from original research papers describing new discoveries, while others come from review articles that summarize what is already known. The system also noted the strength of the evidence, distinguishing between statements that are firmly asserted by the author and those that are merely suggested or speculative.

The true power of Cell IO lies in how these two layers connect. Because both the genetic data and the literature claims use the same standard names for cell types, the database can link them together seamlessly. A user can search for a specific cell type, such as a regulatory T cell, and see a single page that shows both the genetic data and the literature. On one side, the page displays how many experiments have found this cell type and in which tissues it is most common. On the other side, it lists the thousands of sentences from scientific papers that describe what this cell does. The interface allows the user to filter these claims by the type of evidence, the organism, or the direction of the effect, such as whether a gene change increased or decreased a cell's activity.

This connection reveals gaps in our current knowledge. The database shows that while some cell types are well-studied in both genetic data and literature, others appear frequently in genetic experiments but have very little written about their specific functions. For example, a certain type of pineal cell was found in over a million cells across thousands of experiments, yet the database contains only four written claims about what it does. Conversely, some cell types have many written descriptions but lack corresponding genetic data in the current collection. By highlighting these mismatches, the tool helps scientists see where their understanding is incomplete and where future research should be directed.

The database is freely available to anyone with an internet connection, requiring no special registration or software installation. It is designed to be intuitive, allowing a user to start from a cell name, a tissue type, or even a specific gene and navigate to the relevant data and text. The researcher emphasizes that the system is a tool for exploration rather than a final authority. The claims are machine-extracted and linked directly to their original sources, allowing users to verify the information for themselves. The researcher acknowledges that the current collection is limited to specific types of experiments and organisms, and that the automated extraction of text, while highly accurate, is not perfect. However, by providing a unified view of what cells are and what they do, Cell IO offers a new way to navigate the complex landscape of modern biology, turning scattered fragments of knowledge into a coherent whole.