scRM: A single-cell foundation model trained on scRNA-seq reads

Inside every living cell, a bustling factory of life is constantly at work, reading instructions and building proteins. To understand how these factories operate, scientists have long relied on a method that acts like a census taker: they count how many copies of each instruction manual are present. This approach, which produces a simple list of numbers for thousands of genes, has been the standard for studying individual cells for years. However, this census method throws away the actual text of the instructions. It tells you that a factory produced four copies of a specific manual, but it discards the pages themselves, losing details about how those pages were edited, where they were cut, or if they contained typos. For decades, researchers have accepted this loss of detail as the price of being able to study thousands of cells at once, believing that the raw text of the genetic code was too messy and complex to analyze directly.

A team of researchers at Altos Labs and the University of Maryland has now challenged this long-held assumption. They have built a new kind of artificial intelligence that skips the counting step entirely. Instead of asking the computer to look at a list of numbers, they fed it the raw genetic text itself—the actual sequences of letters A, C, G, and T that make up the RNA molecules inside a cell. This new model, which they call scRM, learns directly from the unprocessed data, much like a student learning a language by reading the original books rather than a summary. The researchers found that this approach not only works but allows the model to understand the cell's activity with a level of detail that was previously impossible, all while using a fraction of the computing power required by older methods.

The core of this breakthrough lies in how the data is prepared. In a typical experiment, scientists capture the genetic material from thousands of individual cells. The raw output is a massive stream of short text fragments, each tagged with a unique code to identify which cell it came from. Traditionally, computers would first try to match every single fragment to a reference genome, a process that is slow and requires a pre-existing map of the organism's genes. The new framework bypasses this entirely. It takes the raw fragments, groups them by their cell tags, and assembles them into longer, coherent pieces of text called contigs. These contigs represent the actual mRNA molecules found in the cell, preserving the exact sequence and any variations in how the genes were spliced together. This assembled text is then fed directly into the AI model, which treats the genetic code as a language to be understood, rather than a list to be counted.

The model they built, scRM, is surprisingly small. It contains only 1.2 million adjustable settings, known as parameters. To put this in perspective, other leading models in this field are often hundreds of times larger, with hundreds of millions of parameters, and they are trained on tens of millions of cells. In contrast, the researchers trained their model on just 3,000 cells from mouse neurons, a process that took only one hour on a single high-end computer chip. Despite its small size and limited training data, the model performed remarkably well. When tested on tasks like identifying cell types or predicting how cells respond to genetic changes, it matched the performance of the massive, data-hungry models. It even succeeded in correctly identifying cell types in species it had never seen before, such as humans, flies, and rats, simply because it learned the universal language of the genetic code rather than memorizing a specific list of gene names.

The true power of this approach became evident when the researchers looked at how cells react to specific genetic disruptions. In some cases, a cell might change the way it stitches its genetic instructions together, creating a slightly different version of a protein. The old counting methods, which only see the total number of instructions, cannot detect these subtle structural changes. Because scRM reads the actual sequence, it can spot these differences. In experiments where cells were altered to change how their genes were spliced, the new model detected the changes more accurately than any of the existing large-scale models. This suggests that the raw text of the genetic code contains a wealth of information about cellular function that is lost when data is reduced to simple counts.

This work opens a new door for biological research. By proving that models can be trained directly on raw sequencing reads, the researchers have shown that it is possible to build powerful tools without needing a pre-existing map of the organism's genes. This is particularly important for studying viruses, bacteria, or rare organisms that do not have well-mapped genomes. The framework is also efficient enough that it could allow scientists to train custom models on their own specific datasets, rather than relying on general models trained on massive, public collections of data. While the current model is a proof of concept, the researchers believe that scaling this approach could unlock a deeper understanding of how cells work, revealing details about gene activity that have remained hidden behind the veil of simplified counting methods. The study demonstrates that sometimes, to understand the story of life, one must read the original text, not just the summary.