RegimeFormer: A Large Protein Model of Global Perturbation Regimes
RegimeFormer is a large-scale protein model that, through its associated RegimeAtlas database of over 200 million sequences, establishes global perturbation regimes to predict residue-level mutation effects and improve downstream biological modeling, particularly for unseen proteins and families.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
Proteins are the workhorses of life, tiny molecular machines that build cells, fight infections, and carry out the chemical reactions that keep us alive. They are made of long chains of building blocks called amino acids, and the specific order of these blocks determines how the protein folds and what it does. For decades, scientists have known that changing even a single building block in this chain can have dramatic consequences. Sometimes the change is harmless; other times it breaks the machine entirely, leading to disease. In recent years, researchers have developed powerful computer tools to predict these outcomes, but these tools have mostly been limited to studying one protein at a time. They struggle when faced with the vast, unexplored ocean of proteins that exist in nature, many of which have never been seen in a laboratory.
A team of researchers has now built a new kind of map that covers this entire ocean. They created a system that treats the behavior of proteins not as a collection of isolated puzzles, but as a unified landscape with distinct regions. By analyzing more than 200 million protein sequences from every branch of the tree of life, they discovered that proteins fall into different "regimes," or states, that describe how they react to change. Some proteins are fragile, meaning even a small tweak can break them. Others are adaptable, able to absorb changes and even improve their function. The researchers built a massive model that can navigate this landscape, predicting how a specific change will affect a protein, even if that protein has never been studied before. This work provides a new way to understand the rules of life at a molecular level, offering a guide for scientists who want to engineer better proteins or understand how genetic mutations cause illness.
The researchers started by gathering a staggering amount of data. They collected 424 million raw records of protein sequences from public databases, which include everything from human genes to bacteria found in soil and viruses. After cleaning this data to remove errors and duplicates, they were left with 202 million distinct proteins. This collection, which they named RegimeAtlas, represents a nearly complete map of the protein universe known to science. Instead of just listing these proteins, the team used a sophisticated computer model to organize them. They projected these sequences into a continuous space where proteins with similar behaviors are grouped together. This created a global coordinate system for how proteins respond to mutations, revealing that the entire set of proteins is not random but follows a structured pattern.
To make this map useful for specific questions, the team focused on a smaller, highly diverse group of one million proteins to train their main model, which they called RegimeFormer. This model learned to look at a protein and assign it a position on the map based on its "regime." Once a protein's regime is known, the model can predict what will happen if you swap one amino acid for another. It does this by calculating three key things for every single spot in the protein chain: fragility, adaptability, and uncertainty. Fragility measures how likely a spot is to break if changed. Adaptability measures how likely a change is to be tolerated or even helpful. Uncertainty tells the researcher how confident the model is in its prediction.
The team tested this system against real-world experiments where scientists had already measured the effects of thousands of mutations. They found that the model's predictions matched the experimental results with high accuracy. More importantly, the model excelled in situations where other tools usually fail: when looking at proteins that are very different from anything it had seen before, or when dealing with families of proteins that are rare in the training data. This suggests that the model has learned a fundamental rule about how proteins work, rather than just memorizing specific examples. The researchers also checked if their predictions made sense biologically. They found that the spots the model marked as "fragile" were exactly the spots where proteins usually perform their most critical jobs, such as binding to other molecules or catalyzing chemical reactions. Conversely, the "adaptable" spots were often found in areas of the protein that can change shape without breaking.
To prove that this map was not just a mathematical trick, the researchers compared their results with independent data from other fields. They looked at the 3D structures of proteins predicted by other AI systems and found that proteins the model labeled as fragile tended to have less stable structures, while those labeled as adaptable were more stable. They also examined evolutionary history, looking at how proteins have changed over millions of years. They discovered that the spots the model identified as fragile were the same spots that had remained unchanged for eons, indicating that nature has been selecting against changes in those areas for a very long time. This alignment with structural and evolutionary evidence confirmed that the model had uncovered a real biological truth.
The power of this system lies in its ability to scale. While the model was trained on a million proteins, it was applied to nearly 400 million individual spots across the entire atlas. This created a massive library of predictions that scientists can query. Instead of waiting for expensive and time-consuming lab experiments to test every possible mutation, researchers can now use this map to prioritize which changes are worth testing. The team showed that using this map to select mutations for testing was significantly more efficient than using older methods, allowing scientists to find beneficial changes with far fewer experiments.
Beyond just predicting protein behavior, the researchers connected this protein map to how cells react. They linked their protein predictions to data on how cells change their gene activity when exposed to drugs or other stressors. They found that the protein-level insights helped explain why certain cells responded to drugs in specific ways. This bridge from the molecular level to the cellular level suggests that understanding the "regime" of a protein can help predict how a whole cell or organism will behave. The team also built a public tool, called RegimeAtlas Explorer, which allows anyone to search this vast database, visualize the predictions, and see the evidence behind them.
This work represents a shift in how scientists view the protein world. Instead of seeing proteins as isolated entities that must be studied one by one, the researchers have shown that they exist within a global system of behaviors. By mapping this system, they have created a tool that can guide the future of biology, from designing new medicines to understanding the genetic basis of disease. The model does not just predict what will happen; it reveals the underlying logic of why proteins behave the way they do, offering a new lens through which to view the machinery of life.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.