On the emergence of in silico biology for intervention in living systems
This paper argues that advancing *in silico* biology for living system intervention requires moving beyond current data-driven limitations by adopting a systems perspective that integrates cohesive multi-scale representations, causal inference, and iterative computational-experimental validation to bridge the gap between correlation and biological mechanism.
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 This is an AI-generated explanation of a preprint that has not been peer-reviewed.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
The human body is a vast, interconnected system where molecules, cells, and tissues constantly communicate to maintain health or drive disease. For decades, scientists have tried to map these connections, gathering enormous amounts of data on genes, proteins, and cell behaviors. This data is like a library of snapshots, showing what is present at a single moment in time. However, knowing what exists is different from knowing how it works. To truly understand biology, researchers need to predict what will happen if they change a specific part of the system, such as turning a gene on or off, or introducing a new drug. This is the difference between observing a river's flow and predicting how the water will react if a dam is built. The field of computational biology has grown rapidly, using powerful computers to find patterns in this data, but a major hurdle remains: computers are currently very good at describing what they see, but they struggle to understand the cause-and-effect rules that govern life.
A new perspective from researchers at the University of Calgary argues that the path forward requires a fundamental shift in how we build these computer models. The authors, James Colter, Nicole Rosin, Jeff Biernaskie, and Nils Daniel Forkert, propose that the current approach is too fragmented. They suggest that simply feeding more data into larger, more complex artificial intelligence systems will not solve the problem if the models do not reflect the actual structure of biology. Instead, they call for a new era of "in silico biology," where computer systems are designed to learn the causal rules of life, not just the statistical patterns. This means building models that can distinguish between a coincidence and a true cause, and that can predict how a living system will respond to an intervention, even in situations the computer has never seen before.
The researchers identify three specific challenges that must be solved to make this vision a reality. The first is the representation problem. Biological data is often incomplete and comes from many different sources, such as genetic sequences, microscope images, and patient records. These pieces of information are like fragments of a puzzle that do not always fit together neatly. Current methods often try to force these fragments into a single picture by assuming they are related, but this can hide the true rules of how cells function. The authors argue that we need models that can understand the context of these measurements, recognizing that a gene might act one way in a healthy tissue and a completely different way in a diseased one. Without this context, the computer cannot tell the difference between a rule that is always true and a pattern that only appears by chance in a specific situation.
The second challenge is the causality problem. Even if a computer can describe a biological state accurately, it must also be able to predict how that state will change when something is altered. This is the core of medical intervention. If a doctor wants to treat a disease, they need to know that changing a specific target will lead to a cure, not just a temporary shift in symptoms. The paper points out that many current models are trained on observational data, which only shows what happens naturally. These models are excellent at guessing what comes next in a familiar sequence, but they often fail when asked to imagine a scenario that has never been observed. For example, a model might know that a certain protein is high in a sick patient, but it cannot reliably predict what will happen if that protein is lowered, because it has not learned the underlying mechanism that links the protein to the disease. The authors emphasize that without testing these predictions against real-world changes, we cannot be sure the computer has learned the rules of life or just memorized the data.
The third hurdle is the intervention problem, which involves validating what the computer thinks it knows. A model is only useful if its predictions can be tested and confirmed in a living system. The researchers note that many current benchmarks for testing these models are flawed because they do not account for the complexity of real biology. They show that some advanced computer models perform well on simple tasks, like grouping similar cells together, but fail when asked to predict the outcome of a specific drug treatment or a genetic change. The paper highlights that agreement with existing data is not enough; the model must be able to explain why a change happens and predict how the system will compensate for it. For instance, if a cell is pushed to change, it often has backup systems that try to keep it stable. A good model must understand these hidden defenses to predict the true outcome of an intervention.
The authors argue that the solution lies not in making the computer models bigger or more complex, but in designing them to respect the architecture of biology. They suggest that we need to combine different types of learning, using both the raw data and established biological rules to guide the computer. This approach would allow the models to fill in the gaps where data is missing and to understand how different parts of the body interact across different scales, from molecules to whole organisms. The paper suggests that the next frontier is not about finding a single perfect algorithm, but about creating a cohesive system that can move seamlessly between observation, prediction, and experimental testing. By iterating between what the computer predicts and what happens in the lab, scientists can refine their understanding of the rules that govern life.
Ultimately, this work serves as a call to action for the scientific community to rethink how they approach biological modeling. The authors believe that the separation between describing data, understanding cause, and testing interventions has held the field back. They propose that by integrating these three elements into a single, cohesive framework, we can build computer systems that do more than just analyze data. These systems could eventually guide the design of new therapies and help us understand how to steer living systems toward health. The path forward requires a careful balance between the power of artificial intelligence and the deep, contextual knowledge of how biology actually works, ensuring that our digital tools are grounded in the reality of the living world.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.