The period in which artificial intelligence begins to simulate biology, is now upon us.Artificial Intelligence becomes Biology Simulation is here.
The AI revolution is reshaping biology and drug discovery by introducing a new layer of innovation: virtual cells.In the era of AI, virtual cells are emerging as a new front in transforming the fields of biology and drug discovery.
Explore the potential of AI virtual cells to mimic biological cells, forecast cell behavior, speed up drug discovery, and revolutionize healthcare.
The keywords include virtual cells, AI in biology, virtual cell models, AI drug discovery, computational biology, biological foundation models and digital cell simulation.
What if we could simulated a living cell?
The study of biology has been based on laboratory experiments, microscopes, genetic analysis and mathematical modelling for decades to help understand how living organisms work. Scientists can study cells, alter genes, determine the activity of proteins, and study disease processes. To know exactly how a living cell will react to any specific change, however, is very difficult.
A cell is not just a stampede of independent biological parts. It is a dynamic system that comprises thousands of interacting molecules, proteins, genes, metabolic pathways, and regulating mechanisms. These parts continuously communicate and react to changes in conditions.
Today, however, the possibility of creating computational models to predict the behavior of cells in different biological environments has been opened with the aid of artificial intelligence.
The systems can be referred to as virtual cells.
The ultimate vision is to develop artificial intelligence systems that can simulate the impact of altering a gene, adding a drug, modifying a protein or treating a cell with a specific environment in one scenario before conducting the experiment in the lab.
The success of this technology would revolutionize the fields of drug discovery, genetic research, biotechnology and personalized medicine.
But there is a caveat: a virtual cell doesn’t have to be a full-blown digital version of a real cell. The systems today are more correctly seen as special purpose predictive models that have different capacity to represent biological processes.
1. So what is a Virtual Cell?
The computational model of a biological cell used for predictions or simulations of its behaviour is a virtual cell.
Biological models are generally limited to a single process. A mathematical model can be used to explain, for instance, the way a certain protein concentration changes with time or the way a metabolic pathway reacts to a certain chemical reaction.
The purpose of the virtual cells using artificial intelligence is to identify patterns that can be learned from biological data.
These datasets can include:
- Gene expression measurements.
- Single-cell RNA sequencing.
- Abundances and interactions of proteins.
- Chromatin accessibility and epigenetic information.
- Microscopy images.
- Genetic perturbation experiments.
- Drug or environmental effects on cells.
Understanding these relationships between them allows an AI model to predict what could happen when a cell’s state changes.
Let’s take a simple case.
A researcher is interested in finding out what occurs when a specific gene is turned off. The experiment may involve genetic engineering, cell culture, sequencing, and in-depth analysis.
A virtual-cell model could then first be used to predict the genes, pathways or cellular properties that are likely to be altered following that perturbation.
The prediction could then be used by researchers to determine the most informative laboratory experiments to conduct.
The model would NOT be a substitute for the experiment. Rather it would assist scientists to determine what experiments to conduct.
This is very basic to the technology.
2. AI’s Understanding of the Language of Biology.
AI systems of today leverage statistical relationships from massive datasets. The data is mainly text in language models. In biological foundation models, the data can be gene sequences, molecular measurements, protein structure, or cellular observations.
An important method is the use of single-cell RNA sequencing (scRNA-seq).
RNA sequencing gives information on which genes are transcribed (active) and which are not (not active) in a cell. These measurements can be used to determine cell identities, developmental stages, and the response of cells to disease or treatment, as different cells activate different sets of genes.
Consider the cells to be represented by a set of high dimensional vectors:
[
\mathbf{x}=(x_1,x_2,\ldots,x_n)
]
Measured expression level of each component corresponds to the expression level of a specific gene.
This vector is a complex molecular snapshot of the cell for thousands of genes.
Depending on the data and the AI model, the patterns can be learned from millions of such measurements. It might uncover statistical connections among patterns of gene expression and biological states.
Transformer architectures and other deep-learning techniques can then be used to learn representations that can help predict the characteristics or responses of cells.
But gene expression is just one part of biology. A snapshot of the transcriptome does not directly give information about all the protein levels, interactions, metabolism and physical processes within a cell.
Thus, gene expression can be used to simulate a mechanism only partially.
3. From Biological Foundation Models to Virtual Cells
Biological foundation models are Artificial Intelligence (AI) models that are trained on extensive biological data to aid with a variety of downstream applications.
Some models target the sequences of DNA. Others study the structure of proteins, molecular characteristics, gene expression, or single-cell data.
These may be implemented as a virtual-cell system.
A simpler architecture might consist of four layers:
The data integration for biological data in Layer 1.
This system captures and integrates data from gene expression, protein, imaging, genetic perturbation and other experiments.
Layer 2: Representation learning
AI transforms intricate biological data into mathematical representations, identifying patterns and relationships.
Layer 3: Perturbation prediction
The model predicts the changes in the state of the cell if a gene, a drug, an environmental condition or a variable is changed.
Layer 4: Experimental validation
Scientists compare predictions with laboratory results; they find if predictions are incorrect and learn from the results to make better predictions in later experiments.
This process establishes a feedback loop between calculation and real biology.
The aim is not just to identify the cells. This is to foresee the response of biological systems to interventions.
This is a much more difficult problem.
4. Predicting Biological Interventions Is So Difficult
Identification of a biological pattern is one thing, prediction of what should happen when that pattern is altered is another.
Imagine that an AI model discovers that two genes tend to be switching on at the same time. That correlation doesn’t necessarily indicate any cause-and-effect between the two genes.
The effect of both genes may be mediated through a third gene. Or the relationship may only apply to a specific cell type or environment.
This presents a significant problem that is referred to as the difference between correlation and causation.
Virtual-cell systems need to address the following questions:
- What happens if a particular gene is removed?
- What are other genes being affected by that intervention?
- Is there any difference between the effect on healthy cells and diseased cells?
- What is the change in the response with time?
- Do different cell types respond the same to a drug?
These questions can be explored by researchers with perturbation datasets such as experiments in which the activity of genes is turned off or modified.
However, laboratory testing is costly and the potential biological therapies are numerous that have never been tried.
An AI model thus can be used to forecast results that are not seen in the training information it was taught.
These are not always accurate, especially if the intervention is introduced into an unknown pathway or if a different cell response occurs.
The predictive ability of a credible virtual cell is not just whether it can essentially regenerate the data it has been exposed to thus far.
5. The future of drug discovery lies in Virtual Cells.The future of drug discovery is Virtual Cells.
The process of developing a drug is a lengthy and costly process. Before treatment can advance into clinical development, researchers must find targets, find compounds that fit the target, assess the biological activity of the candidate compounds, and investigate any safety concerns.
Several candidates fail to secure some of the marks they hope for because biological systems are more complex than the initial experiments indicate.
One way to do this might be with the aid of AI-enabled virtual cells.
Suppose a scientist is studying a disease linked to a specific pathway of a cell.
A virtual-cell model could be used to:
- Find potential genetic markers or pathways linked to the disease.
- Anticipate what would happen if the targets were altered.
- Make a guess about which cells will respond.
- Make a comparison of the predicted effects of different interventions.
- Identify likely hypotheses to test in the lab.
This may decrease the number of low value experiments and promote the research of biological mechanisms more efficiently.
But taking a guess at how a drug will affect a single cell, however, does not prove to be safe and effective in humans.
There are several organs, immune responses, metabolic processes, and environmental influences in the human body that interact. The treatment that works in the cell culture may not work the same way in an animal or human patient.
Virtual cells should thus be thought of as a tool that can enhance the research process, rather than a substitute for preclinical testing and/or clinical trials.
6. Can AI Build a digital twin of a Human Cell?
A digital twin is a simulation of a physical thing that can be updated with information and model how the thing will act.
In engineering, a digital twin could be a model of a machine, industrial process or physical infrastructure.
A living system with components that interact across multiple spatial and temporal scales would be a much more complex system to represent with a biological digital twin.
A full-fledged cellular digital twin could embody:
- Gene regulation.
- Production and degradation of proteins.
- Metabolic reactions.
- Molecular transport.
- Cell signaling.
- Growth and mortality of cells.
- Environmental interactions.
- alterations in the condition of an individual cell throughout a cell’s lifetime.
It would also have to take into consideration uncertainty and biological variation.
Cells with the same nominal type may be different due to their developmental history, molecular state, or local environment.
This is contrasted with many man-made systems, which are not as fundamentally different as biology.
A realistic DT can thus be achieved using a combination of machine learning, mechanistic equations, molecular simulation and experimental measurements.
AI will not be used to supplant existing biological knowledge, but rather support the incorporation of this knowledge into more predictive computational frameworks.
7. The Data Problem: There isn’t a perfect biology training dataset.
Data quality is one of the largest challenges to virtual cells.
While big AI models tend to require vast amounts of data to be trained, biological measurements are hard to gather and compare.
Instruments, protocols, cell lines, laboratory and environment settings may vary in experiments. Apparently similar experiments can yield different measurements.
Other problems include:
Limited information: Scientists cannot measure all the components of all cells at once.
Biological assays add noise to the measured values (measurement noise).
Limited perturbation coverage: There are many genes and molecular interactions that have not been experimentally tested in all the conditions of interest.
Technical differences between experiments can show up as biological differences, these are called “batch effects”.
Data imbalance: Many cell types and diseases have been studied, while some are less common and less well understood.
Such restrictions are important because a model can detect patterns from experimental procedures instead of the real biological mechanisms.
A strong virtual-cell model must then have more than just a large set of data. Requires carefully designed experiments, consistent measurements, reliable benchmarks, and uncertainty estimates that are transparent.
8. Can a Virtual Cell find Biology that Humans have missed?
Maybe, but the difference between pattern and mechanism is noteworthy.
AI can process high-dimensional biological data and uncover relationships that would be hard to detect the human eye.
For instance, a model could be able to detect a set of gene-expression modifications linked to a specific cellular response. It might also suggest ideas for new areas of research on molecular pathways to explore.
The researchers could then follow those hypotheses with specific experiments.
This establishes a potent research process:
AI Prediction → Experimental Test → New data → Improved model
These systems, over time, could enable scientists to investigate more hypotheses in the field of biology than would be feasible based on trial and error methods alone.
But a hypothesis produced by an AI is not necessarily a discovery of science! It must be reproducible and independently validated and provide evidence beyond mere statistical association that it has a cause and effect relationship.
The most successful systems will be those which can make correct forecasts in real new experimental conditions.
9. The Ethical and Scientific Risks
Virtual cells would speed up research, but would also present risks due to their limitations.
Using the wrong model could lead researchers to drug targets that are ineffective or miss key biological reactions.
A model trained with a small set of data may not work well in other cell types, other populations, and in different disease states.
In addition, there is a danger that the computational predictions could be reported as proven biological truths, but are actually only tentative hypotheses.
Advanced predictive tools also can reduce obstacles to potentially harmful experimentation in some aspects of biotechnology. Appropriate access controls, safety evaluation and oversight are all necessary for responsible deployment.
The researchers should be able to clearly articulate what the model predicts, the evidence that supports the prediction, and what conclusions have not yet been confirmed in the lab.
Predictions are not the only means by which scientific progress can be made; knowing when to not believe the prediction is equally important.
10. So what’s next for virtual-cell technology?
The future of virtual cells will probably be based upon the integration of many types of biological knowledge.
Future systems are not restricted to gene-expression data, but could utilize genomic data, protein structures, protein interactions, microscopy, spatial measurements, and time-series experiments.
A second thrust is to create models that predict the responses of cells to interventions, instead of simply describing their current state.
More efficient quantification of uncertainty, more equitable model comparison, and testing prediction generality across labs and biology will also be important for researchers.
The most promising breakthroughs might lie in the integration of AI with automated laboratory platforms, where computational systems can suggest experiments, interpret the findings and improve their forecasts.
These systems might enable more of an iterative and data-driven approach to biological research.
However, a universal and complete simulation of a living cell is a far more challenging task than predicting specific responses to cells.
Advancements in AI and a better experimental understanding of biology will be key to making progress.
In conclusion, Simulating Life is the next frontier of artificial intelligence.Finally, Simulating Life is the next frontier of AI.
Virtual cells are part of a transformation in the utilization of artificial intelligence by scientists.
Researchers are not just looking at existing biological data to analyze it with AI, they are also pursuing systems that can predict the response of a biological system to changes.
If these predictions are made sufficiently accurate and reliable, virtual cells could aid scientists in studying disease mechanisms, prioritizing drug targets, understanding genetic perturbations and create more informative experiments.
This is a huge challenge.
All living cells have complex processes that are interconnected and change over time, in different environments and between different organisms. There is no magic model that can be taken for granted that will capture all that complexity when trained with limited observations.
It is not a question of getting a perfect digital copy of life in the near term, just a realistic one. It’s an ever-expanding repertoire of confirmed computational models that simulate a particular sort of biological behavior.
In the end, it will only be a matter of time before the question is asked:
Can a machine forecast the action of a living cell when scientists alter an aspect of it—and can that forecast stand the test of reality?
A virtual cell would be one of the most significant tools in computational biology and medicine, if the answer continues to be yes.