From Blueprints to Behavior

A cell is chemistry, and chemistry follows rules. Molecules meet, bind, and break apart at rates that physics sets, and mass and energy are conserved at every step. This suggests an obvious way to build a virtual cell. Catalog every molecule inside the cell, every reaction among them, and the rate of each one; write the whole thing down as equations; and let a computer run it forward in time. Do it well enough and you would have a cell.

This is the oldest and most literal version of the virtual cell dream: to build life from the bottom up, out of its parts. The previous post defined a virtual cell as a model you can run rather than a description you can read, and asked what happened the last time scientists tried to build one that way. The answer is a landmark, and inside it sits the reason the dream is so hard.

Investor? Learn about our virtual cell

The approach described above is called mechanistic (or “bottom-up”) modeling, and it is how systems biology has always worked. Each molecule becomes a variable, each reaction an equation, each rate a parameter, and the system evolves from there. The mathematics is well established: differential equations for the average behavior of many molecules, stochastic simulation for the low-copy events where randomness matters, constraint-based methods for the steady, realistic flow of metabolism.

Where the wiring is known, this works, and it works well. Genome-scale models of metabolism, hundreds of reactions reconstructed from an organism’s cataloged enzymes, can predict how fast a bacterium will grow on a given nutrient and which of its genes are essential, from the network alone. Simulators like this have existed since the late 1990’s, allowing biologists to watch molecules react and diffuse through the geometry of a modeled cell. For decades the field advanced this way, one pathway at a time, each model a careful reconstruction of a single corner of the cell.

The open question was whether anyone could ever do this for the whole cell, all at once.

In 2012, a team at Stanford did it. They built a working simulation of an entire organism, a complete living cell rather than a pathway or a subsystem, and they chose the organism strategically. Mycoplasma genitalium is a minimal parasite of the human urogenital tract, and its genome, at 525 genes, sits near the lower limit of what independent life seems to require. (Your own cells carry roughly forty times as many.) M. genitalium was deliberately chosen for this task because it’s about as simple as a free-living cell gets, and that simplicity is the only reason the project was tractable.

Even so, it was an enormous undertaking. The model held all 525 genes and the molecules they produce, divided into twenty-eight modules (subparts of the overall computer program): one module for DNA replication, one for transcription, one for protein folding, one for metabolism, and so on. Each module used the mathematics that fit it, and all of them ran at once, sharing a single pool of molecules, so that all twenty-eight had to stay consistent with one another at every step. Assembling it meant pulling thousands of parameters from some nine hundred publications, each one a number that an experiment somewhere had measured.

And it worked. Using the molecular rules alone, with nothing about the cell’s higher-level behavior supplied by hand, the model played out the entire life cycle of a single cell from birth to division. It predicted things that live far above the chemistry: how long the cell cycle takes, which genes are essential, how the cell allocates its resources. It matched measurements it had never been tuned to reproduce. In a few cases it even disagreed with the published literature and turned out to be right, catching parameters that had been measured wrong. For the first time, someone could point at a running program and say, with a straight face, that it was a working model of the cell.

This kind of model has virtues that are easy to undervalue, and each of them returns later in the series.

Conservation. Because the model is built from real reactions, mass and energy are conserved automatically. It cannot predict a protein appearing from nothing, or a cell doing work with no energy source, because no such trajectory exists in its equations. Every state it produces is one a real cell could occupy.

Intervention. You can ask it a real question about acting on the cell, such as deleting a gene or blocking a reaction, and it answers by re-running the affected chemistry and following the consequences. It reasons about doing, not about what usually occurs alongside what. That question of interventional cause and effect, as the last post argued, is what a virtual cell exists to answer.

Transparency. Every variable is a molecule and every step is a reaction, so every prediction made carries its reasons with it. You can open the model, follow the chain of cause and effect, and interrogate it step by mechanistic step. A prediction you can argue with is one a scientist can investigate, and that a doctor can evaluate for treatment choices.

These are the properties you would want in any model used to make real decisions in biomedicine, and the bottom-up tradition has all three. They are worth remembering, because the rest of the series is about how to keep these positive features while overcoming the limitations that typically come with them.

Despite the virtual cell success with Mycoplasma spp., we still do not have a whole-cell model of a human cell. Under the approach used in 2012 alone, there could not be. That method hits a wall: missing knowledge.

An incomplete parts list. A mechanistic model has to be told its reactions in advance. For M. genitalium the wiring was almost fully known. A human cell has some twenty thousand genes and layers of regulation that shift with cell type, signal, and moment, and no complete wiring diagram for it exists. You cannot write down a reaction no one has discovered.

Unmeasured rates. Even where the biochemical wiring and stoichiometric rules are known, each reaction needs its rate, and most of those numbers have never been measured in a living human cell. Nor are they fixed constants; they change with conditions depending on cellular state (e.g., diseased versus healthy). One stripped-down bacterium took nine hundred papers. The same accounting for a human cell mostly does not exist, and could not be produced quickly enough to matter.

Labor that does not scale. The M. genitalium model was years of work by a large team, for the simplest cell in biology. Building a human cell the same way, by hand, for a system tens of thousands of genes larger and far more heavily regulated, is not a matter of working harder. The labor and complexity grows faster than any team can manage.

Three problems, but really one. This is the bottleneck of Series 1 in a new guise. There, cancer biology produced complexity faster than we could comprehend it. Here, the cell holds more complexity than we can write down. The bottom-up method needs complete prior knowledge, and for a human cell that knowledge does not yet exist.

It would be easy to take the wrong lesson from this, to decide that mechanism, causality, and physical grounding are luxuries we cannot afford and should give up. That reading gets the diagnosis backwards.

The first virtual cells did not fail because they were mechanistic, causal, and physical. In fact, those are the exact properties that made them trustworthy. A model that cannot violate conservation, that answers an intervention by re-running biology, and that explains itself in the language of molecules is a model worth having. What failed was the method of construction: building the cell by hand, reaction by reaction, from a literature that will never be complete for a system this large.

The mechanistic dream had the right idea about what a virtual cell must be, and the wrong idea only about how to build it. That distinction is the hinge this whole series turns on. The problem was never including mechanism - it was hand-writing the mechanism. What we want is a way to keep everything that made these models trustworthy without entering the rules of life one measured number at a time.

If knowledge has to be entered by hand, you are limited to what has already been written down, and what has been written down is a small fraction of what a human cell does. So the field decided to turn the problem around: instead of telling the model the rules, it would give the model data and let the rules emerge. Rather than draw the cell from a finished blueprint, it would infer the cell from its behavior.

That is the bet that now dominates the field, and it approaches the cell from the opposite direction. We take it up next.

Next in the series: The Cell That Learns Itself

The Virtual Cell series is published by Galen, where we are building the computational layer for understanding and programming living systems.