What I want to understand by simulating the brain

An idea I keep coming back to

I want to put what we know about brain pathways into mathematical models we can actually run. Represent the relationships in matrices, simulate how they change over time, and use AI tools to help us compare those simulations with real observations.

What interests me is whether this could help us ask better questions about mental disorders. Why can similar symptoms emerge for different reasons? Why might a system respond differently at different points in someone's life? What would we need to measure to tell two explanations apart?

There is already a foundation for this. The Virtual Brain combines connectivity information with models of neural activity to simulate signals such as EEG, MEG, and functional MRI. I want to build on that kind of work and make the connection between the biological question, the mathematics, and the evidence easier to follow.¹

What goes into the matrices

A matrix is a way to represent relationships. In a brain network model, each row and column might correspond to a region, with entries describing connection weights. Other arrays can describe transmission delays, inputs, or model parameters. The Virtual Brain uses both connectivity and local models of activity to generate dynamics over time.¹

Putting everything into one giant matrix would not make it a useful model. I would start by choosing a scale and a question. Molecular pathways, receptor behavior, neural populations, and whole brain networks describe different things. Connecting those levels requires explicit assumptions, compatible units, and evidence for the connections.

The equations tell the system how to change. Dynamic causal modelling is one established approach that links models of interacting neural states to measured signals and estimates their parameters. Its causal interpretation depends on the model and experimental design. A correlation matrix by itself cannot tell us which biological process caused a symptom.²

x(t + Δt) = x(t) + Δt · f(x(t), W, θ, u(t)) + ε(t)
y(t) = g(x(t)) + η(t)

Imaging gives the model something to answer to

I would use imaging and other measurements to constrain the simulation. Structural connectivity estimates can help define a network. Recorded activity provides signals a model should be able to reproduce. A forward model then translates simulated neural activity into the kind of signal the instrument actually measures.¹

That last step matters. Functional MRI measures changes related to blood oxygenation, rather than directly reading every neuron's activity or a receptor's state. A model linking neural activity to that signal has to include the measurement process. Otherwise, it is too easy to mistake a plausible looking simulation for a biological explanation.²

I would also want symptom tracking, behavior, sleep, and relevant physiological measurements where a study supports collecting them. Each should retain its own meaning and uncertainty. Missing data should stay visible. An AI system should never silently turn an unmeasured variable into a fact about a person.

Where AI could do useful work

AI can help search a space of possible models. Simulation based inference trains statistical models on simulated data to estimate which parameter combinations are compatible with an observation. Gonçalves and colleagues demonstrated this approach across mechanistic neuroscience models, including neurons and circuits. A useful result can be a distribution of possible parameters, rather than one deceptively precise answer.³

I would use language models to help organize literature and propose code for review, with every biological claim traceable to a source. I would use statistical inference to compare models with data. Those are different jobs, and I would want the interface to make the distinction obvious.

The difficult part is deciding whether the data can distinguish the explanations at all. Different parameter combinations may reproduce similar observations. If that happens, the next useful output is a proposed experiment or an uncertainty report. We should be able to say we do not yet know.³

Mental health diagnostics

AI simulation based mental health diagnostics is a direction I want to investigate. The first task would be narrower: can a model explain or predict a clearly defined feature of a person's symptoms or behavior, using data it has not already seen? That is a question we can design a study around.

I would be careful about turning an interesting brain pattern into a diagnostic label. Marek and colleagues examined data from roughly 50,000 people and found that many associations between brain measures and cognitive or mental health characteristics were smaller and less reproducible than small studies suggested. Their findings concern population association studies; they do not imply that every imaging experiment needs thousands of participants.⁴

Before I would call a tool diagnostic, I would want a defined clinical use, prospective evaluation, independent testing across settings, and evidence that it improves decisions. I would want to know who it fails for, how often it is uncertain, and how its performance compares with existing care. A successful simulation is only one part of that work.

Women's health

Women's health deserves its own research questions in this project. I am particularly interested in biological change over time, including how hormonal variation might interact with brain networks, symptoms, and individual sensitivity. The same modeling approach could support different questions without assuming they all share one mechanism.

In a 2024 study of 60 healthy women, researchers examined functional MRI across three menstrual cycle phases and reported differences in brain network dynamics associated with phase, hormones, and age. The study was observational and used a model free analysis. It did not establish a mechanism for PMDD or a diagnostic test for a disorder.⁵

For me, findings like this make timing a serious design consideration. I would want repeated observations and carefully defined comparison groups. I would also want to distinguish ordinary variation from changes linked to distress or impairment. Women should not be treated as a single biological pattern, and a useful model has to leave room for that variation.

What I would build first

I would begin with a small, documented model and an appropriately licensed research dataset. The first milestone would be reproducing a published result, then testing a specific extension. Every parameter would show whether it was measured, estimated, or assumed. Every run would record its data version, settings, and uncertainty.

The interface would let a researcher inspect a matrix, follow a connection back to its evidence, change an assumption, and see what happens to the simulated signal. I would want it to compare competing models and show where they disagree. That disagreement could help identify the next measurement worth making.

Drug discovery and repurposing could become later research applications. A model might help prioritize a hypothesis about an intervention, but a simulated improvement would still need laboratory and clinical testing. I would not turn a pathway match into a recommendation to take a drug or supplement.

This is the kind of work I want to bring my background in programming and data science into. I want to make difficult biological questions easier to investigate, and eventually help people get better answers about their health. The quality of the evidence has to grow alongside the ambition.