How to build a digital ‘twin’ of the human brain – what existing models overlook

The potential to create personalised digital “twins” of your brain and body is a hot topic in neuroscience and medicine today. These computer models are designed to simulate how parts of your brain interact, and how the brain may respond to stimulation, disease or medication.

The extraordinary complexity of the brain’s billions of neurons makes this a very difficult task, of course, even in the era of AI and big data. Until now, whole-brain models have struggled to capture what makes each brain unique.

People’s brains are all wired slightly differently, so everyone has a unique network of neural connections that represents a kind of “brain fingerprint”.

However, most so-called brain twins are currently more like distant cousins. Their performance is barely any closer to the real thing than if the model were using the wiring diagram of a random stranger.

This matters because digital twins are increasingly proposed as tools for testing treatments by computer simulation, before applying them to real people. If these models fail to capture fundamental principles of each patient’s unique brain organisation, their predictions won’t be personalised – and in worst cases could be misleading.

In our latest study, published in Nature Neuroscience, we show that realistic digital brain twins require something that many existing models overlook: competition between the brain’s different systems.

Our findings suggest that without competition, digital twins risk being overly generic, missing out on what makes you “you”.

Excess of cooperation

The human brain is never static. The ebb and flow of its activity can be mapped non-invasively using neuroimaging methods such as functional MRI. A computer model can be built from this, specific to that person and simulating how the regions of their brain interact. This is the idea of the digital twin.

The brain is often described as a highly cooperative system. Yet everyday experiences such as focusing attention or switching between tasks tells us intuitively that brain systems compete for limited resources. Our brains cannot do everything at once, and not all regions can be active together all the time.

During attention-demanding cognitive tasks, some regions of the brain such as the intraparietal sulcus (IPS) routinely increase activity, while others such as the posterior cingulate cortex (PCC) and medial prefrontal cortex (MPF) decrease activity. At rest, the opposite happens. Fox et al/PNAS, Author provided (no reuse)

Despite this, the vast majority of brain simulations over the past 20 years have not taken these competitive interactions between regions into account. Rather, they have “forced” neighbouring regions to cooperate. This can push the simulated brain into overly synchronised states that are rarely seen in real brains.

In a large comparative study of humans, macaque monkeys and mice, our international team of researchers used non-invasive brain activity recordings to show that the most realistic whole-brain models not only require cooperative interactions within specialised brain circuits, but long-range competitive interactions between different circuits.

To achieve this, we compared two types of brain model: one in which all interactions between brain regions were cooperative, and another in which regions could either excite or suppress each other’s activity. In humans, monkeys and mice, the models that included competitive interactions consistently outperformed cooperative-only models.

Using a large-scale analysis of over 14,000 neuroimaging studies, we found that spontaneous activity in the competitive models more faithfully reflected known cognitive circuits, such as those involved in attention or memory. This suggests competition is crucial for enabling the brain to flexibly activate appropriate combinations of regions – a hallmark of intelligent behaviour.

Visual summary of our study:

When whole-brain models of humans, macaques and mice are allowed to treat interactions between some brain regions as competitive, they consistently do so – generating activity patterns that closely resemble those associated with real cognitive processes. Luppi et al/Nature Neuroscience, CC BY

We concluded that competitive interactions act as a stabilising force, allowing different brain systems to take turns in shaping the direction of the brain’s ebbs and flows without interference or distraction. This ability to avoid runaway activity may also contribute to the remarkable energy-efficiency of the mammalian brain, which is many orders of magnitude more efficient than modern AI systems.

Crucially, models with competitive interactions were not only more accurate but also more individual-specific. This means they were better at capturing the unique brain fingerprint that distinguishes one person’s brain from another’s.

No longer lost in translation?

The fact that our findings hold across humans and other mammals suggests they reflect fundamental principles of how intelligent systems work. In each case, we found models with competitive interactions generated brain activity patterns that closely resembled those associated with real cognitive processes.

This could have major implications for translational neuroscience. Animal models are routinely used to test treatments before human trials, yet differences between species often limit how well these results translate. Around 90% of treatments for neuropsychiatric disorders are “lost in translation”, failing in human clinical trials after showing promise in animal trials.

Combining brain imaging data from human patients with whole-brain modelling could radically change this. A framework that works across species would provide a powerful bridge between basic research and clinical application.

If someone needs intervention in the brain, for example due to epilepsy or a tumour, their digital twin could be used to explore how the patient’s brain activity would change when stimulated with different levels of drugs or electrical impulses. This might significantly improve on existing trial-and-error approaches with real patients, and thus provide better treatments.

The general principles of brain organisation across species also offer a path for understanding how to shape the next generation of artificial intelligence. In the not-too-distant future, we may be able to construct digital twins that are more faithful in reproducing the salient features of the human brain – and potentially, AI models that are more faithful to the human mind.