Anatomists dancing down the rabbit hole

We came for the anatomy. Somehow, we ended up with a Russian mathematician, a statistical blanket and a rather provocative question about the nature of living tissue.

How does a biological system remain coherent while continuously changing?

How do millions of seemingly unpredictable molecular events produce the remarkably reliable tissues we dissect, teach, move and treat? Are these questions that keep you up at night wondering too?

And, more specifically for us fascia nerds: how does the connective tissue matrix adjust its mechanical behaviour in response to the world around it?

These are questions my research has been circling for several years.

They have led me to collaborative work with John Sharkey, Suzanne Scarlata, and recently, Robert Schleip. In our reviews, we follow the rabbit from the stochastic nature of fascia to calcium signalling, hyaluronan homeostasis and our proposal of a myofascial rheostat.

Now we're working on a manuscript about viscoelasticity and mathematical modelling of biological tissues.

But wait. How did we get from fascia to something so... statistical?

First, the journey that brought us here

Our starting point was the relationship between microscopic variability and macroscopic order. You know, how we can all be subject to so much randomness and yet (most of us) walk on two feet and end up with a startlingly similar humanoid shape and a shared sort of humanoid experience.

In our 2025 paper, John Sharkey and I considered fascia as a stochastic morphogenetic system. Big words, yes, but it all gets back to something so simple it almost disappears from view: the tiniest, seemingly random fluctuations give rise to our bodies. Individual molecular and cellular events can be probabilistic, yet the tissues emerging from them possess recognisable architecture and relatively dependable mechanical properties.

Think of raindrops falling unpredictably across a hillside, eventually gathering into recognisable streams and pools.

This picture of probability in fascia is mechanistically grounded in the calcium–hyaluronan (CHA) axis, developed with Suzanne Scarlata in our paper featured in Carla Stecco's special issue of the International Journal of Molecular Sciences.

Most recently, stochasticity and the CHA axis inform what we have termed the myofascial rheostat, whereby each tissue niche operates its own calcium-dependent constraints on HA production.

The rheostat was our attempt to capture a deceptively simple idea: connective tissue does not merely respond to loading. Through cellular signalling, HA synthesis, degradation and clearance, it participates in regulating its own mechanical environment.

Hyaluronan synthesis, signalling and turnover within the fascial extracellular matrix. 

Schematic illustration of hyaluronan (HA) within the collagen-rich extracellular matrix, highlighting four interconnected processes: (a) HA and its molecular structure within the surrounding matrix; (b) HA-mediated cellular signalling involving CD44 and RHAMM receptors; (c) HA synthesis and extrusion through membrane-bound hyaluronan synthase; and (d) enzymatic HA degradation by hyaluronidases. Together, these processes illustrate the dynamic balance between HA production, receptor-mediated signalling and breakdown that contributes to extracellular matrix remodelling, hydration and viscoelastic behaviour.

---

 

Not just the rope, but the water surrounding it

Much of connective tissue biomechanics focuses, understandably, on collagen fibres: their alignment, stiffness, crosslinking and capacity to transmit force.

In this suite of papers, we (here I refer to the previously mentioned collaborators) are particularly interested in the hydrated environment in which those fibres exist.

Hyaluronan (HA) influences matrix hydration, lubrication, molecular interactions and viscoelastic behaviour. Its properties are dynamic, shaped not just by how much HA is produced, but by molecular weight, organization, degradation and clearance.

The question became less about how much HA we make, and more about how well we regulate its continual turnover. How does the body "know" when to make HA? When to unmake it? How does it even do that?

And that brought us to an intriguing problem. We had a biological feedback mechanism, but could we identify the variables that actually mediate the relationship between a cell and the mechanical environment it inhabits? Because all of this happens "spontaneously" outwith the nervous system, outwith any "system" we might otherwise try to name.

So, to answer that question, we needed to leave the anatomy laboratory for a moment.

1. A Russian mathematician who counted poetry

Stay with me here. A coin toss is independent of the one before it; in other words, whether you end up with heads or tails is random chance. Language is not. If the letter Q appears (in an English word, at least), the letter U becomes considerably more likely to follow.

Meet Andrey Markov, a Russian mathematician (1856–1922) interested in the probability of dependent events.

---

In 1913, Markov famously analysed 20,000 letters from Pushkin's Eugene Onegin, examining how vowels and consonants followed one another.

His contribution was to show how sequences containing dependencies could still be described probabilistically. And we can take this gambling from language to physicality.

A walk through a Highland glen

Imagine a walker navigating three familiar places:

• S — Shelter: a small bothy.

• C — Crossing: a burn (stream) cutting through the glen.

• R — Ridge: a woodland path climbing towards higher ground.

From the crossing, she might head towards the ridge, return to the shelter or remain where she is. 

Each possible move has a probability. The arrows in a Markov chain represent these potential transitions.

Here is the interesting part: in a simple Markov model, knowing where our walker is now is sufficient to estimate where she might go next. We need not reconstruct her entire walking history.

This is the Markov property.

A three-state Markov chain. S (Shelter), C (Crossing) and R (Ridge) represent possible locations, while directional arrows indicate the probabilities of moving between them or remaining in the same state.

 

---

Of course, our real walker remembers things. Perhaps she is tired, has forgotten her lunch or knows a storm is approaching. Those factors would naturally influence her next decision.

The Markov property is therefore a modelling assumption, not a universal law. If you were to observe a hundred such walkers, we would naturally see patterns emerge, because people get their feet wet crossing burns and that leads to a set of normative behaviors (stop for a rest, turn back, etc). To represent more complicated behaviour, we may need to include more information in the current state.

 There is an important implication here for anatomists. Biological tissues have memory too. Loading history, molecular organization and previous injury can and do influence what happens next in living organisms. We must therefore choose our state variables carefully enough to represent those influences.

And Markov gave us the resulting Markov chain... he did not himself invent the Markov blanket.

That came decades later.

2. The blanket came from statistics, not anatomy

By the 1980s, artificial intelligence researchers were grappling with complicated networks of dependent variables.

Judea Pearl's work on Bayesian networks helped establish the Markov blanket as a way of identifying a statistically sufficient neighbourhood around a variable. This would have monumental implications in medicine with global benefits to healthcare at population scale.

Imagine a clinician weighing hundreds of possible factors in a diagnosis: genes, symptoms, blood results, age, environmental exposures and medical history. Actual implementations of causal Bayesian networks (CBNs) help clinicians make decisions spanning oncology, pediatrics, neurology, and critical care to support diagnosis, counterfactual treatment evaluation, and clinical decision support.

At their heart, CBNs ask: Must every variable be known?

Not necessarily.

In oncology diagnosis, hybrid networks combine deep learning image classification with Bayesian networks to model causal dependencies in radiology (X. et al., 2023). Diagnostic networks developed for ocular oncology integrate patient demographics, staging, and tumor characteristics to achieve diagnostic accuracy of 99.4% in identifying ocular malignancies (Revathi et al., 2025).

Similarly, automated systems construct clinical networks from electronic medical records using greedy search algorithms to discover causal disease-symptom dependencies for stroke care (Shen et al., 2018).

In short: some variables may contain all the information needed to understand the condition within the chosen model. Once those are known, the remaining variables provide no additional predictive information. The rest, as they say, is noise.

That collection of essential variables is the Markov blanket.

The room, the window and the draught under the door

Imagine sitting in a room while the weather changes outside... it is October and the chilly wind is picking up!

A window transmits light and warmth. Shutters allow you to influence conditions. You can interact with the weather without having direct access to every atmospheric process.

The window-and-shutters analogy. Sensory and active exchanges connect an internal system to its external environment, illustrating the principle of mediated interaction.

But here's the crucial distinction: the window isn't a Markov blanket simply because it separates inside from outside.

It becomes a candidate blanket when the information passing through it is sufficient to explain the relationship between the room and the world outside. Remember that chilly wind?

What if there's a draught beneath the door?

Then the window hasn't captured everything. Information is getting through another route, an icy note that we can feel. So the window isn't the only location for info about the outside to inform your experience of sitting in the room... 

In statistical terms, the proposed blanket (window) has failed to establish conditional independence.

That is the central principle we will return to when discussing fascia.

3. Friston asks: what keeps a living thing itself?

Enter neuroscientist Karl Friston.

Turns out that this question is surprisingly anatomical.

A living cell and a newly dead cell contain much the same chemical ingredients. Yet one maintains its organization, continually exchanging matter and information with its surroundings. The other begins to lose that organization (thanks, entropy).

How does a living system maintain its characteristic states in an unpredictable environment?

Friston's Free Energy Principle explores whether biological self-organisation can be understood through processes that keep organisms within viable states.

I find analogies indispensable when it comes to statistics, so you'll have to bear with it as we go from hillwalking to sitting in a draughty room to sailing on a loch. Picture a small sailing boat crossing a choppy loch. The sailor cannot control the wind, but can continually adjust the sails and rudder to remain afloat.

The Markov blanket offers a statistical way of describing the relationship between the boat and its environment without treating them as one undifferentiated system.

And one important clarification for biologists: the variational free energy discussed in this theory is an information-theoretic quantity, not cellular ATP or Gibbs free energy.

4. Where biology is using Markov blankets

These applications vary in maturity. Neuroscience has a substantial theoretical and computational literature, whereas some multiscale biological and ecological applications remain exploratory.

For our work, morphogenesis is especially interesting, and it remains underexplored in the fascia literature. 

We began with the question of how stochastic events at microscopic scales can contribute to predictable tissue architecture. Nested Markov blankets offer one possible mathematical language for investigating that transition.

But the existence of nested biological blankets must be demonstrated, not assumed, and that means a deep dive into the experimental evidence. And at this level (blog), it means another visualization to keep things digestible.

5. Returning to the fascia: meet our hypothetical fasciacyte

Imagine an HA-producing cell embedded in the hydrated extracellular matrix of the crural fascia.

Its owner starts walking uphill. Maybe this is you?

The mechanical environment changes as you hike your way up the pasture terrain and into the forest. Membranes deform in response to the changing terrain. Mechanosensitive channels respond. Intracellular calcium signals fluctuate. Downstream pathways influence HA synthesis and matrix interactions.

Meanwhile, changes in extracellular HA composition, hydration and organisation affect what the cell experiences next. How your old Achilles injury may rear its ugly head as you approach your picnic spot. That niggling ache around your ankle heavily influences your gait, and compromises your walking posture, which influences your next steps, how soon you stop, your mood, and on and on.

We have returned to our figure-of-eight rheostat.

Could the CHA axis have a Markov blanket?

Here is our proposed translation.

The proposed mapping of the CHA axis onto external, sensory, internal and active states.

 

---

Notice something important.

External does not mean outside the body.

External means outside the system we have chosen to model. For a fasciacyte, its surrounding collagen, HA and interstitial fluid are external, although they remain very much inside the body.

From the myofascial rheostat to a statistical blanket

This is where our original figure-of-eight becomes useful.

Conceptual figure-of-eight mapping of the CHA axis. Pale sage represents the local extracellular environment (η), teal the sensory states (s), cream and gold the internal states (μ), and lavender the active states (a).

The figure-of-eight depicts continuous feedback. The Markov blanket formalism asks whether particular sensory and active variables statistically mediate the relationship between intracellular regulation and extracellular mechanical conditions.

The sensory and active states together form the candidate blanket.

But drawing a figure-of-eight does not establish that the statistical relationship exists.

That is our next challenge.

6. Why use a Markov blanket when we already have feedback?

This is the question a sceptical reviewer should ask.

We already have homeostasis, systems biology, mechanotransduction and nonlinear feedback models.

What does a Markov blanket add?

It asks which variables carry enough information to account for the interaction between a system and its environment.

That could help us identify a reduced set of variables for modelling tissue viscoelasticity, rather than attempting to represent every possible molecular interaction.

It also creates the possibility of testing the model.

For example, we could examine how experimentally modifying Ca²⁺ signalling alters the relationship between mechanical loading and HAS2 expression. If substantial loading-dependent regulation persists through alternative pathways, our proposed sensory partition might need expanding.

The point is not to insist that calcium is the only route through which cells detect mechanical conditions. It is to identify a sufficiently complete set of mediating variables.

From mathematics to anatomy, in one table

7. Is the blanket actually real?

There is a lively scientific and philosophical debate here.

A Pearl blanket is fundamentally a statistical tool. A Friston blanket can carry stronger claims about biological autonomy and physical organisation.

These are not necessarily equivalent.

We are using the Markov blanket instrumentally. We are not proposing a new anatomical boundary or asserting that fascia possesses a privileged, experimentally demonstrated statistical interface.

We are asking whether the formalism can help us build a better, testable model.

Back from the Rabbit Hole

So how did we, as anatomists, end up here?

Because the anatomy kept asking questions that static descriptions alone could not answer.

Collagen provides architecture. Hydrated HA contributes a dynamic mechanical and signalling environment. Cells continually respond to their surroundings and participate in modifying them.

Our rheostat brought these ideas together as a physiological feedback system that can help us visualize and predict tissue states such as viscoelasticity.

The Markov blanket now gives us a more precise question: can we identify a set of measurable variables that captures the relationship between cellular regulation and extracellular mechanical conditions?

What is the absolute common denominator here?

Because a blanket is not proven simply by drawing a boundary around a cell... we may argue that drawing such lines and layers is useful in dissection but only gets us so far probabilistically.

It earns its place when its proposed statistical relationships produce predictions we can test, and potentially falsify. In this case, we are asking: What if calcium dependence is that essential variable, and its relationship with HA is the axis that stochastically (opportunistically) determines viscoelasticity?

Perhaps statistics isn't so far from the anatomy laboratory after all.

I see this journey as another way of asking how living tissues manage to remain themselves while continually becoming something new.