Computational tissue engineering and why the field is still experimental

In 2024 the Nobel Prize in Chemistry went in part to the people behind AlphaFold, a model that predicts the shape of a protein from its sequence. Weather forecasts have gained about a day of skill per decade: by 2015, a six-day forecast was as good as a five-day forecast had been ten years earlier (Bauer et al., 2015). Engineers test new wings in a computer before they build them.

In tissue engineering, we still make the scaffold, seed the cells and wait to see what happens. Computational tools are everywhere in the field, and I build some of them, but none of them lets us skip the experiment. This post is my attempt to explain why, and what would have to change.

The short answer is not computing power. The fields where simulation is trusted share a few things that tissue engineering does not yet have.

Even those fields are less settled than they look from outside. A review of aerospace simulation commissioned by NASA found that reliable use was “confined to a small but important region of the operating design space”, because flows that separate from a surface are still hard to predict (Slotnick et al., 2014). A trusted simulation is always trusted for a particular use, and checked against reality for that use.

What the term covers

“Computational tissue engineering” is not one method. I find it useful to think of three layers.

The physics layer uses equations that are well established elsewhere: elasticity for how a scaffold deforms, the Navier–Stokes equations for how medium flows through it, diffusion for how oxygen and nutrients reach the cells. Finite element and computational fluid dynamics models live here.

The biology layer describes what cells do: whether they divide, migrate, die or turn into bone or cartilage. Models here, such as mechano-regulation rules and agent-based simulations of individual cells, are built from observed patterns rather than from first principles.

The data layer skips mechanisms and learns from examples: machine learning models trained on experiments, and optimisation methods that decide which experiment to run next.

These layers have to cover an enormous range. Physiology spans about nine orders of magnitude in space and fifteen in time, from molecules to organs and from microseconds to a lifetime, which is why it needs a hierarchy of models rather than one model of everything (Hunter, 2004).

Where computation already earns its place

It would be wrong to say that models do not help. They help most where the physics dominates and the question is narrow.

Perfusion bioreactors are a good example. The shear stress that flowing medium puts on cells inside a scaffold cannot be measured directly, but it can be computed from the scaffold’s geometry, and that calculation was being done on micro-CT images of real scaffolds twenty years ago (Porter et al., 2005). Combining such simulations with a rule for how bone cells respond to shear lets researchers choose a flow rate for a given scaffold without a series of trial-and-error cultures; in one study the best rates depended mainly on scaffold geometry and ranged from 0.5 to 5 mL per minute (Zhao et al., 2018).

Models also make search cheaper. A simplified model of tissue growth in a bioreactor ran about 100,000 times faster than the full three-dimensional one, fast enough to optimise feeding schedules on the computer (Mehrian et al., 2018). In bioprinting, Bayesian optimisation found good printing settings for gelatin-based inks in between 4 and 47 experiments per ink, from a search space of thousands of combinations (Ruberu et al., 2021).

What these successes have in common is that the computer narrows the options and the experiment still decides. That is valuable. It is not the same as prediction you can rely on without checking.

The physics depends on the scaffold you actually made

Even the physics layer is less solid than it looks, because the model describes the scaffold that was designed, not the one that came out of the printer.

Printed strands sag, fuse and vary in width; struts in laser-melted metal come out rough, with cracks and trapped voids (Khalil et al., 2024). In five samples of one commercial scaffold with a nominally regular design, simulations based on micro-CT scans of each sample gave local fluid velocities and shear stresses up to five times higher than the design file predicted, and the same location differed by up to two-fold from one sample to the next (Campos Marin and Lacroix, 2015). Hydrogels add swelling, slow creep and degradation on top of that, so their geometry keeps changing after printing.

Easy maths, hard numbers

Some of the most useful calculations in the field are simple. How far into a block of cells can oxygen reach before the cells in the middle run out? With a few textbook assumptions, the answer is one short formula, and every student who designs a thick construct should try it.

The difficulty is not the equation. It is the inputs. How much oxygen a cell consumes depends on the cell type, its state and its surroundings, and published values for cells in culture range from below 1 to more than 350 attomoles per cell per second (Wagner et al., 2011). Because the depth depends on the square root of the inputs, a tenfold uncertainty in one of them still moves the answer about threefold, which is the difference between a construct that survives and one with a dead core. The formula is not the weak point; we just do not know the numbers to put into it precisely enough.

Cells do not follow an equation we know

The physics layer at least has agreed equations. The biology layer does not. There are no accepted first-principles laws for when a stem cell becomes bone, cartilage or scar tissue, so models of those decisions are built from rules that fit observations.

That makes them hard to test. In bone, for example, the links between mechanical loading, cell proliferation, blood supply and nutrient transport are still poorly understood, and every extra rule added to a model makes it harder to prove wrong (Giorgi et al., 2016). When several published mechano-regulation rules were run in the same fracture-healing model, all of them reproduced normal healing (Isaksson et al., 2006). A model that fits the data is not yet a model that explains it.

Agent-based models, which simulate cells one by one, face the same problem from another direction. A recent model of muscle regeneration needed 52 parameters that could not be related to any known physiological measurement and had to be tuned against a few time courses of tissue-level data (Haase et al., 2024).

And the system does not stand still. A scaffold is designed to disappear.

The data are small, uneven and optimistic

If the rules are unknown, the obvious alternative is to learn them from data. This is where I work, and it is where the limits are clearest to me.

First, the datasets are small. AlphaFold was trained on structures from the Protein Data Bank, which now holds more than 260,000 experimentally determined structures, all described in a standard format (RCSB PDB, October 2026). A large tissue-engineering dataset has a few thousand rows assembled by hand from papers; the MLATE dataset has 2,646.

Second, the inputs are not what they seem. Natural materials vary: two batches of Matrigel measured in one study had a stiffness of about 400 Pa and a third had about 840 Pa (Aisenbrey and Murphy, 2020). Cells vary too: the age and sex of the donor change how mesenchymal stromal cells grow and behave (Siegel et al., 2013), and these cells change gradually with every passage rather than only at the end (Wagner et al., 2008). A row in a dataset that says “alginate, 2%, mesenchymal stem cells” hides all of this. Part of the variability is a choice, though: a carefully controlled synthesis of gelatin methacryloyl gave batches that were almost identical (Zhu et al., 2019).

Third, the record is optimistic. Failed formulations are rarely published. We have no direct measurement of this for tissue engineering, but other preclinical fields show its size: in animal studies of stroke, about one experiment in seven was never published, and the missing studies made treatments look nearly a third more effective than they were (Sena et al., 2010).

Fourth, reporting is not standardised. Bio-nano research has proposed a minimum reporting checklist (Faria et al., 2018). Bioprinting has a standard guide for bioinks, ASTM F3659-24, but it does not yet give quantitative criteria for judging printability (Martín Compaired et al., 2025).

All of this affects how models are tested. When the test data come from the same papers as the training data, a model can score well by recognising the lab rather than understanding the material.

Where the field’s attention is

I wanted to know how much of the field’s work is computational, so I counted papers in PubMed.

Two things stand out. Computational modelling has stayed a small share of tissue-engineering papers for twenty years, between about 1.7% and 2.8%, creeping up since 2016. Machine learning was almost absent until about 2018 and has grown quickly since, passing modelling for the first time in 2025, but much of that growth is in reviews rather than new models. The field is talking about machine learning faster than it is testing it.

What “credible” would mean

Other fields have worked out how to decide when a model can be trusted, and their answers are instructive. The ASME V&V 40 standard (2018), which the US Food and Drug Administration recognised in 2019, does not ask whether a model is good in general. It asks how much a decision relies on the model and how bad a wrong decision would be, and it scales the evidence required to match (ASME). The FDA’s guidance on computational modelling in medical device submissions, final since November 2023, builds on the same idea, but it covers physics-based and mechanistic models, including mechanistic models of physiology, and explicitly leaves out standalone machine learning (FDA, 2023).

There are precedents for simulation standing in for experiments. A simulator of type 1 diabetes was accepted by the FDA as a substitute for animal trials in preclinical testing of insulin-control algorithms (Kovatchev et al., 2009), and an entirely computer-simulated imaging trial reached the same conclusion as a clinical trial for detecting breast masses (Badano et al., 2018). Both rest on well-characterised physics or physiology and large amounts of data. Since December 2022, US drug law also counts computer modelling among the “nonclinical tests” that can support an application, alongside cell-based assays and animal tests (Public Law 117-328, Sec. 3209). I have not found a documented case of a regulator accepting in silico evidence in place of experiments for a tissue-engineered product.

So when I ask how credible a tissue-engineering model is, I think of a ladder.

What would make it less experimental

None of this means the field should model less. It means computation and experiment have to be joined more tightly, so that every experiment also improves the model.

That is already happening. A robotic system that used Bayesian optimisation to improve how stem cells are turned into retinal pigment epithelium tested 143 conditions, out of about 200 million possible combinations, in 111 days, and improved the pigmentation score by 88% over the starting protocol (Kanda et al., 2022). Digital twins of bioreactors, mechanistic models that run alongside a culture and are updated from its measurements, are an active idea, although still mostly at the stage of concepts and prototypes (Möller and Pörtner, 2021).

What would move the field further, in my view:

• defined materials and scaffolds characterised as built, not only as designed;
• a minimum reporting standard for biofabrication studies, in the spirit of the bio-nano checklist;
• shared datasets that follow the FAIR principles (Wilkinson et al., 2016) and include formulations that failed;
• blind community benchmarks, like the CASP experiment that has tested protein structure prediction on unpublished structures since 1994 (CASP);
• models that report their uncertainty, and credibility frameworks that cover data-driven models as well as physics-based ones.

Tissue engineering is still experimental because living tissue is more variable, less described and less measured than air over a wing. Computation will not remove the experiment. It will, more and more, choose which experiment to run.

References

1. Aisenbrey EA, Murphy WL. Synthetic alternatives to Matrigel. Nature Reviews Materials 2020;5:539–551.
2. ASME. V&V 40-2018: Assessing credibility of computational modeling through verification and validation: application to medical devices. 2018.
3. Badano A, et al. Evaluation of digital breast tomosynthesis as replacement of full-field digital mammography using an in silico imaging trial. JAMA Network Open 2018;1(7):e185474.
4. Bauer P, Thorpe A, Brunet G. The quiet revolution of numerical weather prediction. Nature 2015;525:47–55.
5. Campos Marin A, Lacroix D. The inter-sample structural variability of regular tissue-engineered scaffolds significantly affects the micromechanical local cell environment. Interface Focus 2015;5(2):20140097.
6. Colom A, et al. Oxygen diffusion and consumption in extracellular matrix gels: implications for designing three-dimensional cultures. Journal of Biomedical Materials Research Part A 2014;102(8):2776–2784.
7. De Boodt S, et al. Bi-modular flow characterization in tissue engineering scaffolds using computational fluid dynamics and particle imaging velocimetry. Tissue Engineering Part C: Methods 2010;16(6):1553–1564.
8. Faria M, et al. Minimum information reporting in bio–nano experimental literature. Nature Nanotechnology 2018;13:777–785.
9. FDA. Assessing the credibility of computational modeling and simulation in medical device submissions. Guidance, 17 November 2023.
10. Giorgi M, Verbruggen SW, Lacroix D. In silico bone mechanobiology: modeling a multifaceted biological system. WIREs Systems Biology and Medicine 2016;8(6):485–505.
11. Haase M, et al. Agent-based model demonstrates the impact of nonlinear, complex interactions between cytokines on muscle regeneration. eLife 2024;13:RP91924.
12. Hunter P. Computational physiology and the Physiome Project. Keynote, Basel Computational Biology Conference, 2004.
13. Isaksson H, et al. Comparison of biophysical stimuli for mechano-regulation of tissue differentiation during fracture healing. Journal of Biomechanics 2006;39(8):1507–1516.
14. Jumper J, et al. Highly accurate protein structure prediction with AlphaFold. Nature 2021;596:583–589.
15. Kanda GN, et al. Robotic search for optimal cell culture in regenerative medicine. eLife 2022;11:e77007.
16. Khalil M, et al. Simulation and physical validation of metal triply periodic minimal surfaces-based scaffolds for bioengineering applications. Engineering Modelling, Analysis and Simulation 2024;2(1). doi:10.59972/fn8161af.
17. Kovatchev BP, et al. In silico preclinical trials: a proof of concept in closed-loop control of type 1 diabetes. Journal of Diabetes Science and Technology 2009;3(1):44–55.
18. Lovett M, Lee K, Edwards A, Kaplan DL. Vascularization strategies for tissue engineering. Tissue Engineering Part B: Reviews 2009;15(3):353–370.
19. Martín Compaired P, García-Gareta E, Pérez MÁ. An experimental workflow for bioprinting optimization: application to a custom-made biomaterial ink. International Journal of Bioprinting 2025;11(3):397–415.
20. Mehrian M, et al. Maximizing neotissue growth kinetics in a perfusion bioreactor: an in silico strategy using model reduction and Bayesian optimization. Biotechnology and Bioengineering 2018;115(3):617–629.
21. Möller J, Pörtner R. Digital twins for tissue culture techniques: concepts, expectations, and state of the art. Processes 2021;9(3):447.
22. Place TL, Domann FE, Case AJ. Limitations of oxygen delivery to cells in culture: an underappreciated problem in basic and translational research. Free Radical Biology and Medicine 2017;113:311–322.
23. Porter B, et al. 3-D computational modeling of media flow through scaffolds in a perfusion bioreactor. Journal of Biomechanics 2005;38(3):543–549.
24. Protein Structure Prediction Center. CASP: Critical Assessment of protein Structure Prediction. predictioncenter.org.
25. Rafieyan S, et al. MLATE V3: an open-source multi-tissue AI framework for data-driven optimization of 3D-printed and bioprinted scaffolds. Biofabrication 2026.
26. RCSB Protein Data Bank. rcsb.org, accessed 25 October 2026 (261,139 experimentally determined structures).
27. Royal Swedish Academy of Sciences. The Nobel Prize in Chemistry 2024. Press release, 9 October 2024.
28. Ruberu K, et al. Coupling machine learning with 3D bioprinting to fast track optimisation of extrusion printing. Applied Materials Today 2021;22:100914.
29. Sena ES, et al. Publication bias in reports of animal stroke studies leads to major overstatement of efficacy. PLoS Biology 2010;8(3):e1000344.
30. Siegel G, et al. Phenotype, donor age and gender affect function of human bone marrow-derived mesenchymal stromal cells. BMC Medicine 2013;11:146.
31. Slotnick J, et al. CFD Vision 2030 study: a path to revolutionary computational aerosciences. NASA/CR-2014-218178, 2014.
32. United States Congress. Consolidated Appropriations Act, 2023 (Public Law 117-328), Section 3209: animal testing alternatives. 29 December 2022.
33. Wagner BA, Venkataraman S, Buettner GR. The rate of oxygen utilization by cells. Free Radical Biology and Medicine 2011;51(3):700–712.
34. Wagner W, et al. Replicative senescence of mesenchymal stem cells: a continuous and organized process. PLoS ONE 2008;3(5):e2213.
35. Wilkinson MD, et al. The FAIR Guiding Principles for scientific data management and stewardship. Scientific Data 2016;3:160018.
36. Zhao F, van Rietbergen B, Ito K, Hofmann S. Flow rates in perfusion bioreactors to maximise mineralisation in bone tissue engineering in vitro. Journal of Biomechanics 2018;79:232–237.
37. Zhu M, et al. Gelatin methacryloyl and its hydrogels with an exceptional degree of controllability and batch-to-batch consistency. Scientific Reports 2019;9:6863.