How do you grow blood vessels in a lab?
Human tissue grown in a dish stops at about the size of a grain of rice, and the reason is diffusion.
Oxygen and nutrients reach cells by wandering in from the surrounding liquid. They make it roughly 150 microns — two hair-widths — before running out.1 Past that, cells starve. So an organoid doesn’t stop growing when it hits the limit; it reorganizes into a living rind around a dead middle, new cells appearing at the surface at the same rate cells in the centre die. Section most organoids past a certain size and you will find that dead middle, which is called a necrotic core.
Folkman and Hochberg measured the ceiling in 1973, growing cell clumps in soft agar with generous, frequently replaced medium:
[the spheroids] eventually reached a dormant phase at a diameter of approximately 3–4 mm and a population of approximately 10⁶ cells
Three to four millimetres. A million cells. That number has been sitting there since 1973, and nothing grown in a dish has beaten it without plumbing.
Core concepts
If you want to grow a liver, you need liver cells. A heart needs heart cells. A brain needs brain cells. That part is obvious, and it is the part everyone gets right.
The surprise, once you start doing this, is how far that is from sufficient. A real tissue is not a pile of its headline cell type. It is that cell type plus a handful of others nobody outside the field has heard of, and the uncomfortable thing we keep rediscovering is that you cannot drop any of them. Leave one out and the tissue doesn’t come out slightly worse. It comes out broken in a specific, reproducible way.
Here is the vocabulary. Everything after this section is written assuming you have it.
Endothelial cells are what a blood vessel is made of. Not a lining on the inside of a pipe — at the smallest scale there is no pipe. A capillary is a single endothelial cell, rolled into a tube and sealed to its own edge, with a hole down the middle just wide enough for a red blood cell to squeeze through, sometimes deforming as it goes. One cell. Slightly larger vessels use two or three arranged like staves in a barrel. That is the entire wall.
The hole has a name, the lumen, and it is the hard part. Endothelial cells will find each other and line up into long connected strings within about a day, in almost any conditions, with no encouragement at all. What they will not do on their own is make those strings hollow. Left alone they pack against each other and you get a solid cord of cells. Nothing flows through solid. Most of the work in this field is the work of persuading a string of cells to open up down the middle and stay open.
Pericytes are the second cell type, and the obvious question is why you need one. If a vessel is an endothelial cell wrapped around a hole, why not just make more endothelial cells?
Because a tube made only of endothelium takes itself apart. This is close to a law in culture: endothelial cells alone build a network, hold it four to seven days, and disassemble it. Naoto Koike’s 2004 paper is the founding demonstration — endothelial cells implanted alone regressed, and the same cells implanted alongside structural precursor cells produced vessels still working roughly a year later. Pericytes sit on the outside of the tube, wrapped around it, and they are the difference between a vessel and a temporary arrangement. They also set diameter: in one titration, vessels went from 120 microns wide with no pericytes to 33 microns with them. Endothelial cells and pericytes are not two options. They are a pair.
Fibroblasts are the general-purpose structural cell of connective tissue — the ones that build and maintain the scaffolding your organs sit on, and the ones that show up to make a scar when you cut yourself. They are not part of the vessel. They matter because they are loud. Fibroblasts secrete the signals that tell endothelial cells to grow toward something, and in most systems, without them, you get no network at all.
Two of those signals are worth knowing by name, because the rest of this is unreadable without them.
VEGF — vascular endothelial growth factor — is the one everybody knows. A cell short of oxygen releases VEGF, and VEGF means, approximately, grow a blood vessel toward me. It is the central currency of the field. Nearly every protocol adds it, most at 50 nanograms per millilitre, and a surprising fraction of what follows is people discovering that adding more of it does not do what you would expect.
HGF — hepatocyte growth factor — is the quieter one. Named for the liver cells it was discovered acting on, but it works as a general “come here and start moving” signal, and it is one of the main things fibroblasts are shouting. Fibroblasts express it at roughly fifty times the level pericytes do, which is much of why the two are not interchangeable.
Flow is the last one. Vessels in a body are constantly scraped by fluid moving past them, a force called shear stress, and endothelial cells can feel it. Shear is one of the signals telling a vessel to stay open and stay sealed.
I want to be careful here, because flow gets overstated. In kidney organoids it appears to be close to mandatory, for reasons I’ll get to. In cardiac and intestinal systems, people routinely keep endothelium alive in completely static culture with nothing but a growth factor left in the medium. Why the kidney should be so much more insistent about shear than the gut is not obvious to me — it is an organ whose entire job is filtering flow, so perhaps its cells are simply more attentive to it, but that is a story rather than an answer. Treat flow as one available lever, not a requirement.
Which is the frame I’d want anyone starting here to hold on to.
You are not engineering a vascular network. You are trying to convince cells to do something they already know how to do. Every cell above has the full program for building vasculature written into it, and has had for six hundred million years. Tissue needs blood vessels in order to get big; the cells know this; they are not withholding. The reason it isn’t happening in your dish is that you haven’t given them a reason — no partner cells, no gradient, nothing moving. Every technique below is best read as a different answer to the question of what would count as a good enough reason.
The four things you can do
The literature looks like forty unrelated techniques with forty acronyms. It isn’t. There is one question — where does the endothelium come from? — and four answers.
MAKE: it was already there and we lost it
Grow kidney organoids with any of the standard protocols, stain them around day seven, and you find something peculiar: there are already blood vessel cells inside. Nobody put them there. They arose from the same starting stem cells as everything else, because that is what happens in an embryo — tissue and vasculature develop together out of a shared pool of precursors. This isn’t a subtle or contested finding; it shows up across the standard kidney protocols and people have been reporting it for years.
Then stain the same organoids at day twenty-five and they’re gone. If you hadn’t looked at day seven you would never have known they were there.
That reframes the problem completely, and it gets glossed over more than it should. This is not a construction problem. It is a retention problem. The endothelium showed up on schedule, looked around, found no pericytes and nothing moving and no reason to be a blood vessel, and disassembled itself. Every hour spent working out how to add vessels to a kidney organoid is an hour not spent asking why the ones it already made left.
Kimberly Homan’s 2019 experiment is the instructive one, because of what she did not add. No endothelial cells. Kidney organoids went into a chip and had fluid pushed over them, hard, for ten days. The vessel cells stayed. Vessel area went up roughly five-fold and vessel length roughly ten-fold against a low-flow control, and endothelium invaded the kidney’s filtering structures in over 60% of cases against 10–20% without flow.
The lever was mechanical. Given a reason to be a blood vessel, they remained blood vessels. This is also the clearest answer to why the kidney is so insistent about flow, and it’s the reason I’d point anyone working on kidney at a chip before anything else.
The same paper contains the single most useful counterintuitive result I know of in this area, and it concerns VEGF.
They added VEGF at high dose, expecting more vessels, and got fewer. So they went the other way and added a drug that blocks VEGF — bevacizumab, which you may know as Avastin — and that made things worse too. Both directions were worse than leaving it alone.
The organoid was already producing its own VEGF, in a gradient: more in some places, less in others. The cells were navigating by the slope. Flooding the dish flattens the slope. Blocking it flattens the slope. The information the cells were using was never in the concentration, and both interventions destroyed it. If you take one thing from this essay into your own protocol, take this one — it generalizes well beyond kidney, and it is the most common way I see a reasonable-looking experiment fail.
What MAKE buys you is identity, and this is not a technicality. Blood vessels are not interchangeable. Kidney endothelium is perforated with windows for filtering. Liver endothelium is leaky in a different way. Brain endothelium is welded shut tightly enough to form the blood-brain barrier, which is why most drugs cannot get into your head. Those differences are learned during development, by being present while the surrounding organ is deciding what it is. A vessel cell that grew up alongside kidney tissue is a kidney vessel cell. A bought one is not, and cannot become one.
What it costs is quantity. These populations run somewhere between 0.6% and 4% of all cells, which is not much to build a network from.
MIX: put them on the inside
Here is the requirement that separates methods that work from methods that photograph well.
Vessels have to end up inside the tissue. Vessels wrapped around the outside of an organoid are close to worthless — the cells in the middle are still 400 microns from the nearest lumen, still starving, still dying on schedule. It is entirely possible to produce a gorgeous, dense, publishable vascular network that does nothing whatsoever for the thing it is wrapped around. Interior placement is the whole game, and most of the difficulty in this field is that sprouts arriving from outside tend to stop at the surface.
MIX solves it by not letting the vessels arrive from outside at all. They start in the middle, because everything starts mixed together.
Coat a plate with Matrigel — the mouse-tumour-derived protein gel most organoids are grown in — and let it set. Deliberately soft; the softness turns out to matter enormously.
Take three tubes of counted cells: liver precursor cells, endothelial cells, and mesenchymal stem cells, a general-purpose structural cell from bone marrow. Mix them, pipette the mixture onto the gel, go home.
The ratio is 10 : 7 : 2. Ten parts tissue, seven parts vessel, two parts structural, from Takashi Takebe’s 2013 Nature paper. It survived into the later all-stem-cell version, where it was re-derived by screening combinations rather than inherited. Note how much endothelium that is — nearly a third of the mixture.
Come back the next morning and the flat lawn of cells has pulled itself into a lump. By 48 hours it’s a three-dimensional bud you can pick up with forceps. Nobody assembled it.
The obvious explanation is wrong, and this is the most useful piece of physical intuition in the area. The cells are not crawling toward each other. The Takebe lab measured the edge of the condensing lump moving at about 1,200 microns per hour — far faster than any of these cells can migrate — and measured the stiffness gradient of the gel as far too shallow to be steering anything. The mesenchymal cells aren’t crawling, they’re contracting. The sheet squeezes shut like a net being pulled closed, and the endothelial cells are passengers, carried inward as the tissue compacts.
That passenger status is the entire value of the method. You get interior placement overnight, as a side effect of a mechanical process you didn’t have to design.
The cost is the mirror image of MAKE’s benefit. The endothelial cells in most of these protocols, including the original, are HUVECs — human umbilical vein endothelial cells, harvested from donated umbilical cords. They are vein cells from a newborn being asked to stand in for capillaries in an adult liver.
There are alternatives, and the tradeoff isn’t simply price. Rodent endothelial cells are cheaper than HUVECs and grow faster; they’re just not human, which rules them out for anything you intend to translate. Endothelial cells derived from induced pluripotent stem cells are the other direction — you can make them from the same donor as the tissue, which is a real advantage, and they are slower and more expensive to produce. HUVECs dominate because they sit at a particular local optimum: human, cheap, robust, and available in a vial tomorrow. Everyone knows what they are giving up.
MEET: build them separately, introduce them later
Grow the tissue and the vasculature in two different dishes, under two different sets of conditions, each optimized for what it is, and only put them together at the end.
The argument is real: much of the difficulty in co-culture comes from the two cell types wanting different food, and keeping them apart through the fussy developmental period means neither has to compromise while it’s still deciding what to be.
We do this two ways. We can grow a blood vessel organoid — a ball of pure vasculature, endothelium and pericytes and basement membrane and nothing else, following Reiner Wimmer’s 2019 protocol — and then fuse it to a tissue organoid by placing the two touching in a single drop of gel and waiting a day or two for them to become one object. Or we do it on a chip: three parallel channels, the middle filled with gel, the outer two lined with endothelial cells. We drop an organoid into a well in the middle channel and watch, daily, as vessels sprout out of the side channels toward it.
They do sprout toward it. We see first contact around day three and most of the sprouts across by day seven. In the good cases we can push a fluorescent tracer into one channel and watch it come out the other having passed through the organoid.
And then there’s the failure mode, which is where the interior-placement problem shows up in its purest form.
The vessels arrive at the organoid. They wrap around it. They form a dense, photogenic halo hugging the outside surface. And they stop at the surface. Nothing goes in.
Mia Rambøl’s study on pancreatic islets pinned this down carefully. She measured vessel coverage in two places: inside the islet region, and across the rest of the chip. Coverage in the islet region went up — from about 12% to 20% — while coverage elsewhere on the chip didn’t change. So the islets genuinely were attracting vessels; that part of the biology worked exactly as intended. But when she looked at where those vessels actually were, they were all in the shell around each islet. There were essentially none inside. The 12-to-20% rise was entirely a thickening halo.
She also tried the obvious fix of adding more endothelial cells, and it didn’t help: the extra cells just made bigger clumps on the outside. Her summary line is one of the more bracing things I’ve read in a methods paper: “The successful generation of a vascularized and perfused islet-on-a-chip model has not been seen to date.”
The halo happens because sprouting is shout-and-follow, and all the shouting is coming from the surface. Yuji Nashimoto’s group fixed it about as directly as possible: put the shouting cells inside. They build the spheroid with fibroblasts in it, so the organoid itself becomes the loudest thing in the dish and the sprouts have a reason to keep going once they arrive. Fibroblasts outside gets you a halo. Fibroblasts inside gets you penetration.
The limitation, which the authors state themselves, is that this only works if your organoid can accommodate fibroblasts. If it can’t, you’re back to a halo and you need some other source of shouting in the middle. Nobody has a general way to do that.
BORROW: stop building a circulation and go find one
Put the organoid into a living animal and let the animal plumb it.
What makes this interesting isn’t that it works. It’s how much the timing depends on what you hand the animal.
Implant a naked organoid with no vessels of its own, and the host has to grow entirely new vessels into it from scratch. That process is angiogenesis, and it’s slow. In Abed Mansour’s 2018 study, brain organoids implanted into mouse cortex showed vessels beginning to invade at 7–10 days, with the earliest documented blood flow at 30 days.
Implant an organoid that already has vessels in it — a MIX bud — and the host’s vessels find yours and splice into them end to end. That process is inosculation, and it’s fast. Takebe’s liver buds connected to the host circulation in 48 hours.
Same class of surgical site. Two days against a month. The cleanest isolation of the variable is Xin Chen’s 2009 study using plain fibrin constructs, where the pre-vascularized version had red-cell-carrying vessels at day five and the identical non-vascularized control took until day fourteen.
Inosculation is plugging in a cable. Angiogenesis is laying one. During the month it takes to lay one, the middle of your tissue is dying.
Notice this describes two verbs used together, which is the real takeaway. The four aren’t competitors. Nearly every strong result in the field is MIX then BORROW, or MAKE then BORROW: get vessels into the tissue while it’s in a dish, then hand it to something with a heart.
Three things that will trip you up
”Perfusable” means four different things
Read this literature for a week and you’ll meet the word several hundred times, doing duty for four experiments that differ enormously in what they demonstrate.
Tier A. You add fluorescent dye to the liquid the organoid sits in and later find dye inside the vessels. The tubes are open to the outside. Nothing flowed.
Tier B. You inject under pressure, or set up a height difference so liquid is pushed through. Now the tubes are continuous and hold pressure.
Tier C. Sustained flow driven by a pump or a pressure head, with measurable velocities.
Tier D. Host blood, red cells, arterial pressure, in an animal.
All four get called perfusion, and a 2026 review in Regenerative Therapy notes there’s no standardised benchmark for the term. So supply one yourself: find the actual figure in the paper and assign it a tier. It takes ten seconds and it changes your read of a good third of the literature.
The important corollary is that you very often don’t need Tier D, or even Tier C. The reason to build vasculature is usually to stop the middle of your tissue from dying and to make the model resemble the thing you’re modelling. Open lumens carrying oxygen and nutrients into the interior accomplish that. Sustained pumped circulation is a much harder engineering problem that a lot of experiments do not actually require, and I’ve watched people burn quarters chasing it because a reviewer used the word “perfused” without specifying which kind.
The vessels come out too fat
A real capillary is under 10 microns across, because that’s set by the size of a red blood cell. Self-assembled networks in a dish come out at 25 to 130 microns — three to thirteen times too wide. They’re small veins being asked to model capillaries.
There’s a genuine tension underneath this. Zhengpeng Wan’s 2022 study mapped the parameter space and found that within a plain self-assembly protocol, the two things you want pull against each other: push the conditions toward narrow and you get narrow vessels, around 21 microns, that nothing will flow through; push toward flow and you get 74 microns.
I want to be careful not to overstate that into an impossibility claim, because it isn’t one. It is a real constraint on unassisted self-assembly, and getting both usually means adding something the plain protocol doesn’t have — staged seeding, templating, matrix engineering, mural cell titration. Several groups, including ours, have gotten further on this than the self-assembly literature would suggest. The honest framing is that vessel calibre is tunable, that the naive protocol makes you choose, and that if you need capillary-scale vessels that actually carry something you should expect to do work beyond mixing cells in a gel.
The food fight
Cells are fed a defined cocktail called medium, and every cell type has a preferred recipe. The recipe that keeps endothelial cells happy and the recipe that makes your organoid develop correctly are frequently in direct opposition.
One group working on kidney organoids reported that standard endothelial growth medium abolished nephron formation entirely. Not reduced — the tissue they were trying to build did not form. A group working on pancreatic islets found the opposite: endothelial medium cut cell death about six-fold compared to standard islet medium.
There’s no general answer, only a titration you have to run yourself. The highest-value cheap control in this whole area is to grow your organoid alone, in the co-culture medium — not the co-culture in organoid medium. That single arm catches the failure mode where the vessels look great and the tissue has been quietly ruined.
There’s a further layer of absurdity. “Endothelial cell medium” is not one thing. Pull the published compositions from five major suppliers and VEGF varies about ten-fold between them. Vitamin C varies seventy-five-fold. One supplier’s heparin content is zero and another’s is 90 micrograms per millilitre. “Cells were maintained in endothelial cell medium” carries substantially less information than it appears to.
So where does this leave us
The achieved list is not nothing, and I think the field undersells it.
We can make human endothelial cells in quantity from stem cells. We can get them to build connected networks with genuinely hollow lumens, confirmed by electron microscopy rather than hopeful staining. We can get pericytes to wrap them and a proper basement membrane laid down. We can get those networks to splice into a living animal’s circulation and carry its blood, sometimes within two days. We can grow a ball of pure human vasculature and use it to model diabetic vessel damage well enough that it reproduces a specific molecular mechanism and responds to a specific drug — a disease model, not just plumbing. And with printed channels we can hold tissue alive at near-native cell density in pieces measured in cubic millimetres rather than fractions of one.
What nobody has done is put all of it into one system at once: right size, right leakiness, right shear, self-organizing, stable for months, in a dish, with something pumping. Every piece has been demonstrated. Nobody has the full set.
I think the reason is mostly timing rather than technique. In an embryo the tissue and its blood supply are the same age. They differentiate side by side, trading signals the whole way, and the vasculature learns what kind of vasculature to be by being present while the organ decides what kind of organ to be. The plumbing isn’t installed in the building; they’re co-authored, over months. What we do instead is introduce them as adults and hope they get along — a vein cell from a newborn’s cord, a clump of liver-ish tissue, a dish, a protein that means grow. Sometimes it works.
That framing predicts the tradeoffs cleanly. MAKE keeps identity because it preserves the shared childhood. MIX gets quantity and interior placement because it gives that childhood up in exchange for physically putting cells where they need to be. MEET protects both parties during development at the cost of a difficult introduction. BORROW works because you’ve handed the problem to a system that’s been solving it for six hundred million years.
But I want to end on the thing I think matters more than any of it, which is that the full set is the wrong target.
Nobody is trying to build a human. If you were, you would indeed need all of it at once, and you would be attempting something absurd. An organoid is a model of part of a human, built to answer a specific question. George Box put the general version of this better than anyone has since: “essentially, all models are wrong, but some are useful.” He wrote “all models are wrong” in a 1976 paper and the full line appeared in his 1987 book with Norman Draper. It is not a hedge. It is the entire operating principle of this field.
So the question is never “is my organoid complete.” It’s which parts can I leave out and still get a trustworthy answer to the thing I’m asking?
Sometimes that answer is: all of them. If your drug acts on a target inside a single cell and you’re measuring whether that cell lives or dies, you may need no vasculature whatsoever, and adding it would be an expensive way to make your assay noisier.
Sometimes you need vessels but not flow. If the problem is that your tissue is big enough to have a dying middle, open lumens reaching the interior solve it, and a pump adds nothing.
Sometimes you need the vessels themselves, because they’re the subject. Vascular disease, barrier permeability, anything about how a drug crosses out of blood into tissue — here the vasculature isn’t infrastructure, it’s the experiment, and it has to be right.
And sometimes you need the whole thing, because you’re building tissue to put into a person, and a person is not a model.
Most people I talk to about this are, without quite noticing, trying to answer a question in the first or second category using a system designed for the fourth. The useful discipline is to write down what you’re actually asking before you decide how much plumbing to build. The field’s hardest problem is a physical constant and its reference implementation took evolution most of a geological era to debug. There is no reason to solve more of it than your question requires.
Corrections welcome. Several numbers here disagree with numbers you’ll find in the review literature; in each case I went to the primary source, and said so where I couldn’t get there.
Footnotes
-
The micron figures in this area are squishier than they’re usually presented, and I went looking. You’ll often see “100–200 microns” credited to Carmeliet and Jain’s 2000 Nature review; I could not verify that sentence in that paper. You’ll also see “organoids are limited to 400 microns” credited to Folkman, whose actual published number is 3–4 millimetres, and I could not find a primary source for 400 microns anywhere. The figure I’d stand behind is ~150 microns, from Grebenyuk and Ranga’s 2019 review in Frontiers in Bioengineering and Biotechnology, with 150–200 microns as the working range other groups quote. None of this changes the argument. It does mean confident micron figures in this corner of the literature are worth a click. ↩