A Year to Grow a Brain Organoid, an Hour to Simulate It

Siyu He is a postdoc in biomedical data science at Stanford and lead author of two recent papers, Squidiff and CORAL, both supervised in a group that also produced James Zou's Virtual Lab work. She tells Nathan Labenz, host of *The Cognitive Revolution*, that the point of both projects is the same: use AI on current biological measurements — single-cell RNA sequencing in the first paper, spatial transcriptomics in the second — to understand cellular systems, disease mechanisms, and responses to drugs. The grand challenge Labenz sketches is causal and multi-scale: sequence to protein shape is getting better, but how that adds up to a cell, a tissue, and an organism is still a long way off. He's watching this layer because that is where the frontier sits.

Squidiff's name is a running joke with a method underneath. He's first model paper was Starfish; he wanted a family of sea-animal names, and a diffusion model suggested squid — also a backronym for stimuli-related quantitative inference of transcriptomics. CORAL is the sequel, another marine name. Labenz calls it a cinematic universe in the making. He hopes to collect more. The science the names sit on is less cute. Single-cell sequencing is destructive. You cannot watch the same cell tomorrow. You grow a new batch and wait.

That wait is the motivation. He works as a wet-lab person as well as a modeler, and getting from culture to a sequencing file is painful: cells have to grow, the pipeline takes months, it is expensive, and some experiments cannot be run at all. Organoids — engineered tissues from induced pluripotent stem cells, brain or vessel, used for disease modeling — often take months. Some brain organoids need a year or more. A mistake means starting over. A generative model that can spin a digital transcriptome, and then manipulate the conditions that produced it, is how he wants to take that task off the bench.

Almost every cell in the human body shares the same DNA. What makes them different is transcription: DNA to RNA, RNA to protein, protein folding into the structures AlphaFold tries to predict. Gene expression — which genes are being transcribed, and how much — is the readout of that variance. One of He's postdoc advisors, Stephen Quake, put it in a line the field repeats: the cell is a bag of RNA. Complexity can be read from gene expression. Protein expression can distinguish cells too, but current protein assays are noisier, so most people start with the transcriptome.

Single-cell RNA sequencing is the technology that made that bag countable. Tissue is dissociated into individual cells, and each cell is profiled for on the order of 30,000 to 60,000 genes. The starting object for everything that follows is a vector of that many intensities, assembled into a cell-by-gene matrix that can run to thousands, millions, even billions of cells depending on the project. It is high-dimensional, high-throughput, and the reason quantitative biology and AI models suddenly have something to chew on. The transcriptome is not a complete story. Epigenomics, chromosome state, protein, mitochondria, and other layers differ across cells. It is, however, the data type you can currently get at whole-genome scale, which makes it the highest-leverage place to start.

Once you have a transcriptome, measured or generated, the usual analysis is the same. Dimension reduction and clustering group similar states; each cluster typically corresponds to a cell type. Marker genes — labels the community already agrees on — tell you what the cluster is. Cancer-related markers showing up is how you call a cancerous stage, the molecular version of a pathologist staining tissue with antibodies. Co-expression of similar gene groups is what lets that aggregation happen. The 30,000 numbers are not the clinical sentence; they are the evidence you compress into one.

Squidiff is not shipped as a pretrained model. Users bring their own data. About 5,000 cells is enough to train in roughly 15 minutes, which is also about where underfitting or overfitting starts if you go much smaller. He used small, specific datasets rather than a grab-bag of everything, with metadata for cell type, tissue, and disease stage, mixing public releases with new data from collaborating wet labs. High-quality, consistently annotated data is one of the field's main bottlenecks. The next step he wants is a foundation model that pools organoid types, then real human tissue.

An early manuscript called the compressor a VAE. He corrects it on the air: it is a semantic encoder wired to a conditional DDIM. Two reasons he did not reach for a Transformer, even though that would be Labenz's first instinct for "next transcriptome given a perturbation." Transformers in this area usually want a sequence, which in practice means ranking genes, so both input and output become ranks instead of the original continuous counts; most foundation models then use the embedding for classification, which is a different job from generating a full transcriptome. Diffusion models are more flexible on complex distributions, and gene-expression distributions are extremely complex. When he started, nobody had really used diffusion to generate single-cell transcriptomes. That blank space is what he wanted to test.

The important training difference is not the name of the encoder. Latent-diffusion recipes usually train a VAE first for dimension reduction, then attach a denoiser. Squidiff trains a unified model, so semantic extraction and noise learning share one loss. Labenz notes that this is all end to end, unlike the usual mashup. The denoising network is an MLP with residual connections that fold in timestep and semantic features — not a U-Net — because images are two-dimensional and a cell is a one-dimensional vector. Expression matrices are sparse, full of zeros, so the team keeps highly variable genes and drops the rest. The noising process itself is simple, not the specialized corruption protein folders need. Semantic features are a unified condition: cell type, disease stage, experimental condition, manipulable in latent space rather than gene space.

The in-silico experiment is word2vec arithmetic. To ask what stimulus Y would do to hard-to-culture cell type X, run or reuse an experiment on easy cell type Z, subtract the semantic embeddings of treated and untreated Z, and add that delta to X. The diffusion model then paints a full transcriptome for a cell type you never touched. He agrees that is exactly the mechanism, and also that it is an assumption. Biological processes are complex and usually not linear. Latent variables are more linear, smoother, and more structured than gene-expression space, so a linear approximation can stand there while the diffusion model absorbs the remaining nonlinearity. Relying on the output without validation is still risky.

Validation comes in three designs, each holding out a known answer. For differentiation, real data run from day-0 iPSCs to day-3 mesoderm; training sees only the endpoints, and days 1 and 2 are the test set, generated by interpolating the semantic vectors and scored against the real expression. For gene perturbation, the team has experiments on gene A, gene B, and A plus B, then holds out the combination and asks whether vector addition predicts the joint effect. For vascular organoids cultured from day 0 to day 11, you cannot sample in the middle without destroying the sample, so they measure only the ends and interpolate. They found transient cell states that had not been reported, then checked public time-series data from other labs and saw consistent findings. That external match, He says, made the model's ability to find new things much stronger.

Labenz maps the architecture onto MindEye, the fMRI-to-image project that also built a diffusion prior and then steered it. The lineage is the same one that, in 2021, had the Twitter account RiversHaveWings generating bizarre art, and that a couple of years later makes photographs hard to tell from the real thing. Project that curve onto biology, he argues, and you are looking at a paradigm shift: simulations accurate enough to upend how the field queues wet-lab work. On cost, dissociation to a sequencing file takes at least a week; a new condition in the model is about an hour. Graduate students, postdocs, and bench space are roughly fixed, so progress is whether you ran the right experiment. A cancer patient's cells are the sharper version of the same bet: better to simulate which drug is likely to work, and what the side effects look like, than to try at random.

The cases they have tested so far work, He says; they have not seen a clean breakdown of the linear trick. What they have seen is that intermediate-state predictions are weaker than the endpoints. Differentiation is driven by several growth factors, each a different direction, and the current model packs them into one approximate straight line. If each factor had its own learned semantic vector, the trajectory would be less of a cartoon. A next version is already planned: borrow from video diffusion, train on time-series data, and learn the developmental nonlinearity instead of interpolating it.

The sharper limit is generalization across drugs. If the model has seen drug A and drug B, those vectors can be transferred onto any cell type. An unseen drug C has no semantic vector, so the model cannot say anything. Reviewers asked about this. The planned variant adds an adapter that encodes the drug's molecular structure and dosage and concatenates that embedding with the cell's semantic variable. Then even an untrained drug C can be queried if you have the compound and the dose. Organoids as drug-screening platforms are the obvious customer for that upgrade: vectors per growth factor to predict development, plus response to drug components.

Labenz asks whether the next steps — scale, a separate encoder for interventions, time series — are a simple matter of programming, the software-engineer phrase for "we will hit snags but we will get there." He says it depends on the community, not just this project: biologists for mechanisms, machine-learning people for models, statisticians, even artists to make the work land. Because the application is healthcare, accuracy is a responsibility to patients. A lot of experimental and clinical testing has to happen before anything is clinically available. The lab is where you can be more flexible. Do not trust the model totally, he says earlier; use it for intuition about which way a project should point.

CORAL attacks a different scale gap. Single-cell RNA-seq's obvious shortage is that dissociation throws away location. Cells are not loners; neighbors and communication are how tissue texture happens. Spatial omics, about seven years old and now producing more data than analysis tools, measures on the slice instead. The catch is a resolution mismatch He analogizes to photography: one grayscale high-resolution portrait, one colorful mosaic taken at a slightly different moment. Can you make a colorful high-resolution image? Spatial transcriptomics is the mosaic — whole-genome 30,000 to 60,000 genes, low spatial resolution. Spatial proteomics is the sharp grayscale — high spatial precision, only a limited number of proteins. You also cannot run both assays on the same slice. Adjacent sections shift, like two photos taken from slightly different spots.

Deconvolution is the main task: pull a mixed low-resolution spot back into single-cell transcriptomes. Around that, CORAL integrates the two modalities, models cell–cell interaction with a graph neural network, organizes spatial regions into functional domains, and looks at spatial variability inside the tissue. Neighbor count K is a hyperparameter; He says there is no need to make the graph too long, because the point is local communication. Labenz hears a smoothness prior in the graph-Laplacian loss and worries about bones, which do not fade into muscle. He answers that it depends on tissue type. Bone and blood vessels have clear boundaries; tumors often do not, because cancer cells infiltrate normal tissue. Molecular patterns can show structure morphology misses.

The loss is a balance. Graph Laplacian regularizes toward smoothness; reconstruction and KL divergence pull toward the data. Single-cell-level measurements are sharp enough that bone and vessel boundaries stay crisp, and you even see inner structure of the organ. Labenz puts the whole problem in a larger frame: science keeps failing at cross-scale integration, from quantum mechanics to the split between micro and macroeconomics, and biology may be the hardest setting of that exam. CORAL is one joint. On synthetic data, the worry is circularity — if you can generate the data, do you already know the answer? Squidiff used Splatter; CORAL uses a collaborator's simulator, now a professor at the University of Connecticut, published in *Nature Biotechnology*, which samples designed spatial expression. Ground truth for domains and interactions exists in the simulation, which is generated by a process independent of CORAL, so there is no loop. Validate there, then train separately on real tissue and check with domain knowledge, because real tissue has no labels.

Labenz is less excited, in this conversation, about a fuzzy Manhattan Project for superintelligence than about unlocking the data these models starve for. A lot of what they need sits in medical-record systems behind privacy rules that have a point and also a cost. If the most compelling promise of superintelligence is curing disease and longer lives, he asks, why not subsidize the data so people like He can make the discoveries directly? He points at the projects already running: Google DeepMind's virtual-cell effort, aiming at a unified model across cell types, organs, and scales; the Chan Zuckerberg Initiative's Billion Cells program to gather enough cells for a similar foundation model; digital twins that would let you try a choice on a copy of yourself. People are building biological versions of GPT and virtual doctors. It is an exciting one-to-two-year window, he says, and a little scary — but this kind of focused work feels like pure upside.

He does not think Squidiff or CORAL is how AI gets out of control. They answer specific questions. Collaboration is the constraint he keeps returning to, not cluster size: biologists, ML people, statisticians, even artists. He is optimistic that AI will make healthcare development faster and lives better. At least for now, he does not see a path where this work destroys the world. Labenz's closing reminder is the bitter lesson in reverse. Not every useful system is a bigger Transformer on a bigger cluster. Carefully designed architectures still do things frontier language models are not being asked to do, and both of these papers are evidence.