Unlocking Cells' Secrets: Diffusion, Deconvolution, & Discovery with Siyu He of Squidiff & CORAL
Summary
Squidiff turns a cell’s 30,000–60,000-gene expression state into a software object that researchers can perturb before committing scarce wet-lab capacity. The system jointly trains a semantic encoder and a DDIM diffusion model; Siyu He explicitly corrects the original description—“it’s not actually the VAE”—and reports that a roughly 5,000-cell dataset can train in about 15 minutes. The architecture illustrates how techniques developed for image generation can transfer into continuous biological state modeling.
The near-term return is better experiment selection, not autonomous biology. Some organoids take months to culture, with brain organoids requiring “even a year or more,” while obtaining single-cell RNA data can take at least a week; Squidiff might generate another condition in roughly an hour. Siyu He cautions against trusting the model completely: its value is triaging growth factors, drugs, and gene perturbations so the fixed wet-lab pipeline runs more informative experiments.
Squidiff’s load-bearing trick is semantic vector arithmetic, and its load-bearing risk is that biology is not linear. A perturbation direction learned by comparing baseline and perturbed cells in one cell type can be transferred to another, much like the familiar man-to-woman embedding direction; diffusion then expands the manipulated latent state into a full transcriptome. Siyu He calls this an approximation: the latent space might be smoother and more linear than gene-expression space, but multiple growth factors and directions can lead to the same final state.
The reported validation is stronger than a plausibility demo but still exposes where the model needs more information. The team withheld day-one and day-two states from a day-zero-to-day-three differentiation series, withheld combined gene perturbations while training on their components, and compared inferred organoid intermediates with public time-series findings. The tested cases generally worked, yet intermediate-state predictions were “not as good as we thought,” preserving an important distinction between useful signal and ground truth.
Generalization to unseen drugs requires conditions to become first-class model inputs. Base Squidiff cannot predict drug C if training contains only drugs A and B because it has no semantic vector for C; a proposed variant adds an adapter encoding the drug’s molecular structure and dosage. A separate video-generation-like direction would learn full time series rather than drawing a straight line between endpoints.
CORAL addresses a complementary bottleneck: molecular richness and spatial resolution usually arrive in different, misaligned measurements. Single-cell RNA sequencing destroys tissue location; spatial transcriptomics can cover tens of thousands of genes but at lower spatial resolution, while spatial proteomics offers higher spatial resolution across far fewer proteins—and the measurements may come from adjacent, shifted slices. CORAL aims to combine them into the biological equivalent of a “colorful high-resolution image.”
CORAL’s graph architecture deconvolves coarse tissue measurements into cell-level states while modeling neighborhood effects and boundaries. A graph neural network models each cell and a configurable number of nearest neighbors, enabling cell-cell interaction analysis, functional-domain discovery, and spatial-variability mapping. Reconstruction and KL-divergence objectives support fidelity, while a graph smoothness term regularizes local smoothness; the balance matters because tumors may infiltrate gradually, whereas bone and blood-vessel boundaries can be sharp.
Data quality and validation infrastructure remain the strategic bottlenecks behind any virtual-cell platform. Synthetic data from established simulators supplies known ground truth for benchmarking, and the results can then be compared with real tissue data and biological knowledge; Siyu He insists there is “no loop.” Siyu He cites virtual-cell, Billion Cells, biological-model, and digital-twin efforts and looks ahead to their development over the next one to two years, but says progress and clinical credibility require biologists, machine-learning researchers, statisticians, high-quality data, experiments, and clinical testing working together.
Deep dive
1. The transcriptome is a powerful state readout, not a complete cell
Siyu He begins with biology’s central dogma: cells largely share the same DNA, but transcription converts different portions of that DNA into RNA, which is then translated into proteins with distinct structures and functions. That makes gene expression a major source of observable cellular variation.
One of Siyu He’s postdoctoral advisers, Stephen Quake, supplied the memorable shorthand: “The cell is a bag of RNA.” Single-cell RNA sequencing turns that bag into a vector containing expression measurements for roughly 30,000–60,000 genes, giving quantitative biology an unusually rich, genome-scale starting point.
Researchers interpret those vectors through dimensionality reduction, manifold methods, and clustering. Cells with similar co-expression programs group into states or types; highly characteristic marker genes then connect those clusters to biological interpretations, including whether a population exhibits known cancer-related signatures.
Siyu He keeps the scope bounded: transcriptomics omits information in epigenomic and chromosomal changes, protein abundance, mitochondria, and other cellular processes. It is attractive because it offers whole-genome coverage with current technology, not because it is the whole story of a cell.
2. Destructive, slow experiments create the opening for virtual transcriptomes
Single-cell sequencing begins by dissociating tissue into individual cells, so obtaining the molecular readout destroys the sample and removes its spatial organization. Researchers cannot repeatedly measure the same sample through every intermediate stage; collecting a sample for sequencing destroys that sample.
Siyu He’s practical motivation came from wet-lab experience. Culturing an organoid can take months, some brain organoids take “even a year or more,” and an error may force the experiment to restart. Sequencing is expensive and, from dissociation through usable RNA files, can require at least a week.
Squidiff asks whether generative AI can create “virtual or digital transcriptomes” corresponding to chemical stimuli, gene up- or down-regulation, differentiation, or other conditions. It could also examine states that experiments cannot readily capture at all.
Labenz’s framing makes the resource constraint explicit: scientific progress depends partly on whether finite lab space, trained personnel, and culture time are allocated to the right experiments. Siyu He’s answer is not to eliminate measurement, but to supply fast predictions and useful intuition that steer costly experiments toward more informative conditions.
3. Generated expression states remain useful because existing biology can interpret them
Squidiff’s final output is a synthetic transcriptome with the same broad form as experimental single-cell RNA data. Researchers can therefore apply familiar analyses—cell annotation, clustering, marker-gene inspection, pathway analysis, and comparisons between conditions—rather than inventing an entirely new interpretation layer.
Similar states should express similar groups of genes, enabling a predicted cell to be placed among known populations. Marker genes recognized by the biological community then provide the bridge from a 30,000-dimensional output to questions such as which cell type emerged, which signaling pathways activated, or whether a cancerous state persisted.
That interpretability is conditional rather than automatic. A generated vector is valuable when its predicted marker programs and pathway relationships survive comparison with withheld experiments or external biological findings; a transcriptome that merely appears statistically realistic is not sufficient evidence of a biological mechanism.
4. Squidiff starts with modest, task-specific datasets rather than a universal model
The underlying data are cell-by-gene matrices supplemented by metadata: annotated cell types, tissue of origin, disease stage, and other labels or experimental conditions. Siyu He draws on publicly released community datasets and collaborations with wet labs that can produce new samples or perform targeted validation.
Data volume varies dramatically by project—from thousands of cells toward millions or billions across the field—but current Squidiff instances train on specific subsets. Siyu He reports that about 5,000 cells can train in roughly 15 minutes, while warning that too few cells can leave a model underfit or overfit.
A major challenge is data quality and missing information, despite public sharing. Public datasets may omit conditions or metadata needed to align experiments, making it hard to develop consistent models.
Siyu He’s scale-up path is to aggregate related systems: a foundation model spanning many organoids, or a separate model for real human tissues. The current contribution is a flexible method that researchers train on their own relevant data, not a finished universal checkpoint.
5. The architecture is a jointly trained semantic encoder plus DDIM
Labenz’s natural challenge is “Why not a Transformer?” Siyu He explains that many biomedical Transformers treat genes as ranked sequences and learn embeddings for classification or prediction, whereas Squidiff directly models continuous expression values and aims to generate transcriptomes from a complex distribution.
Siyu He also corrects a consequential mistake in the original manuscript and the host’s setup: “It’s not actually the VAE; it’s a semantic encoder.” Squidiff’s two main components are that encoder and a DDIM diffusion model, with the encoded semantic variable conditioning a denoising process that begins from Gaussian noise.
Unlike latent-diffusion systems that train an autoencoder separately for dimensionality reduction and later connect it to a generator, Squidiff trains its encoder and diffusion process together. The process learns both noise and semantic information while generating realistic expression data for individual cells.
The semantic representation is intended to combine more than a cell-type label. Disease stage, experimental conditions, and other related information can occupy the same latent space, allowing researchers to manipulate conditions without directly editing tens of thousands of gene measurements.
6. Transcriptomes require different diffusion mechanics from images
Image diffusion operates over two spatial dimensions and commonly uses U-Nets. A Squidiff sample is instead a one-dimensional vector of gene-expression values, so the denoiser uses a multilayer perceptron with residual connections that incorporate diffusion time and semantic conditioning.
The noising procedure itself is comparatively simple, but single-cell expression matrices are sparse, with many zeros. The team filters genes that are not useful and focuses modeling capacity on highly variable genes rather than treating every measured gene as equally informative.
Siyu He’s reason for choosing diffusion is its ability to represent complex distributions. Gene-expression data can have complex patterns, and the diffusion model has the capacity to represent them.
Labenz sees a broader architectural lesson: procedures popularized through image and video generation are finding direct applications in biology. He regards this as another example of AI systems helping with problems where humans struggle to reason over the full amount of information.
7. Semantic arithmetic transfers perturbations across cell types
The core in-silico experiment begins with a contrast. If researchers possess baseline and perturbed examples of cell type Z, the encoder can estimate a semantic direction for that treatment; applying the direction to cell type X asks how X might respond without first culturing and treating X.
Labenz compares the operation to embedding arithmetic in which the vector from man to woman can be added to king to approach queen. Here, subtraction estimates the effect of a stimulus and addition transfers it to another starting state; the conditioned diffusion model converts that altered latent point into a full expression profile.
Siyu He accepts the analogy but emphasizes its assumption: biological processes are complex and usually nonlinear. The case is only that semantic space might be smoother and more structured than raw gene-expression space, while the diffusion process can represent nonlinear biological processes.
Differentiation supplies an intuitive example. Given iPSC measurements at day zero and more mature mesoderm cells at day three, the system uses semantic interpolation to infer day-one and day-two states. A related experiment combines learned effects from separate gene perturbations A and B to predict A-plus-B.
8. Holdout tests support the method while exposing trajectory error
For the differentiation test, the model trains on day zero and day three while real day-one and day-two measurements are held out. It generates those intermediate transcriptomes from semantic interpolation, then compares the predictions with the withheld experimental measurements.
The perturbation evaluation follows the same logic: experiments for gene A, gene B, and their combination exist, but the combination is withheld during training. The model must construct the A-plus-B state from the individual semantic directions rather than use the observed combined result.
In a more realistic organoid case, collecting every day would destroy the samples, so researchers may collect data at day zero and at the final day, then use interpolation to generate intermediate stages. Siyu He also describes comparisons with publicly available time-series data: the team identified novel cell states, studied genes triggered by differentiation, and found consistent results against experimental data from other publications.
Labenz presses for failures rather than only successes. Siyu He says the tested cases generally worked, but intermediate predictions were “not as good as we hoped”: multiple growth factors and directions can lead to the same final state. Having semantic variables for each individual growth factor could improve the trajectory; the current interpolation remains an approximation.
9. Unseen drugs and time dynamics define the next Squidiff variants
Base Squidiff can transfer a known drug direction between cell types, but it cannot infer drug C from training that includes only A and B; no semantic vector for C exists.
The proposed remedy is a drug adapter that encodes the drug’s molecular structure and dosage, then combines that embedding with the cell’s semantic variable. A previously unseen drug could thereby condition generation if its relevant components and dosage are provided.
A second variant takes inspiration from video generation. Instead of learning endpoint states and assuming a scaled direction between them, it would train on time-series data and represent development over time, potentially capturing more of the data’s nonlinearity.
Siyu He sees immediate applications in selecting organoid growth-factor recipes and screening how engineered tissue responds to drug compounds. For a cancer patient, the longer-term ambition is to compare candidate drugs against the patient’s cellular state and predict likely usefulness and possible side effects before a long wait for the outcome.
10. Clinical usefulness requires experiments, accountability, and mixed teams
The operational contrast is compelling but bounded: obtaining a new single-cell dataset may take a week or more, while generating another conditioned state might take about an hour. That speed can provide “some intuition” about a project’s direction, but Siyu He explicitly says researchers should not “trust the model completely.”
Labenz distinguishes discovery from deployment. In a lab, erroneous predictions can be tested with additional experiments; in patient care, accuracy and consequences matter, so experimental testing, preliminary validation, and clinical testing must precede use in real patients.
Asked whether the roadmap is merely a “simple matter of programming,” Siyu He emphasizes that progress depends on community collaboration: biologists understand disease mechanisms, machine-learning researchers build models, statisticians evaluate them, and other specialists help make the work more rigorous.
Siyu He’s confidence is therefore conditional. High-quality data, domain expertise, and external validation remain inseparable from model improvement, especially because these systems are being applied to health care and medicine.
11. CORAL reconstructs the spatial context destroyed by single-cell sequencing
CORAL begins from the information Squidiff’s primary modality lacks. Dissociating tissue preserves molecular measurements for individual cells but removes their locations, even though cells interact with one another and their neighbors and placement within tissue can matter for development and disease.
Spatial transcriptomics and related spatial technologies retain location but introduce resolution tradeoffs. Spatial transcriptomics may measure 30,000–60,000 genes at lower spatial resolution, while spatial proteomics may resolve locations more sharply but measure only a limited number of proteins.
The modalities may also come from adjacent rather than identical tissue slices, so structures can be shifted between measurements. Siyu He’s analogy is two photographs of the same person: one high-resolution but grayscale, the other colorful but mosaic-like, captured from a slightly different position.
CORAL’s goal is the equivalent of a colorful, high-resolution image. It integrates the incomplete modalities and deconvolves lower-resolution measurements into predicted cell-level profiles, enabling tissue analysis at a resolution the original measurement did not directly provide.
12. Graphs, synthetic ground truth, and real tissue complete CORAL’s evidence chain
CORAL uses a graph neural network to model individual cells and their interactions. Each cell connects not only to the immediately adjacent cell but to a configurable number of nearest neighbors; increasing that number expands the modeled neighborhood, though Siyu He primarily focuses on nearby cells and their communication.
Cell-level latent features support identification of “functional domains,” tissue architecture, interaction patterns, and spatial variability. Tumors may infiltrate surrounding normal tissue without a crisp morphological edge, while bone or blood vessels may show sharp boundaries; molecular patterns can also reveal structures invisible from morphology alone.
CORAL balances those regimes through a compound objective. Reconstruction losses support accuracy and can help preserve genuine sharp edges, while graph smoothness regularization encourages plausible local smoothness. The aim is to balance smoothness and sharpness rather than require every tissue boundary to blur.
Labenz’s synthetic-data pushback is worth keeping: if a simulator can generate the answer, “don’t we in some sense already know what we need to know?” Siyu He replies that simulators such as Splatter and a collaborator’s scDesign provide known domains, interactions, and distributions for controlled validation—not the discoveries ultimately claimed from real tissue.
Synthetic data is used along with experimental data to validate the model. Because the synthetic data has ground truth, the team can test whether CORAL identifies domains, interactions, and higher-resolution structure. The synthetic-data generation process is not the same as the model being evaluated, so Siyu He stresses that “there’s no loop.”
Labenz closes by proposing biological data as a more concrete national investment than an undefined superintelligence or trillion-dollar compute buildout. He points to privacy-limited clinical records and expensive measurement; Siyu He cites virtual-cell work, the Chan Zuckerberg Initiative’s Billion Cells Project, biological models that answer questions about cells, virtual doctors, and digital twins, while remaining “very optimistic” about AI accelerating health care over the next one to two years.