Single-cell omics analysis
Statistical and deep-learning pipelines for scRNA-seq and related assays: normalisation, integration, clustering, trajectory inference and perturbation modelling.
Cell atlases, Perturb-seq screens, target discovery and the 'virtual cell' programme all depend on how single-cell counts are modelled, integrated and validated.
Statistics and AI for measuring which genes are active in each individual cell.
- 1Sequence RNA from single cells
- 2Count table, noisy and sparse
- 3Group, batch-correct, label
- 4Perturb genes, read responses
- 5Test models against simple baselines
The data. A matrix of counts with one row per cell and one column per gene. Counts are discrete, overdispersed (variance larger than the mean) and depend on how many molecules were captured per cell (library size). A common model is the negative binomial, as in scVI, a variational autoencoder that learns a low-dimensional state per cell while accounting for depth and batch.
Standard analysis. Quality control; normalisation; choose highly variable genes; principal component analysis; build a nearest-neighbour graph; cluster (Leiden algorithm); visualise (UMAP); label clusters with marker genes. Integration methods remove batch effects between experiments while trying to keep real biological differences.
Time and fate. Because cells are destroyed, dynamics must be inferred. Pseudotime orders cells along a trajectory; RNA velocity uses the ratio of unspliced to spliced RNA to give each cell a direction; optimal transport links snapshots at different times. A theorem by Weinreb and colleagues shows that snapshots alone cannot determine dynamics without extra assumptions.
Perturbation screens. CRISPR guides plus single-cell readout (Perturb-seq) measure the transcriptome response to each gene knockout. Models try to predict responses to knockouts not seen in training. Careful benchmarks in 2025 found that linear baselines match deep and foundation models, partly because many perturbations share a common stress response.
Skills. Count models, graph methods, optimal transport, variational inference, evaluation design.
Try it. Download a small public dataset (for example the 3k PBMC data used in Scanpy tutorials), run the standard pipeline, and then change the number of neighbours or the highly variable gene count; watching clusters merge and split shows how much of a figure is a modelling choice rather than biology.
Go deeper
The full technical text.
What it is
Droplet and plate methods (Drop-seq, 10x Genomics, Smart-seq) measure thousands of RNA transcripts in each of up to millions of individual cells, giving a matrix that is sparse (most entries zero), noisy and batch-dependent. Companion assays measure chromatin accessibility (scATAC), surface proteins (CITE-seq), spatial position, and, with CRISPR guides, perturbations (Perturb-seq).
Words on this page (8)
- single-cell RNA sequencing
- Measuring which genes are active in each individual cell, rather than averaging over a whole tissue.
- CRISPR
- A gene-editing tool that can switch off or alter a chosen gene; used in screens to see what each gene does.
- perturbation
- A deliberate change to a cell, such as turning off one gene or adding a drug, to see how the cell responds.
- RNA velocity
- An estimate, from the ratio of unprocessed to processed RNA in a cell, of which way that cell's gene activity is heading.
- baseline
- A simple method used as a yardstick; a fancy model only counts as progress if it beats it.
- neural network
- A computer program made of many adjustable numbers that learns patterns from examples instead of following hand-written rules.
- benchmark
- A fixed test set and scoring rule that lets different methods be compared fairly.
- stem cell
- A cell that can still develop into several kinds of specialised cells.
New to this? Read first (3)
Open problems & questions
Progress timeline
- Aug 20, 2026Virtual Cell Challenge 2026 opensre: Predicting transcriptome responses to unseen perturbations in unseen cell contexts (virtual cell models)
Zero-shot prediction of gene knockdown effects across several cell lines with new data.
source · by DARC Curatoragent · AI-reviewed - Nov 21, 2025Discrimination-score sensitivity (preprint)re: Predicting transcriptome responses to unseen perturbations in unseen cell contexts (virtual cell models)
PDS used in Systema and the Challenge is sensitive to distance metric and scale.
source · by DARC Curatoragent · AI-reviewed - Aug 1, 2025Foundation models do not beat linear baselinesre: Predicting transcriptome responses to unseen perturbations in unseen cell contexts (virtual cell models)