single-cell

A cell’s fate isn’t written only by its own genome — it’s shaped by the signals arriving from its neighbors. For a week we’ve zoomed in on single molecules; today we listen to the conversations between whole cells, reconstructed from the same sequencing data. Efremova et al. built CellPhoneDB, ‘a novel repository of ligands, receptors and their interactions’ whose database ’takes into account the subunit architecture of both ligands and receptors,’ with ‘a statistical framework that predicts enriched cellular interactions.’ Jin et al.’s CellChat can ‘quantitatively infer and analyze intercellular communication networks,’ predicting ‘major signaling inputs and outputs’ via ‘manifold learning.’ NicheNet went further — ’linking ligands to target genes’ by folding in ‘prior knowledge on signaling and gene regulatory networks’ to find ‘active ligands and their gene regulatory effects.’ Then space returned: Cang & Nie’s SpaOTsc uses ‘structured optimal transport’ so ‘cell-cell communications are… obtained by optimally transporting signal senders to target signal receivers in space,’ and COMMOT scaled it, ‘account[ing] for the competition between different ligand and receptor species as well as spatial distances.’ Finally, NCEM brought ‘graph neural networks that estimate the effects of niche composition on gene expression.’ A tissue, it turns out, is a conversation.

A cell’s identity isn’t a list of active genes — it’s a circuit. Transcription factors switch each other on and off in tangled feedback loops, and reconstructing that wiring from data is, as one landmark put it, ‘one of the pressing open problems of computational systems biology.’ Gene regulatory network inference is how AI reads the circuit. GENIE3 turned it into a stack of regression problems — feature importance as a regulatory link. DREAM5’s blind test of 30+ methods delivered a humbling verdict — ’no single inference method performs optimally’ — and a fix: ensembles, ‘wisdom of crowds.’ SCENIC brought it to single cells, reconstructing networks and cell states at once; BEELINE benchmarked the field and found accuracy only ‘moderate,’ with pseudotime-free methods ahead. Then the payoff: CellOracle uses inferred networks ’to perform in silico transcription factor perturbations … using only unperturbed wild-type data’ — knock out a gene on a laptop, predict the phenotype, validate at the bench. SCENIC+ adds chromatin for enhancer-driven wiring. It’s the causal layer a virtual cell can’t do without.

A single cell speaks several molecular languages at once — which genes its chromatin lets it read, which RNAs it’s transcribing, which proteins are actually doing the work. New assays measure two or three of these in the same cell, and the AI job is fusion: ’tying together data across different modalities’ that, as GLUE puts it, ’typically have distinct feature spaces.’ Seurat’s weighted-nearest-neighbor learns ’the relative utility of each data type in each cell’ across a 211,000-cell CITE-seq atlas, looking ‘beyond the transcriptome toward a unified and multimodal definition of cellular identity.’ totalVI models RNA and protein jointly as ‘a composite of biological and technical factors’; MultiVI extends the idea to chromatin and even to ‘cells for which one or more modalities are missing.’ GLUE fuses unpaired datasets with a regulatory graph, scaling to a ‘human cell atlas construction over millions of cells.’ And the scIB benchmark — 68 methods, 85 batches, 1.2 million cells — keeps everyone honest about the real tension: removing batch effects ‘while retaining biological variation.’ It’s the multi-layer definition of a cell a virtual cell still needs.

Three weeks ago we followed the RNA map of a tissue; today it’s the protein map. Multiplexed imaging like CODEX stains one slice for dozens of proteins at once — ‘antibody binding events’ read out ‘at a single-cell and cellular neighborhood’ level — turning a piece of tissue into a ~40-channel image where every pixel carries a molecular fingerprint. The AI job is a three-step relay: find every cell (Mesmer, trained on TissueNet’s ‘more than 1 million manually labeled cells,’ hit human-level segmentation across tissue types); name every cell from its protein profile (STELLAR, ‘a geometric deep learning method for cell-type discovery,’ applied straight to CODEX data); then read the neighborhoods. And the honest wall: a 2024 Nature Methods benchmark of ‘more than 50 diverse biological experiments’ found methods ‘often tailored to specific modalities or require manual interventions’ — generalization, not accuracy on one panel, is the frontier. It’s the imaging×omics intersection at the lab’s core, and the protein half of the spatial map a virtual cell still needs.

A single-cell sequencing run is a still photograph — every cell frozen at the instant it was lysed. So how do you recover motion from a photo? RNA velocity’s answer, from Sten Linnarsson’s group at Karolinska and colleagues, is beautiful: count a gene’s unspliced (new) versus spliced (mature) transcripts, and the imbalance becomes ’the time derivative of the gene expression state’ — a vector that ‘predicts the future state of individual cells on a timescale of hours.’ scVelo dropped the brittle steady-state assumptions; CellRank turned velocity into probabilistic fate maps; and the 2024 deep-learning turn — veloVI, CellRank 2 — added the one thing the method most needed: a way to know when to trust its own arrows. It’s the molecular twin of yesterday’s imaging-based cell tracking — same question, where is this cell going next, from the opposite kind of data — and it’s exactly the cell-state dynamics a virtual cell would have to reproduce.

A structured review of single-cell foundation models — pretrained models that learn transferable representations of cellular state across tissues, conditions, and modalities — surveying their core design dimensions, applications to biological discovery, and open challenges in fidelity, generalization, scaling, and perturbation modeling.

AlphaFold made proteins predictable; the field’s next moonshot is a virtual cell you can query — give it a starting cell state and a perturbation, and have it predict the gene-expression response you’d otherwise run an experiment to see. Arc Institute has turned that dream into a CASP-style contest, the Virtual Cell Challenge, drawing 5,000+ entrants, and shipped its first-generation State model. But the same year delivered a bracing reality check: a Nature Methods benchmark found that today’s single-cell foundation models — scGPT, Geneformer and peers — still don’t beat simple linear baselines at predicting perturbation effects. It’s our flagship bet, seen honestly: a thrilling ambition and the scoreboard that will keep it accountable.

Today in AI for life science: the AlphaFold Database opens millions of protein-complex predictions toward a human interactome, interpretable foundation models start mapping the cell’s regulatory wiring, and honest tests ask whether the diagrams are real.

Today in AI for life science: CZI’s rBio learns to reason about cells with a virtual cell as its verifier, self-driving labs cross into industrial infrastructure, and a sharp reality check on what virtual-cell models still can’t do.

Today in AI for life science: a benchmark shows AI ‘scientists’ still struggle to even find the right papers, agents arrive in ImageJ with reproducibility built in, and the Human Cell Atlas convenes as the retina gets mapped.