Lab Newsletter — September 29, 2026: Predicting the Perturbation

All month we’ve built toward one question, and today we ask it directly: if I perturb this cell — silence a gene, add a drug — what will it become? A cell answers by rewiring which genes it expresses, and the dream of the virtual cell is to predict that answer before doing the experiment. It’s the single most load-bearing task in cell modeling: state + perturbation → new state. Today’s digest follows the models learning to compute it — a distinct lineage from the foundation models we covered earlier: these are purpose-built response predictors.

🧫 The data engine: measure responses at scale

You can’t learn a mapping you can’t observe. Dixit et al. (Cell, 2016) built the observation tool, Perturb-seq, “combining single-cell RNA sequencing … and CRISPR-based perturbations to perform many such assays in a pool.” It “accurately identifies individual gene targets, gene signatures, and cell states affected by individual perturbations and their genetic interactions” — turning “what does perturbing gene X do to the whole transcriptome?” into a high-throughput, readable measurement. Every predictor below trains on data like this.

🗺️ The phenotype landscape: genetic-interaction manifolds

Genes don’t act alone — combinations do surprising things. Norman et al. (Science, 2019) asked “how cellular and organismal complexity emerges from combinatorial expression of genes,” and built “high-dimensional landscapes of cell states (manifolds)” from Perturb-seq profiles of strong genetic interactions. Navigating that manifold let them order regulatory pathways and classify interactions (even find “suppressors” and unexpected synergies) — the geometric picture of perturbation space that later models learn to interpolate and extrapolate across.

🧮 The first leap: predict what you haven’t seen

Fitting the training data is easy; the prize is generalization. Lotfollahi, Wolf & Theis (Nature Methods, 2019) named the gap plainly — “no generalization of predictions to phenomena absent from training data (out-of-sample) has yet been demonstrated” — and closed it with scGen, “a model combining variational autoencoders and latent space vector arithmetics.” scGen “accurately models perturbation and infection response of cells across cell types, studies and species,” learning “features that distinguish responding from non-responding genes and cells.” Predicting a response as a direction in a learned latent space: elegant, and it worked.

🎚️ Composition: dose, drug, cell type, time

Real screens vary many factors at once. Lotfollahi et al. (Molecular Systems Biology, 2023) built the compositional perturbation autoencoder (CPA) for exactly that, starting from the hard truth that “an exhaustive exploration of the combinatorial perturbation space is experimentally unfeasible.” CPA “combines the interpretability of linear models with the flexibility of deep-learning approaches” to “in silico predict transcriptional perturbation response at the single-cell level for unseen dosages, cell types, time points, and species” — and, validated on new data, “can predict unseen drug combinations.” Decompose the factors, recombine them: predict the cell you never measured.

🕸️ The truly novel: multigene perturbations via a knowledge graph

The combinatorial wall is steepest for multigene perturbations. Roohani, Huang & Leskovec (Nature Biotechnology, 2024) scaled it with GEARS, which “integrates deep learning with a knowledge graph of gene-gene relationships to predict transcriptional responses to both single and multigene perturbations.” The headline capability: GEARS “is able to predict outcomes of perturbing combinations consisting of genes that were never experimentally perturbed,” with “40% higher precision than existing approaches.” Biological priors (which genes relate to which) let the model reach genuinely unseen combinations — priors plus learning, the lab’s recurring recipe.

🎛️ Disentangle, then shift

A final angle: if you can cleanly separate what a cell is from what was done to it, you can recombine them. Piran et al. (Nature Biotechnology, 2024) built biolord, “a deep generative method for disentangling single-cell … data to known and unknown attributes.” By “virtually shifting cells across states,” biolord “generates experimentally inaccessible samples, outperforming state-of-the-art methods in predictions of cellular response to unseen drugs and genetic perturbations.” Disentanglement as the route to controllable, counterfactual cell states.

🧫 Why it’s our kind of problem

This is the virtual cell in one sentence: state + perturbation → new state. Everything the lab is building toward — simulate a cell, then ask it what a drug, a knockout, or a signal would do — reduces to this prediction. Two themes ring loud. First, the real bottleneck is generalization, not fitting: Norman, CPA, and GEARS all invoke the combinatorial explosion — you can never measure every perturbation × cell × dose, so the entire value is extrapolating beyond the training grid. That makes honest out-of-distribution benchmarks and open perturbation datasets as important as any architecture — the open-data stance the lab keeps returning to. Second, the winning designs inject structure and priors — knowledge graphs, disentanglement, compositionality — rather than trusting scale alone. Predicting a cell’s answer to a question we never asked it: that’s the assay of a working virtual cell, and these are the first drafts.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (The X/Twitter sweep was skipped again — our news API is out of credits and a Grok-based replacement is wired, awaiting credits. Anchors were verified via NCBI E-utilities and Europe PMC.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.