Perturbation-response model that predicts differentially expressed genes and fold changes for unseen cell lines, perturbations and datasets.
No providers recorded yet. Browse all providers
A single-cell perturbation screen measures thousands of cells per knockdown, yet no individual cell is ever observed both before and after the perturbation. Models that try to generate the post-perturbation expression of each cell must therefore separate the perturbation's effect from ordinary cell-to-cell heterogeneity and batch effects, while the metrics researchers care about, such as which genes are differentially expressed, are population-level quantities. PIE (Perturbation is Everything) predicts those population-level outcomes directly. For a (context, perturbation) pair it outputs, for every measured gene, a probability of differential expression, a log2 fold change and a change in mean expression.
PIE was developed at Arc Institute by Rishi Verma, Abhinav Adduri, Beatrice Bevilacqua, Basak Eraslan, Dave Burke, Hani Goodarzi (also UCSF) and Yusuf Roohani, and released as a bioRxiv preprint in October 2026. The same group built State, which models sets of individual cells; PIE is a separate model built around a different learning target.
Generalization comes from describing the experiment rather than memorizing it. A cell line is represented by its Cellosaurus entry, and a perturbed gene by its NCBI Gene record, ESM-2 protein embedding, STRING network embedding and DepMap knockout profile. Compounds are represented by their PubChem entry, SMILES and measured L1000, PRISM and JUMP activity profiles. An unseen cell line or gene is therefore encoded the same way as a seen one.
PIE follows the encode–process–decode design of Perceiver IO. A set of 512 learned latents cross-attends to the knowledge-source tokens, passes through self-attention, and is then queried independently by each gene, giving a per-gene feature vector that feeds three task heads. The trunk has model width 768, eight attention heads and 4/6/4 encoder, process and decoder blocks. Training starts from random initialization with AdamW in bf16, for 5,000 steps in within-dataset runs and 10,000 in cross-dataset runs; the knowledge-source embeddings stay frozen.
The paper evaluates two settings, and the distinction matters. In the within-dataset setting, one PIE model is trained per fold of the Replogle–Nadig CRISPRi dataset (K562, RPE1, HepG2, Jurkat), holding out one cell line. Averaged over folds, AUPRC for differentially expressed genes is 0.424 for unseen cell lines against 0.344 for Rhaister-O, and 0.304 for unseen perturbations against 0.094 for TabICL. In the cross-dataset setting, a single checkpoint trained on Tahoe-100M, the Jiang et al. cytokine CRISPRi screen, X-Atlas/Orion and the Arc Virtual Cell Challenge 2025 H1 data is applied to Replogle–Nadig with no training on its response labels and with shared cell lines excluded. It reaches AUPRC 0.378 for perturbations seen in training (Rhaister-O 0.283) and 0.252 for unseen perturbations (context mean 0.102). Baselines also include State and STACK. Shapley ablations attribute most of the unseen-context gain to the same-perturbation evidence and the perturbation knowledge sources.
PIE is aimed at prioritizing experiments: estimating which genes a knockdown or compound will
move in a cell line before it is screened, or transferring results from existing atlases to a
new line. Predictions come from a single forward pass with no fine-tuning; a new context needs
its control cells plus the original training assets, which supply the response evidence. The released cross-dataset checkpoint evaluates on Replogle out of the
box, and pie infer writes predictions for new contexts prepared from their control cells.
PIE argues for reframing perturbation modeling around condition-level effects and rich external descriptions of the system rather than larger cell-level generators, and its modular inputs make that argument testable through ablation. Performance drops sharply when context and perturbation are both unseen: within-dataset AUPRC falls to 0.239, and the gain over having no inputs at all is only 0.010. It is also weaker for perturbations with small effects. All results are computational benchmarks in a preprint that has not been peer reviewed. Code is released under CC BY-NC-SA 4.0 and checkpoints under a custom Arc non-commercial license, so commercial use requires a separate agreement.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.