ExPert infers causal perturbations from cell states across biological and clinical contexts
- 1Northwestern University
Abstract
Pooled single-cell CRISPR screens have driven substantial effort toward models predicting the response to a perturbation. We posed the inverse question: which perturbation produced an observed cell state? Here we developed ExPert (Extracting Perturbations), which learns one prototype per perturbation in a latent space shared by 12 cellular contexts, trained on 8,964 perturbations and 19.5 million cells, and ranks 19,895 candidate genes per query, over half absent from training. ExPert more than doubled held-out-context recall over the best existing method with 1,000-fold less compute, and transferred without retraining to drugs, protein degradation and combinations. In alcohol-associated liver disease it implicated mRNA export, unchanged in expression, and we experimentally confirmed nuclear poly(A) retention and isoform switching. In primary mitochondrial disease it recovered the affected complex from blood, localized the signal to monocytes, and flagged a misdiagnosed control. Perturbation atlases can thus be queried in reverse, from state to likely driver.
Related
- PIE: Generalizing perturbation effects across unseen perturbations, contexts and datasets
- VOC·BIO Library: An Integrated Platform for Published Aphid GC-MS Evidence, Biological Contexts, Cheminformatics and Exploratory OBP Docking
- Eukaryotic AUG initiation contexts consist of a broad range of nucleotide patterns associated with distinct biological functions
- Reference-guided pseudotime inference across species and biological contexts
- Signature Recontextualization: Mapping perturbational signatures across biological contexts
BibTeX
@misc{schwartz2026expert,
title = {{ExPert infers causal perturbations from cell states across biological and clinical contexts}},
author = {L. Schwartz and J. L. Schlamowitz and B. Kuznets-Speck and C. W. Johnson and D. Das and E. I. Grody and T. Bai and H. Sun and I. Senthilkumar and G. Primavera and J. Jung and B. Kuusik and R. A. Grant and E. Rojas and A. M. Abdrabou and B. Peiffer and Z. Sun and A. M. Diehl and C. Marr and N. S. Chandel and D. S. Mithal and S. E. Weinberg and M. Adli and A. Kalsotra and E. Prashnani and Y. Goyal},
howpublished = {bioRxiv},
year = {2026},
month = nov,
doi = {10.64898/2026.11.02.762549},
url = {https://www.biorxiv.org/content/10.64898/2026.11.02.762549v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - ExPert infers causal perturbations from cell states across biological and clinical contexts
AU - L. Schwartz
AU - J. L. Schlamowitz
AU - B. Kuznets-Speck
AU - C. W. Johnson
AU - D. Das
AU - E. I. Grody
AU - T. Bai
AU - H. Sun
AU - I. Senthilkumar
AU - G. Primavera
AU - J. Jung
AU - B. Kuusik
AU - R. A. Grant
AU - E. Rojas
AU - A. M. Abdrabou
AU - B. Peiffer
AU - Z. Sun
AU - A. M. Diehl
AU - C. Marr
AU - N. S. Chandel
AU - D. S. Mithal
AU - S. E. Weinberg
AU - M. Adli
AU - A. Kalsotra
AU - E. Prashnani
AU - Y. Goyal
PY - 2026
DA - 2026/11/02
PB - bioRxiv
DO - 10.64898/2026.11.02.762549
UR - https://www.biorxiv.org/content/10.64898/2026.11.02.762549v1
AB - Pooled single-cell CRISPR screens have driven substantial effort toward models predicting the response to a perturbation. We posed the inverse question: which perturbation produced an observed cell state? Here we developed ExPert (Extracting Perturbations), which learns one prototype per perturbation in a latent space shared by 12 cellular contexts, trained on 8,964 perturbations and 19.5 million cells, and ranks 19,895 candidate genes per query, over half absent from training. ExPert more than doubled held-out-context recall over the best existing method with 1,000-fold less compute, and transferred without retraining to drugs, protein degradation and combinations. In alcohol-associated liver disease it implicated mRNA export, unchanged in expression, and we experimentally confirmed nuclear poly(A) retention and isoform switching. In primary mitochondrial disease it recovered the affected complex from blood, localized the signal to monocytes, and flagged a misdiagnosed control. Perturbation atlases can thus be queried in reverse, from state to likely driver.
ER -