Cell-state conditioning predicts the expression response to an unseen stimulus but not stimulus-dependent genetic effects

Second-generation sequence models can be conditioned on a cell state supplied as a single-cell snapshot rather than selected from a fixed catalogue, but whether that conditioning carries an experimentally applied perturbation has not been tested. We conditioned such a model on interferon-β-stimulated and unstimulated monocytes from a published experiment, using cells disjoint from those providing the measurement, and asked whether the predicted expression response tracks the measured one and whether the same conditioning identifies expression QTL specific to the stimulated state. Predicted and measured responses agree at Spearman ρ = 0.616 across 119 genes (95% CI 0.486 to 0.716), and at 0.607 after all 119 target genes are deleted from the conditioning input; across fifteen conditions the model's response scales with how much each condition's biology genuinely overlaps the target response (slope 0.73, Pearson 0.92), and a near-neighbour monocyte subset receives 0.295 against a measured overlap of 0.336. For stimulation-specific genetic effects we find no reliable evidence: the pre-specified primary arm moves from p = 0.0019 to 0.443 once both conditions must carry a measurable predicted effect, and the separation localises to the resting-condition baseline. Neither model generation's catalogue contains a perturbed sample, 0 of 4,430 tracks and 0 of 2,799 conditions, although both carry activation-state labels, so the gap is experimental perturbation rather than cell state. Two incidental findings apply beyond this model: credible-set membership mislabels over half of putatively condition-specific loci, and gating a ratio statistic on its larger term manufactures signal.

Authors

Institutions

Publication Details

Journal
bioRxiv (Cold Spring Harbor Laboratory)
Published
2026-11-17
DOI
https://doi.org/10.64898/2026.11.16.767041
Primary Topic
Genetic Associations and Epidemiology
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Cell-state conditioning predicts the expression response to an unseen stimulus but not stimulus-dependent genetic effects

Jianxuan Sun, Yifan Huang, Maimaiti Abudureyimu, Shaogang Wang et al.
3 links
bioRxiv (Cold Spring Harbor Laboratory)
Genetic Associations and Epidemiology
preprint

Cell-state conditioning predicts the expression response to an unseen stimulus but not stimulus-dependent genetic effects

Jianxuan Sun, Yifan Huang, Maimaiti Abudureyimu, Shaogang Wang, Qidong Xia, Haodong He, Zhihua Wang, Jingyu Xu, Yuxuan Yang
preprint en

Abstract

Second-generation sequence models can be conditioned on a cell state supplied as a single-cell snapshot rather than selected from a fixed catalogue, but whether that conditioning carries an experimentally applied perturbation has not been tested. We conditioned such a model on interferon-β-stimulated and unstimulated monocytes from a published experiment, using cells disjoint from those providing the measurement, and asked whether the predicted expression response tracks the measured one and whether the same conditioning identifies expression QTL specific to the stimulated state. Predicted and measured responses agree at Spearman ρ = 0.616 across 119 genes (95% CI 0.486 to 0.716), and at 0.607 after all 119 target genes are deleted from the conditioning input; across fifteen conditions the model's response scales with how much each condition's biology genuinely overlaps the target response (slope 0.73, Pearson 0.92), and a near-neighbour monocyte subset receives 0.295 against a measured overlap of 0.336. For stimulation-specific genetic effects we find no reliable evidence: the pre-specified primary arm moves from p = 0.0019 to 0.443 once both conditions must carry a measurable predicted effect, and the separation localises to the resting-condition baseline. Neither model generation's catalogue contains a perturbed sample, 0 of 4,430 tracks and 0 of 2,799 conditions, although both carry activation-state labels, so the gap is experimental perturbation rather than cell state. Two incidental findings apply beyond this model: credible-set membership mislabels over half of putatively condition-specific loci, and gating a ratio statistic on its larger term manufactures signal.

bioRxiv (Cold Spring Harbor Laboratory)
Tongji Hospital (CN), Huazhong University of Science and Technology (CN)
Genetic Associations and Epidemiology
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon