AnnFlux: object-conditioned neural stochastic differential equations for single-cell perturbation dynamics
Year: 2027
Type: dataset
Abstract: AnnFlux is a single-cell perturbation modeling framework for predicting transcriptional responses to genetic and chemical perturbations. This resource contains the processed data, trained models, reference knowledge, and derived outputs used to evaluate AnnFlux across diverse perturbation settings. It covers the Replogle, Frangieh, SciPlex3, Norman, ComboSciPlex, Wessels, and Cook datasets, including single-perturbation, combination-perturbation, cell-holdout, time-course, and ablation analyses. The resource is organized into complementary components: Inputs and Models: Canonical processed input datasets, train/test fold definitions, trained AnnFlux and ablation-model checkpoints, and perturbation, pathway, and prior-knowledge reference resources. Cell-level Predictions: Predicted gene-expression profiles for all reported datasets, folds, experimental settings, and random seeds. In each 'pred.h5ad' file, the AnnData 'X' matrix contains predicted expression values and 'obs["perturbation"]' identifies the corresponding perturbation. Prediction files and their associated per-perturbation evaluation tables are grouped by dataset. Observed Test Data: Ground-truth test observations and control-cell pools are provided in a losslessly deduplicated representation. For each dataset, 'observed_union_<dataset>.h5ad' stores each distinct observed row once, while `observation_maps.parquet` preserves the source file, row order, and evaluation role required to reconstruct every original 'obs_test.h5ad'. The included reconstruction script was validated against all 142 source observation files, comprising 2,558,014 rows, with no mismatches in expression values, shape, gene order, or perturbation order. Analysis and Evaluation: Per-perturbation evaluation results, benchmark summaries, pathway analyses, latent-space analyses, perturbation-magnitude analyses, prior-knowledge-network analyses, time-course results, and visualization source data used in the AnnFlux study. This resource supports reproduction of the reported benchmark and distributional evaluations, reuse of trained AnnFlux models, and downstream analysis of the published cell-level predictions. See the associated AnnFlux source-code repository.
Institution: Korea University
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7204172457