AI Paper Atlas

Research map / Label-efficient and robust learning / Recognition and long-tail

scDEBART: Predicting in silico Single-Cell Perturbation Responses via Large-Scale Differential Expression Learning

Jieun Sung, Wankyu Kim

ICML 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Dense features and foundation priors

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add pretraining on 6.28M differential expression profiles (ablation)scDEBART · Norman K562, top-50enrichment factor: 12.693 -> 14.901Improved
add pretraining (ablation)scDEBART · Norman K562, top-50cosine similarity: 0.825 -> 0.852No clear effect
add perturbation-response cell filtering in fine-tuning (ablation)scDEBART · Norman K562, top-50enrichment factor: 14.349 -> 14.901No clear effect
add cell filtering (ablation)scDEBART · Norman K562, top-50RMSE: 0.463 -> 0.494Mixed
increase pretraining corpus from 50% to 100% (ablation)scDEBART · Norman K562, top-50enrichment factor: 12.51 -> 14.33Improved
increase pretraining corpus size (ablation)scDEBART · Norman K562cosine similarity: 0.772, 0.779, 0.775 (25%, 50%, 100%)No clear effect
inject uncertainty-proportional noise into scVI logFC targets at fine-tuning (ablation)scDEBART · Norman K562, top-50enrichment factor: 15.992 -> 15.453 (gamma 0.5), 15.370 (gamma 1.0)Hurt
apply scVI denoising to pretrained scGPT inputsscGPT · Perturb-seq top-50 cosine similaritycosine similarity: 0.45 -> 0.70Improved
apply scVI denoising to non-pretrained linear regression and MLPLR, MLP · Perturb-seq top-50 cosine similaritycosine similarity: 0.62 -> -0.27 (LR), 0.63 -> -0.27 (MLP)Hurt
use simple MLP/LR instead of GEARS/scGPT on raw countscount-based baselines · Perturb-seq top-50 cosine similaritycosine similarity: 0.14 (GEARS), 0.45 (scGPT) -> 0.63 (MLP), 0.62 (LR)Improved
restrict pretraining to high-detection, high-confidence DE genesscDEBART · top-50 predicted genesmedian non-zero proportion: baselines 0.09-0.28 -> 0.60 (perturbed), 0.73 (control)Improved
fine-tune scDEBART on genetic perturbations then apply to drug perturbations without extra trainingscDEBART · SCIPLEX K562, top-20 DEGscosine similarity: 0.04 (10 nM) -> 0.30 (10000 nM)Mixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy