#12 · ML / AI methods
Biological priors constrain deep learning models of omics perturbation response
The dominant thread is learned representation of omics data constrained by biological structure. ORBIT learns asymmetric gene program influence from observational scRNA-seq using intervention-consistent objectives, and DoFormer adapts the do-operator within attention to predict unseen gene perturbations without DAG specification. PRIM uses prior knowledge to predict perturbation-induced expression changes from control-cell expression, while PPI-Net embeds patient RNA-seq profiles in STRING PPI networks and Reactome pathway hierarchies. A GNN explainer benchmark on TCGA BRCA RNA-seq compares four attribution methods over a PPI network and introduces a shell-based hub score. The H&E-based item predicting apoptosis-cascade protein expression with patch-level and cell-level ViTs forms a separate strand that does not share the same structural prior. What is new is the explicit coupling of network or causal priors to model objectives, paired with comparative evaluation, rather than embedding alone.
Published papers
Incorporating valuable prior knowledge to improve deep learning prediction of genetic perturbation responses.
PRIM, a prior-guided deep learning model that uses control-cell expression to predict perturbation-induced expression changes, outperforms existing methods on…
Predicting protein cascade expression from H&E images.
Predicts apoptosis-cascade protein expression from H&E whole-slide images in TCGA-BRCA using patch-level ViT and a cell-level ViT (CellRPPA) against RPPA data.
Preprints
ORBIT: Learning Gene Program Co-Activation Structure for Cell-Type-Stratified Pathway Rewiring Analysis in Single-Cell Transcriptomics
ORBIT, a self-supervised transformer learning asymmetric gene program influence from observational scRNA-seq via intervention-consistent objectives, applied to…
PPI-Net connects molecular protein interactions to functional processes in disease
PPI-Net is a hierarchical graph attention network that embeds patient RNA-seq profiles in STRING PPI networks and Reactome pathway hierarchies, reporting gains…
Graph neural network explanations reveal a topological signature of disease-associated hubs in biological networks
Systematically benchmarks four GNN explainers (SA, IG, GNNExplainer, LRP) on TCGA BRCA RNA-seq over a PPI network and introduces a shell-based hub score with…
DoFormer: Causal Transformer for Gene Perturbation
DoFormer is a causal multimodal Transformer that adapts the do-operator in attention to predict unseen gene perturbations from single-cell data without DAG…
Topic history
| Date | Rank |
|---|---|
| #12 | |
| #1 | |
| #33 | |
| #1 | |
| #1 | |
| #2 | |
| #3 | |
| #5 | |
| #3 | |
| #5 |