OpenPerturb
Interpretable AI for context-dependent genomic perturbation response.
OpenPerturb is an ongoing project on interpretable multimodal AI for genomic perturbation response. The central goal is to predict how genomic perturbations alter transcriptomic and cell-state outcomes as a function of pre-perturbation cellular context, while also supporting biologically grounded interpretation.
Project overview
Modern functional genomics increasingly relies on perturbation assays that measure how cells respond to targeted interventions. However, these responses are strongly context dependent: the same perturbation can have different downstream effects across cell states, chromatin environments, and biological conditions. Exhaustive experimental measurement of all relevant perturbation-context combinations is not realistic.
This project explores multimodal AI methods that integrate pre-perturbation transcriptomic state, chromatin context, and perturbation information to model downstream responses. A parallel goal is to develop structured latent representations that help identify reusable regulatory programs rather than producing only black-box predictions.
Current research themes
- Context-dependent prediction of perturbation response across unseen cellular settings
- Open benchmark design for multimodal perturbation modeling and cross-dataset evaluation
- Structured latent-factor models for biologically grounded interpretation
- Reusable software, model checkpoints, and evaluation protocols for the broader community
Prior work
This direction builds on prior work adapting modern sequence models to difficult genomic prediction tasks.
-
NextVir: Enabling classification of tumor-causing viruses with genomic foundation models.
-
XVir: A Transformer-Based Architecture for Identifying Viral Reads from Cancer Samples.
Status
This page describes an active project area rather than a completed public platform. Content will be updated as benchmark, software, and model releases become available.