Perturb-MARS: Reading mouse experiments through a human lens
2 days ago
- #cancer_research
- #AI_in_biology
- #spatial_biology
- Perturb-MARS combines Perturb-Map (multiplexed in vivo perturbation platform) and TARIO-2 (foundation model trained on human cancer tissue) to interpret mouse experiments through a human biological lens.
- Perturb-Map enables testing hundreds of genetic knockouts in a single mouse with spatial resolution, reducing cost, variability, and scale compared to traditional animal studies.
- TARIO-2 generalizes from human to mouse H&E, predicting human gene expression from mouse tumor morphology, effectively projecting mouse data into human biological coordinates.
- The system validates predictions by showing t-SNE organization aligns with known immune infiltration effects (e.g., cold/hot tumors) and literature on specific knockouts like B2M or CTNNB1.
- Perturb-MARS facilitates patient stratification and combination therapy exploration by analyzing how genetic backgrounds respond to treatments, identifying synergies or antagonisms with drugs like anti-PD-1.
- It supports active-learning loops, using mouse experiments to verify model predictions, improving generative models of human cancer biology beyond associative data alone.
- The approach leverages H&E stains for scalability and cost-effectiveness, aiming to build a simulator for predicting tumor responses to interventions at single-cell resolution.