Fascinatingly, the relationship between BMP and natural retinoic acid is well established in NORMAL DEVELOPMENT, tipping cells between apoptosis and differentiation. This mechanism can seemingly be preserved in neuroblastoma CANCER CELLS and exploited to kill them! 🤯🤯 10/n

Remarkably, using single-cell RNA-seq and IF we found that BMP signaling activity is dramatically higher in disseminated neuroblastoma cells in patient bone marrow. This seems to explain the unusual clinical observation of site-specific drug activity! 🤯 9/n

Retinoic acid has other seemingly unique activities. For example, while it has little activity against established bulky tumors in vivo, it has been shown to clear metastatic cells from the bone marrow in patients. Could this paradox stem from differential BMP activity? 8/n

We then showed that amplifying BMP signaling activity (e.g. by adding exogenous BMP to cells) can potentiate the apoptosis/senescence cell fate choice upon exposure to retinoic acid. Blocking BMP tipped the cells towards differentiation, and a comparatively weaker response. 7/n

We performed a genome-wide CRISPR screen in a retinoic acid-treated hyper-responsive neuroblastoma cell line, finding that BMP signaling genes (ACVR1, BMP, SMAD, etc) were heavily affecting response to retinoic acid. 6/n

No. The hyper-responsive cell lines are predominantly undergoing either apoptosis or senescence, before (or in parallel with) differentiation... 5/n

However, we now have data from much larger preclinical screens. Using these data, we observed that, curiously, some cell lines behave as extreme outliers in their response to retinoic acid. Are these cells (arrows below) mega-differentiating!? 4/n

Spurred by these observations, clinical trials in the 90s showed a dramatic improvement in neuroblastoma patient survival when retinoic acid was added as post-chemo maintenance therapy – seemingly confirming that differentiation is effective 3/n

The idea that retinoids are effective because of differentiation stems from the 1980s when these compounds were shown to make cancer cells differentiate in vitro. Differentiation is particularly striking in some pediatric neuroblastoma cell lines. 2/n

Performance highlights: On simulated 1M-cell data, CSI-GEP recovered all ground-truth GEPs—no missed or spurious programs. On real datasets (e.g., neuroblastoma, mouse brain atlas), it picked up critical cell-state programs that other methods completely missed. 5/6

CSI-GEP outperforms the popular methods we tested (including cutting-edge GPT-based ones) in recovering biological signal in both simulated and real datasets. It’s GPU-based, so it’s fast and scalable to millions of cells. 4/6

How does it work? We apply consensus-NMF repeatedly at multiple dimensionalities (k). Critically: we abandon the accepted notion that there exists some “optimal” value of k: We look for GEPs that consistently reappear in these repeated runs—that’s the true biological “signal”.

Key idea: Instead of forcing single-cell RNA-seq data into hard clusters or unreliable 2D projections, CSI-GEP uses soft clusters to simultaneously capture cell-type and pathway activities, providing a more reasonable model for the structure of most single-cell RNA-seq datasets.