Thread by @elhamazizi

We are thrilled to share #Starfysh ⭐️ an auxiliary deep generative model for multi-modal analysis and integration of spatial transcriptomic (ST) datasets and histology images, and its application to heterogeneous #breastcancer tumors! 😃 tinyurl.com/hffdhvx7 1/

Incredibly fortunate to have a passionate, collaborative & talented team that didn't slow down with all the challenges of the past few years 🙏: @SiyuHe7 @YinuoJin6, @AchilleNazaret @shi_lingting and a fruitful and ongoing collaboration with @GeorgePlitas and Sasha Rudensky. 2/

Using ST, we wanted to find out if our previously observed continuous phenotypic expansion of intratumoral immune states (doi.org/10.1016/j.cell…) could be explained by the #spatialdynamics of immune cells and exposure to different environmental signals and nutrient supply? 🧐 3/

However, like others, we found dissecting refined immune (e.g. Treg) & patient-specific tumor cell states in ST is not straightforward when a matched single-cell reference is not available. Also, how can we integrate heterogeneous ST datasets? ➡️ We developed a new tool! 🧲4/

Starfysh jointly models ST and histology to decompose 🧲 cell states, using anchors⚓️ (empirical prior) constructed by archetypal analysis and any curated genesets 👉 Can detect context-specific states (e.g. patient tumor cells) and find shared immune states across samples 5/

For #machinelearning enthusiasts, a specially structured #variationalinference is used with cell state proportions as auxiliary variables ➕ a variational information bottleneck approach (PoE) for histology integration 💻 6/

Testing Starfysh on data simulated from mixtures of #singlecell RNA with different assumptions, showed anchors successfully decompose and pull apart 🧲data 👇hence the name ⭐️🐠 (with Y because data is not FISH 😄). 7/

Starfysh shows comparable performance to methods that use a reference and does much better in characterizing fine-grained cell states. Plus, it takes < 5min to run Starfysh on a typical ST dataset on a laptop (64 GB RAM)! ⏰ 8/

Also, integration of associated histology helps with the improved estimation of cell density and decomposition of some cell types (e.g. myeloid, tumor) 9/

Dissecting the #spatial #heterogeneity of #TNBC tumors, revealed a transition to mesenchymal states determined by location and environmental stimuli. Interestingly, the tumor state transition coincided with shifts in immune cell states: activation of T cell and myeloid cells! 10/

We integrated 14 ST datasets to find what is special about the #spatialdynamics of aggressive chemo-resistant #metaplastic breast tumors with worse prognosis? Intratumoral regions showed enriched EMT, hypoxia & metabolic remodeling ▶️ likely selecting immunosuppressive Tregs. 11/

We defined “spatial hubs” with unique compositions of cell states. In particular, a stromal hub shared across breast cancer subtypes, showed varying spatial patterning. This niche was concentrated in hypoxic regions with possible adapting angiogenesis in #metaplastic tumors! 12/

Another advantage of identifying sample-specific cell states is the ability to quantify inter- and intra-tumor heterogeneity ⏩ Confirmed greater heterogeneity among #metaplastic tumors compared to other TNBC & ER+ subtypes. 13/

More details in the @biorxivpreprint #preprint! Thanks to all colleagues including @theleonglab @blei_lab @joselmcfaline for valuable contributions!
#Starfysh code & tutorials are available at github.com/azizilab/starf… We look forward to your feedback! 😊 14/

@threadreaderapp unroll